Date post: | 29-Apr-2023 |
Category: |
Documents |
Upload: | khangminh22 |
View: | 1 times |
Download: | 0 times |
1
ESTABLISHMENT OF ARTIFICIAL NEURAL NETWORK FOR SUSPENSION SPRING FATIGUE LIFE PREDICTION USING STRAIN AND ACCELERATION DATA
Von der Fakultät für Ingenieurwissenschaften, Abteilung Maschinenbau und Verfahrenstechnik der
Universität Duisburg-Essen
zur Erlangung des akademischen Grades
eines
Doktors der Ingenieurwissenschaften
Dr.-Ing.
genehmigte Dissertation
von
YAT SHENG KONG aus
Klang, Malaysia
1. Gutachter: Prof. Dr.-Ing Dieter Schramm 2. Gutachter: Prof. Ir. Dr. Shahrum Abdullah Tag der mündlichen Prüfung: 19.02.2019
ESTABLISHMENT OF ARTIFICIAL NEURAL NETWORK FOR SUSPENSION
SPRING FATIGUE LIFE PREDICTION USING STRAIN AND
ACCELERATION DATA
KONG YAT SHENG
THESIS SUBMITTED IN FULFILMENT FOR THE DEGREE OF
DOCTOR OF PHILOSOPHY
FACULTY OF ENGINEERING AND BUILT ENVIRONMENT
UNIVERSITI KEBANGSAAN MALAYSIA
BANGI
2019
DECLARATION
I hereby declare that the work in this thesis is my own except for quotations and
summaries which have been duly acknowledged.
08 April 2019 KONG YAT SHENG
P78155
ACKNOWLEDGEMENT
Author would like to express his sincere gratitude and appreciation to thesis
supervisor Prof. Ir. Dr. Shahrum Abdullah and Prof. Dr.-Ing. Dieter Schramm for
providing valuable guidance, supervision and mentoring as well as encouragement for
carrying this research.
Also, author would like to express my high appreciation to my co-supervisor
Prof. Dr. Mohd Zaidi Omar and Associate Prof. Dr. Sallehuddin Mohd. Haris for
providing valuable comment, support and guidance throughout the course and
research.
Author would like to thank all post graduate students of UKM structure
integrity group for their help, friendship, and creating a pleasant working environment
throughout his years in UKM.
Importantly, author would like to express his heartfelt thanks to his beloved
parent for their lifelong encouragement.
Last but not least, author gratefully acknowledge financial support provided by
MyBrain 15 and DAAD 2016. Without their support, author could not finish his
research in time.
v
ABSTRACT
This study presents establishment of multiple input prediction model for automotive
coil spring fatigue life estimation to shorten automotive suspension design process.
Automotive suspension design is a lengthy work where any changes of the design lead
to repetition of the entire process. It was hypothesised that the established model could
be used to predict the spring design fatigue life without using any strain measurements.
To initiate this model establishment, five sets of strain and acceleration measurement
across different road conditions were collected and used for validations. To include
spring stiffness as a parameter, a quarter car model was generated to obtain the force
time histories from spring and vertical vibration of vehicle mass. In addition, artificial
road profiles of road classes “A” to “D” were also generated for the quarter car
simulation. Through adjusting the spring stiffness in the quarter car model, the spring
and vehicle responses were varied. The simulated force time histories were used to
predict springs’ fatigue life while acceleration time histories were used to calculate
ISO 2631 ride-related vertical vibration. Subsequently, multiple linear regression
approach was applied to determine the relationship between vehicle body frequency,
ISO 2631 ride-related vertical vibration and spring fatigue life. The obtained
regression had shown significance to the spring fatigue life with coefficient of
determination of 0.8320. Reciprocally, multiple linear regression models were also
used to predict the ISO 2631 ride-related vertical vibration with a coefficient of
determination at 0.8810 and mean squared error values below 0.3430. To optimise the
prediction results, artificial neural network was trained for the fatigue and vibration
prediction purposes. The architectures of the artificial neural network were designed
in terms of number of neurons and hidden layers to achieve a higher coefficient of
determination of 0.9926 and lower mean squared error of 0.0824. For vibration
prediction, the vehicle body frequency and spring fatigue life has shown a significant
coefficient of determination to the ISO 2631 weighted vertical vibration, reaching
0.9579 with mean squared error of 0.0004. Based on the experimental strain and
acceleration results, the predicted fatigue lives of multiple linear regression models
were correlated well with the experimental results with coefficient of determination
value of 0.9275. Meanwhile, the maximum difference of vibration prediction to
experimental value using multiple linear regression models was only 18%. For
artificial neural network predictions, the fatigue lives were mostly distributed within
1:2 or 2:1 life correlation and vibration prediction results were within 12%. For a good
prediction, the target correlation value was above 0.80 to demonstrate a good fitted
curve and the difference below 20%. The trained artificial neural network has shown
outstanding capability in fatigue life or ride-related vertical vibration predictions. In
this research, the main novelty was the trained artificial neural network for spring
fatigue life or ride-related vertical vibration predictions which serve to reduce some
procedures of automotive suspension design. The outcome of this study can be used to
provide a new knowledge towards the field of fatigue research as well as vehicle ride
dynamics. This research contributes to automotive industries especially in suspension
spring design where the analysis of fatigue and ride-related vibration are provided.
iv
PENUBUHNAN RANGKAIAN NEURAL BUATAN UNTUK JANGKAAN
HAYAT LESU SISTEM AMPAIAN PEGAS DENGAN DATA
TERIKAN DAN DATA PECUTAN
ABSTRAK
Kajian ini membentangkan pembangunan model ramalan pelbagai input untuk
jangkaan hayat lesu pegas gegelung automotif supaya proses reka bentuk sistem
ampaian automotif disingkatkan. Proses reka bentuk sistem ampaian automotif adalah
memakan masa dan segala perubahan reka bentuk akan mengakibatkan pengulangan
proses tersebut. Hipotesis kajian ini ialah model yang dibangunkan boleh digunakan
untuk meramalkan hayat lesu reka bentuk pegas gegelung dengan tanpa menggunakan
sebarang isyarat terikan. Sebagai permulaan, lima isyarat terikan dan pecutan yang
dicerap daripada keadaan jalan yang berbeza telah digunakan sebagai kes kajian.
Untuk merangkumi kekakuan pegas sebagai parameter, model kereta sukuan dibina
untuk memperolehi isyarat beban pada pegas dan pecutan menegak kenderaan.
Sebagai tambahan, profil jalan buatan kelas "A" hingga "D" juga dijana untuk
simulasi kereta sukuan. Dengan melaraskan kekakuan pegas dalam model kereta
sukuan, tindak balas pegas dan badan kenderaan telah berubah. Isyarat beban simulasi
yang dijana digunakan untuk meramalkan hayat lesu pegas gegelung manakala isyarat
pecutan digunakan untuk mendapatkan indeks getaran menegak yang berkaitan
dengan ISO 2631. Seterusnya, kaedah regresi linear berganda telah digunakan untuk
menentukan hubungan antara frekuensi badan kereta, getaran menegak ISO 2631 dan
hayat lesu pegas gegelung. Perkaitan yang diperolehi telah menunjukkan nilai pekali
penentuan setinggi 0.8320. Sebaliknya, model regresi linear berganda yang digunakan
untuk meramal getaran menegak ISO 2631 mempunyai pekali penentuan minimum
0.8810 dan min kuasa dua ralat di bawah 0.3430. Untuk mengoptimunkan prestasi
ramalan, rangkaian neural buatan telah dilatih untuk tujuan ramalan yang sama. Seni
bina rangkaian neural buatan telah direka dari segi bilangan neuron dan lapisan
tersembunyi untuk mencapai pekali penentuan setinggi 0.9926 dan min kuasa dua
ralat tertinggi 0.0824. Untuk ramalan getaran menegak, frekuensi kereta dan hayat
lesu pegas telah menunjukkan pekali penentu yang tinggi terhadap getaran menegak
ISO 2631, nilai 0.9579 telah dicapai dengan min kuasa dua ralat serendah 0.0004.
Ramalan hayat lesu model regresi linear berganda mempunyai korelasi yang baik
dengan keputusan eksperimen dengan pekali penentu setinggi 0.9275. Ramalan
getaran menegak menggunakan model regresi linear berganda adalah dalam julat 18%.
Bagi rangkaian neural buatan, ramalan hayat lesu kebanyakan berada dalam korelasi
hayat julat 1:2 atau 2:1 manakala perbezaan hasil ramalan getaran menegak kepada
ujikaji adalah dalam julat 12%. Untuk ramalan yang baik, nilai korelasi sasaran ialah
sekurang-kurangnya 0.80 bagi penyesuaian lengkung yang baik dan perbezaannya di
bawah 20%. Rangkaian neural buatan yang terlatih telah menunjukkan keupayaan
yang baik dalam meramalkan hayat lesu pegas atau penunggangan automotif. Dalam
kajian ini, hasil utama ialah rangkaian neural buatan yang terlatih untuk ramalan hayat
lesu pegas atau kesan menunggang supaya prosedur reka bentuk sistem ampaian
automotif dapat dikurangkan. Hasil kajian ini dijangka memberikan pengetahuan
ilmiah baharu terhadap bidang penyelidikan lesu dan penunggangan dinamik
kenderaan. Kerja ini menyumbang kepada industri automotif yang terbabit dengan
reka bentuk sistem ampaian di mana analisis kelesuan dan getaran yang berkaitan
dengan reka bentuk pegas gegelung kenderaan telah dirangkumi.
vi
TABLE OF CONTENTS
Page
DECLARATION ii
ACKNOWLEDGEMENT iii
ABSTRACT iv
ABSTRAK v
TABLE OF CONTENTS vi
LIST OF TABLES ix
LIST OF FIGURES xii
LIST OF ABBREVIATIONS xx
CHAPTER I INTRODUCTION
1.1 Overview of Fatigue Life Assessment 1
1.2 Durability of Automotive Suspension System 3
1.3 Problem Statement 6
1.4 Research Objectives 7
1.5 Scope of Works 8
1.6 Research Hypothesis 9
1.7 Significance of Research 9
CHAPTER II LITERATURE REVIEW
2.1 Introduction 11
2.2 Automotive Suspension System 12
2.2.1 Types of Automotive Suspension Systems 12
2.2.2 Mechanism of Suspension Components 14
2.2.3 Coil Spring Design and Characteristics 20
2.3 Durability Analyses 23
2.3.1 Introduction to Durability Analysis 24
2.3.2 Types of Loading Signals 27
2.3.3 Cycle Counting Methods 34
2.3.4 Fatigue Life Analysis 36
2.3.5 Linear Damage Cumulative Rule 42
2.4 Ride Related Vibration Analysis 44
2.4.1 Introduction to Ride Analysis 44
vii
2.4.2 Road Roughness as Excitation Sources 47
2.4.3 Vehicle Suspension System Modelling 50
2.4.4 International Ride Standards 54
2.5 Data Analysis 56
2.5.1 Introduction to Data Analysis 56
2.5.2 Regression Analysis 59
2.5.3 Neural-Network based Regression Analysis 63
2.5.2 Validation Analysis 67
2.6 Summary 70
CHAPTER III METHODOLOGY
3.1 Introduction 71
3.2 Durability Characterisation of the Suspension System 72
3.2.1 Perform Analysis on Spring 72
3.2.2 Collection of Strain and Acceleration Signals 77
3.2.3 Characterisation of Fatigue and ISO 2631
Vertical Vibration 82
3.3 Establishment of Multiple Linear Regression 89
3.3.1 Generating Artificial Road Profiles 89
3.3.2 Extract Data from Simulated Quarter Car Model 92
3.3.3 Perform Spring Stiffness Sensitivities 97
3.3.4 Establishment of Multiple Linear Regression for
Fatigue Life 99
3.4 Predicting the Spring Fatigue Life 101
3.4.1 Design the Artificial Neural Network Architecture 101
3.4.2 Optimise the Spring Fatigue Life 107
3.5 Validation of the Prediction 108
3.4.1 Validation using Normality Test 108
3.4.2 Validation using Experimental Data 111
3.6 Summary 113
CHAPTER IV RESULTS AND DISCUSSION
4.1 Introduction 114
4.2 Determining the Durability Characteristics 115
4.2.1 Finite Element Analysis of Coil Spring 115
4.2.2 Observing Signal Characteristics 117
4.2.3 Regression Analysis for fatigue life 130
4.3 Data For Modelling 136
viii
4.3.1 Artificial Road Profiles 136
4.3.2 Quarter Car Model Simulation Results 140
4.3.3 Spring Stiffness Sensitivity Analysis 151
4.3.4 Fatigue Life and Vibration Data 153
4.4 Multiple Input Regression 155
4.4.1 Analysis for Life Regression 155
4.4.2 Analysis for Vibration Regression 167
4.5 Implementation of Artificial Neural Network for Prediction 175
4.4.2 Fatigue Life Prediction 175
4.5.2 ISO 2631 Vertical Vibration Prediction 199
4.6 Prediction Validation 208
4.7 Summary 217
CHAPTER V CONCLUSION
5.1 Conclusion 218
5.1.1 Establishment of the Durability Relationships 218
5.1.2 Optimisation of the Durability Relationships 219
5.1.3 Validation of the Durability Relationships 219
5.2 Research Contribution 220
5.3 Recommendations 221
REFERENCES 222
APPENDICES
A Spring Failure Report by TheAAA 250
B Mesh Sensitivity 251
C Optimised Multilayer ANN architecture 253
D Artificial Road Simulated Time Histories 256
E Measured Road Simulated Time Histories 263
F Fatigue Life and Vibration Data 271
G MSE Data For Single Layer ANN 274
H MSE For Three Layer ANN 275
I Data for Error Histogram 278
J Curve Fitting For ANN 281
K Weights and Biases for ANN 285
L List of Publications 289
ix
LIST OF TABLES
NO. TABLE PAGE
Table 2.1 Chronology of published automobile suspension
durability analysis
26
Table 2.2 Literature of types of VAL in automotive
applications
33
Table 2.3 Chronology of vehicle ride analysis 45
Table 2.4 Objective ride and preferred countries 54
Table 2.5 Researches on generation of MLR-based model 64
Table 3.1 Measured geometrical properties of the spring 75
Table 3.2 Monotonic and cyclic properties of the spring
material
76
Table 3.3 Vehicle speed during signal collection 81
Table 3.4 Guide for assessing the effects of vibration on
comfort
86
Table 3.5 Frequency-weighting curves for principal
weighting
87
Table 3.6 Suitability of fit for coefficient of determination
value
89
Table 3.7 Parameter for ISO 8608 road profile generation 91
Table 3.8 Classification of road roughness proposed by
ISO 8608
91
Table 3.9 k value for ISO 8608 road roughness
classification
91
Table 3.10 Function of each block in quarter car model 94
Table 4.1 Range of the measured strain time histories 118
Table 4.2 Statistics for the measured strain time histories
at different type of roads
119
Table 4.3 Statistics for the measured acceleration time
histories at lower arm
120
Table 4.4 Statistics for the measured acceleration time
histories at top mount
121
Table 4.5 RMSE between sprung and un-sprung mass
acceleration time histories
130
x
Table 4.6 ISO 2631 vertical accelerations and spring
fatigue lives
132
Table 4.7 Regression predicted and experimental fatigue
lives
136
Table 4.8 Statistical parameters of artificial generated road
classes
139
Table 4.9 Spring stiffness parameter sensitivity analysis 152
Table 4.10 Predicted fatigue life and ISO 2631 weighted
acceleration from class “A” road
154
Table 4.11 Predicted fatigue life and ISO 2631 weighted
acceleration from highway road
154
Table 4.12 F-test for vibration-life regression analysis 160
Table 4.13 t-test of vibration-life datasets for various
approaches
160
Table 4.14 Standardised coefficients of independent
variables
161
Table 4.15 Coffin-Manson MLR-based fatigue life
predictions
165
Table 4.16 Morrow MLR-based fatigue life predictions 165
Table 4.17 SWT MLR-based fatigue life predictions 165
Table 4.18 F-test of fatigue-vibration datasets for various
approaches
169
Table 4.19 t-test of fatigue-vibration datasets for various
approaches
169
Table 4.20 Coffin-Manson MLR predicted ISO 2631
vertical vibrations
173
Table 4.21 Morrow MLR predicted ISO 2631 vertical
vibrations
173
Table 4.22 SWT MLR predicted ISO 2631 vertical
vibrations
174
Table 4.23 R2 value for all approaches with single hidden
layer ANN
180
Table 4.24 MSE for various approaches and datasets with
single hidden layer ANN
180
Table 4.25 R2 value for all approaches with two hidden
layers ANN
184
Table 4.26 MSE for various approaches and datasets with
two hidden layers ANN
184
xi
Table 4.27 R2 value for all approaches with three hidden
layers ANN
188
Table 4.28 MSE for various approaches and datasets with
three hidden layer ANN
188
Table 4.29 The best performance ANN architecture for
vibration-life predictions
191
Table 4.30 MLR and ANN predicted Coffin-Manson
fatigue lives
191
Table 4.31 MLR and ANN predicted Morrow fatigue lives 192
Table 4.32 MLR and ANN predicted SWT fatigue lives 192
Table 4.33 MSE for different trained vibration prediction
ANN
200
Table 4.34 MLR and ANN Coffin-Manson predicted ISO
2631 vertical vibration
203
Table 4.35 MLR and ANN Morrow predicted ISO 2631
vertical vibration
204
Table 4.36 MLR and ANN SWT predicted ISO 2631
vertical vibration
204
Table 4.37 Difference between experimental and MLR
vibration-life prediction for Coffin-Manson
datasets
211
Table 4.38 Difference between experimental and MLR
vibration-life prediction for Morrow datasets
211
Table 4.39 Difference between experimental and MLR
vibration-life prediction for SWT datasets
211
Table 4.40 Difference between experimental and ANN
vibration-life prediction for Coffin-Manson
datasets
215
Table 4.41 Difference between experimental and ANN
vibration-life prediction for Morrow datasets
216
Table 4.42 Difference between experimental and ANN
vibration-life prediction for SWT datasets
216
xii
LIST OF FIGURES
NO. FIGURE PAGE
Figure 2.1 Solid axle dependent suspension 13
Figure 2.2 Types of independent suspension system: (a)
Macpherson strut, (b) double wishbone, (c)
multi-link, (d) trailing arm
14
Figure 2.3 Schematic diagram of a quarter car model
under road excitation
15
Figure 2.4 Schematic diagram of spring: (a) uniaxially
loaded spring, (b) free body diagram
16
Figure 2.5 Mechanism of a damper: (a) rebound, (b)
compression, (c) hysteresis loop
17
Figure 2.6 Schematic diagram of a tyre model based on a
linear spring damper with SAE coordinate
system
18
Figure 2.7 Mechanism of a lower arm: (a) initial
condition, (b) when wheel is hitting a bump
19
Figure 2.8 Chronology of spring materials development 22
Figure 2.9 Accident investigation due to leaf spring
failure: (a) scene of accident, (b) damaged leaf
spring
25
Figure 2.10 V cycle fatigue design for automotive
suspension system
25
Figure 2.11 “Five Box Trick” durability model 27
Figure 2.12 Classification of signals 28
Figure 2.13 Representation of CAL in different forms: (a)
time series, (b) peak-valley reversals
29
Figure 2.14 Representation of VAL in different forms: (a)
time series, (b) peak-valley reversals
31
Figure 2.15 Strain signals at different mean values: (a)
SAEBKT with zero mean value, (b) SAESUS
with negative mean value, (c) SAETRN with
positive mean
32
Figure 2.16 Rainflow cycle counting algorithm: (a) – (f)
sequence of the method, (g) cycles derived
from Rainflow cycle counting method
35
Figure 2.17 Typical S-N curve of steel 37
Figure 2.18 A typical ɛ-N curve of steel 38
xiii
Figure 2.19 Procedures for Palmgren-Miner damage
summation
43
Figure 2.20 Vehicle ride dyanmic system 44
Figure 2.21 Example of profilometer measured road
roughness profile
47
Figure 2.22 PSD as a function of spatial frequency of
various classes of road
49
Figure 2.23 Input and output of a linear vehicle system 50
Figure 2.24 Schematic diagram of a 2 D.O.F quarter car
model
52
Figure 2.25 Group of data mining technique 57
Figure 2.26 A basic model of a single node ANN 65
Figure 2.27 An example of multilayer feed forward ANN
architecture
66
Figure 2.28 Residuals normality assessment graph: (a)
error histogram, (b) scatter plot
68
Figure 2.29 Schematic diagram of scatter band for fatigue
life using generated data
69
Figure 2.30 Schematic diagram of correlation between
prediction and experimental data
70
Figure 3.1 Part 1: Process flow of the research
methodology
73
Figure 3.2 Part 2: Process flow of the research
methodology
74
Figure 3.3 CAD of the coil spring 75
Figure 3.4 Experimental setup for strain and acceleration
signals collection
80
Figure 3.5 Highway road conditions: (a) road preview, (b)
route map
81
Figure 3.6 UKM campus road conditions: (a) road
preview, (b) route map
81
Figure 3.7 Hilly road conditions: (a) road preview, (b)
route map
81
Figure 3.8 Residential road conditions: (a) road preview,
(b) route map
82
Figure 3.9 Rural road conditions: (a) road preview, (b)
route map
82
xiv
Figure 3.10 Flowchart for statistical analysis using
Glyphworks®
83
Figure 3.11 Flowchart for spectrum analysis using
Glyphworks®
84
Figure 3.12 Flowchart for nCode Glyphwork®-based strain
life fatigue assessment
85
Figure 3.13 Procedures for ISO 2631 vibration assessment 86
Figure 3.14 Procedures for ISO 8608 road profile
generation
90
Figure 3.15 Setup of quarter car model in SimulationX®:
(a) diagram view, (b) 3D view
94
Figure 3.16 CAD of Macpherson un-sprung mass
components: (a) rim, (b) brake caliper, (c) disc
brake, (d) hub, (e) knuckle, (d) lower arm
95
Figure 3.17 Procedures for quarter car model simulation 96
Figure 3.18 Spring fatigue life prediction using nCode
DesignLife®
99
Figure 3.19 Process flow for establishment multiple linear
regression
100
Figure 3.20 Flowchart to determine single hidden layer
ANN architecture with the lowest MSE
103
Figure 3.21 Flowchart to determine two hidden layer ANN
architecture with the lowest MSE
104
Figure 3.22 Flowchart to determine three hidden layer
ANN architecture with the lowest MSE
106
Figure 3.23 ANN architecture for optimised Coffin-
Manson vibration-life predictions
108
Figure 3.24 Validation of the established multiple linear
regression
109
Figure 3.25 Validation of the ANN predictions 110
Figure 3.26 Process flow for validation of the ANN
predictions
112
Figure 4.1 Stress distribution of the coil spring under
axial loading
116
Figure 4.2 Stress distribution of the coil spring under
combination of axial and torsional loading
116
Figure 4.3 Fatigue life contour of the coil spring 117
xv
Figure 4.4 Strain time histories of spring collected from
various road conditions: (a) highway, (b)
campus, (c) hill, (d) residential, (e) rural
118
Figure 4.5 Time histories of the measured acceleration
from lower arm under various roads: (a)
highway, (b) campus, (c) hill, (d) residential,
(e) rural
120
Figure 4.6 Time histories of the measured acceleration
from top mount under various roads: (a)
highway, (b) campus, (c) hill, (d) residential,
(e) rural
121
Figure 4.7 Damage histogram of spring strain time history
under highway road using various strain life
approaches: (a) Coffin-Manson, (b) Morrow,
(c) SWT
124
Figure 4.8 Damage histogram of spring strain time history
under UKM campus road using various strain
life approaches: (a) Coffin-Manson, (b)
Morrow, (c) SWT
125
Figure 4.9 Damage histogram of spring strain time history
under hilly road using various strain life
approaches: (a) Coffin-Manson, (b) Morrow,
(c) SWT
126
Figure 4.10 Damage histogram of spring strain time history
under residential road using various strain life
approaches: (a) Coffin-Manson, (b) Morrow,
(c) SWT
127
Figure 4.11 Damage histogram of spring strain time history
under rural road using various strain life
approaches: (a) Coffin-Manson, (b) Morrow,
(c) SWT
128
Figure 4.12 PSD of collected strain and acceleration time
histories for various roads: (a) highway, (b)
campus, (c) hill, (d) residential, (e) rural
131
Figure 4.13 Correlation of spring fatigue life and ISO 2631
vertical vibration in power form
133
Figure 4.14 Correlation of spring fatigue life and ISO 2631
vertical vibrations in linear form
135
Figure 4.15 Correlation analysis for fatigue life using
various fatigue approaches: (a) Coffin-
Manson, (b) Morrow, (c) SWT
137
Figure 4.16 Classification of measured road profile
according to ISO 8608
138
xvi
Figure 4.17 ISO 8608 road profile in form of spatial
frequency for generated road classes: (a) class
A, (b) class B, (c) class C, (d) class D
140
Figure 4.18 ISO 8608 road profile in form of temporal
frequency for various road classes: (a) class A,
(b) class B, (c) class C, (d) class D
141
Figure 4.19 Simulated force time histories under road class
“A” for different spring stiffness: (a) k14, (b)
k16, (c) k18, (d) k20, (e) k22, (f) k24, (g) k26,
(h) k28, (i) k30, (j) k32
143
Figure 4.20 Simulated acceleration time histories under
road class “A” for different spring stiffness: (a)
k14, (b) k16, (c) k18, (d) k20, (e) k22, (f) k24,
(g) k26, (h) k28, (i) k30, (j) k32
146
Figure 4.21 Simulated force time histories under highway
road for different spring stiffness: (a) k14, (b)
k16, (c) k18, (d) k20, (e) k22, (f) k24, (g) k26,
(h) k28, (i) k30, (j) k32
148
Figure 4.22 Simulated acceleration time histories under
highway road for different spring stiffness: (a)
k14, (b) k16, (c) k18, (d) k20, (e) k22, (f) k24,
(g) k26, (h) k28, (i) k30, (j) k32
150
Figure 4.23 Spring design variants with different bar
diameter
152
Figure 4.24 Multiple linear regression for fatigue life
prediction using various approaches: (a)
Coffin-Manson, (b) Morrow, (c) SWT
157
Figure 4.25 Normal Probability-Probability plot of
vibration-life regression standardised residual
for various approaches: (a) Coffin-Manson, (b)
Morrow, (c) SWT
162
Figure 4.26 Error histogram of vibration-life regression
standardised residual for various approaches:
(a) Coffin- Manson, (b) Morrow, (c) SWT
163
Figure 4.27 Scatter plot of vibration-life regression
standardised residual for various approaches:
(a) Coffin- Manson, (b) Morrow, (c) SWT
164
Figure 4.28 Correlation analysis of MLR fatigue life: (a)
Coffin-Manson, (b) Morrow, (c) SWT
166
Figure 4.29 Response surface plot for ISO 2631 vertical
vibration prediction using various approaches:
(a) Coffin-Manson, (b) Morrow, (c) SWT
168
xvii
Figure 4.30 Normal Probability-Probability plot of
vibration regression standardised residual for
various approaches: (a) Coffin-Manson, (b)
Morrow, (c) SWT
171
Figure 4.31 Error histogram of vibration regression
standardised residual for various approaches:
(a) Coffin-Manson, (b) Morrow, (c) SWT
172
Figure 4.32 Scatter plot of vibration prediction regression
standardised residual for various approaches:
(a) Coffin-Manson, (b) Morrow, (c) SWT
172
Figure 4.33 MSE of trained neural network with single
hidden layer for various approaches: (a)
Coffin-Manson, (b) Morrow, (c) SWT
178
Figure 4.34 Curve fitting of trained Coffin-Manson
vibration-life ANN approach with single
hidden layer for various datasets: (a) all, (b)
training, (c) validation, (d) ANN test
179
Figure 4.35 Error histogram of single hidden layer
vibration-life ANN for various approaches: (a)
Coffin-Manson, (b) Morrow, (c) SWT
181
Figure 4.36 MSE of trained vibration-life ANN with two
hidden layers for various approaches: (a)
Coffin-Manson, (b) Morrow, (c) SWT
183
Figure 4.37 Curve fitting of trained Coffin-Manson
vibration-life ANN with two hidden layers for
various datasets: (a) all, (b) training, (c)
validation, (d) ANN test
184
Figure 4.38 Error histogram of two hidden layers
vibration-life ANN for various approaches: (a)
Coffin-Manson, (b) Morrow, (c) SWT
185
Figure 4.39 MSE of trained Coffin-Manson vibration-life
ANN first hidden layer with various number of
neurons: (a) 1, (b) 2, (c) 3, (d) 4, (e) 5, (f) 6, (g)
7, (h) 8, (i) 9, (j) 10
187
Figure 4.40 Curve fitting of trained Coffin-Manson
vibration-life ANN with three hidden layers
for various datasets: (a) all, (b) training, (c)
validation, (d) ANN test
189
Figure 4.41 Error histogram of three hidden layers
vibration-life ANN for various approaches: (a)
Coffin-Manson, (b) Morrow, (c) SWT
190
xviii
Figure 4.42 Correlation curve for Coffin-Manson target
fatigue life using various approaches: (a)
ANN, (b) MLR
194
Figure 4.43 Correlation curve for Morrow target fatigue
life using various approaches: (a) ANN, (b)
MLR
195
Figure 4.44 Correlation curve for SWT target fatigue life
using various approaches: (a) ANN, (b) MLR
196
Figure 4.45 Linear regression analysis for prediction and
target Coffin-Manson fatigue life using
various approaches
197
Figure 4.46 Linear regression analysis for prediction and
target Morrow fatigue life using various
approaches: (a) ANN, (b) MLR
198
Figure 4.47 Linear regression analysis for prediction and
target SWT fatigue life using various
approaches: (a) ANN, (b) MLR
199
Figure 4.48 Curve fitting of trained vibration prediction
ANN for various approaches: (a) Coffin-
Manson, (b) Morrow, (c) SWT
202
Figure 4.49 Error histogram of three hidden layers
vibration prediction ANN for various
approaches: (a) Coffin-Manson, (b) Morrow,
(c) SWT
203
Figure 4.50 Correlation between Coffin-Manson predicted
and target vertical vibration using various
approaches: (a) ANN, (b) MLR
206
Figure 4.51 Correlation between Morrow predicted and
target vertical vibration using various
approaches: (a) ANN, (b) MLR
207
Figure 4.52 Correlation between SWT predicted and target
vertical vibration using various approaches: (a)
ANN, (b) MLR
208
Figure 4.53 Correlation curve for three vibration-life
regressions: (a) Coffin-Manson, (b) Morrow,
(c) SWT
210
Figure 4.54 Difference between prediction and experiment
ISO 2631 vertical vibration for various
approaches
212
xix
Figure 4.55 Correlation curve of three vibration-life ANN
prediction and target fatigue life using various
approaches: (a) Coffin-Manson, (b) Morrow,
(c) SWT
214
Figure 4.56 Difference between prediction and experiment
ISO 2631 vertical vibration for various strain
life approaches
216
xx
LIST OF ABBREVIATIONS
r.m.s Root mean square
FFT Fast Fourier transform
PSD Power spectral density
VAL Variable amplitude loadings
D.O.F Degree of freedom
S-N Stress-life
b Fatigue strength exponent
c Fatigue ductility exponent
ɛ-N Strain-life
ɛ’f Fatigue ductility coefficient
σ’f Fatigue strength coefficient
E Modulus elasticity
L Spring coil diameter
d spring bar diameter
Di Accumulated fatigue damage
R Stress ratio
σ Stress
V Vehicle speed
fc Centre frequency
m Sprung mass
Aw Frequency weighted acceleration
Wk ISO 2631 vertical weighting factor
r Pearson coefficient
R2 Coefficient of determination
ISO International Standardisation Organisation
MLR Multiple linear regression
ANN Artificial neural network
MSE Mean square error
Sf Tangent sigmoid activation function
VIF Variance inflation factor
ANOVA Analysis of variance
xxi
RMSE Root mean square error
SWT Smith-Watson-Topper
a Regression coefficient
br Regression exponent
NCM_power Power regression predicted Coffin-Manson fatigue life
NMorrow_power Power regression predicted Morrow fatigue life
NSWT_power Power regression predicted SWT fatigue life
Wa ISO 2631 vertical weighted acceleration
NCM_linear Linear regression predicted Coffin-Manson fatigue life
NMorrow_linear Linear regression predicted Morrow fatigue life
NSWT_linear Linear regression predicted SWT fatigue life
k ISO 8608 proposed road waviness
QCM Quarter car model
Pn ANN prediction outcome
f Sigmoidal activation function
NCM_MLR Multiple linear regression predicted Coffin-Manson fatigue life
NMorrow_MLR Multiple linear regression predicted Morrow fatigue life
NSWT_MLR Multiple linear regression predicted SWT fatigue life
Suspension natural frequency
ISO 2631 vertical acceleration input
P-P Probability-Probability
Wij Weights for ANN
Bij Biases for ANN
ks Spring stiffness
SAE Society of Automotive Engineers
1
CHAPTER I
INTRODUCTION
1.1 OVERVIEW OF FATIGUE LIFE ASSESSMENT
Mechanical failures have caused many injuries and property losses. However, the
mechanical failures can only be minimised rather than eliminated. Large numbers of
properly designed mechanical components and structures have successfully minimised
the mechanical failure of mechanical structures, such as aircraft (Zhu, Zhang & Xia
2015), turbines (Song et al. 2017), automotive dampers (Coutinho, Landre, &
Magalhaes 2016). There is a category of mechanical failure that is caused by continuous
loads, either constant or variable, causing the loss of material ductility, cracks, and
failure of a component (Ye, Su & Han 2014). Such failure is known as fatigue failure.
At least half of the mechanical failures are due to fatigue (Dowling 2004; Ye,
Su & Han, 2014). Fatigue failure does not only cause loss of property but also leads to
catastrophic accidents that may cost lives. Hence, it is important to understand the
chronology and recent development of fatigue analysis to perform a durable product
design. The concept of fatigue failure was properly addressed in 1850 when a German
engineer, August Wöhler, first introduced a stress-life (S-N) curve using smooth
cylindrical specimens under constant loadings. The proposed S-N curve has proven as
a great step towards understanding the materials fatigue properties (Liu et al. 2016).
Another significant contribution was the invention of a force measurement tool to
measure service loads of a freight passenger vehicle. A zinc plate with lever mechanism
was designed and attached to the freight vehicle for axle deflection measurement. The
project successfully measured the deflection signal of the freight vehicle for 22,000 km
between Breslau and Berlin. This measurement has led to the future invention of strain
2
gauges. A strain gauge is an electronic device used to measure the strain value of a
component under dynamic conditions and regarded as very useful for fatigue life
analysis (Xia & Quail 2016).
During the operating conditions, most of the mechanical components are
exposed to vibration loadings in which the loadings were randomly set at varying mean
stresses (Repetto & Torrielli 2017). Such conditions require the use of a strain gauge to
measure the strain while identifying the mean stresses. W. Gerber in 1874 proposed a
parabolic method to model the boundary line for mean stress corrections in order to
measure the mean stresses. In 1910, Basquin transformed the finite life region of
Wöhler’s curve into log form. Gerber’s model was then modified by J. Goodman and
C.R. Soderberg, as reported by Zhu et al. (2016). The mean stress correction models
were normally applied together in S-N fatigue analysis. However, the S-N method
neglected the difference of stress-strain behaviour between monotonic and cyclic
loadings which deviated from the real fatigue behaviour.
Due to the shortfall, the Bauschinger effect was introduced by J. Bauschinger
which highlighted the changing material’s elastic limit after repeated cyclic loading (Hu
et al. 2016). The Bauschinger effect also mentioned the cyclic hardening and softening
underwent by the material under cyclic loading which led to the development of the
Coffin-Manson model in 1954. The Coffin-Manson model is a strain life model which
explains the fatigue crack growth in terms plastic strain. The establishment of Basquin
model, Bauschinger effects and Coffin-Manson model have further contributed to the
development of the Morrow model in 1968 and the Smith-Watson-Topper model in
1970 which considered the non-zero mean stress effects.
The earliest cycle counting algorithm for fatigue life prediction was first
introduced in 1955. Nevertheless, the most popular cycle counting method was the
rainflow cycle counting method developed by M. Matsuishi and T. Endo in 1968. This
robust method enables the reduction of complex varying spectrum into a set of simple
stress reversals. The simplified stress reversals were adopted into Palmgren-Miner
linear cumulative fatigue damage rule in the estimation of components’ fatigue life
when subjected to complex loading. Fatigue testing of a component using the repeating
3
strain time histories is time consuming as the loadings need to be continuously applied
on the specimen until failure occurs (Stanzl-Tschegg 2014). The fatigue test could be
accelerated by transforming the strain time histories into frequency domain using a
transfer function (Mrnik, Slavic & Boltezar 2016).
The importance of fatigue analysis has prompted many industries to perform
fatigue studies in their product designs. Industries which utilised fatigue analyses
include automotive (Coutinho, Landre & Magalhaes 2016), aviation (Zhu, Zhang & Xia
2015), and offshore industry (Song et al. 2017). Nevertheless, it was proven difficult to
have a general model (Ye, Su & Han 2014) due to the complexity of fatigue phenomena.
In aviation industry, the aircrafts are exposed to aerodynamics or turbulences while
offshore components are subjected to vibrations caused by ocean waves. Meanwhile,
the analysis for automotive vehicles is often subjected to variable driving conditions
such as the users and the driving environments. Dynamic driving environment may end
with failure caused by fatigue during their operations and these dynamical properties
are heavily influenced by its suspension system design.
1.2 DURABILITY OF AUTOMOTIVE SUSPENSION SYSTEMS
Durability of an automotive component is influenced by many factors namely the
manufacturing process and design. Most of the automotive mechanical components
were treated with heat treatment or shot peening process to prolong its fatigue life before
being released to the market (Fragoudakis et al. 2014). The treatment is necessary
because a passenger car is regarded as a complex dynamical system that travels on the
ground. Response of the vehicle depends mostly on the collision and contact of tyres
towards the road surfaces which may change rapidly. The dynamic interaction induces
a certain amount of vibration to the vehicle components and causes the components to
fail. Hence, the relevant manufacturing process associated with fatigue failure causing
vibrations must be considered during the design stage of automotive components
(Moon, Chu & Yoon 2011).
Sources of vibration in a vehicle originate from engine, driveline and road
excitations (Pandra 2016). Road induced vibration is the most critical cause for
4
component fatigue (Dimitris 2016). In such a case, isolation of road induced vibration
is advantageous. Thus, a system for automotive vehicles was established to isolate the
road vibrations, known as suspension system. A suspension system does not only serve
to filter road disturbances and provide ride comfort, but also reduces the probability of
fatigue failure among the supported automotive components. In general, a suspension
system consists of a set of spring damper which supports the car body and allows the
wheel to translate in vertical direction. Although the tyre is also a part of the suspension
system, it is regarded as a tear and wear component in which periodic change of the
component is required (Ganjian, Khorami & Maghsoudi 2009). Constant inspection on
the tyre draws more attentions on spring fatigue analysis because the spring is expected
to last until the end of its durability targeted warranty period for at most six years
(Huang, Huang & Ho 2017).
When the wheel travels across uneven terrain, the spring is compressed or
expanded to absorb the vibration energy. Hence, the spring will interactively react to
the road profiles. During the spring compression, the spring experiences principal stress
in various directions due to its circular geometry. The complex stress state complicates
the fatigue analysis process. In order to analyse the coil spring fatigue characteristics,
experimental testing using a fatigue testing machine was performed (Pyttel et al. 2014).
The research investigated the effects of material state and spring wire diameter of coil
springs on fatigue life. In the research, a total of 1400 springs were utilised to study the
spring fatigue parameters. The spring fatigue test is a destructive test considering the
spring to be destroyed during the experiment. Hence, the experiment was deemed as
very cost ineffective and the number of samples could be reduced with a good fatigue
simulation tool.
Fatigue simulation is a very convenient tool to predict spring fatigue life when
all the relevant information for fatigue prediction are available. The information
includes material cyclic properties, loading time histories and a fatigue model. Kamal
and Rahman (2014) conducted a fatigue prediction of automotive coil spring using
strain life approach with a SAESUS loading time history to obtain an approximate
fatigue life prediction. The analyses also provide the fatigue results of the spring under
different stress states. Sedlák et al. (2014) investigated the complex stress state of a
5
spring in relation to material phase composition and fatigue resistance. However,
realistic loading time histories are needed to obtain a good fatigue simulation result of
a spring. When the actual measurement is not available, simple life testing for
components is applied by using constant amplitude loading (CAL) or block cycle testing
test that are similar to the test conducted by most automotive component manufacturers
on their products (Giannakis, Malikoutsakis & Savaidis 2016).
Understanding the importance of time history measurement towards spring
fatigue analysis, Putra (2016) proposed a novel acceleration to strain conversion using
a simulation approach. The strain time histories of a spring could be obtained from
acceleration measurements while acceleration time histories could be converted into
strain time histories. Strain time histories were usually used to predict the fatigue life of
spring while acceleration time histories were used to obtain an indicator for vehicle ride
comfort. Pawar Prasant and Saraf (2009) performed an analysis on vehicle ride and
durability based on a new system to measure a road profile. In their research, they
pointed out that both vehicle durability and ride were mainly based on the road profile
excitations.
Road excitation does not only cause fatigue to a vehicle coil spring, but also
affects the ride quality. Rough road conditions contribute to less comfortable vehicle
ride. However, it is impossible to change the road conditions in order to improve the
vehicle ride. Hence, the suspension system is designed to compensate with the
vibrations caused by the road conditions (Mitra & Benerjee 2015). A quarter car model
was constructed to simulate the vehicle ride. Vehicle vertical acceleration amplitudes
that are generated by a series of spring stiffness and damping coefficient values were
determined in order to propose an optimised parameter. Thite (2012) proposed a method
to refine a vehicle quarter model for ride analysis. A simplified equation was developed
to estimate the damping effects towards vehicle ride.
Since road profiles are the key component to vehicle ride and durability, efforts
to reconstruct the road defects and roughness were performed using artificial neural
networks (Ngwangwa et al. 2014). Three sprung mass vertical accelerations for ride
analyses were used to train a neural network model for a road profile estimation. A set
6
of measurement data was used for the neural network model validation. On the other
hand, Paraforos, Griepentrog and Vougioukas (2016) developed a sensor frame to
acquire road and field profiles in absolute geo-referenced coordinate for agricultural
machine hitch fatigue life prediction. Based on these two recent literature studies,
different modelling methods were proposed to predict vehicle ride and durability using
a generated road profile with consideration of the vehicle suspension systems. Although
the modelling methods are different, the concept consistently aims to provide faster
solution to durability or ride analysis.
1.3 PROBLEM STATEMENTS
Nowadays, car buyers do not only demand for a durable vehicle system but also request
for excellent riding characteristics. Automotive industries have been focusing much
attention to satisfy the customers’ needs in the quest to sustain their competency.
Henceforth, a good automotive suspension design is compulsory to achieve this target.
Unfortunately, the automotive suspension design process is a time-consuming process
which entails selecting the appropriate vehicle level target, system architecture, hard
points, bushing rates, suspension load analysis, spring rates, shock absorber
characteristics, structure integrity on each component, and analysis on the vehicle
dynamics of the design results (Saurabh et al. 2016). Each of these steps involves many
human efforts, time and resources to ensure the suspension design is sufficient to
withstand the repeated road vibrations without premature failure while providing good
vehicle ride characteristics.
Vehicle vertical vibration contributes the most to vehicle ride performance
(Shao, Xu & Liu 2018). Recent literature highlighted the potential redundancy caused
by the separate suspension analyses on the durability and vertical vibration analysis of
an automotive. Upon the completion of suspension structure integrity analysis, the
effects of suspension design towards vehicle ride dynamic were analysed to ensure they
provide good ride characteristics. Therefore, Scheiblegger et al. (2017) utilised hydro
mounts in a car simulation model to obtain durability loads for ride and durability
analyses. The proposed method has a drawback as the modelling is complex and every
tuning on the suspension generates a new transfer function. These processes need to be
7
repeated until both the fatigues strength of the suspension components and ride
dynamics are balanced to achieve optimal design. This repeating process is time
consuming and produces little engineering value. After all, there is still the possibility
of unavailability or error in the simulation or measurement that may hinder analysis of
the desired results (Gu 2017). In a worst-case scenario, it may cause a delay of product
launching and affect the reputation of the company.
It is significant to have a robust and simple method to quickly perform the design
analysis and serve as a guideline to reduce unnecessary works. In addition, all the
valuable durability test data for old vehicle variants were wasted after investing lots of
workmanships and monetary costs. The problem that arises now is how to allow the
information from actual vehicle generation to be used in the next vehicle generation so
that the development process would not be repeated. Eventually, the complete solution
for this issue has not yet been found. Although the road load analysis was extensively
conducted, the ride dynamics were not considered in the integrity structure which were
closely related to suspension components. Hence, there is a need to determine the
relationship between ride comfort and durability of an automobile suspension design
for design assistances. To the best of the author’s knowledge, no solution was offered
to solve this matter which leads to the contribution of this research.
1.4 RESEARCH OBJECTIVES
The aim of the current research is to establish an automotive spring durability
relationship for vehicle ISO 2631 vertical vibration prediction with the consideration of
road surface profiles. Several specific objectives have been determined to achieve the
aim of the research. The specific objectives are as follows:
1. To determine a multiple input durability relationship that predicts the vehicle
ISO 2631-1 weighted vertical vibration using quarter car model simulations with
different spring stiffness variants.
2. To establish the fatigue life predictions using multilayer perceptron artificial
neural network through the determination of the architecture with the lowest
mean square error.
8
3. To validate the durability characteristic with ISO 2631-1 weighted vertical
vibration using actual vehicle measurement data.
1.5 SCOPE OF WORK
This study focuses on mathematical analysis of developed spring durability relationship
for automotive ISO 2631 vertical vibration. The scope is defined according to the three
objectives:
Objective 1: A computer aided software is used to model the coil spring while a
finite element (FE) software is applied to determine the stress-strain
characteristics of the coil spring. Strain signals are measured at the
point with the highest stress of the actual spring while acceleration
signals are measured at the vehicle lower arm using two different
data acquisition systems. The strain signals are used to calculate the
spring fatigue life using strain life models while acceleration signals
are used to obtain ISO 2631 vertical vibration. A multi component
kinematic quarter car model is created and simulated using road
input. Altering the spring stiffness provides a series of spring
variants associated with fatigue and ISO 2631 vertical vibrations.
The parameters are graded using power and multiple linear
regression approaches. The goodness of relationship between the
parameters and normality of the residuals are analysed using
statistical approaches.
Objective 2: To optimise the fatigue life prediction, an artificial intelligence is
used with the neurons ranging from one to ten and the hidden layers
up to three. A looping process is implemented to find the optimised
artificial intelligent in which the artificial intelligent with the lowest
mean squared error (MSE) was determined. The goodness of fitting
for artificial intelligent is analysed using statistical parameters
including the coefficient of determination and residuals. In addition,
9
the normality of the residual is tested using a suitable statistical
approach.
Objective 3: The regression relationships and artificial intelligence are validated
using strain and acceleration measurement data from an actual
vehicle running on different roads. The fatigue life prediction
relationship is subsequently validated via a conservative fatigue life
correlation method. Meanwhile, the ISO 2631 vertical vibrations are
validated using experimental acceleration data. Errors of the
outcome are analysed to ensure the accuracy of the regression as the
results contribute to the overall body of knowledge.
1.6 RESEARCH HYPOTHESES
Prior to the research works, several hypotheses related to the objectives have been
identified as follows:
Hypothesis 1: If the suspension frequencies, ISO 2631 vertical vibrations and
spring fatigue lives are related to each other, then the established
regressions possess high coefficients of determination and low
mean squared error values.
Hypothesis 2: If a suitable artificial intelligence was applied, then the performance
of predictions in terms of mean square error shall be improved.
Hypothesis 3: If a set of independent validation data fits nicely into the created
solutions, the solutions could be used to estimate the spring fatigue
behaviour with acceptable accuracy.
1.7 SIGNIFICANCE OF RESEARCH
This research contributes new understandings to several branches of mechanical
engineering such as durability, vehicle dynamics, artificial intelligence or machine
10
learning, mechanical design, and engineering mathematics. This research presents a
development of a durability relationship for automotive vertical vibrations according to
ISO 2631 standard. Firstly, the durability of the automotive spring was assessed using
different road conditions. The road profile measurement data are precious and
significant for a realistic fatigue life prediction. Without these measurements, the
fatigue assessment of springs could only be performed using constant amplitude (CA)
tests which are much deviated from the actual environment. Secondly, the relative
vehicle ISO 2631 vertical vibration was also performed. It is also important to know the
quality of the current suspension design in providing road excitation filtering.
In addition, this research bridges the gap between spring durability and vehicle
ride dynamics using an artificial intelligence modelling method. A series of neural
network architectures have been proposed throughout the modelling process. The
application of neural network to find these parameters is novel and serves as a new
contribution to knowledge. For application, the process to design a suspension spring
with good durability and vertical vibration characteristics is time consuming and
requires lots of efforts. Therefore, this prediction of durability and ride characteristic
for automotive suspension is crucial to shorten the time frame because no simulation
model is needed. The established relationship usage is quite user friendly and the results
could be interpreted immediately, hence the prediction results are always robust. The
generated prediction relationships with acceptable accuracy are the novel and main
contribution of this research.
11
CHAPTER II
LITERATURE OVERVIEW
2.1 INTRODUCTION
This chapter presents the relevant theory and research trends according to the current
research scope and topic that are related directly to the literature of automotive
suspension system. The mechanisms of different suspension systems were reviewed by
focusing on coil spring designs and characteristics which served as the main case study
component of this research. As coil spring design plays a key role in vehicle component
durability and automobile ride, the discussions concentrate on durability analysis which
is the main parameter of this research. The trends of coil spring failure from 2009 to
2012 reported by the British motoring company, the automotive association is listed in
Appendix A. The first discussion deliberates on the various types of loading signals
that contribute to automobile fatigue life, followed by cycle counting methods, fatigue
life analysis and damage cumulative rule. The suitable fatigue life models for
determining automobile suspension component analysis are also highlighted.
In the following section, ride related vertical vibrations of an automobile
involving road roughness excitations, simulation models and vibration standards were
reviewed. A contemporary design of coil spring significantly affects the automobile
vertical vibration and its fatigue life because of varying suspension mechanisms. In
order to have an insight on the direct relationship between spring fatigue life and vertical
vibration to the mechanism of automotive suspension system, the following section
focuses on the data analysis method which serve as a connecting medium to bridge those
parameters. The data analysis section studies the trends of recent methods in regression
modelling and data mining techniques which also includes neural network analysis.
12
Validation of the regression residuals such as normality test and scatter band approach
are also literally studied. Lastly, the summary of this chapter is provided in Section 2.6.
2.2 AUTOMOTIVE SUSPENSION SYSTEM
2.2.1 Types of Automotive Suspension Systems
Developments of automotive suspension system have drastically increased over the
years. Continuous developments have been conducted on all types of suspension
systems. The suspension systems could be classified into three classes which are active,
semi-active and passive. An active suspension system consists of coil spring and damper
which could dissipate or add energy to the system in a controllable way while semi-
active suspension systems allow variable damping with a fixed spring characteristic
(Pionke & Bocik 2011). Nevertheless, the most widely used suspension system is the
simplest passive suspension system due to its simplicity (Rizvi et al. 2016) whereby a
passive suspension system consists of a pair of spring and damper with predefined
characteristics. For the advance suspension system, Mercedes Benz utilised a
suspension system with continuous variable damping while the Porsche Panamera
implemented air springs for varying spring and damper characteristics (Aboud, Haris,
Yaacob 2014).
In general, passive suspension systems could be categorised into three groups
which are dependent, semi-dependent and independent. Wheel motion of a dependent
suspension system depends on the motion of its partner on the other side of the wheel
as both the right and left wheel are attached to a same solid axle. When one wheel of
the dependent suspension hits a bump, its upward movement causes a slight tilt of the
other wheel. One of the most well-known dependent suspensions is Hotchkiss
suspension system as shown in Figure 2.1 (Jazar 2014). Advantages of this dependent
passive suspension are its ability to assure constant camber that is suitable for heavy
load vehicles. For a semi-dependent suspension, rigid connections between a pair of
wheel are replaced by a compliant link. An example of this type of suspension system
is a trailing twist axle suspension. The main difference between dependent and semi-
13
dependent suspension systems is the number of compliant links which are connected to
the left and right wheels.
For an independent suspension system, wheels on the same axle could move
independently of each other. Independent suspension systems are beneficial in
packaging and provide greater design freedom compared to dependent suspension
system. The most well-known examples of independent suspension system are
Macpherson strut, double wishbone, multilink and trailing arm as shown in Figure 2.2.
A Macpherson strut consists of one strut with a combination of spring and shock
absorber and connects the wheel to the frame of a vehicle. The Macpherson suspension
system is recognised for its lightweight and compact size as well as lower cost because
it comprises lesser components (Purushotham 2013).
Figure 2.1 Solid axle dependent suspension system
Source: Jazar 2014
Double wishbone and multilink suspension systems provide greater
performance and adjustability compared to the Macpherson suspension system
(Purushotham 2013). The double wishbone suspension system usually possesses two
lateral unequal length control arms along with a coil spring and shock absorber as shown
in Figure 2.2(a). On the other hand, the multilink suspension utilises a five-link
mechanism with five connecting links as shown in Figure 2.2(c). As the number of
connecting point increases, the design parameters also increase and prompt the dynamic
and kinematic complexity of the suspension system which is actually good for handling
properties (Yarmohamadi & Berbyuk 2013).
14
However, a higher number of links may increase the cost due to the additional
components. Trailing arm is widely applied in heavy vehicle such as military trucks due
to the existence of solid axle in the vehicle (Naeem & Jagtap 2017) and usually applied
by connecting an axle to pivot point as shown in Figure 2.2(d). Selection of suitable
suspension system type is based on vehicle type, performance target and allocated costs.
Selecting an appropriate suspension system for vehicle design requires thorough
understanding on the mechanism of suspension.
(a) (b)
(c) (d)
Figure 2.2 Types of independent suspension system: (a) MacPherson strut, (b) double wishbone,
(c) multi-link, (d) trailing arm
Source: Jazar 2014
2.2.2 Mechanism of Suspension Components
The mechanism of a suspension system is used to specify the kinematics of the wheel
in vertical and lateral movements when the vehicle is travelling on a road. When the
wheels of the vehicle travel across different terrains, uneven road surfaces will cause
the wheel to move up and down, as shown in the schematic diagram in Figure 2.3. In
Figure 2.3, Z is the vehicle body response, c is the damper’s damping coefficient, and
Zg is the road surface excitation. Movements of the wheels are perpendicular to the road
15
surface. Movement of the wheel is carried on a reflection which further stretch or
compress the coil spring of the suspension system. The spring absorbs the energy and
then releases it gradually. The remaining excitation loads are then transmitted to the
vehicle body and create a body dynamic response. This mechanism is the same for most
suspension systems in filtering road disturbances.
Figure 2.3 Schematic diagram of a quarter car model under road excitations
Source: Blundell & Harty 2004
When the wheel moves, the attached coil spring moves relatively to the wheel.
The relationship between spring and wheel displacement is known as motion ratio (MR)
(Uberti et al. 2015). Bhatt (2010) suggested to maintain the MR value of a vehicle at
one to ensure the sustainability of the vehicle balance during roll for the frontal
suspension. Since the tyre was modelled as linear spring element and connected in series
to the coil spring and hence, the ride rate (kr) could be defined as follows (Cao, Rakheja
& Su 2008):
kr = kskt
(ks+ kt) (2.1)
A coil spring is also known as helical spring and compressed when pressure is applied.
The mechanism of spring compression is shown in Figure 2.4. When the spring is
compressed, the wire is subjected to a direct and torsional shear (T) as shown in Figure
2.4(b). For a coil spring, the fundamental of Hooke’s law is used as follows (Kreyszig
2011):
Zg
Time (s)
Z
Sprung mass Mass
s
Vehicle body response
c
Spring
and
damper Ground input
kr
16
𝑘 = 𝐹
𝑥 (2.2)
where k is the spring stiffness, F is the applied force, x is the spring displacement.
Hooke’s law is widely used in coil spring design and it exists for more than 300 years
(Choube 2016).
(a) (b)
Figure 2.4 Schematic diagram of spring: (a) uniaxially loaded spring, (b) free body
Source: Sequeira, Singh & Shetti 2016
A spring returns the mass to its original position to achieve equilibrium. When
sudden impact is imposed, the spring bounces and vibrates until all the energy is spent.
A suspension system built on springs alone is extremely bouncy and bad for ride.
Implementation of a shock absorber reduces the vibration of the coil spring because the
kinetic energy is turned into heat that could be dissipated through hydraulic fluid (Dixon
2007). Mechanism of a shock absorber could be divided into two, rebound and
compression as shown in Figure 2.5(a) & (b). The damping is usually resulted by
viscous effects. The viscous damping of a shock absorber is normally proportional to
velocity and can be defined as follows:
d = F
x (2.3)
where d is the damping coefficient and is the velocity. As shown in Figure 2.5(c), the
damping curve of a shock absorber consists of hysteresis effects due to the valve
17
compressing the oil on the damper (Cossalter et al. 2010). The positive damper force
indicating the compression of the damper while the negative damper force shows the
rebound condition of damper (Schramm, Hiller & Bardini 2014).
(a) (b)
(c)
Figure 2.5 Mechanism of a damper: (a) rebound, (b) compression, (c) hysteresis loop
Source: Skagerstand 2014; Schramm, Hiller & Bardini 2014
Besides spring and shock absorber, tyre is also an important part of automotive
suspension system, especially in the vehicle dynamic properties. Simulating the ride
properties of a vehicle requires knowledge of tyre properties. The required properties
for vehicle ride simulations are vertical stiffness and damping. Normally, the tyre is
modelled as a linear spring damper system as shown in Figure 2.6 where mt is the wheel
mass, kz is the tyre vertical stiffness, cz is the vertical damping coefficient of the tyre, P
is the contact point, δz is the vertical deflection of the tyre, Xsae is the longitudinal axis
and Zsae is the vertical axis according to the SAE coordinate system.
18
A linear spring and damper tyre model is sufficient to simulate most of passenger
vehicles (Koch et al. 2010). The tyre of a passenger car is quite lightly damped and the
running vehicle motions are usually dominated by wheel hop (Blundell & Harty 2011).
When a heavy vehicle is considered, the linear tyre model needs to be extended into
nonlinear model due to large deflection. Suspension tyre could be modelled using either
Kelvin-Voigt-type or Maxwell model. Kelvin-Voigt-type model consisted of a
Newtonian damper and Hookean elastic spring connected in parallel while Maxwell
model can be represented by a viscous damper and elastic spring connected in series.
Hackl et al. (2016) proposed that the Kelvin-Voigt model worked well with tyre under
fixed a fixed operating point but not for manoeuvre.
Figure 2.6 Schematic diagram of a tyre model based on a linear spring damper with SAE
coordinate system
Source: Blundell & Harty 2011
The control arm is a hinged suspension link between the chassis and the upright
suspension that carries the wheel. The end of the control arm to the chassis is attached
by a single pivot bushing that controls the position of the end in a single degree of
freedom (D.O.F) and maintains the radial distance from the inboard as shown in Figure
2.7. When the tyre is travelling across uneven terrain, the lower arm swings in a constant
radius and brings the movement of the wheel into an arc. The control arm plays an
important rule for vehicle stability and ride comfort where the movement of wheel is
controlled by the control arm (Mahmoodi-Kaleibar 2013). Meanwhile, Genetic
19
algorithm (GA) was proposed to find the optimised design of suspension lower arm in
terms of length to minimise the vehicle lateral force. The weight of the control arm has
also lead to some improvement on vehicle dynamics. Kim et al. (2014) proposed a
design of carbon fibre reinforced on the composite lower to reduce the suspension
weight while maintaining the strength.
(a) (b)
Figure 2.7 Mechanism of a lower arm: (a) initial condition, (b) when wheel is hitting a bump
Source: Skagerstand 2014
The purpose of suspension components is usually the same for most of the
passenger cars. However, different setup of suspension systems provides different
dynamic characteristics for the automobile. The most favourable suspension systems
are Macpherson and double wishbone suspension systems as listed in many vehicle
specifications (Kavitha et al. 2018). As reported by Purushotham (2013), Macpherson
suspension system was applied on high performance car, such as Porsche 911 and BMW
3-series. Meanwhile, many locally manufactured car also applied the Macpherson
suspension system as their vehicle front suspension system. Therefore, many studies
aimed to analysis the Macpherson suspension system design on vehicle dynamics
through simulations (Reddy et al. 2016; Kavitha et al. 2018).
20
2.2.3 Coil Spring Design and Characteristics
Both Macpherson and double wishbone suspension systems utilise a coil spring
as the main component to absorb impact. Coil spring designs are usually dependent on
the stiffness target for the designated suspension natural frequency. Meanwhile, the
maximum and minimum (Fmax, Fmin) forces acting on a spring during wheel travel are
also considered in which the force mean and amplitude (Fm, Fa) loadings could be
defined as follows (Valsange 2012):
Fm= (Fmax + Fmin)
2 (2.4)
Fa= (Fmax − Fmin)
2 (2.5)
The mean (τm) and amplitude (τa) shear stress could then be obtained as follows
(Valsange 2012):
τm = ks (8FmD
πd3
) (2.6)
τa = 𝑊𝑤𝑎ℎ𝑙 (8PaD
πd3
) (2.7)
where WWahl is the Wahl factor and ks is spring design constant and could be obtained
from the equation as below (Gandomi 2015):
ks = 1 + 0.5
C (2.8)
WWahl = 4C − 1
4C − 4 +
0.615
C (2.9)
where C is the spring index as below (Gandomi 2015):
21
C = L
d (2.10)
where L is the inner diameter of the spring and d is the spring wire diameter.
Stress distributions of coil spring designs are important for design perspective
where the design is validated. Analysis of stress distribution of coil spring has been
widely adopted in spring industry. The approximate spring design stressing for different
spring material is shown in Figure 2.8. In a review by Rathore & Joshi (2013), the FEA
method well fits the stress analysis of different kinds of coil springs. Khurd et al. (2016)
performed an FEA analysis to validate the stress levels of different coil spring designs.
Under certain assumptions, the stress and strain possessed a linear relationship known
as Hooke’s law. The linear relationship is maintained by the material within its elastic
limit. If the proposed conditions are satisfied, the following linear relationship of stress
and strain could be applied:
σ = Eε (2.11)
where E is the material modulus elasticity, σ is the stress and ε is the strain. Under linear
elastic condition, the spring could return to its original position after the load is removed.
A component design is not always linear and within elastic limit and gives rise
to the need to create a non-linear stress-strain relationship for plastic design application,
such as car body and components (Evin et al. 2014). In 1943, W. Ramberg and W.R.
Osgood proposed an equation to describe the nonlinear behaviour between stress and
strain. The original form of Ramberg-Osgood equation could be written as follows:
ε = σ
E + (
σ
K')
1
n'
(2.12)
where K’ is the cyclic strain coefficient and n’ is the cyclic strain hardening exponent.
The first part of the model defines the elastic relationship between stress and strain
while the second part defines the plasticity of material. The incremental reversals with
22
respect to a reference turning point are calculated based on Masing’s model as follows
(Lee, Barkey & Kang 2012):
∆𝜀 =∆𝜎
𝐸 + 2 (
∆𝜎
2𝐾′)
𝑛′
(2.13)
Most of the automotive components failed after damage accumulation in plastic limit
due to strain hardening effects (Upadyaya & Sridhara 2012).
Figure 2.8 Chronology of spring materials development
Source: Prawoto et al. 2008
The transition of elastic to plastic properties is known as yield. In case of simple
stress state analysis, the principle stress theory was utilised for component design
especially for brittle material. However, von Mises yield criterion is always applied
when dealing with complex stresses for ductile material (Aygül, Al-Emrani &
Urushadze 2012). The von Mises yield criterion could be written as follows (Juvinall &
Marshek 2017):
σi = √3𝐽2
=√(σ11 - σ22)2 + [σ22 - σ33]2 + [σ33 - σ11]2 + 6(σ12
2 + σ232 + σ31
2)
2
(2.14)
23
where J2 is the Cauchy stress tensor component which was named after Augustin-Louis
Cauchy in 1827, as reported by Barulich, Godoy and Dardati (2017) as below:
σ = [
σ11 σ12 σ13
σ21 σ22 σ23
σ31 σ32 σ33
] = [
σx τxy τxz
τyx σy τyz
τzx τzy σz
] (2.15)
The von Mises stress criterion is still widely used in automotive applications such as
analysis of twin helical springs (Kaoua et al. 2011) and normal automotive coil springs
(Lavanya, Rao & Pramod Reddy 2014). This method is still popular because many
experiments have been conducted to prove its accuracy for stress estimation (Coppola,
Cortese & Folgarait 2009).
When the stress of a component exceeds the ultimate tensile strength of the
material, the component will fail instantly. This type of failure is known as static failure.
Nevertheless, literature on the failure of automotive springs (Zhu, Wang & Huang 2014;
Vukelic & Brcic 2016; Shevale & Khaire 2016) claimed that the failure of automotive
coil springs was never due to static issues. Their investigations showed that the failure
of a coil spring was due to repeated cyclic loading as a coil spring is always able to
withstand a single applied load but the loadings toward automotive suspension systems
were repeatedly. Hence, durability analysis was usually performed to estimate the
fatigue life of the design in order to prevent the coil spring from failing before the
desired lifespan.
2.3 DURABILITY ANALYSIS
Durability analysis is a crucial element in automotive industry where fatigue failure
could cause catastrophic losses. In case of ground vehicle, fatigue failure of suspension
components could lead to a road accident. As published by the Institute for Traffic
Accident Research and Data Analysis (ITARDA) Japan in 2005, one of the most
common road accident cases was due to fatigue failure of a spring. Usually when the
driver saw a car coming from the opposite direction, the driver would shift to the left to
let the car pass. However, the truck may tilt to the left and deprive the driver of control,
and eventually rolled over on the passenger side. The scene of accident is shown in
24
Figure 2.9(a). The accident was caused by fatigue and break down of the leaf spring.
The broken leaf spring is shown in Figure 2.9(b). Such accident may cause monetary
loss due to the damage of the vehicle and goods, as well as injuries of the driver. Hence,
study of fatigue of automotive suspension component is compulsory to prevent potential
failure and promote safety.
2.3.1 Introduction to Durability Analysis
Acknowledging the significance of fatigue assessment, automotive industries and
engineering researchers put lots of efforts into fatigue area to improve the component
designs. V cycle was implemented on automotive suspension component fatigue design
as shown in Figure 2.10. The V cycle aims to convert the customer idea into vehicle
design requirements, before breaking down the requirements into subsystems and
component targets. The design of the suspension components was then validated using
experimental results. Once a good understanding on fatigue requirement was obtained,
the component design could be optimised to reduce unnecessary material and
maintenance costs (He et al. 2010).
Computer aided engineering (CAE) durability analysis is usually performed
using a commercial software to predict the fatigue life of a component based on
geometry, material properties and service loads as input. These processes are known as
“Five Box Trick” durability model as shown in Figure 2.11 (Karthik, Chaitanya &
Sasanka 2012). The geometry is used to determine the critical region of stress
concentration and fatigue life distribution. Besides, the correct boundary conditions
reflect the actual specimen setup for experimental fatigue testing and provide
meaningful analysis. Material properties and service loads are used in conjunction with
geometry for fatigue analysis. Selection of an appropriate fatigue analysis algorithm
leads to a more accurate result.
Throughout the years, many durability models have been applied in automobile
suspension fatigue design. A chronology of published studies on the durability analysis
of automobile suspension components in high impact journals is illustrated in Table 2.1
to facilitate better understanding on the trends of the analysis. Based on this chronology,
25
numerous researches are found to use virtual simulation to predict fatigue life of
automobile suspension components. Recently, a few simulations were assisted with
metaheuristic or evolutionary algorithm to perform fatigue design optimisation of
automotive suspension components (Yildiz & Lekesiz 2017; Huang, Luo & Yi 2013).
The objective was to reduce component weight while maintaining the fatigue
characteristic.
(a)
(b)
Figure 2.9 Accident investigation due to leaf spring fatigue: (a) scene of accident, (b) damaged
leaf spring
Source: http://www.itarda.or.jp/itardainfomation/english/info59/next06.html
Figure 2.10 V cycle fatigue design for automotive suspension system
Source: Thomas, Bignonnet & Perroud 2005
26
Table 2.1 Trends of published automobile suspension component durability analysis
Author Year Contribution Component
Luo, Huang
& Zhou 2018
Applied Multi-Gaussian fitting and long short-
term memory neural network for automotive
suspension durability analysis Torsion beam
Yildiz &
Lekesiz 2017
Performed fatigue optimisation of an
automobile lower arm using hybrid charged
system search algorithm Lower arm
Kulkarni,
Ranjha &
Kapoor
2016
Assessed the durability and life cycles for a
damper of an electric vehicle using measured
strain signals Damper
Darban,
Nosrati &
Djavanroodi
2015
Applied multiaxial fatigue algorithm to predict
fatigue life of a stranded wire coil spring Coil spring
Nam, Shin &
Choi 2014
Evaluated the fatigue life of vehicle subframe
using nonlinear suspension model Vehicle subframe
Huang, Luo
& Yi 2013
Evaluated fatigue life of an automobile torsion
beam suspension using radial basis function
neural network simulated loading time histories Torsion beam
suspension
Klemenc,
Janezic &
Fajdiga
2012
Predicted strain-life curves of automotive
spring steel with combination of material
influential parameter, such as testing
temperature and specimen diameter using
hybrid neural network
Coil spring
Zonfrillo &
Rossi 2011
Predicted strain-life curves of spring steel with
combination of material influential parameter
like testing temperature and specimen diameter
using fuzzy approach Coil spring
Knapcyzk &
Maniowski 2010
Optimising the durability of suspension
bushing with a dynamic criterion using multi-
criteria goal function
Bushings of 5 rod
suspension
Bae et al. 2005
Investigated the shot peening effects on
automobile leaf spring’s fatigue strength Leaf spring
Nakano et al. 2000
Developed high strength spring materials with
consideration of corrosion effects Coil spring
Qian &
Fatemi 1995
Investigated the durability effects of ion-
nitriding process of spring steels using strain
block loadings
Spring steel
Takechi 1990
Investigated the effects of non-metallic
inclusion of automotive steels in affecting their
fatigue life
Spring steel
Tanaka et al. 1985
Predicted the fatigue life of spring steel under
fretting condition Spring steel
27
Figure 2.11 Durability model-based “Five Box Trick”
Source: Karthik et al. 2012
Meanwhile, integrations of artificial intelligence into fatigue analysis are also
observed. Those works mainly focused on applying data mining techniques on fatigue
data, such as artificial neural network (ANN) or fuzzy approach for fatigue life
prediction purpose. Apart from that, methods to determine road loadings for fatigue life
prediction are also of huge interest. Those researches aim to construct reality road
loading profiles to simulate fatigue life of automotive suspension components.
2.3.2 Types of Loading Signals
Signal is a function of independent variables that carry information in the form
of service load. It is important to understand the details of every type of signal,
especially in durability and automotive ride analysis. In general, signals are divided into
two categories known as analog and digital. A digital signal is a discrete time signal
generated by digital modulation while an analog signal is a continuous signal which
represents physical measurements (Gazi 2017). As proposed by Lee et al. (2004), the
strain signal measured using a strain gauge is in analog form and processed into digital
signal using analog-to-digital (A-D) converter for computer reading purpose. The
correct recording method of the loading signal leads to accurate analysis of durability
and ride.
In order to study the dynamic properties of a vehicle, the characteristics of the
measured signal are first to be statistically determined. Fatigue behaviour of an
automotive component depends on the strain signal behaviour (Zakaria et al. 2014). In
the first step, the signal is classified into a few classifications. The classifications are
Material properties Geometry Service loads
Durability model
Fatigue life
28
displayed in an organisation chart as shown in Figure 2.12. A deterministic signal
possesses a complete specific function of time and does not vary over time which can
be completely represented by a mathematical equation, e.g.
y = Asin(2πfot+ θ) (2.16)
where A is the amplitude, fo is the frequency and θ is the phase angle. Deterministic
signals can be further classified into periodic and non-periodic. Periodic signals
complete a pattern within a measurable time frame while non-periodic signals do not
repeat the pattern over time.
Figure 2.12 Classification of signals
Source: Norton & Karczub 2007
A simple periodic signal contains only a single frequency while a complex
periodic signal contains multiple frequencies. A nonperiodic signal can also be
interpreted as quasi-periodic signal and can be defined as a periodic signal that has
infinite lengthy period. A transient signal is a signal which lasts within a limited period.
For fatigue cyclic loading, there are two major categories of signal type, known as
constant amplitude loading (CAL) and variable amplitude loading (VAL). CAL is also
a type of periodic loading signal. Two example forms of CAL are shown in Figure 2.13.
29
The CALs repeat under the same pattern over time. In automotive industry,
manufacturers rely on the CALs to predict components’ fatigue life (Karthik, Chaitanya
& Sasanka 2012) as the equivalent CALs are easier to implement for component testing
in lab.
(a) (b)
Figure 2.13 Representation of CAL in different forms: (a) time series, (b) peak-valley reversals
Source: de Jesus & da Silva 2010
Non-deterministic signals are classified into two types which are stationary and
non-stationary. Non-deterministic signals represent random physical phenomena that
could not be described by means of mathematical rules. Variable amplitude loading
(VAL) is a type of non-deterministic loading histories which is widely applied in
automotive industries. Non-deterministic loading signals could be stationary or non-
stationary. Many strain time histories for fatigue analysis exhibit non-stationary
behaviour (Capponi et al. 2017). Karlsson (2007) applied the stationary road load data
to generate cyclic load for automobile fatigue life prediction. Meanwhile, Chaari et al.
(2013) applied nonstationary loadings to study dynamic response of an automobile
planetary gear. In the actual automobile operating conditions, most of the loadings are
non-stationary because road surfaces always consist of potholes or bumps (Lennie et al.
2010).
For cyclic loading, it is important to know the amplitude and mean stress of the
signal. In terms of CAL cyclic loading, the stress amplitude is expressed as follows:
σ = Δσ
2 (2.17)
30
and the stress range, Δσ is as follows:
Δσ = σmax − σmin (2.18)
where 𝜎𝑚𝑎𝑥 is the maximum stress and 𝜎𝑚𝑖𝑛 is the minimum stress. The mean stress,
𝜎𝑚𝑒𝑎𝑛 is defined as follows:
σmean= σmax − σmin
2 (2.19)
For the CAL fatigue application, alternating loads with mean stress are usually applied
on the component. Stress ratio (R) is used to represent the mean stress where it could be
defined as the ratio of minimum stress to maximum stress in one cycle of loading in a
fatigue test as follows:
R = σmin
σmax
(2.20)
Different stress ratios were often used to obtain fatigue life of high strength steel for S-
N curve construction (Gao 2017).
For CAL, it is easier to determine the stress ratio for fatigue analysis while
observing other parameters toward fatigue. Kovaci et al. (2016a) determined the effects
of two parameters which were plasma nitriding time and temperature on fatigue crack
growth of nitride steels under CAL. They have also pointed out that the VAL was less
significant for their studies since the applied loading was an independent variable.
However, most of the loading signals in real life have involved stress amplitudes that
change in irregular manner and exhibit variable characteristics. Fatigue life of a
component under VAL is totally different from CAL.
VAL is a cyclic load oscillating between the variable maximum and minimum
amplitudes that consist of no specific pattern as shown in Figure 2.14. Mottitschka et al.
(2010) conducted an experimental investigation to study the effects of fatigue life using
CAL and VAL. The results revealed that the fatigue life of a component is lower under
31
the VAL when compared to CAL at equivalent stress range. For VAL, a low-high-low
block loading with different R was applied to the specimen. For detailed VAL and load
interaction effects on fatigue crack, an experiment to determine the fatigue crack growth
under overload (OL), underload (UL), OL + UL and UL + OL was performed. When
an OL was applied, the total fatigue crack life was increased which is known as crack
retardation. When an UL was applied, the total life was decreased, and crack
acceleration occurred (Dore & Maddox 2013). The application of VAL for crack
initiation and fatigue crack growth is quite different because the higher tensile stresses
initiate cracks while it prevents fatigue crack growth in FCG. Load sequence effects of
low-high-low block loading have generated a “memory” effect on the material (Lu &
Liu 2011). If the loading sequence is different, the “memory” effect will be different.
These effects cause the lower amplitude cycles to be more damaging when they were
applied after high cycle amplitude.
(a) (b)
Figure 2.14 Representation of VAL in different forms: (a) time series, (b) peak-valley reversals
Source: de Jesus & da Silve 2010
In the engineering fields, VALs are considered as part of the design requirement.
All the steps of the design procedure should be consistent with each other in terms of
damage. For efficiency, robustness and capitalisation purpose, VAL could be converted
into a damage equivalent loading because VAL analysis and experiments are time
consuming and expensive (Thomas, Bignonnet & Perroud 2005). SAE fatigue design
and evaluation (FD&E) committee has conducted a test to provide a set of basic data to
determine the validity and fatigue life of various automobile component. Three VAL
load histories were obtained from vehicle suspension component, bracket, and
transmission system, respectively, as shown in Figure 2.15. The suspension VAL
32
(SAESUS) was obtained from the bending moment of a suspension system component
driven over an accelerated durability course by Wetzel in 1977, as reported by Oh
(2001). The SAE bracket loading history was obtained from a vibration with nearly
constant mean load over a rough road while the transmission was measured from the
transmission torque measured on a tractor engaged in front end loader where drastic
change of mean has been observed.
VAL was widely used in automotive industry for fatigue characteristic studies
of automotive suspension components (Facchinetti 2018; Anes et al. 2017; Kulkarni,
Ranjha & Kapoor 2016). Loading time histories applied in automotive industry exist in
different formats such as stress, strain or force. Literature trend of VAL in automotive
suspension component fatigue design over 30 years is tabulated into Table 2.2 to
highlight the usage of VAL in automotive suspension systems. Based on the trend, a
few researches focused on converting the VAL into CAL for automotive component
fatigue analysis. Some of the researches tried to use the outdoor-measured VAL in
indoor testing using equivalent spectrum. Different forms of loading time histories were
used in the study. This shows that suitable VAL is very significant in automotive
suspension durability analysis.
(a)
(b)
(c)
Figure 2.15 Strain signals at different mean values: (a) SAEBKT with zero mean value,
(b) SAESUS with negative mean value, (c) SAETRN with positive mean value
Source: C.S. Oh 2001
33
Table 2.2 Trend of types of VAL in automotive suspension applications
Authors Year Contribution Types of loading
Facchinetti 2018 Comparing the VAL and CAL loading
spectra for automotive chassis component Force VAL
Anes et al. 2017
Analysed the fatigue life of automotive
suspension steel and transformed the VAL
into CAL using stress scale factor Stress CAL & VAL
Kulkarni,
Ranjha &
Kapoor 2016 Predicted the fatigue life of a damper in
electric vehicle using strain life approach Stress VAL
Matteti, Molari
& Mertua 2015
Converted the acceleration VAL of a
wheel hub into displacement VAL and
applied on a four-poster test rig.
Force & Displacement
VAL
Zhao et al. 2014
Assessed fatigue life of rear axle
suspension with consideration of loading
below fatigue limit
Strain & Force VAL
Kowfie &
Rakhbar 2013 Derived a new damage summation model
based on the applied cyclic stress Strain VAL
Karthik,
Chaitanya &
Sasanka 2012
Predicted the fatigue life of a leaf spring
using SAE strain signals Strain VAL
Saoudi,
Bouzara &
Marceau 2011
Predicted the fatigue life of lower arm
using a quarter car model simulation Force VAL
Grujicic et al. 2010
Integrated first objective reliability method
and strain life approach to optimise fatigue
life of an automobile lower arm Force VAL
Ledesma et al. 2005
Developed a accelerated durability test rig
for commercial vehicle suspension
systems Force VAL
Fu & Cebon 2000
Proposed a new model for calculating the
probability density function of stress range
for bi-modal spectral densities using a case
study of trailing arm suspension Strain VAL
Yang et al. 1995
Designed a new force transducer and
correlated with axle strain measurements
for chassis applications Force VAL
Ramamurti &
Sujatha 1990
Applied random vibration concept to a
finite element model of bus chassis using
correlation of left and right tracks Strain and Acceleration
VAL
Dabell 1985
Discussed on the importance of service
loads on automotive suspension prototype
development Force & Strain VAL
34
2.3.3 Cycle Counting Methods
VAL time history varies according to time and behaves randomly. Hence, it is not easy
to define the cycle counts of a long VAL signal. A few cycle counting techniques were
proposed to summarise the irregular-load-versus time histories by providing the number
of time cycles across numerous sizes occurrences. The cycle counting method aims to
obtain equivalent constant amplitude cycles and consists of level crossing counting,
peak counting, simple range counting, range-pair counting, and Rainflow counting.
However, each method defines the cycle differently. Cycle counts could be applied in
all kinds of loading time history including force, stress, strain, acceleration, and torque
(ATSM E1049-85 2005).
Single parameter algorithms analyse a single parameter such as difference
between points to regenerate a simple cycle. Examples of single parameter cycle
counting methods are level crossing counting method, peak cycle counting, and simple
range cycle counting method. Range pair and Rainflow cycle counting methods are
classified as two parameter methods (ASTM E1049-85 2005). Among these methods,
Rainflow cycle counting method, developed by M. Matsuishi and T. Endo in 1968,
ranks as the most popular method. Downing and Socie developed the Rainflow
technique into vector-based using three-point criteria in 1982 (Marsh et al. 2016). The
procedures to perform Rainflow cycle counting are illustrated in Figure 2.16 in which
the loads were initially rearranged from the maximum or minimum peak, whichever is
greater in absolute magnitude.
Rainflow cycle counting provided the most acceptable results compared to other
methods (Roshanfar & Salimi 2015). This method could determine the cycles in a
complex loading sequence. Rainflow cycle counting was popular because it required
less storage and could be used for real-time monitoring. Besides, the accuracy of
Rainflow cycle counting has been verified through a wide range of industrial
applications (Marsh et al. 2016). The small cycles are individually extracted as shown
in Figure 2.16 (c) & (d). The large cycles are extracted at the end of the process. When
the cycles are extracted from the load time history, the mean value of each cycle are
also determined. However, Anes et al. (2014) pointed out that Rainflow cycle counting
35
technique is only acceptable for uniaxial loading but may yield a poor result under
multiaxial application. Nevertheless, Rainflow cycle counting technique is still widely
applied in automotive applications such as steering (Ligaj 2011) and suspension upper
arm due to its simplicity (Mrad et al. 2016). After extracting the cycles, this information
is used as input to the fatigue models to calculate the fatigue life of the components.
(a) (b)
(c) (d)
(e) (f)
Range Mean Counts Events
8 1 1 C – D
6 1 1 H – I
4 1 1 E – F
3 -0.5 1 A – B (g)
Figure 2.16 Rainflow cycle counting algorithm: (a) – (f) sequence of the method, (g) cycles
derived from Rainflow cycle counting method
Source: ASTM E1049-85 (2005)
36
2.3.4 Fatigue Life Analysis
Fatigue life could be defined as the number of stress cycles of a specified character that
a specimen sustains before failure of a specified nature occurs, as recited by Chetan,
Khushbu & Nauman (2012). Fatigue life assessment methods are generally divided into
two approaches which are high cycle fatigue (HCF) and low cycle fatigue (LCF).
Fatigue life occurring at above 1×104 cycles is usually called as HCF. LCF usually
occurs at condition below 1×104 cycles. For LCF, the total fatigue life of a component
includes the crack properties such as crack initiation and propagation. When the fatigue
life occurred beyond 1 × 107 cycles, it is defined as very high cycle fatigue (VHCF)
regime (Korhonen et al. 2017). Some components like turbine blades, rotating
components are designed using HCF because the components are expected to sustain
high numbers of cyclic load or high frequency vibrations (Tafti 2013).
In order to prevent failure due to fatigue, periodic maintenance of a structure
needs to be performed constantly (Pantazopoulus et al. 2014). This preventive method
is suitable for costly products such as aircraft or crane where the replacement cost of
these components is extremely expensive. Meanwhile, it is not feasible to perform
regular checking on small mass-produced components such as automobile engines,
steering part, and suspension components because lots of efforts are required. Hence,
an easier way to avoid fatigue failure is through replacing the cracked components after
the designated service period (Jiang & Murthy 2011). If the components are designated
to be replaced, the components are considered as fail when a crack is initiated. However,
a small crack opening of spring steel is harmless. Therefore, Takahashi et al. (2011)
investigated on crack length effect on fatigue failure of spring steel SUP 9A. They
proposed that any crack length below 0.2 mm could be rendered as harmless. In their
paper, defects smaller than 0.2 mm were non-damaging to the fatigue limit after
extensive number of experimental testing for different crack size.
Crack initiation is caused by stress concentration at a point and a fatigue feature
determined by means of the stress life (S-N) approach was used. The concept of S-N
was introduced by A. Wӧhler and still widely applied by current researchers (Sonsino,
2007). A typical S-N curve for steel is shown in Figure 2.17. Main critical parameters
37
for S-N approach are mean and amplitude stress. An earlier empirical model by W.
Gerber in 1874 was proposed to include mean stress effect on high cycle regime. The
shortfall of this proposed model is its ignorance of tension and compression. J.
Goodman modified the model into another more conservative form in 1914. C.R.
Soderberg in 1930 reformulated the Goodman model using yield strength, 𝜎𝑢𝑡𝑠 .
Soderberg model is very conservative and expects no fatigue failure nor yield to occur
(dos Santos, Auricchio & Conti 2012). As for Gerber, Goodman and Soderberg model,
the loading condition is only applicable to tensile loadings.
Figure 2.17 Typical S-N curve of steel
Source: ASTM E1049-85 2005
After the mean stress correction, the stress was used to predict the fatigue life of
the component. For Wӧhler’s S-N approach, the life prediction model is assumed to
follow a power relationship:
σ = σf'(2Nf))b (2.21)
where 𝜎𝑓′ is the fatigue strength coefficient while b is the fatigue strength exponent.
Goodman model was more conservative when loading sequence was tensile
predominant (Karthik, Chaitanya & Sasanka 2012). When it is a zero-mean loading, the
Gerber model is more conservative. Although the mean stress effect has been described,
the S-N method still has a drawback in terms of amplitude and mean stress at different
levels. Nevertheless, S-N approach is suitable for components which require higher
safety factor.
38
Mean stress effect is not capable to be estimated accurately with using solely
ultimate tensile strength. This has been proven by Dowling (2004) through the testing
on steel and aluminium specimens using S-N approaches. Using the ultimate tensile
strength of material to estimate the fatigue life of a component is far from sufficient.
However, S-N approach has a limited capability where the accuracy is good for
components with only high fatigue life, at approximately 105 life cycle (Ghafoori et al.
2015). This approach emphasises on the nominal stress of the specimen and it does not
consider the materials behaviour under time-variable loading, such as cyclic hardening
or softening in the middle cycle fatigue. Hence, a model to predict the localised fatigue
life with the consideration of cyclic hardening effects is in need. The approach which
considers the cyclic strain hardening effects is strain life (ε-N) method.
A ε-N approach is much more accurate to assess the fatigue life of a component
compared to traditional S-N approach. This is because ε-N approach considers plastic
deformation that exists in localised region where fatigue cracks begin (Ince & Glinka
2011). The example ε-N curve for steel is shown in Figure 2.18 where the ε-N model is
labelled as “Total” in the curve and derived from the plastic and elastic data. ε-N method
is applicable to ductile materials with low cycle fatigue of components within 103 cycles.
This method permits a more rational and handling of mean stress effects through
employing a local mean stress rather than nominal stress. There is a commonality
between S-N and ε-N approaches as both methods do not include crack growth.
Figure 2.18 A typical ε-N curve of steel
Source: Williams, Lee & Rilly 2003
39
ε-N approaches have the possibility to replace S-N approach due to its greater
accuracy (Fajdiga & Sraml 2009). In applying the ε-N approach, the S-N curves are
obtained from fatigue test under complete reversed cyclic loading between constant
strain limits using strain-controlled fatigue test standard which is ASTM E606. It is
significant to note that the there is a force-controlled fatigue test standard ASTM E46.
The force-controlled fatigue test serves to obtain the fatigue strength of metallic
material in the regime where the elastic strain is predominant throughout the test. For
the strain control, cyclic total strain is measured where the plastic strain is also known.
The number of cycles to failure is counted until there is a substantial crack on the
specimen. Cyclic strain-controlled fatigue test is recommended because the material
with the stress concentration of a component is subjected to cyclic plastic deformation
even the component behaves elastically during cyclic loading.
For the elastic linear relation of ε-N approach, the model is based on a
relationship from O.H. Basquin in 1910, as mentioned by Lee et al. (2005). In 1954,
S.S. Manson expressed a power relationship for low cycle regimes in terms of the plastic
strain range as follows:
εp = εf'(2Nf)
c (2.22)
where 𝜀𝑝 is the plastic strain, 𝜀𝑓′ is the fatigue ductility coefficient, c is the fatigue
ductility exponent.
After the plastic strain range relationship was proposed, the elastic and plastic
strains were combined to obtain a correlation between the fatigue life and strain known
as Coffin-Manson relationship. The Coffin-Manson relationship is the first ε-N
approach which could be written as follows:
εp = σf
'
E(2Nf)
b + εf
'(2Nf)c (2.23)
The original relationship performs the fatigue life assessment with an assumption of
zero-mean stress. Again, the mean stress effect is crucial for fatigue life estimation
40
whether using S-N or ε-N approach because the mean stress is part of the fatigue failure
criteria (Dowling 2004; Ghaffori et al. 2015).
The Coffin-Manson relationship was assessed and compared with other fatigue
models. Most of the comparisons show that this relationship could provide acceptable
results (Runciman et al. 2011). Due to the simplicity of this relationship, its application
has been expanded to ductile cast iron (Ricotta 2015). Gribbin et al. (2016) revealed
that when the plastic strain is small, Coffin-Manson relationship cannot fit the data well.
Hence, a bilinear model was used to improve the data fitting process. Thomas (2012)
proposed that this relationship provides the best means for predicting the high strength
steel for automotive chassis and suspension. In real life application, the components are
loaded with different mean stresses where the limitation of Coffin-Manson relationship
exhibits. Nevertheless, further development of ε-N models has led to a solution for this
drawback. In fatigue analysis, the tensile or positive mean stress tends to reduce fatigue
life while compressive or negative mean stress tends to increase fatigue life (Chiou &
Yang 2012; Taheri, Vincent & Le-roux 2013; Khan et al. 2014; Bruchhausen et al.
2015). Therefore, some automotive components were induced with compression tensile
stress to enhance fatigue life. An instance is the shot peening process that is often
applied on automotive component such as gears to prolong fatigue life of the component
(Zhang et al. 2016).
There is a model that could be used for applications involving mean stress effects
in 1968, as reported by Manson and Halford (2006), known as Morrow strain-life model.
This model stated that the ε-N curve could be adjusted using added mean stress effects.
J.D. Morrow added the mean stress, 𝜎𝑚𝑒𝑎𝑛 into the Coffin-Manson relationship and
could be rewritten as follows:
ε = σf
' − σmean
E(2Nf)
b + εf
'(2Nf)c (2.24)
The mean stress, 𝜎𝑚𝑒𝑎𝑛 is positive when the loading is tension and negative when the
loading is compression. Nevertheless, Coffin-Manson relationship correlated better
than Morrow model (Ertas & Sonmez 2008) in terms of spot weld joint fatigue analysis.
41
K.N. Smith, P. Watson and T.H. Topper proposed another mean stress corrected
ε-N model in 1970. This model is known as Smith-Watson-Topper (SWT) model. This
model is widely used as reported by Ince and Glinka (2011). The parameters considered
for fatigue prediction are the maximum stress and strain amplitude per cycles. These
SWT parameters were obtained from ε-N fatigue test data observed at various mean
stresses and stress or strain amplitudes. The mathematical model of SWT could be
written as follows:
σmaxε = (σf
')2
E(2Nf)
2b + σf
'εf
'(2Nf)
b + c (2.25)
where σmax is the maximum stress. SWT model has proven to be more robust compared
to Morrow model where the fatigue life prediction is higher and more conservative after
a wide range of materials were tested using this approach (Ince & Glinka 2011). Cui
(2002) suggested that the Morrow model is only acceptable for steel while SWT model
is better for general use.
For fatigue analysis, there is a group based on continuum damage mechanics in
which fatigue life was predicted through computing a damage parameter cycle by cycle
(Santechchia et al. 2016). In terms of fatigue crack analysis, the fatigue crack length of
specimen was usually measured. The recent interest of fatigue crack analysis includes
adding parameter for simulation. For example, Meneghetti et al. (2016) added mean
stress effects as additional parameter to heat energy-based approach for fatigue crack
simulation of steel specimens.
Controversy appears during the selection of a suitable model for automotive
component fatigue life prediction, especially when the component is made of steel.
Abdullah (2005) suggested that the SWT was more conservative for tensile loading
sequence while Morrow yields greater correlation during compression loading when
compared to SWT predicted fatigue life of a steel automotive lower arm. Ince & Glinka
(2011) mentioned that the SWT model gives better results and more flexibility for
fatigue life prediction for most of the applications. Although the Morrow and SWT
models consider mean stress effects, both of them are still applicable for zero mean
42
stress fatigue analysis (Karthik, Chaitanya & Sasanka 2012). When the mean stress is
zero, the Morrow model is the same as Coffin-Manson relationship. Each approach
gives certain interest range or mean strain amplitude to the fatigue damage. Then, the
total fatigue damage is required for the final fatigue life analysis. Hence, a method to
sum up the fatigue damage is required and proposed. The most acceptable damage
summation method is known as the Palmgren-Miner linear damage cumulative rule.
2.3.5 Linear Damage Cumulative Rule
Cumulative fatigue damage theory is obtained through summation of normal and shear
energy of peak valley in each block loading. A. Palmgren introduced the concept of
fatigue damage summation in 1924 while M.A. Miner proposed a linear cumulative
fatigue damage criterion in 1954 (Petaś, Mróz & Doliński 2013). This rule is known as
the Palmgren-Miner linear cumulative damage rule. The procedures for Palmgren-
Miner damage summation are shown in Figure 2.19. The linear damage summation rule
proposed by Palmgren-Miner and the damage of one cycle Di is expressed as:
1i
fi
DN
= (2.26)
where Nfi is the number of constant amplitude cycles to failure. A linear cumulative
fatigue damage calculation under VAL is defined in the equation:
1ii
fi
nD
n = = (2.27)
where Di is the cumulative fatigue damage, ni is the number of applied cycles, and nfi is
the number of constant amplitude cycles to failure. The failure occurs when the
summation of individual damage value caused by each cycle reaches the value of one
(Upadhyaya & Sridhara 2012).
Different range of damage summation indicates varying confidence probability.
The critical damage summation values were 0.08, 0.3 and 1 which show confidence
probability of 90, 50 and 10 %, respectively in 2002 (Beretta & Regazzi 2016).
43
Selection of damage summation value depends on the desired design target. However,
the failure occurs mostly at the damage summation of one (Fiedler & Vormwald 2016).
This proposed Palmgen-Miner rule consists of a few limitations. Under spectrum
loading, component failed during the portion of the duty cycle when the load is below
the endurance limit. Secondly, the Palmgren-Miner rule also underestimates the damage
because the damage of lower load is neglected. Thirdly, the Palmgren-Miner rule
neglected the load sequence effects. Experimental results showed that the damage
summation at failure are very different when the load is from high to low or low to high
(Shamsaei et al. 2010).
Therefore, many damage accumulative theorems have been proposed since 1945
as reported by Gao et al. (2014) and could be generally classified into six types including
linear damage summation, nonlinear damage curve, life curve modification to account
load sequence effects, crack growth concept based, continuum damage mechanics and
energy-based methods. Despite the shortcomings of Palmgren-Miner linear damage
summation rule, it is still dominantly used in design due to its simplicity and acceptable
accuracy (Sun, Dui & Fan 2014). Fatigue analysis using VAL and Palmgren-Miner rule
is usually consumed lots of time and many experimental works. Hence, accelerated
fatigue test or fatigue assessment in frequency domain has been introduced to shorten
the time domain fatigue analysis. The frequency domain fatigue analysis aims to
provide similar prediction results with the time domain.
Figure 2.19 Procedures for Palmgren-Miner damage summation
Source: Woo 2017
44
2.4 RIDE RELATED VIBRATION ANALYSIS
When an automobile travels at a certain speed, it experiences a broad spectrum of
vibrations. These vibrations are transmitted to the passengers by tactile or visual
excitations, commonly known as the term of “ride”. The spectrum of vibrations is
divided according to frequency, with frequency ranging from 0 to 25 Hz is classified as
ride (Rafael 2007). On the other hand, frequency ranged between 25 to 20,000 Hz is
classified as noise. The maximum of 25 Hz is the frequency limit of simpler vibrations
to all motor vehicles. Vibration environment is one of the most important criteria by
which people judge the design and construction criteria “quality” of an automobile.
Hence, the ride quality of a vehicle is judged mainly through the performance of the
suspension design. However, being judgemental is subjective in nature and give rises
to one of the greatest difficulties in developing objective engineering methods which is
to deal with ride as a performance mode of a vehicle.
2.4.1 Introduction to Ride Analysis
Subjective assessment of vehicle ride involved human response and was usually
conducted using questionnaires (Cossalter, Roberto & Rota 2010). Subjective ride
assessment is difficult to perform because it requires human as judges. Hence, objective
ride standards were developed to assess the vehicle ride based on human perception to
vibrations. Objectively define the vehicle quality requires measuring the vibration of a
vehicle through an automobile model in virtual environment. The standard procedure
to study the objective ride of a vehicle is shown in Figure 2.20. The first element to
study the vehicle ride is the excitation source where it plays the role as the input for
analysis. Subsequently, the vehicle dynamics responses are obtained and processed to
get the ride rating.
Figure 2.20 Vehicle ride dynamic system
Source: Wong 2008
Excitation
sources (road
roughness)
Vehicle
dynamic
response
Vibration Ride
perception
45
In order to study the trends of objective ride analysis, the trend of previous
research for the past 30 years related to this vehicle ride was tabulated in sequence into
Table 2.3. From the trend analysis, automotive ride analysis is found to be closely
related to automotive suspension system design. Numerous researches proposed to
optimise the vehicle ride by altering the suspension system design. The most important
finding in vehicle ride analysis is the road input for simulation model. The road profiles
are obtained either through measurement using sensors on actual vehicle or artificially
generated vehicle according to ISO 8608. Meanwhile, surface roughness of the road
profile could be classified into different levels. As for the ride indicator, different ride
standards or root mean square (r.m.s) vertical acceleration were used to quantify the
ride
The aim of the vehicle ride analysis is to optimise the suspension parameters for
better perceptions. The optimisation of ride was assisted by metaheuristic algorithms
such as genetic algorithm (Seifi, Hassannejad & Hamed 2016). The ride standard was
also used to determine the suitable suspension parameter like damping in affecting the
ride quality of a vehicle. Meanwhile, Abdelkareem et al. (2018) improved the ride
performance of a truck through optimised leaf spring design. The constructed vehicle
simulation model was also a critical element to perform the simulation for ride analysis.
Based on the trend analysis, quarter car model was widely used in performing the ride
simulation (Hassaan 2015; Burdzik & Koniczny 2013). Seifi, Hassannejad & Hamed
(2016) proposed that quarter car model is sufficient for ride analysis but not sufficient
for road holding and rollover capability assessments.
Table 2.3 Trend of vehicle ride analysis for the past 30 years
Authors (Year) Objective Road
excitation Ride indicator
Abdelkareem et al.
(2018)
To improve the truck ride
performance through
optimisation of leaf spring design
Artificial
generated road
profile
r.m.s. value of seat
body acceleration
Loprencipe & Zocalli
(2017)
To analyse the effects of ride
standards to a real road profile
Profilometer
measurements
ISO 2631, Michigan
ride quality index,
ASTM ride number
Continue…
46
Continued…
Seifi, Hassannejad &
Hamed (2016)
To optimise the suspension
system in terms of ride and
handling using GA
Deterministic
road profile
(Sine waves)
ISO 2631
Hassaan (2015)
To simulate ride comfort of an
automobile using quarter car
model under bump excitation
Simple
harmonic bump ISO 2631
Reza-Kashyzadeh,
Ostad-Ahmad-
Ghorabi & Arghavan
(2014)
Investigated the road roughness
effects on a semi-active
suspension system
Road class A to
E Vertical acceleration
Burdzik &
Konieczny (2013)
Investigated ride of a passenger
car under different damping
parameters
On-road test ISO 2631
Velmurugan,
Kumaraswamidhas &
Sankaranarayarasamy
(2012)
Reduced whole-body vibration of
trailer cabin Class B
Driver seat
acceleration
Barbosa (2011) Simulated frequency response of
a half vehicle model Class A Vertical acceleration
Pang et al. (2010) Optimised ride of a heavy truck
suspension system Class B ISO 2631
Goncalves &
Ambrosio (2005)
To optimise the ride, handling
and rollover parameter of a
vehicle
Profilometer
measurements ISO 2631
Stone & Demetriou
(2000)
To simulate vehicle ride and
handling performance during
braking and cornering events
Static
displacement Heave displacement
Cherry & Jones
(1995)
To analyse the application of
fuzzy logic in controlling vehicle
ride performance
Tyre force Sprung mass
acceleration
Elmadany (1990)
Optimised the ride performance
of a vehicle through load
levelling system
Derived road
surface based
on vehicle
velocity
Sprung mass
acceleration
Kamawura & Kaku
(1985)
To evaluate the effects of road
roughness on vehicle ride
Displacement
PSD
Acceleration at
vehicle centre of
gravity
47
2.4.2 Road Roughness as Excitation Sources
There are a few sources of excitation from which the vehicle is induced. These sources
generally fall into two major classes, known as road roughness and on-board sources.
The on-board sources arise from rotating components and include the tyre/wheel
assemblies, the driveline and the engine, while road roughness encompasses everything
from potholes resulting from localised surface failures to random deviations. Roughness
is defined as the elevation profile along the wheel tracks where the vehicle passes and
possibly obtained through measurement from a profilometer. An example of road
roughness measurement using a profilometer is shown in Figure 2.21. with international
roughness index (IRI) value. The IRI were divided into measurements from inner wheel
path (IWP) and outer wheel path (OWP).
Road profiles were fitted into category of “broad-band random signals” and
therefore, could be described either by the profile itself or its statistical properties. One
of the most useful representations is the power spectral density (PSD) function because
of randomness nature of the profile. Reza-Kashyzadeh et al. (2014) presented a quarter
car model with different road roughness in terms of PSD function to simulate the vehicle
vertical vibration response for automotive suspension component durability analysis.
Due to the randomness of road roughness, the instantaneous value of the function could
not be estimated in a deterministic manner.
Figure 2.21 Example of profilometer measured road roughness profile
Source: Abulizi et al. (2016)
48
However, some properties of random functions could be determined statistically
such as the mean or the r.m.s value. It could be determined by averaging the value while
the frequency content of the function could be obtained through Fourier transform. In
addition, road surface elevation is determined as stationary in which the statistical
properties derived from a portion of the surface profile could be used to define the entire
section of the road surface (Wong 2008). This meant that a small sampling of road
profile data could be repeated to represent the whole road profile. Since the mean value
of the signal remained constant when computed over different segment of samples, the
random road surface functions are also considered as ergodic.
Road roughness measurement requires a profilometer and an accelerometer
attached on a vehicle (Loprencipe et al. 2017). Considering the expensive cost to build
the experimental setup for rough roughness measurement, ISO committee proposed a
standard road roughness classification known as ISO 8608 (2016). The degree of
roughness, Sg(Ω), of a road could be obtained using range and geometric mean using
the equation as follows:
Sg(Ω) = CspΩ-N
(2.28)
where Csp and N are constants, Ω is the spatial frequency. Over the years, various
organisations attempted to classify the road roughness. ISO proposed a road roughness
classification (classes A – H) based on the PSD, as shown in Figure 2.22 for application
of vehicle dynamic analysis.
The mean value of road properties is often used to study the response of a vehicle
to road roughness. However, different road types with different surface qualities have
different spectral qualities (Mohammadi 2012). Cantisani & Loprencipe (2010)
established a speed related relationship between International Roughness Index (IRI) of
road and ISO 2631 vehicle vertical ride, Awz, model under a pavement road. This model
restrained to vehicle speed of 80 kmph due to Reference Average Rectified Slope
(RARS). RARS suggested that vehicle speed of 80 kmph is the best numerical index.
The wavelength of a vehicle with speed of 80 kmph is detectable by most road users
and covers all road frequencies.
49
The measurement of rough roughness required a profilometer attached to a
vehicle. Attaching profilometer for road roughness analysis is not convenient for
research purpose because it incurs extra cost and involves complex operation. Hence,
Du et al. (2014) developed a model to measure road roughness with a Z-axis
accelerometer and a global positioning system (GPS). They utilised a linear time
invariant (LTI) quarter vehicle model and regarded the pavement as a continuous
surface. A regression model to predict the IRI was proposed using PSD amplitude of
acceleration, LTI system and PSD of pavement roughness (Du et al., 2014). The model
was then expanded using multiple linear regression (MLR) to include both wheels. Du,
Liu and Wu (2016) proposed a speed correction factor to enhance the feasibility of the
regression model for different vehicle speeds.
Figure 2.22 PSD as a function of spatial frequency of various classes of road
Source: ISO 8608 2016
Acceleration road profile PSD could be converted to displacement or shifted
from acceleration to displacement (Du, Liu & Wu, 2016). Displacement PSD is
obtained through second integration of acceleration suggested as follows:
50
Sa(ω) = (ω
v)
4
Sx(ω) (2.29)
where Sa(ω) is the acceleration PSD, ω is the angular frequency, Sx is the displacement
PSD. This conversion aims to provide a more feasible input to vehicle simulation model
because the input for a vehicle simulation usually involves displacement (Loprencipe
& Zocalli 2017; Seifi, Hassannejad & Hamed 2016).
2.4.3 Vehicle Suspension System Modelling
Considering the unneglectable road surface irregularities, the ideal way to control
vehicle vibration response is through suspension isolation, particularly the motion of
vehicle body and axle (Wong 2008). In this case, the vehicle suspension system is
usually modelled as a linear system with a direct linear relationship, surface
irregularities and excitation input. The typical modelling process flow for a simulation
model is shown in Figure 2.23. The vehicle system is characterised by its transfer
function and converts the road surface irregularities input to vehicle vibration output.
In addition, a linear quarter car model was usually used in such ride analysis. Yu et al.
(2013) utilised a quarter car model to design an adaptive real-time road profile
estimation observer considering the vehicle vertical dynamics.
Figure 2.23 Input and output of a linear vehicle system
Source: Wong 2008
In terms of suspension modelling, the vehicle was modelled into a single D.O.F
system with nonlinearities. Surface irregularities were expressed as input while vehicle
vibrations were expressed as output (Naik & Singru 2011). In a basic level analysis, a
quarter car model consisted of a sprung mass supported by primary suspension applied
at each wheel. Dynamic behaviour of this system is the first level of isolation from the
road roughness. These dynamic characteristics are largely defined by ride rates (RR) as
follows (Shim & Velusamy 2007):
Surface
irregularities
(Sg(f))
Vehicle system
H(f)
Vehicle
Vibrations
Sv(f) = |H(f)|2Sg(f)
51
RR = KsKt
Ks + Kt
(2.30)
where Ks is the spring stiffness and Kt is the tyre stiffness. Implications of ride rate affect
four main vehicle modal dynamic characteristics like pitch, roll, yaw and bounce. A
conventional automobile has body bounce frequencies ranging from 1.0 to 1.5 Hz. In
this recommended frequency range, the road holding and ride properties are the most
compromised aspects.
In the absence of damping, the general bounce natural frequency of each
suspension system of the passenger car could be determined using the equation:
fn =
1
2π√
RR × g
W (2.31)
where fn is the bounce frequency, g is the gravity acceleration, and W is the corner
weight of the vehicle. When damping exists, the resonance occurs at damped natural
frequency, fd as follows:
fd = f
n√1- (ξs)
2 (2.32)
where fn is the undamped natural frequency, ξs is the damping ratio which is obtained
using the equation below:
ξs = Cs
√4KsM (2.33)
where Cs is the suspension damping coefficient. In order to maintain a good ride, the
suspension damping ratio usually falls on the range of 0.2 to 0.4. Since there is small
difference between the damped and undamped natural frequency, undamped natural
frequency is usually used to characterise vehicle (Wong 2008).
52
For a quarter car model, the suspension is modelled by using two simple
equations of motions as illustrated in Figure 2.24. The equation of motion for 2 D.O.F.
vehicle model is written as below (Satishkurma et al. 2014):
MsX2 + Cs(X2- X1) + Ks(X2 - X1) = 0 (2.34)
MuX1 + Kt(X2 - w) - Ks(X2 - X1) − Cs(X2 - X1) = 0 (2.35)
where Ms is the sprung mass, Mu is the un-sprung mass, 𝑋2, 2, 2 are the sprung mass
displacement and derivatives respectively, 1 , 1 , 𝑋1 are the un-sprung mass
acceleration and their derivatives respectively, Kt is the tyre stiffness, and w is the road
input. Application of quarter car model in vehicle ride analysis is very common. Sharma
et al. (2013) modelled a 2 D.O.F. car suspension system under state space representation
for vehicle ride study. State space suspension model is more widely used for active
suspension controller design than traditional transfer function method because the
applicability of nonlinearity. For example, Popovic et al. (2011) considered the
nonlinearity of actuator in the controller design of a vehicle suspension by using the
state space quarter car model.
Figure 2.24 Schematic diagram of a two D.O.F quarter car model
Source: Sathiskumar et al. 2014
It is known that increasing D.O.F of a simulation model leads to greater accuracy
but simultaneously causes complex model and heavy computation. Quarter car model
53
usually consists of two D.O.F while half car model has four D.O.F. Full vehicle models
are usually consisted of at least seven D.O.F. (Krauze & Kasprzyk 2016). For a half
vehicle model, bounce and pitch motion of the vehicle body are considered through
integrating bounce and pitch governed equation into the model (Shafiqur Rahman &
Kibria 2014). For vehicle model with bounce and pitch motion, lower front frequency
puts the bounce centre close to rear axle and the pitch centre forwards near the front
axle. This tuning process is known as the “Olley tuning” which could improve the ride
quality (Fu et al. 2013).
Full vehicle model is the most complex but complete model to study. The
number of D.O.F of vehicle model increases according to vehicle motion. A traditional
way to mathematically model a full vehicle is tedious and consumes lots of time and
effort (Ning, Zhao & Shen 2013). Over the years, many algorithms have been developed
to make multibody dynamics analysis more accessible. Nowadays, many commercial
multibody dynamics software have been introduced to help designers to promptly
perform vehicle modelling. Commercial multibody dynamics software serves better in
current fast pace automotive industry due to its effectiveness and ease of use. Ning,
Zhao and Shen (2013) performed a kinematic and dynamic analysis of vehicle
suspension using ADAMS/Car to seek for better ride comfort. Virtual analysis was
continuously conducted until the design met the expected ride index before processed
to prototype making.
Other than ADAMS, SimulationX is a lumped network simulation with
equation-based model objects for simulation. Due to low requirements in terms of
computing power, SimulationX is suitable to analyse complex dynamic systems
involving substantial number of components (Dor & Vielhaber 2015; Sim et al. 2017).
Sim et al. (2017) developed a LQR control algorithm for agricultural tractors hydro-
pneumatic cab suspension to increase the ride comfort. SimulationX was used to model
the hydro-pneumatic system that is installed at the left and right side of the vehicle to
provide suitable load and stabilise the tractor. Jin, Yu and Fu (2016) published a
research on vehicle ride driven by in-wheels motor. A road test was performed on the
selected car model with different added weight in un-sprung mass to represent the
additional weight to corroborate the 11 D.O.F simulation model. With the road
54
roughness excitation and constructed simulation model, the vehicle vibrations are
measured and suspension parameters are tuned to improve vehicle ride.
A frontal quarter car model has been used to conduct vehicle ride dynamics
analysis and the vehicle mass ratios and suspension parameters were tuned to improve
the vehicle ride (Unaune, Pawar & Mohite 2011). The choice is based on the simplicity
to model the quarter car model and sufficiency of front quarter suspension for simple
vehicle ride analysis. Sentil Kumar & Vijayarangan (2007) used a quarter car model to
optimise the vehicle ride through proportional-integral-derivative (PID) controller. The
ride performance of passive suspension was compared to active suspension in terms of
ride. A comparison between quarter, half and full vehicle model for vehicle ride was
also performed (Mahala, Gadkari & Anindya 2009). The research showed that the full
vehicle model yielded the highest accuracy followed by half and quarter vehicle model.
However, the research paper stated that the trends of the responses were similar and
acceptable for all three models (Mahala, Gadkari & Anindya 2009). The acceleration
responses obtained from quarter car, half car and full car model were in the same range.
2.4.4 International Ride Standards
Once the PSD function for acceleration of the vehicle was obtained, further analysis is
required to relate to it to potential ride comfort criterion. Numerous ride comfort criteria
have been proposed such as vertical vibration described in the Ride and Vibration Data
Manual J6a from the SAE, ISO 2631, BS 6841, Society of German Engineers VDI 2057,
Average Absorbed Power (AAP), NATO Reference Mobility Model. Each standard is
used by specific country or region as shown in Table 2.4. Every developed objective
ride standards provides different weightings to quantify the ride, but the concept is
virtually similar where the amplitude and frequency are considered.
Table 2.4 Objective ride and preferred countries
Standards Country Use
ISO 2631 Europe
BS 6841 United Kingdom
VDI 2057 Germany and Austria
AAP / NATO Reference Mobility Model United States of America
Source: Els 2005
55
AAP standard was developed by the US Army Tank Automotive Command in
1966. Under vibration, the human body behaves elastically and generates restoring
forces that are related to displacement. The generated restoring forces dissipate the
energy until they are imparted. Time rate of this energy absorption process is named as
absorbed power. The AAP standard computes the energy in frequency domain. It
proposed the frequency range of interest from 1 to 80 Hz. This standard emphasised the
frequency range from 4 to 5 Hz because the range is in resonance with human tissues.
Zhao and Schindler (2014) utilised AAP and ISO 2631 standard to calculate the ride
quality of a crawler chassis for suspension design. The need of off-road vehicle for ride
quality was also emphasised.
British Standards BS 6841 proposed the concept of vibration dose value (VDV)
in 1987. The frequency range of interest is from 0.5 – 80 Hz where the frequency of 0.5
Hz is to compensate the motion sickness. This standard modifies the frequency
weighting of Z-axis so that the results are closer to experimental findings. For the ride
comfort, they proposed the r.m.s of the weighted signal. This ride index is easily
comparable because the value is linear to the ride quality. Paddan et al. (2012) utilised
the BS 6841 standard to analyse the seat backrest towards ride perception. VDI standard
was initially published by Society of German Engineers in 1963 where it was the first
standard to quantify ride comfort. VDI standard was created using many human data
under sinusoidal vibrations with certain frequency and amplitude. In 1979, the VDI
adopted the tolerance curves of ISO 2631 but maintained the K-factor for subjective
comparison.
VDI standards was applied by Hohl in 1984 to assess vehicle ride, as reported
by Els et al. (2007). However, there is very limited recent publications on the use of
VDI standard. ISO 2631 standards extended the frequency range below 0.1 Hz in 1997.
The r.m.s value remained as the basis for ride assessment. Generally, the method of
assessment resembles the standard BS 6841. The main difference between ISO 2631
and BS 6841 is the vertical weighting factors. The ISO standard included the frequency
range for motion sickness in frequency range of 0.1 to 0.5 Hz that is not considered in
BS 6841 standard. Literature review has shown that the current trend for objective ride
56
rating of a vehicle mostly adopts the ISO 2631 standard. ISO 2631 was used to assess
the ride comfort of many vehicles because of its comprehensiveness and ease of use.
For example, it was used to assess the ride comfort of compact wheel loaders
(Zhao & Schindler 2014), motorcycles (Chen et al. 2009), passenger cars (Hassaan
2015), mining vehicles (Eger et al. 2008), bus (Sekulic et al. 2016), and agricultural
tractor (Loutridis et al. 2011). Compact wheel loaders, mining vehicles and agricultural
tractors are off-road vehicles which are used in harsh terrain with higher chance of
exposure to vibrations compared to normal passenger vehicle. Long-time exposure of
vibration is detrimental to operator’s health. ISO 2631 standard also illustrates the limit
of exposure to vibration. For passenger vehicle, the ISO 2631 standard has improved
the vehicle ride because passengers nowadays do not only demand for durability, but
also a greater ride vehicle.
For vehicle vibration, a complete vehicle model has many D.O.F. Specifically,
there is a seat suspension between the driver and the vehicle. Moreover, the random
input to the vehicle is always more than one because the passenger car has four wheels.
Each wheel possesses one random input. Interaction of the random inputs with each
other becomes important to determine the output of the system. Hence, cross-spectral
densities are essential. For cross spectral, time lag between the front and rear wheel
should be taken into consideration (Feng et al. 2013). The rear time delay could replicate
the response of the front. Understanding the front suspension state in principles allows
the rear suspension to also react to road profiles.
2.5 DATA ANALYSIS
2.5.1 Introduction to Data Analysis
Fatigue life of a product is scattered and has an acceptable variations range of 100%
(Shamsaei & MeKelvey 2014; Karolczuk 2016). Hence, it is not easy to interpret the
fatigue life in a direct manner. Better understanding on the fatigue characteristic of a
component requires data analysis to statistically define the fatigue life. For deeper
analysis, a specific field to identify the correlations and patterns within the data known
57
as “data mining” is introduced. The list of data mining techniques is shown in Figure
2.25. In general, there are six branches of data mining techniques. Application of data
mining techniques to build models and predict outcomes is called “Machine Learning”.
Machine learning is an application of artificial intelligence that provides the system the
capability to learn and improve without explicitly programming. Statistical theories
form the language of machine learning problems which are amenable to solve.
In the scope of durability, data analysis is an essential element due to the
randomness of the fatigue process (Schneider & Maddox 2003). Additionally, samples
for fatigue life test are often destructed and could not be practically used. In case of
fatigue test, the small standard coupon that is used to determine the fatigue behaviour
of material is generally known as specimen. For a fatigue loading time history, the
sample size is known as the number of data points in a measurement. The sample size
for fatigue analysis is usually high and depends on the sampling rate. When the
sampling size is large, fatigue data are characterised using statistical method because
value of a single point does not bring any significant meaning.
Figure 2.25 Groups of data mining technique
Source: Ngai et al. 2011
The eigenvalues of the observed data representing the statistical nature of a
sample can be classified into two categories: (1) the central position of data and (2)
dispersion of the data. Examples of central position data include the mean value while
dispersion of the data represent the standard deviation (𝜎𝑆𝐷) or variance. Mean and
Data mining
Classification
Clustering
Outlier detection
Prediction
Regression
Visualisation
58
standard deviation of strain signals were used in automotive component fatigue analysis
to determine when the strain signals were random and stationary (Kosobudzki 2014).
The mean value of a sample can be defined as follows:
x = 1
n∑ xi
n
i=1
(2.36)
where n is the sample size. The sample mean represents the centre position of the data.
Dispersion of a set of datasets is analysed using 𝜎𝑆𝐷. 𝜎𝑆𝐷 of less than 30 sample
is derived as follows:
σSD = √∑ (xi- x)2n
i=1
n - 1 (2.37)
where is the sample mean. Meanwhile, for any sample of more than 30, the 𝜎𝑆𝐷 is
obtained as below:
σSD = √∑ (xi - x)2n
i=1
n (2.38)
𝜎𝑆𝐷 is a crucial index to determine the dispersion of data. Greater value of 𝜎𝑆𝐷 indicates
larger dispersion of the observed data. However, Wang, Chen & Zhou (2016) proposed
that r.m.s was more frequently used in fatigue analysis because r.m.s shows the
vibration energy of every fatigue cycle. The r.m.s value of datasets is defined as below:
r.m.s = √∑ (xi)
2ni=1
n (2.39)
Yan, Wang and Zhang (2014) suggested the commonly used statistical
parameters in statistical analysis including skewness (λ) and kurtosis (γ). When λ and γ
exist in a sample data, the data are non-Gaussian distributed. λ is a measure of the
59
asymmetry of the probability distribution of a real-valued random variable about its
mean (Rizzi, Behnke & Przekop 2010). λ is expressed in the central of moment of the
probability density function (PDF) as follows:
λ = ∑ (xi - x)3n
i=1
n(r.m.s)3 (2.40)
γ denotes the “tailedness” of the probability distribution of a real-valued random
variable (Rizzi, Behnke & Przekop 2010). The γ is defined as follows:
γ = ∑ (xi - x)4n
i=1
n(r.m.s)4 (2.41)
Crest factor (CF) is often used in characterisation of vibration signal and indicator for
vehicle ride. It could be defined as follows:
CF = |𝑋max
r.m.s| (2.42)
where Xmax is the maximum amplitude of a time history.
2.5.2 Regression Analysis
Regression analysis is one of the data mining techniques involving a set of statistical
processes to estimate the relationships among variables. In case of a simple single
independent variable to dependent variable, the regression method is known as simple
linear regression. This type of regression method is considered as linear. Simple linear
regression is commonly seen in engineering applications, such as spring stiffness
calculation (Ryu et al. 2010), and loads and cycles to failure modelling for automobile
chassis (Thomas 2012). For the spring stiffness analysis, the dependent variable was the
load applied on spring while the independent variable was the spring displacement.
Load and cycles to failure were also modelled using linear regression after applying the
natural logarithm on fatigue life.
60
In 1880s, F. Galton originated the concept of correlation which was the
fundamental to linear regression, as mentioned by Dodge (2010). Linear regression is
used to fit with a linear curve. However, introduction of mathematical approximations
indicates the presence of uncertainty that occurs and causes deviations from the true
value. Hence, the standard “goodness of fitting” measure for a regression-type model
was introduced. This statistic indicator is called the coefficient of determination (R2).
The R2 indicates the proportion of variance in the dependent variable that is predictable
from the independent variable. Graphically, the measurement is used to determine how
well the regression line represents the data set.
For simple linear regression, the R2 value of a regression could be obtained using
the equation as follows:
R2 = 1 - ∑ (X1 - X2)2n
i=1
∑ (X1 - x)2ni=1
(2.43)
where X1 is the vector of true value and X2 is the vector of n predictions. The R2 value
ranges from 0 to 1. Higher R2 values indicate greater fitness of the regression with the
data. When the prediction is poor, uncertainty or error shall reflect a significance value
of the prediction. In statistics, mean squared error (MSE) could be used to measure the
average squares of the error. The MSE is defined as below:
MSE = ∑ (X1 - X2)2n
i=1
n (2.44)
As R could also provide negative value, root squaring the R value will obtain positive
value with more weight and smaller value.
When a set of parameters possesses linear relationship, least squares method
could be applied to obtain linear regression. Least square method has been used to
determine the fatigue characteristic and ultrasound (Padzi, Abdullah & Nuawi 2014).
The function of least squares method could be written as below:
61
y = bL + aLx (2.45)
where bL is the intercept value and aL is the slope. The aL and bL are obtained using the
following solutions:
aL=∑ xi
ni=1 y
i -
1
n∑ xi
ni=1 ∑ y
ini=1
∑ xi2n
i=1 - 1
n(∑ xi
ni=1 )2
(2.46)
bL = ∑ yi
ni=1
n -
aL ∑ xini=1
n (2.47)
A prerequisite to apply least squares linear fitting method is the linearity of data.
However, engineering components tend to exhibit both linear and non-linear
relationships. For nonlinear regression, the preferred methods are logarithmic, power,
exponential, polynomial and moving average (Skrobachi 2007). A logarithm is suitable
when the data grow rapidly, followed by a period where the period continues to increase
in slower pace. Polynomial regression is applied when the data fluctuates. When the
data values increased or decreased at very high rate, exponential regression model
should be applied. Exponential value could not be created when there is zero or negative
value without any intercept. Apart from these, moving average model uses average
value to smoothen out any fluctuating data to show pattern or trend.
Power regression model is suitable for data that increase at a specific rate and
commonly used in fatigue application because the fatigue nature of materials is in this
form. S-N curve proposed by O.H. Basquin in 1910 was based on this power function.
Meanwhile, power law was used to fit stress-strain behaviour of a steel under different
strain rate loading to represent flow (Gupta & Kumawat 2017) in which the obtained
fatigue data appear to have a power-law character. The power function shall be used to
represent the data and the function as found in Zielesny (2014) can be written as follows:
y = apxbp (2.48)
where ap and bp are obtained using the following equation (Zielesny 2014):
62
ap = exp [1
n∑ ln y
i - bp
∑ ln xini=1
n
n
i=1
] (2.49)
bp = ∑ ln xi ln(y
i) -
1
N(∑ ln xi)(∑ ln y
i)
(∑(ln xi)2) - 1
N(∑ ln xi)2
(2.50)
where i is the data number.
Amsterdam and Grooteman (2016) studied the fatigue stress crack length of
aluminium alloy using Paris law with power exponent to model the crack growth
variation. Carpinteri and Paggi (2009) proposed a study of relationship existing between
Wöhler’s and Paris’s representation of fatigue. According to dimensional analysis, the
Wöhler’s and Paris’s equation provides a rational interpretation to most of the empirical
power law criteria used in fatigue. Both the linear and power functions are only
applicable when a single independent variable is used to predict the value of a dependent
variable. When the independent variable is more than one, a method for multiple input
is required for the model construction and MLR should be applied.
In 1896, F. Galton and K. Pearson systemised the correlation analysis and
established a theory of correlation for three variables. This was the fundamental idea of
MLR in which MLR was used to model the predictor with one or more independent
variables. The general form of MLR is as follows (Stojanovic et al. 2013):
yi = β
o + xiβ1
+.. .+ xikβk + ei i = 1,2, ..., n (2.51)
where 𝛽𝑜 is a constant, xk is the regression coefficient, βk are the variables, and ei are
the residuals. The MLR is applied for prediction, explanation or theory building
purposes for the parameters. Dang et al. (2014) correlated vehicle handling with other
46 metrics, shifting each dataset to zero mean value in order to obtain good results. Du
et al. (2014) also applied this method to determine the IRI road roughness index with
left and right wheel acceleration measurements. The IRI is determined as the
63
dependence parameter while the left and right wheel accelerations mean values were set
as the independent parameters.
The application of MLR in the field of fatigue is not common because fatigue
usually exhibits nonlinear behaviour such as in forms of power (Amsterdam &
Grooteman 2016). Limited research publication has been found in fatigue life analysis.
A list of recent applications of MLR in creating fatigue life prediction model is shown
in Table 2.5. These research results focused on fatigue life prediction using
experimental data, either through geometry of specimen or loading conditions. The
advantage of using MLR models distributes in its ability to detect outliers. However,
this method has some limitations such as linear relationship, randomness and equal
variance variables (Kamer-Aimu & Marioara 2007).
2.5.3 Neural Network-based Regression Analysis
Soft computing is a collection of algorithms that are employed to solve a very complex
problem. The algorithms are like fuzzy logic, artificial neural network (ANN),
evolutionary algorithms and hybrid systems (Ibrahim 2013). Fuzzy system is a method
that is traceable using the fuzzy set theory that was developed by L. Zadeh in 1965. A
mathematical framework was provided to describe the vagueness of variables by
mapping a given input to output using fuzzy logic which is known as adaptive neural
fuzzy inference system (ANFIS). As for fatigue application, ANFIS was used to model
the S-N curve of polyester under CALs (Vassilopoulos, Georgopoulos, &
Dionysopoulos 2008). However, the performance of ANFIS was less accurate as
compared to ANN (Khademi et al. 2016; Tayyebi & Pijanowski 2014; Mas & Ahlfeld
2007; Elbayoumi, Yusof & Ramli 2015; Tosun, Aydin & Bilgili 2016).
ANN is a versatile methodology that can accurately approximate non-linear
functions, developed by W. McCulloch and W. Pitts in 1943 and inspired by human
brains neurons (Pospíchal & Kvasnička 2015). Neural networks resemble the human
brain as learning and knowledge are stored within the connected neurons. A basic model
for the node in ANN is shown in Figure 2.26 where xi is the input, wi are the weights, f
is the transfer function and y is the output. ANN is also divided into two categories
64
which are single and multi-layer. Single layer ANN consists of only one hidden layer
while multilayer ANN consists of two or more hidden layers.
Table 2.5 Researches on generation of fatigue related MLR-based model
Authors Year Research contributions
Dong,
Garbatov &
Soares
2018
Established a fatigue crack related notch of a cruciform weld joint using
weld leg and slit lengths
Mayén et al. 2017
Generated a crack prediction model using stress amplitude and number
of cycles as input.
Jin et al. 2016
Generated a multiple linear prediction model to correlate damage,
vibration and temperature of a highway bridge.
Rothleutner et
al. 2015
Created a shear stress model for hardened steel using fatigue life, carbon
content and normalised effective case depth as independent variables
(radius).
Bižal,
Klemenc &
Fajdiga
2014
Established a fatigue life model for aluminium using different porosity
level and manufacturing properties like die temperature and pressure.
Kanoje,
Sharma &
Harsha
2013
Established a natural frequency model of locomotive wheel using rim
geometry and loadings. The natural frequencies were then correlated to
fatigue life.
Kannappan &
Dhurai 2012
Modelled a relationship between tensile strength properties of fibre
reinforced composite with temperature, pressure and time as
independent parameter.
LeBozec &
Thierry 2010
Modelled corrosion cyclic test of automotive aluminium alloy with
parameters like concentration of salt, drying level, humidity cycle,
frequency of salt spray and temperature.
Shell,
Buchzeit &
Zoofan 2005
Modelled residual fatigue life of aluminium alloy with non-destructive
evaluation metrics.
Lu, Behling &
Halford 2000
Modelled total strain range as dependent variable with heating hold time
and cyclic life relation for stainless steel alloy SS409 which was widely
used in automotive exhaust system.
Atkins &
Gibeling 1995
Modelled a relationship of creep strain according to stress and
temperature for Al-SiC metal matrix composite.
Zhang, Bui-
Qouc &
Gomuc
1990
Modelled fatigue life of stainless steels with heating temperature and
strain rate effects under cyclic loadings.
Vinokur et al. 1985
Developed a multiple linear regression model to predict intensity of
wear with temperature and time conditions of tempering and hardening
65
Figure 2.26 A basic model of a single node ANN
Source: Livinstone 2008
An example of a multilayer neural network is shown in Figure 2.27 (Borah,
Sarma & Talukdar 2015). This example of multilayer neural network has two hidden
layers with four neurons in the first and second hidden layers. There are many types of
ANN available in the supervised and unsupervised learning. In general, the neural
network could be classified into two main types, namely the feed forward and feedback
ANN (Baykal & Yildirim 2013). Feed forward type is a type of straight forward ANN
with no loop, with an example of feed forward ANN is illustrated in Figure 2.27.
Architecture of feed forward ANN could be in single or multiple layers. Feed forward
type of neural network is also a static type ANN which has no feedback.
There is another type of ANN known as recurrent or feedback ANN. Recurrent
or feedback type of neural network is a type of ANN where the output provides feedback
to the neuron (Borah, Sarma & Talukdar 2015). This type of neural network is a type of
dynamic neural network where the output depends on the current input. One example
of a recurrent neural network is the Hopfield neural network where all the neurons were
fully interconnected (Hsu 2012). Another famous neural network architecture is the
radial basis function network invented by D. Broomhead and D. Lowe in 1988. The
radial basis function network is similar to feed forward neural network with exception
of its application of radial basis function as the activation function. When more than
one hidden layer is applied, the learning process is known as deep learning
(Schmidhuber 2014). Deep learning could be implemented in both feed forward and
feedback neural networks. ANN provides a very good prediction results because
numerical optimisation was applied to iteratively improve the regression (Hanke 2010).
This technique is used to optimise least square problems.
66
Figure 2.27 An example of multilayer feed forward ANN architecture
Source: Borah, Sarma & Talukdar 2015
Another one of the important parameters to improve the performance of an ANN
is the activation function. The activation function of ANN transforms the activation
level of a neuron into an output signal. There are various available activation functions
such as uni-polar sigmoid function, bipolar sigmoid function, hyperbolic tangent
function, radial basis function, conic section function (Karlik & Vehbi Olgac 2011).
Comparison between these activation functions has been widely performed. The bipolar
sigmoid function is the most applied function because of its computational efficiency.
Application of ANN is vast in engineering application. Many ANNs are involved in
recent advance applications such as adaptive robotics and control, handwriting
recognition, speech recognition, keyword spotting, music composition, attentive vision,
protein analysis, stock market prediction and many other sequence problems. Recurrent
neural network was particularly used in the recognition of handwriting (Graves 2011),
image, and speech (Graves, Mohamed & Hinton 2013). A recurrent neural network is
heavier than a feed forward neural network in terms of computing because it provides
feedback to the hidden or input layers to correct the weights and biases. Therefore,
recurrent neural network is suitable for the abovementioned complex analysis.
Application of feed forward neural network is common in fatigue applications
due to the straightforward data (Figueira Pujol & Andrade Pinto 2011). Al-Assadi, Kadi
and Deiab (2007) demonstrated the ability of different neural networks in predicting
fatigue life. The input parameters include volume fraction, tensile modulus, tensile
strength, applied load parameters, probability of failure, and statistical parameters of
fatigue life. The output was the logarithm of the fatigue life cycles. Kang et al. (2006)
67
utilised a feed forward neural network together with the critical plane method to predict
the multiaxial fatigue life of an automotive subframe. The results indicated that feed
forward ANN was appropriate in searching critical locations. In addition, neural
networks were applied to estimate the fatigue crack growth rate. ANN was also applied
to address the stochastic aspects of fatigue phenomenon. For example, Janežič,
Klemenc and Fajdiga (2010) implemented a feed forward neural network to model
cyclic stress-strain scatter using arbitrary selection of temperature, size of cross-section,
content of alloying elements, and loading rate as the inputs.
2.5.4 Validation Analysis
Once the regression models are generated, performances of the generated models will
be verified. Performance of a regression was usually analysed using residuals. The
residuals of a good regression prediction should be in normal distribution (Ghasemi &
Zahediasi 2012). There are two types of method to test for data normality, which are
graph assessment method and normality test. For the graph assessment, residuals were
plotted into error histogram and scatter plot (Francisco et al. 2006). An example of
residual histogram and scatter plot is shown in Figure 2.28(a) & (b), respectively. The
commonly used normality test method is Kolmogorov-Smirnov (KS) test. KS test is
widely used because it does not depend on the sample size.
Besides KS test, another well-known normality test is Lilliefors test. Lilliefors
(LF) test is an improvement of KS where the tails of probability distribution were
corrected. Lilliefors test is limited to up to 1000 sample size. The first step in Lilliefors
test is to obtain the individual z-scores (zi) for every member of the sample. The zi is
obtained as follows:
zi = xi - x
σSD
(2.52)
Subsequently, the next step is to calculate the test statistic which is empirical density
function based on the zi. The formula for Lilliefors test statistic is as follows:
68
Residual
LF = maxx|Dz(x) - Lz(x)| (2.53)
where Dz(x) is the theoretical standard normal distribution function, Lz(x) is the
empirical distribution density function of the zi value. The last step is to determine the
critical value for the test. When the number of sample is greater than 30, the critical
value (CV) for significance level, α = 0.05 (95 % confidence level) is obtained as
follows:
CV = 0.886
√n (2.54)
The null hypothesis (H0) is rejected when the test statistic is greater than the critical
value. The hypothesis for Lilliefors test is defined as below:
H0: Data are normally distributed
H1: Data are not normally distributed
(a) (b)
Figure 2.28 Residuals normality assessment graph: (a) error histogram, (b) scatter plot
Source: Francisco et al. 2006
In a regression analysis, the prediction value is always required to be in normal
distribution to ensure the value would not deviate too much from the real data. Razali
and Yap (2011) studied normality of regression and concluded that Lilliefors test is
Fre
quen
cy
Reg
ress
ion s
tand
aris
ed r
esid
ual
Regression standarised predicted value
69
suitable to test for normality among regression residuals. Normality test examines the
goodness of regression. However, scatter band is always applied to examine fatigue
results. Scatter band is often used in fatigue study to describe the difference of two result
sets, such as experimental and simulation results. Shamsaei et al. (2010) utilised this
method to compare Palmgren-Miner and Fatemi-Socie critical plane predicted fatigue
lives. For illustration purpose, a generated schematic diagram for this scatter band is
shown in Figure 2.29. Both the axes are logarithmic and the reference line is labelled as
1:1. The upper bound is labelled with 2:1 while the lower bound is labelled with 2:1
(Karolczuk 2016). When a datum locates beyond the 1:2 or 2:1 boundary from the
reference line, the datum is considered to have poor correlation.
Figure 2.29 Schematic diagram of scatter band for fatigue life using generated data
For standard validation of test data, correlation between experimental and
predicted data is usually performed. The correlation between predicted data and
experimental data is determined using R2 value as shown in Figure 2.30. R2 value above
0.50 is considered as acceptable. R2 value above 0.80 is considered as good while any
value above 0.90 is considered as very good. A good correlation indicates the prediction
results are closely associated with the experimental data. Meanwhile, Huang and Griffin
(2014) used this correlation method in vibration analysis to study how good the
vibration model prediction as compared to experimental results. A good validated
mathematical model is applicable for prediction so that conclusions could be drawn.
Log – Experimental fatigue lives (cycles)
Lo
g –
Pre
dic
ted
fat
igue
lives
(cycl
es)
70
Figure 2.30 Schematic diagram of correlation between prediction and experimental data
2.6 SUMMARY
This chapter discusses the related research works that had been done by researchers in
the fields of fatigue and automotive ride. The focus of this chapter is on the development
of durability and vehicle ride in terms of statistical and mathematical modelling. The
discovery of fatigue had begun since 1837 while the vehicle ride standard was first
introduced by Society of German Engineering in 1963. The literature trends have also
shown extensive publications of vehicle ride analysis in recent years. Both durability
and ride analysis were directly linked to the automotive suspension design. As for the
vehicle ride and durability analysis, regression and neural network analysis are
frequently found in the literature and illustrates the significance of the data analysis
approaches in automobile ride and durability analysis.
Based on the findings of this literature study, it is found that the fatigue life
estimation in automotive suspension design still requires some improvements because
the spring fatigue life analysis and automotive ride were previously conducted
independently. Fatigue life prediction process in automobile spring design consisted of
a scientific gap in the aspect of spring design’s effects towards vehicle ride which
required additional efforts during the spring design phase. Generation of mathematical
models to shorten the automotive suspension design process has become the motivation
of the current research works. In Chapter III, details of the methodology employed in
this study are discussed. A comprehensive presentation of the specific methods used to
develop an original mathematical model as the novelty of this study is fully described.
Experimental data
Pre
dic
ted
dat
a
71
CHAPTER III
METHODOLOGY
3.1 INTRODUCTION
This chapter explains the methodology employed to achieve the aim of this research
which is the establishment of regression-based durability relationship for automobile
vertical vibration. In Figures 3.1 and 3.2 are the illustrations of the process flow of this
current study, dividing the methodologies according to the three objectives. In order to
achieve the first objective which was characterisation of durability relationship, coil
spring design from a local automobile was analysed with collected strain and
acceleration time histories. Statistical analysis was performed to understand the spring
responses under different road conditions. In obtaining additional datasets, artificial
road data and quarter car models were employed to generate new datasets. The
simulation outcomes were used to predict spring fatigue life and ISO 2631 vertical
vibration. Subsequently, a new regression relationship was obtained with vehicle body
frequency of quarter car model, spring fatigue life and ISO 2631 vertical vibration as
the parameters. This proposed regression was used to predict fatigue life of spring.
For the second objective, the steps involved the establishment of the regression-
based durability relationship which led to the novelty of current research. This step
included utilisation of artificial neural network (ANN) to train the datasets. Based on
the literature review carried out by author, ANN was suitable to perform regression
analysis for linear and nonlinear data. In this analysis, ANN performed high
dimensional and multimodal search space to determine the optimised iterations with a
few sets of weights and biases. In addition, the improvised ANN architecture led to
better performance in terms of MSE and coefficient of determination (R2) value.
72
Therefore, different ANN architectures with varying hidden layers and neurons were
analysed. Through these processes, the output of the ANN predictions was the
optimised fatigue life of coil spring.
Lastly, the third objective was to validate the regression and ANN-based
durability regressions. Five sets of strain and acceleration measurement data from an
actual vehicle were processed and put into the generated fatigue life prediction
regression and ANN. The accuracy of the prediction was analysed using statistical
methods such as R2 value and conventional fatigue life scatter band. The conventional
fatigue life scatter band approach was found to be suitable for fatigue life prediction
between simulation and experimental. Meanwhile, the normality of the prediction
residuals was assessed according to Lilliefors test. The obtained R2 value for the
regression and experimental results indicated the level of correlation between
simulation and experimental results.
3.2 DURABILITY CHARACTERISATION OF THE SUSPENSION
SYSTEM
3.2.1 Perform Analysis on Spring
Macpherson suspension system was used as the case study to establish the durability
regression because this type of suspension system had a simple architecture and
required lesser space which is suitable for the application of most passenger cars
(Mahmoodi-Kaleibar et al. 2013). Thus, a frontal coil spring of a local passenger car
was used to perform the regression establishment. The Macpherson suspension system
suits well for this type of vehicle application because of the simplicity of suspension
design when compared to a double wishbone suspension system (Mahmoodi-Kaleibar
et al. 2013). Both the Macpherson and double wishbone suspension systems possess
good strength and durability. However, Macpherson type suspension system is simple
and requires less components such as control arms.
73
Figure 3.1 Process flow of the research methodology: Part 1
OBJECTIVE 1:
Determine a multiple-linear regression
analysis with fatigue life as the response
Spring stiffness
sensitivities
OBJECTIVE 2:
Establish the spring fatigue life
prediction using ANN
Collect and analyse the strain and acceleration
signals on various road conditions:
a) highway
b) hilly
c) residential
d) UKM campus
e) Rural (gravel)
Characterise durability of the suspension system
Force time history Acceleration time histories
Generate artificial data according to
Road class “A” to “D”
class “A” to “D”
Continue to
Figure 3.2
Vehicle body
frequency Spring fatigue life ISO 2631 vertical
vibration
Determine multiple linear regression for fatigue life prediction
Perform CAD and CAE
on spring geometry
Extract data from simulated quarter car model
simulation
74
Figure 3.2 Process flow of the research methodology: Part 2
Analyse the prediction outcomes using
statistical method
OBJECTIVE 3:
Validate the regression-based and ANN
durability predictions
Input experimental data into
regression-based and ANN durability
relationship
Normality test
Continued
from Figure
3.2
Determine the number of neurons Determine the number of layers
Set number of neurons: 1:10 Set number of layers: 1: 3
Analysis of number of neurons and layers
Determine the suitable architecture for
fatigue life prediction
Obtain the suitable fatigue life
NOVELTY:
Integrated vibration-life relationships
Scatter band
Validate established relationship
Design the ANN architecture
Suitability of fitting
Note:
bold letter: objectives
bold & italic letter: novelty
75
The selected automobile suspension system was from a Proton SAGA BLM
1300 cc with kerb weight of 1085 kg which consisted of a set of spring and damper with
the spring geometry listed in Table 3.1. These data were obtained from measurement
and observation on an actual spring. The spring stiffness was calculated as 20,000 N/m
while the spring index was 9.833. Optimum spring index ranged from 4.00 to 20.00
where the spring index out of this range would lead to extra cost and tolerance during
manufacturing (Taktak et al. 2014). The weight of this coil spring was determined at
2.64 kg, with maximum shear stress of a helical spring usually occurred in the spring
inner surface. For this spring, the Wahl shear stress correction factor was calculated to
be 1.15 in which this value was applied in spring shear stress design (Chiu et al. 2016).
The geometry design of the coil spring as shown in Figure 3.3 was produced
using a computer aided design (CAD) software by means of the CATIA software
package. The geometry of this spring was created and used as input to determine the
stress distribution using FEA approach in SIMULIA ABAQUSTM software.
Subsequently, the FEA stress-strain model was applied for fatigue life analysis using
the nCode DesignLife®.
Figure 3.3 CAD of the coil spring
Table 3.1 Measured geometrical properties of the spring
Geometry MacPherson spring values
Spring free length (mm) 385
Mean coil diameter (mm) 130
Material diameter (mm) 12
Number of active coils 6
76
In terms of material, the SAE 5160 carbon steel was given and this material is
known as the most common spring steel with high yield strength and elasticity (Lopez-
Garcia et al. 2016). Based on the carbon content, the spring steel was classified as low-
medium carbon steel with the carbon content ranged from 0.3 to 0.6 % (Rasyidi &
Pratiwi 2015). Monotonic and cyclic mechanical properties of this spring steel are listed
in Table 3.2. The cyclic properties of this carbon steel were obtained through fatigue
test of heat treated specimens and documented in nCode material database, as reported
by Bhanage and Padmanabhan (2015).
Initial design process of a coil spring involved calculation of spring stiffness
including measurements of material properties, boundary and loading conditions for
target stiffness (Arjun & Peter 2014). Once these parameters were determined, the study
continued with static stress analysis to prevent design static failure based on the FEA
procedures. Initially, the coil spring was meshed with three-dimensional (3D)
hexahedra (Hex) elements. This type of element was selected because 3D Hex elements
provided more accurate results compared to 3D tetrahedrical and two-dimensional (2D)
quadrilateral elements (Anitua et al. 2009). A mesh sensitivity analysis was conducted
as listed in Appendix B. Considering the bar size of the coil spring, the element size of
2 mm was suitable for the meshing because the model consists of 6 elements across the
thickness level (Gokhale et al. 2008). In this case, the spring model consisted of 9227
nodes and 7170 3D Hex elements. In the case of boundary conditions, the top of the
spring was fixed where no movement was allowed and the loads were applied at the
opposite of the spring.
Table 3.2 Monotonic and cyclic properties of the spring material
Properties Values
Ultimate tensile strength (MPa) 1,584
Modulus of elasticity (GPa) 207
Yield strength (MPa) 1,487
Fatigue strength coefficient (MPa) 2,063
Fatigue strength exponent -0.08
Fatigue ductility exponent -1.05
Fatigue ductility coefficient 9.56
Cyclic strain hardnening exponent 0.05
Cyclic strength coefficient (MPa) 1,940
Poisson ratio 0.3
Source: Bhanage & Padmanabhan 2015
77
Considering the weight of fully loaded passenger car is about 1,200 kg, the static
load on each suspension is about 300 kg. The assumption that the loads were evenly
distributed in all four wheels was made (He et al. 2010). Hence, an axial force of 3000
N was applied on the bottom of the spring. For torsional loading, same amount of torque
load was also applied on spring material as proposed by Akiniwa et al. (2008). The
torque load was applied on the same spot of coil spring but in X-axis according to SAE
coordinate system to simulate bending moment because the wheel is moving in arc
direction (Ryu et al. 2010). All the values and directions of force, material properties,
and mesh qualities could affect the accuracy of the FEA results (Gokhale et al. 2008).
Extra cares were needed when meshing the spring CAD. On the other hand, linear static
analysis was conducted in the calculation due to its simplicity to determine the stress
distribution (Putra 2016). From the static analysis, results of the spring FEA results
indicated the maximum stress region of the spring where the critical location for strain
time histories measurement was indicated. Furthermore, the effects of axial and
combination of axial and torsional of the spring in stress level were analysed because
the loads of spring could come from various directions due to the arc movement of
wheel (Liu et al. 2008).
3.2.2 Collection of Strain and Acceleration Signals according to Road Surface
Once the critical region of the spring was identified, the measurements of strain time
histories on the vehicle were conducted using a set of strain gauges. In automotive
industry, the required strain time histories were obtained by driving the vehicle
equipped with prototype components for durability evaluation known as a
“development mule” (Sashikumar et al. 2017). The desired loading time histories of the
prototype vehicle components were collected for driving behaviour and road surface
effects analyses. In this work, the required loading signals were strain time histories of
the coil spring and acceleration time histories of the vehicle sprung and un-sprung mass.
The strain data collection process could be generally divided into two stages which were
instrument installation and signal collection stages.
The equipment for the strain measurement were a 2 mm uniaxial strain gauge, a
SoMAT eDAQ data acquisition system, and a computer with TCE software package
78
(SoMAT 2002). Meanwhile, the applied uniaxial strain gauge was a rectangular, planar
strain gauge. This type of rectangular uniaxial strain gauge was used when the direction
of principal strains was unknown (Vishay 2005a). For this type of strain gauge, the
gauge factor of 2.07 ± 1.0% was obtained as listed in the purchase packaging. In
addition, the gauge resistance was 120 Ω and this resistance range is the lowest among
the commercial strain gauge because the strain gauge with lower gauge resistance
possess lower errors (He, Yi & Sun 2016). These constants defined the properties of
strain gauge sensors where the generated voltage from Wheatstone bridge were scaled
according to these constants. These constants were also part of the inputs to SoMAT
software.
Meanwhile, the accelerometers used in the analysis were piezoelectric type
from PCB PIEZOTRONICSINC. The measurement range and frequency of the
accelerometer were required to be higher than the vehicle vibration level to capture the
road induced vibration. The selected accelerometers had the sensitivity of 1.02
mV/(m/s2), measurement range of ± 4900 m/s2 and frequency range of 0.5 – 10 kHz.
The vibration level of an automobile was usually below 8 m/s2 and frequency of 500
Hz (Burdzik 2014). Hence, this type of accelerometer was suitable in automobile
vibration data collection because it was able to capture all the required information of
the suspension vibrations. In this case, the accelerometer was attached at the top mount
for ride assessment because the generated quarter car model did not take the seat into
consideration (Ahmed et al. 2015; Phalke & Mitra 2016).
Attachment of a strain gauge to the coil spring was the main factor that directly
affected the accuracy of the collected data (Vishay 2005b). The surface of the
attachment point must be flat and clean. Prior to the measurement, the desired surface
of the coil spring was scrubbed with a sand paper of grade 400 to ensure polished surface
(ASTM E112-96 2004). As for the attaching material, a specific glue cynoacrilite (CC-
33A) was used to attach the strain gauges. The adhesive glue can withstand heat up to
80 oC as heat dissipation was an important parameter for data collection (Radhakrishna
& Gurmukhdas 2013).
79
An experimental setup flow for strain and acceleration data collection is shown
in Figure 3.4. The strain gauge sensor was attached at the coil spring hotspot while
another end of the strain gauge was connected to a data acquisition system. Gauge factor
of 2.07 were set into the Somat data acquisition software. On the other hand,
acceleration time histories of the vehicle were measured using single axis
accelerometers. One accelerometer was attached at the lower arm of the vehicle to
collect un-sprung mass vibration while another accelerometer was attached at the top
mount of the suspension strut to collect sprung mass vibration. Both accelerometers
were attached to the components using a strong adhesive material.
The accelerometers were then connected to a data acquisition system NI 9234
and the data were previewed and recorded using NI LabVIEW®. Once the instruments
were installed, the vehicle was driven across five different road conditions to collect
both the strain and acceleration time histories. The selected area for data measurement
were highway, UKM campus, hilly, residential, and rural road surfaces. During the
testing, the vehicle speeds varied across the different road conditions as listed in Table
3.3. Although the driver tried to maintain constant speed for all road conditions, there
were some uncertainties on the road, such as sudden vehicle lane changes or extreme
rough road surfaces which lead to varying vehicle speed.
For the selection of sensor locations, the accelerometer was attached as close as
possible to the top mount to collect the transmitted vibration response while the
accelerometer at the lower arm was collected at the lower arm which close to the damper
strut for un-sprung mass excitations. Meanwhile, the strain gauge attachment point was
determined based on the FEA stress hot-spot referred to Section 3.2.1. The
accelerometers and strain gauge were calibrated according to the calibration certificate
and the initial value was reset into zero before experiment. The proposed vehicle speed
was 80 kmph according to RARS but due to the harsh road conditions, the driver has to
reduce the vehicle speed which lead to the varying vehicle speed in Table 3.3.
During the vehicle test, various road conditions in Malaysia were identified for
data collection with the route as shown in Figure 3.5 – 3.9. The road conditions were
selected based on the common usage of Malaysians. Different road profiles contributed
80
to distinct fatigue behaviour of the automotive components. For example, Ozmen et al.
(2015) collected data from a harsh road using different leaf spring to develop a new in-
lab fatigue test bench. In addition, Kubo et al. (2015) investigated the influence of shock
absorbers on leaf spring fatigue behaviour using pothole tracks. These literatures
justified the significance of road profile effects on automotive durability analysis
Figure 3.4 Experimental setup for strain and acceleration signals collection
Accelerometer
81
Table 3.3 Vehicle speed during signal collection
Selected Road Vehicle speed (kmph)
Highway 80 – 100
UKM campus 60 – 80
Hilly 40 – 80
Residential 40 – 50
Rural 30 – 40
(a) (b)
Figure 3.5 Highway road conditions: (a) road preview, (b) route map
(a) (b)
Figure 3.6 UKM campus road conditions: (a) road preview, (b) route map
(a) (b)
Figure 3.7 Hilly road conditions: (a) road preview, (b) route map
82
(a) (b)
Figure 3.8 Residential road conditions: (a) road preview, (b) route map
(a) (b)
Figure 3.9 Rural road conditions: (a) road preview, (b) route map
3.2.3 Characterisation of Fatigue and ISO 2631 Vertical Vibration
After data collection, both collected strain and acceleration signals were characterised
to seek for the current spring design durability and performance. In order to understand
the characteristics of time histories, statistical properties of the signal were analysed
because the fatigue and vibration time histories were generally non-stationary in nature.
In this case, four commonly used statistical parameters in fatigue analysis namely the
mean value, standard deviation (SD), root mean square (r.m.s), and kurtosis were
applied and listed in Equations 2.36, 2.37, 2.38 and 2.39. The mean value of a strain
signal identified the vibrations experienced by the coil spring, either in tensile or
compression depending on the positive or negative signs of the mean value.
83
The statistical parameters were obtained using the nCode GlyphWorks®
software with the process flow as shown in Figure 3.10. In addition, the acceleration
signals were analysed in the frequency domain utilising the fast Fourier transform (FFT)
and power spectral density (PSD). Frequency spectrum setup in nCode GlyphWorks®
was used to determine the FFT and PSD of the signals and the flowchart for spectrum
analysis is shown in Figure 3.11. Frequency domain analysis was performed to
determine the energy content inside a time history where these energy contents were
related to vehicle ride and spring fatigue life.
Figure 3.10 Flowchart for statistical analysis using Glyphworks®
The fatigue analysis of the spring using nCode GlyphWorks® was conducted as
shown in Figure 3.12. The GlyphWorks® strain life fatigue assessment predicted the
fatigue life of the spring using material cyclic properties and strain variable amplitude
loadings (VAL) as inputs. All three Coffin-Manson, Morrow and Smith-Watson-
Topper (SWT) approaches were used to predict the fatigue life of the coil spring because
these models considered different mean stress effects. The cyclic hardening effects were
also considered when applying the Ramberg-Osgood relationship and Masing’s model
which could be referred to Equations 2.12 and 2.13. The Ramberg-Osgood and
Start
Input strain or acceleration time
histories
Select statistic function block
Calculate the statistic parameters
Display the results
Stop
Are the results within
acceptable range?
Yes
No
84
Masing’s model considered the total strain (elastic and plastic) for fatigue life
assessment.
Figure 3.11 Flowchart for spectrum analysis using Glyphworks®
Applying the strain life approaches in this analysis is suitable because coil spring
was a small component. When a crack was initiated on the spring, the spring was
considered as failed and replacement must be immediately done. In addition, application
of strain life approaches to predict fatigue life was common in automotive components
such as spring (Karthik et al. 2012), disc brake (Pevec et al. 2014), and truck cab (Fang
et al. 2015). Although both stress and strain approaches are considered localisation
effects, stress-life approaches had a very high safety factor and was not suitable for
components that could be easily replaced. Hence, strain life approach was selected to
perform fatigue life prediction for coil springs.
After analysing the fatigue life of spring using measured strain time histories,
the ISO 2631 vertical vibration of the vehicle associated with the coil spring was
obtained using Matlab® R2015a software package. The procedures to obtain the ISO
Start
Input strain or acceleration time
histories
Select spectrum analysis block
Calculate the PSD of time histories
Display and save the results
Stop
Is the PSD range
positive?
Yes
No
85
2631 objective vertical vibration are shown in Figure 3.13. Firstly, the collected
acceleration signals were transformed into frequency domain PSD and a series of
discrete frequencies within the range of interest was selected. Determining the mean
square value of acceleration at a given frequency (fc) requires integration of the obtained
PSDs over a one-third octave band. In other words, the centre frequency band of 0.89 –
1.12 fc was applied where the r.m.s value of acceleration at each centre frequency, fc
was obtained. The Sv(f) was the amplitude of the acceleration PSD for vehicle sprung
mass while ar.m.s was the r.m.s acceleration.
Figure 3.12 Flowchart for nCode Glyphwork®-based strain life fatigue assessment
Subsequently, the ISO 2631 weighting factors were obtained by multiplying the
acceleration amplitude of one third centre frequencies with the amplitude of ISO 2631
proposed vertical weighting factors. In Figure 3.13, Aw is the frequency weighted r.m.s
acceleration, Wi is the weighting factor for the ith one-third octave band, and ai is the
r.m.s acceleration for the ith one-third octave band. The obtained Aw were used to assess
the comfort level of the vehicle using the guide proposed by ISO 2631 as listed in Table
3.4 while the applied weighting factors proposed by ISO 2631 (1997) are listed in Table
3.5. According to ISO 2631, the frequency of interest for human ranged from 0.1 to 80
Start
Input strain time history
Calculate fatigue damage and life
Display and save the results
Stop
Select strain-life function block
Is the fatigue life > 1
and damage < 1?
Yes
No
86
Hz. In this case, the vertical weighting factors, wk from Table 3.5 was applied in the one
third octave band to obtain the vibration index.
Figure 3.13 Procedures for ISO 2631 vibration assessment
Table 3.4 Guide for assessing the effects of vibration on comfort
Range of weighted acceleration Comfort level
Less than 0.315 m/s2 Not uncomfortable
0.315 – 0.63 m/s2 A little uncomfortable
0.5 – 1 m/s2 Fairly uncomfortable
0.8 – 1.6 m/s2 Uncomfortable
1.25 – 2.55 m/s2 Very uncomfortable
Greater than 2 m/s2 Extremely uncomfortable
Source: ISO 2631-1 1997
Measured acceleration
Transform the acceleration time histories into PSDs
Transform PSD into one-third octave band
𝑎𝑟𝑚𝑠 = ቌ න 𝑆𝑣ሺ𝑓ሻ𝑑𝑓
1.12𝑓𝑐
0.89𝑓𝑐
ቍ
12
Apply ISO 2631 vertical weighting factors
𝑎𝑤 = ሺ𝑊𝑖𝑎𝑖ሻ2
𝑖
൩
12
Apply integration
Obtain vertical vibration index
Analyse the vibration index
Start
Stop
87
Table 3.5 Frequency-weighting curves for principal weighting
Frequency, Hz Wk
0.100 31.2
0.125 48.6
0.160 79.0
0.200 121
0.250 182
0.315 263
0.400 352
0.500 418
0.630 459
0.800 477
1.000 482
1.250 484
1.600 494
2.000 531
2.500 631
3.150 804
4.000 967
5.000 1039
6.300 1054
8.000 1036
10.000 988
12.500 902
16.000 768
20.000 636
25.000 513
31.500 405
40.000 314
50.000 246
63.000 186
80.000 132
Source: ISO 2631-1 1997
The weighting factors of ISO 2631 (1997) were regarded as the most
comprehensive standard in which the lowest frequency was 0.1 Hz. The frequency from
0.1 to 0.5 Hz covered the motion sickness of the passengers. Upon the completion of
the ride related vertical vibration and fatigue analysis, linear regression was performed
to determine the relationship between these two parameters under this fixed suspension
system configuration so that the fatigue life could be expressed in terms of ISO 2631
vertical vibration. The effective method to perform these tasks was through a simple
regression method.
There were many existing types of regression approaches for engineering
application. Nevertheless, power law was commonly applied in fatigue analyses of
materials (Amin Yavari et al. 2016). A power law is a functional relationship between
88
two quantities in which a relative change in one quality results in a relative change of
the other quantity as a power of another, independent of the initial size of those
quantities. There were also a few types of nonlinear power curve characterisations, such
as polynomial, exponential, cubic and approximate cubic. Still, power law is the most
commonly used for fatigue analysis (Carrillo et al. 2013).
The relationship between fatigue damage and ISO 2631 vertical vibration was
determined using power law because the fatigue level of material was nonlinear in
nature. The applied power law regression possessed the relationship between two
parameters in the form as follows (Amitrano 2012):
𝑦 = 𝑎𝑥𝑏 (3.1)
where a and b are coefficients with fitted least squares and defined as below:
b = n ∑ (ln xi - ln y
i)n
i=1 − ∑ ሺln xiሻni=1 ∑ (ln y
i)n
i=1
n ∑[ሺln xሻ2] − ሺ∑ ln xሻ2 (3.2)
a = ∑ሺln yሻ − b ∑ሺln xሻ
n (3.3)
Here, n is the number of data, x is the spring fatigue life, and y is the ISO 2631 vertical
acceleration.
In determining the goodness of fitting of the model, coefficient of determination
(R2) was used. The R2 was used to determine the proportion of the variance in the
dependent variable that was predicted from the independent variable and defined as
follows:
R2= [∑ ln xi ln y
i-
1n
∑ ln xi ln yi]
2
[- ሺ∑ ln xiሻ2
n] (ln y
i)
2-
(∑ ln yi)
2
n൩
(3.4)
89
Apart from R2, MSE was used to find the line of best fit. The classification of goodness
of fit using R2 value is listed in Table 3.6. Adopting the power law curve fitting, the
dependent variable was the weighted ISO 2631 vertical acceleration while the
independent variable was the fatigue life. The data were fitted using Matlab® curve
fitting tool. These predictions were limited to a single spring fatigue life and ISO 2631
weighted vertical acceleration without the consideration of suspension parameters
effects. In actual, different coil spring designs exhibit different effects in spring
durability and vehicle vertical vibration characteristics.
Table 3.6 Suitability of fit for coefficient of determination value
Obtained R2 value Goodness of fitting
0.90 Very good
0.80 Good
0.60 Acceptable
Source: Sivák and Ostertagová 2012
The novelty of this research was to determine the relationship between spring
fatigue life and ISO 2631 vertical vibration. Nevertheless, the suitability of the fitting
was important because it affected the consistency of predictions. This proposed power
law regression only took a single parameter which was the ISO 2631 vertical vibration
as the independent variable. In consideration of the spring design, a vehicle quarter car
model was constructed to extract the effects of spring stiffness. A regression method
with multiple inputs were used to build the relationship between these parameters.
3.3 ESTABLISHMENT OF MULTIPLE LINEAR REGRESSION
3.3.1 Generating Artificial Road Profiles
To simulate a quarter car model, realistic road profiles were required to simulate actual
road conditions. However, the measurement data of road conditions was very limited
due to time constraints. In agreement with ISO 8608 road profiles, it was possible to
generate road profiles with stochastic representation in the form of PSD of vertical
displacements. Fundamental idea of the ISO 8608 was spatial frequency, road profile
and PSD. ISO 8608 proposed the description of the road roughness profile through the
calculation of PSD of vertical displacement as a function of spatial frequency. For
90
comparing different road roughness profiles, a classification based on their PSD was
conducted by calculating the conventional values of spatial frequency.
A total of four classes of road roughness were generated according to ISO 8608
as shown in Figure 3.14 and the parameters are defined in Table 3.7. The road profiles
were regarded as simple harmonic function and geometric mean values of the road
roughness profile were used to formulate the road profile according to road roughness
and waviness functions as listed in Table 3.8 and 3.9, respectively. The random phase
angle was assumed to follow a uniform probabilistic distribution within the 0 - 2π range.
The road profile was then converted into time domain with consideration of vehicle
speed. The waviness function was programmed into Matlab to construct the different
classes of road profile under a vehicle speed of 80 kmph because this speed was the
standard value for road roughness (Kropac & Mucka 2007).
Figure 3.14 Procedures for ISO 8608 road profile generation
Determine the required road class
Generate the PSD according to degree of roughness
as listed in Table 3.8
𝐺𝑑ሺ𝑛ሻ = 𝐺𝑑ሺ𝑛𝑜ሻ. ൬𝑛
𝑛𝑜൰
−2
Start
Stop
Applied the waviness function with k value in Table 3.9
ℎሺ𝑥ሻ = ξ𝛥𝑛. 2𝑘 . 10−3. ቀ𝑛𝑜
𝑖. 𝛥𝑛ቁ . 𝑐𝑜𝑠ሺ2𝜋. 𝑛𝑖. 𝑥 + 𝜑ሻ
𝑁
𝑖=0
Convert spatial to temporal domain
𝑆𝑔ሺ𝑡ሻ = 𝑆𝑑ሺ𝑑ሻ
𝑉
91
According to the ISO 8608 standard, four of ISO 8608 road classes (Class A to
D) were generated for sprung mass acceleration and spring force time histories
extraction. These road classes were selected due to their suitability for automobile
application based on reasonable roughness, as reported by Agostinacchio et al. (2014),
Patil et al. (2016), and Koulocheris et al. (2016). Road classes “A” and “B” were
classified as smooth runway and highway while road class “C” was gravel road. The
class “D” was the rough runway which had a very rough surface profile (Balmos et al.
2014). The researches have reported that road classes of E and above were too harsh for
automobile analysis. Hence, road classes A to D were generated and applied for quarter
car model simulations.
Table 3.7 Parameter for ISO 8608 road profile generation
Parameter Definition
Gd(no) Degree of roughness value as listed in Table 3.7
no Spatial frequency
n Road length data point
h(x) Amplitude of the vertical displacement
Ai Amplitude which obtained from the mean square value of the component for the
spatial frequency
𝜑 Phase angle
ni Generic spatial frequency
x Abscissa of road length from 0 to L, Δn = 1/L, N = L/B
B Sampling frequency
k Constant value depending on ISO 8608 road classification as depicted in Table 3.8
Sg(d) Road profile in terms of road length
V Vehicle speed
Source: ISO 8608 2016
Table 3.8 Classification of road roughness proposed by ISO 8608
Road Class Degree of roughness
Lower limit Upper limit Geometric mean
A 8 32 16
B 32 128 80
C 128 512 320
D 512 2048 1280
Source: ISO 8608 (2016)
Table 3.9 k values for ISO 8608 road roughness classification
Road class k value
A 2
B 3
C 4
D 5
Source: Agostinacchio et al. (2014)
92
3.3.2 Extract Data from Simulated Quarter Car Model
After obtaining the generated artificial road profile, a vehicle suspension quarter model
was constructed using a modelling methodology based on object orientation and
equations solving. Computer based modelling is a powerful tool to perform analysis
across multiple disciplines. One of this computer-based modelling commercial software
is SimulationX®. SimulationX® is a multibody dynamics commercial software
developed by ITI GmbH. This software was used to model the quarter car model
through a simple graphical user interface. Function blocks of the component were
dragged and connected in a diagram view. Furthermore, movement of the vehicle
system was visualised in a 3D view.
For practical purpose, a vehicle quarter car model was constructed to study the
dynamic interaction between vehicle and road roughness profile. The generated quarter
car model was including the kinematic properties because the compression of spring
was not directly perpendicular to the ground as shown in Figure 3.15 (Balike, Rakheja
& Stiharu 2013). Variation of the mass values, spring stiffness and damping properties
of the quarter car model made it possible to model any type of ground vehicle: car, truck
or bus using these equations of motion (Agostinacchio et al. 2014). The vehicle’s
response along a road section with uniformly distributed roughness was determined by
the study of the model forced oscillations. Due to the oscillations, the vertical force
exchanged with the road was not constant with time but in sinusoidal variation. For this
simulation, SimulationX® was selected because of its capability to simulate these
stochastic road profiles as the amplitude changed drastically.
For the visualisation of the quarter car model in SimulationX®, a block diagram
view of the linear three-dimensional (3D) quarter car model was built and shown in
Figure 3.15. From Figure 3.15(a), a total of 11 main blocks inside SimulationX®
interface were connected to build the quarter car model for mimicking the selected
Macpherson suspension system. Functions of the blocks are tabulated into Table 3.10.
Block 1 was a signal source with curve sets as input. It defined any types of excitation
profiles in the simulation. In this case, the generated ISO 8606 road profiles were
applied in this block. Block 2 was a spatial spring-damper element. The spring force Fs,
93
the damper force, Fd, and the internal force, Fi were computed for each spatial direction.
In this analysis, the Block 2 represented the tyre and a stiffness value of 105,000 N/m
was assigned (Wu et al. 2015). Block 3 was a set of cylinder element which was
visualised as a cylinder, hollow cylinder or pipe. This component was used to model
the lower arm in 3D view as shown in Figure 3.15(b). Movement of the wheel was
visualised across the simulation period to ensure the correct mechanism of the
suspension model. Block 4 and 5 were cylinder elements which represented the tie rod
attachments. Block 6 was a linear translation mass which modelled the knuckle.
Block 7 was also a spring element which was implemented with elastic
behaviour for coil spring model while Block 8 is force-damping element for damper
model. Block 8 represented the spring and damper of the suspension system in which
the damping coefficient was 2000 Ns/m for compression and 6000 Ns/m for rebound
(Jugulkar et al. 2016). The internal forces Fi, the spring force Fs, and the damper force
Fd were computed using the following formulas:
𝐹𝑠 = 𝑘. 𝛥𝑥 (3.5)
𝐹𝑑 = 𝑏. 𝛥𝑣 (3.6)
𝐹𝑖 = 𝐹𝑠 + 𝐹𝑑 (3.7)
where b is the damping coefficient, k is the spring stiffness, x is the displacement, and
v is the velocity. The change of potential energy in this element was calculated using
the equation as follows:
𝑃𝑝 = 𝐹𝑠 . 𝛥𝑣 (3.8)
The power loss of the damper Pl is calculated as follows:
𝑃𝑙 = 𝐹𝑑 . 𝛥𝑣 (3.9)
94
(a)
(b)
Figure 3.15 Setup of quarter car model in SimulationX®: (a) diagram view, (b) 3D view
Table 3.10 Function of each block in quarter car model
Blocks Function
1 Input the road profile
2 Input the tyre stiffness
3 Define the lower arm connectivity
4 Define the tie rod part 1
5 Define the tie rod part 2
6 Define the knuckle
7 Input the spring stiffness
8 Input the damping coefficients
9 Define the damper rod
10 Define the damper tube
11 Input the sprung mass
95
Block 9 and 10 were two cylindrical elements which modelled the damper road
and tube for visualisation purpose. In this case, all the cylindrical elements served to
define the weights of un-sprung mass. The total wheel mass of the suspension system
considered the geometry of rim, brake calliper, disc brake, hub, knuckle and lower arm
using CAD with material density as shown in Figure 3.16. The mass of the components
in Figure 3.16(a) – (e) was the input to Block 6 while Figure 3.16(f) was the input to
Block 3. With the density of the material as the input, mass of these components was
obtained. On the other hand, sprung mass of 350 kg was applied into Block 11. Two
spherical joints and one revolute joint were defined in the model to connect the knuckle,
lower arm and tie rod. The prismatic joints which allowed only translation movement
were assigned on sprung mass and tie rod. Upon completion of quarter car model setup,
the model was ready for a dynamic simulation.
Figure 3.16 CAD of Macpherson un-sprung mass components: (a) rim, (b) brake calliper, (c) disc
brake, (d) hub, (e) knuckle, (f) lower arm
(a) (b)
(c) (d)
(e) (f)
96
After determining the quarter car model parameters, the simulations were
performed as shown in Figure 3.17. During the simulation, the model was defined in
forms of ordinary differential equation (ODE) and solved using a numerical algorithm
(Roman et al. 2014). SimulationX® offered the solution of ODE using backward
differentiation formula (BDF) for transient simulation (ITI 2009) while transient
simulation was performed in time domain. BDF solvers were suitable for non-stiff or
stiff model with eigenfrequencies in wide range (Zheng & Zhou 2014). This algorithm
is a predictor-corrector method with automatic control of step size and order. With this
kind of implicit multistep method, the current value was extrapolated from every state
variable (1 ≤ k ≤ 6).
Figure 3.17 Procedures for quarter car model simulations
Select the suitable function blocks
Connect the function blocks according to suspension
system configuration
Start
Stop
Input the road profile, tyre stiffness, sprung and un-sprung mass
weights, spring stiffness and damping coefficient
Request force and acceleration output from
spring and sprung mass
Define the road-wheel contacts
Run transient simulation
Check the kinematic of the suspension system
Extract and save the time history data
97
The model was simulated for a total duration of 90 s to ensure the signals were
sufficient to provide information on ISO 2631 vertical vibration and fatigue analysis.
The acceleration time histories of the vehicle body were extracted based on different
spring stiffness and damping coefficient. The optimal suspension stiffness for passenger
car ranged from 1 to 1.5 Hz (Shirahatt et al. 2008) while the suspension stiffness for
race car ranged from 2 to 2.5 Hz, as proposed by Sun et al. (2014). However, the
suspension natural frequencies for racing purpose were built to the range of 5 to 7 Hz.
In this case, the quarter car mass for this passenger car with Macpherson suspension
system was 300 kg. The original simulation of quarter car model was based on the actual
spring stiffness but this was restricted to a fixed value. In order to incorporate the spring
design into fatigue analysis, spring stiffness parameter sensitivities analysis was
required to determine the effect of spring design towards fatigue life and ISO 2631
vertical vibration.
3.3.3 Perform Spring Stiffness Sensitivities
Changing spring design usually involved spring geometry such as bar diameter, outer
diameter or number of active coils. The simplest way to obtain different spring stiffness
was by changing the bar size as different outer diameter and number of active coils was
usually done to revise the allocated space for suspension system in the chassis.
Changing the spring diameter increased the spring width while changing the number of
active coil increased the spring length under fixed material properties. Hence, changing
the bar size was the best option to find different spring stiffness.
In this analysis, spring stiffness parameter sensitivity analysis was performed to
seek for the effects of spring stiffness towards fatigue life and ISO 2631 vertical
vibration. The difference between the new designed springs were required to be as close
as possible to the calculations. Zhang et al. (2007) proposed that deviation below 10 %
is good while deviation below 20 % is acceptable. The spring stiffness was subsequently
decided based on the natural frequency of the suspension system using the modal
equation as below (Chen et al. 2014):
98
𝑓𝑛 = 1
2𝜋√
k𝑟
𝑚 (3.10)
where kr is the ride rate, m is the distributed vehicle mass on the spring, and fn is the
vertical vibration of vehicle body frequency. The kr value was obtained based on the
spring stiffness of FEA and constant tyre stiffness calculations as stated in Equation 2.1.
The original spring stiffness of the case study vehicle was 20,000 N/m and the
calculated natural frequency of 1.3 Hz was obtained. A series of spring variants for
passenger car spring stiffness were designed from 14,000 to 32,000 N/m to achieve the
designated frequency range of 1 to 1.5 Hz in the quest to examine the effects of spring
stiffness or natural frequency towards the ISO 2631 vertical vibration and spring fatigue
life. The option to alter the spring stiffness was either through using different spring
materials or distinctive design. In the current spring industry, most of the springs were
made of spring steel and alternative materials are not feasible for commercial use
(Abidin et al. 2013). Hence, the spring design changes were considered to achieve
various stiffness levels.
After the bar sizes for different spring stiffness were obtained, the CAD models
of spring for every size were prepared and processed with FEA. The FEA results of the
springs were used for fatigue life prediction. The setup of fatigue life prediction using
nCode DesignLife® and force time histories from the quarter car model are shown in
Figure 3.18. In Figure 3.18, there were 6 blocks used in setting up the fatigue simulation.
Block 1 was an FE input in which the FEA stress-strain simulation results of the springs
were applied. In the FE simulation, static force was applied so that the loadings were
scaled in DesignLife® for fatigue analysis. The scalar loading for scaling is shown in
block 2 as the force time histories were obtained from the quarter car model simulation.
Block 3 was a function block to perform strain-life fatigue analysis.
During the fatigue analysis, all three uni-axial strain life fatigue approaches
namely the Coffin-Manson, Morrow, and SWT were applied. Block 4 was the FE output
from the strain life fatigue simulation. This block contained the fatigue information of
each node of the spring models. The simulation results were visualised in Block 5.
99
Block 6 was the data value displaying the fatigue simulation results. Meanwhile, the
acceleration time histories of the vehicle sprung mass with different spring stiffness
were also extracted from quarter car model and processed into ISO 2631 vertical
vibration. For feasibility of model construction, the spring stiffness was converted into
suspension natural frequency based on the vehicle mass using Equations 2.1 and 3.10
so that the established model could be used for all types of ground vehicle. Once all the
required parameters were ready, the modelling process was subsequently conducted.
Figure 3.18 Spring fatigue life prediction using nCode DesignLife®
3.3.4 Establishment of Multiple Linear Regression for Fatigue Life
Multiple linear regression (MLR) is a type of regression approach with several
independent variables. In order to establish a relationship with more than one
independent variable, MLR method is one of the options to deal with multiple
independent variables as it is available in IBM® SPSS® Statistic software package. This
approach was different from the power law regression which only involved one
parameter. Applying the MLR method required a few assumptions (Krzanowski 2010).
Firstly, the dependent variable was measured on a continuous scale. The continuous
measurement involved any observation that fell anywhere on a continuum, such as
displacement or time. Secondly, two or more independent variables were continuous or
categorical. Categorical is also known as discrete as the data are obtained through
100
counting. Subsequently, a linear relationship needed to be established between the
dependent variable and the independent variable. Performing the MLR analysis
involved entering all 90 datasets and input into SPSS® data editor. The details for MLR
generation is shown in Figure 3.19.
Figure 3.19 Process flow for establishment of multiple linear regressions
Initially, the dependent and independent variables were determined. The
regression with the fatigue life as the dependent variable is known as “vibration-life”
while the regression with ISO 2631 vertical vibration as the dependent variable is
known as “life-vibration” regression. As for the vibration-life regression, the dependent
variable was spring fatigue life while the independent variables were vehicle body
Input the suspension frequencies, spring fatigue
lives and ISO 2631 vertical vibrations into SPSS®
Define the dependent and independent variables
Start
Stop
Request regression coefficients, model fit, scatter
plot, normal probability plot, error histogram
Run the multiple linear regression analysis according least
square estimate
𝐵 =
ێێێێێۍ𝛽0
𝛽1
.
.
.𝛽𝑘ۑ
ۑۑۑۑې
= ሺ𝑋′𝑋ሻ−1𝑋′𝑌
where B is the least square estimate, X’ is the transpose of X, Y
is the predicted value
Use the model summary to fit the regression into form:
𝑦𝑖 = 𝛽𝑜 + 𝑥𝑖𝛽1+. . . + 𝑥𝑖𝑘𝛽𝑘 + 𝑒𝑖 i = 1,2, …,n
where 𝛽𝑜 is a constant, xk is the regression coefficient, 𝛽𝑘 are the
variables, ei are the residuals.
101
frequency and ISO 2631 vertical vibration. As for life-vibration regression, the ISO
2631 vertical vibration was set as the dependent variable while the vehicle body
frequency and spring fatigue life were determined as the independent variables.
Once the multiple linear regressions were obtained, the R2 value for the
regression was examined to ensure the data fitted nicely. The limitation of this approach
was the linearity of the datasets where nonlinear data were not applicable. In addition,
this approach was very straightforward as the regression was obtained using a least
square estimate method (Krzanowski 2010). A more advance artificial intelligence
method known as the artificial neural network (ANN) was proposed to enhance the
prediction outcome.
3.4 PREDICITNG THE SPRING FATIGUE LIFE
3.4.1 Design the Artificial Neural Network Architecture
ANN is a supervised machine learning method to find the best iteration. The matrix
converged until the defined conditions were satisfied. In some circumstance, ANN
approach provided better prediction results than MLR method (Abyaneh 2014). The
difference between ANN and MLR method in determining the spring fatigue and ISO
2631 vertical vibration was tested using the same datasets to train an ANN architecture.
Before the training process, suitable types of ANN were determined. ANN with single
hidden layer was classified as a simple feedforward neural network and trained using
Matlab® ANN toolbox. ANN with more than one hidden layer was known as multilayer
perceptron neural network (Bogdan et al. 2011).
The multilayer perceptron neural network could be divided into feedforward and
hybrid neural network. The major difference between these two types of ANN was the
connectivity to the input neurons. The output neurons of hybrid neural networks were
connected to the front hidden layers and input neurons while feedforward output layer
was only connected to the prior layer. The hybrid neural network was applied because
of optimal time and acceptable accuracy of the method as compared to feedback neural
network (Welch et al. 2009). After determining the suitable type of neural network, the
102
first step to train a neural network was deciding the architecture of the neural network
in terms of neurons and hidden layer. Random selection of a few hidden neurons and
layers might cause either overfitting or under fitting (Ghana Sheela & Deepa 2013).
However, there was no established rule for neuron network architecture design and the
conventional way to determine the suitable neuron was based on trial and error.
There were a few proposals to determine the suitable number of neurons based
on mean square error (MSE). The neuron with the lowest MSE was usually determined
to be the most optimum architecture for a neural network (Salim et al. 2015). In this
analysis, the architecture of different ANN was constructed for MSE analysis. The
“perform” function in Matlab® was applied to the ANN test data in order to call the
MSE for ANN analysis. The MSE function is written as follows:
MSE = 1
n (f
i− y
i)
2
n
i=1
(3.11)
where n is the number of samples, and fi is the estimation of yi. The performance of the
neural network and MSE were plotted against the number of neurons to determine the
most suitable ANN architecture (Salim et al. 2015).
Adjusting the proper number of hidden layers and neurons is an empirical
exercise for every analysis (Lopez-Moreno et al. 2014). Increasing the number of
neurons was not necessary to obtain a better performance (Ghana Sheela & Deepa,
2013). Similar conditions applied to a number of hidden layers. Hence, an analysis to
determine the optimum number of neurons and hidden layers for neural network
architecture was performed in Matlab® environment. A “for” loop was used in Matlab
to continuously search for the optimum ANN architecture with minimum MSE. The
flowcharts for the searching loop were listed in Figures 3.20 – 3.22. Figure 3.20 shows
the loop for simple feedforward ANN with single hidden layer while Figures 3.21 – 22
depict the hybrid ANN with two and three hidden layers, respectively.
ANN with one to three hidden layers were analysed to determine the suitable
architecture with the lowest MSE. In Figure 3.20, the flowchart for determine the single
103
hidden layer ANN for fatigue life prediction is illustrated where nN1
was the number of
neurons to be applied in the hidden layer. A maximum of ten neurons was assigned
because the number of input and output was only three (Heaton 2015), as this number
of neurons provided the suitable results within the acceptable time frame (Salim et al.
2015). In addition, the proposed suitable number of neurons to train an ANN was the
mean value between input and output which indicated two neurons were sufficient in
training the ANN (Heaton 2015). For ANN with one hidden layer, the “for” loop was
applied to calculate the MSE of neuron number from one to ten and the loop was
terminated after reaching ten neurons. The MSE was calculated using Equation 2.44
where the ANN predictions were the output and the fatigue life data used to train the
ANN were the target. With the lowest MSE value, the suitable ANN architecture to
determine fatigue life was proposed.
Figure 3.20 Flowchart to determine single hidden layer ANN architecture with the lowest MSE
In Figure 3.21, an additional parameter nL was defined to add the number of
hidden layers. The input for constructing the two hidden layers ANN were number of
hidden layers and neurons (nL, nN
). The maximum number of neurons in the first and
Start
Set nL
nN1
≤ 10?
Stop
nL = n
L + 1
Yes
No
Determine fatigue life using ANN
Calculate MSE
Suggest suitable ANN
104
second hidden layer (nN2 & n
N1) were set to be ten and the maximum number of hidden
layer (nL) was determined as two. 100 sets of ANN architecture were simulated and the
ANN architecture with the lowest MSE were selected for further analysis. The output
of the loop process was the suitable ANN with two hidden layers for fatigue life
predictions.
Figure 3.21 Flowchart to determine two hidden layer ANN architecture with the lowest MSE
Start
nL = 2
Stop
Set new nN1
Set new nN2
NN2
= nN2
+ 1
NN1
= 10?
NN2
= 10?
NN1
= nN1
+ 1
No
Yes
No
Calculate MSE
Suggest suitable ANN
Set nN1
= 1
Set nN2
= 1
Determine fatigue life using ANN
Yes
105
After the two-hidden layer, ANN architecture was expanded to three hidden
layers. The process to find the lowest MSE ANN architecture with three hidden layers
is listed in Figure 3.22. In this case, the maximum hidden layer (nL) was set as three and
the maximum number of ANN hidden layers of this research was only three because
three hidden layers were able to handle many of the complex analysis with suitable
computational power (Bozorg-Haddad et al. 2016). Each hidden layer was also looped
up to ten neurons with a total of 1000 sets of ANN architectures to obtain a single
durability regression. The MSE for three hidden layers were determined and the lowest
MSE ANN architecture was analysed using statistical method. As the outcome, a most
suitable three hidden layers ANN for fatigue life prediction was accordingly obtained
for Coffin-Manson, Morrow, SWT strain life approaches.
Despite the possibility of better results by increasing the number of layers and
neurons, the computational time was shown to increase exponentially (Lopez-Moreno
et al. 2014). The computational power needed for ANN architectures was heavy because
there were many data and repeated simulations. Thus, a fast and appropriate architecture
as well as a training algorithm was needed. For training algorithm, Levenberg-Marquadt
algorithm was established as one of the quick and accurate algorithms for ANN training
(Cui et al. 2017; Kuruvilla & Gunavathi 2014). Levenberg-Marquadt algorithm
computed the solution through calculation of Jacobian matrix (Cui et al. 2017). Due to
its accuracy, this algorithm was used to train all the required ANN.
During the training process, the weights and biases were randomly generated
according to Nguyen-Widrow layer initialization function. After each training, the
information was applied to adjust the new weights and biases. The iterations converged
until no improvement was obtained. Hence, the hybrid ANN was optimised through this
back-propagation algorithm (Ibrahimy et al. 2013). For the vibration-life ANN,
suspension natural frequency and vertical vibration were used as inputs while the spring
fatigue life was the output. Contradictorily for vibration prediction ANN, the inputs
were vehicle body frequency and spring fatigue life while the output was ISO 2631
vertical vibration. The outcome of the ANN predicted vibration was also analysed using
error histogram and Lilliefors test.
106
Figure 3.22 Flowchart to determine three hidden layer ANN architecture with the lowest MSE
Start
nL = 3
Stop
Set new nN1
Set new nN2
NN2
= nN2
+ 1
NN1
= 10?
NN2
= 10?
NN1
= nN1
+ 1
Yes
No No
No
Calculate MSE
Suggest suitable ANN
Set nN1
= 1
Set nN2
= 1
Set new nN3
Set nN3
= 1
NN1
= 10?
NN3
= nN3
+ 1
Yes
Yes
Determine fatigue life using ANN
107
In the hidden layer, each neuron consisted of a weight and bias value. Before
proceeding to process output, the weighted value of the hidden layer was activated with
a transfer function. A hyperbolic tangent sigmoid (tansig) transfer function was applied
because of the importance of the training speed (Dorofki et al. 2012). The hyperbolic
tangent sigmoid transfer function is suitable for nonlinear function fitting ANN
(Dorofki et al. 2012) as the function is written as follows:
Sf = 2
ሺ1 + e-2nሻ - 1 (3.12)
where Sf is the hyperbolic tangent sigmoid activation function, and n is the mathematical
function of neuron.
After the suitable ANN architecture was determined, it was trained for the
durability predictions. The simulation data were randomised and split into training,
validation and testing datasets so as to facilitate training and testing of the ANN.
Training purposes used 70% of the data while 15% of the data were used for validation
purpose. The remaining 15% of the datasets were used for ANN test. There was no
specified rule for the percentage of data for training and validation purpose, but this
segregation was used for small number of datasets (<100), as mentioned by Chandwani
et al. (2015). Therefore, the separations of datasets based on 70% training, 15%
validation and 15% test were used when the number of datasets was below 100
(Chandwani et al. 2015). In this analysis, the total number of datasets were only 90 and
hence, the ratio of 70:15:15 was applied. The goodness of fitting R2 value of the training
were examined to ensure the trained ANN was good. The R2 of ANN architecture with
the lowest MSE was selected and analysed.
3.4.2 Optimising the Spring Fatigue Life
The ANN design process in Section 3.4.1 aimed to determine the most suitable
architecture in performing the prediction process of spring fatigue life by determining
the lowest MSE. The improvement process for fatigue life using Coffin-Manson ANN
is illustrated in Figure 3.23. The inputs were the suspension frequencies and ISO 2631
108
vertical vibrations while the output was the spring fatigue life. The designed ANN
architecture played the role as the function to optimise the spring fatigue life predictions
by providing the predictions with the lowest MSE. Subsequently, the optimised fatigue
life could be proceeded for validations. For Morrow and SWT fatigue life ANN, the
determined suitable architectures are listed into Appendix C. In addition, the determined
suitable ANN architecture for ISO 2631 vertical vibration predictions are also listed in
Appendix C.
Figure 3.23 ANN architecture for optimised Coffin-Manson vibration-life predictions
3.5 VALIDATION OF THE PREDICTION
3.5.1 Validation using Normality Test
Validating the established multiple linear regression requires accessing the value of R2
which represents the suitability of the data fitting. The validation process was followed
by a series of statistical evaluation methods as listed in Figure 3.24. The key to a
prominent prediction regression was normal distribution of the residuals. For regression
approaches, F-test was performed to test the normality of the residuals to ensure the
parameters have the same standard deviation (Smyth 2002; Palowitch et al. 2016). For
detail analysis, the normality of residuals was interpreted from the normal probability
plot (P-P) and F-test. If the residuals are normally distributed around zero in residual
Spring
fatigue life
ISO 2631-1
weighted
acceleration
Vehicle body
frequency
Input layer Second hidden layer First hidden layer Output layer
109
histogram, the expected value of the error term was zero. Residuals were significant in
MLR analysis because it biased the results by "pulling" or "pushing" the regression line,
thereby leading to biased regression coefficients. Often, excluding just a single extreme
case could yield a completely distinct set of results (Jaccard & Turrisi 2003).
Figure 3.24 Validation of the established multiple linear regression
The t-test was also applied to test the significance of individual independent
variables towards the dependent variable. In addition, multicollinearity was also
checked using tolerance and variance inflation factor (VIF). The proposed tolerance
value should be greater than 0.1 while the VIF value should be less than 10 to avoid
linearity between the independent variables (Mokhtari et al. 2013). When the VIF value
was within this range, the residuals were assumed to have a normal distribution (Ibrahim
Obtained the multiple linear regression
Check the coefficient of determination value
Start
Stop
Check the normality of the regression using normal
Probability-Probability plot
Check the t-test and F-test for parameter
significance
Check the residuals of the predictions
Error histogram Scatter plot
Check the
homoscedasticity
Validate the regression performance
110
2013). For regression validation, the normality of prediction residuals was validated
using F-test, normal P-P plot and VIF from ANOVA. Subsequently, residuals of the
regression predictions were investigated using an error histogram in which the
prediction error range was defined. The residuals were plotted into scatter plots for
homoscedasticity examinations with the standardised predicted value in X-axis while
the standardised residual value in Y-axis.
Different from the regression approach, ANN resorted to perform iteration
process instead of ANOVA analysis. Hence, a different performance validation process
was performed as shown in Figure 3.25. After the ANN training processes were
completed, the trained architecture with the weighting values were obtained in Matlab®
interface. Prior to that, the R2 value for all the separated datasets such as training,
validation and ANN test were analysed and the MSE of the ANN was plotted in a 3D
scatter plot for data visualisation. An error histogram was plotted to observe the
difference of the ANN trained output and target data. Lilliefors test was performed to
examine the normality of the residuals (Razali & Yap 2012). The critical value of
Lilliefors test was obtained and compared to test the statistical value. The null
hypothesis was rejected when the test statistic value was greater than the critical value.
Figure 3.25 Validation of the ANN predictions
Run the ANN analysis for various architecture
Optimise the ANN using separate data
Start
Stop
Examine the R2 value for all datasets
Examine the normality and errors of the predictions
Plot the MSE for all different ANN
111
However, the cross-validation process using the 30% validation and testing data
were only able to indicate how good the data were fitted and used to select the best
trained ANN. The accuracy of the selected ANN was yet to be clarified. This ANN was
then validated using a set of independent experimental data for accuracy and robustness
in predicting the actual fatigue life or ISO 2631 vertical vibration.
3.5.2 Validation using Experimental Data
For validation of the regression, a set of experimental measurement data was applied to
the MLR and ANN approaches. The validation data set was divided into two groups,
which were the strain and acceleration measurement signals from a case study vehicle
and the process is shown in Figure 3.26. The measured strain time histories were used
to calculate the spring fatigue life while the acceleration time histories were used to
obtain the ISO 2631 vertical vibration. For vibration-life validation, the ISO 2631
vertical vibrations were used as input to the MLR and ANN prediction approaches. The
output of the vibration-life approach was the spring fatigue life. The regression-based
relationship compared the fatigue life to experimentally obtained fatigue life using
coefficient of determination and RMSE.
For ANN based predicted fatigue life, a conservation fatigue life correlation
approach known as scatter band method was applied. The conventional scatter band was
constructed with boundary of 1:2 or 2:1 to validate the estimation results in which any
point distributed beyond the boundary was considered as conservative (Kim et al. 2002).
The point distributed beyond the boundary was known as non-conservative. For life-
vibration relationships, the validations were performed using the experimental strain
time histories predicted fatigue life as input. Together with the suspension natural
frequency, the ISO 2631 vertical vibrations were predicted and compared to the
experimental value. The difference between the regression prediction and experimental
results were obtained as follows:
112
Δs = (absሺ𝐴wm − 𝐴weሻ
𝐴we
) × 100% (3.13)
where Δs is the difference in percentage, Awm is the regression predicted ISO 2631
vertical vibration, and Awe is the experimentally obtained ISO 2631 vertical vibration.
Figure 3.26 Process flow for validation of the ANN predictions
For further comparison, root mean square error between experimental and
prediction ISO 2631 vertical vibrations was performed. The formula for RMSE is as
below:
Collected road vibration signals
Collected sprung mass
acceleration time histories
Start
Stop
Input to regression/ANN for
fatigue life predictions
Spring stiffness
ISO 2631 vertical
vibration
Vehicle body
frequency
Check the normality of residuals using
error histogram and Lilliefors test
Plot correlation between predicted and
experimental fatigue lives
113
RMSE = √(Ypred − Yexp)
2
N (3.14)
where Ypred is predicted ISO 2631 vertical vibration, Yexp is experimental ISO 2631
vertical vibration, and N is number of data. The difference between the regression
prediction and experimental results was ensured to be within 20% to indicate good
predictive results (Zhang et al. 2007).
3.6 SUMMARY
This chapter detailed the methodology employed to achieve the three objectives. This
work focused on two main novelties which were generation of multiple input
regression-based durability relationship and establishment of artificial neural network
(ANN) durability approach for predictions. The strain-life durability approaches were
fundamental for fatigue life prediction and combined with ISO 2631 vertical vibration
which indicated the vehicle ride level using multiple linear regression (MLR). A deep
learning hybrid artificial neural network was then proposed to optimise the regression
relationships. For validation, the conventional scatter band and error estimation method
were performed for both approaches.
In the following Chapter IV, the results which reflect the objectives and
novelties of this research are presented. The results are presented according to the
sequence of the three objectives which were characterisation of ISO 2631 vertical
vibration and spring fatigue life, establishment of multiple input regression-based
durability relationships, predictions using ANN and validations of the results. The
results were discussed to provide justifications to the novelty of this research.
114
CHAPTER IV
RESULTS AND DISCUSSION
4.1 INTRODUCTION
This chapter presents the details of the results according to the three research objectives.
The first objective was to establish a regression-based vibration-life relationship for
spring fatigue life and ISO 2631 vertical vibration predictions. In this research, the
preliminary results were the finite element analysis (FEA) of the coil spring under static
compression and torsion loading which were followed by the performance of global
statistical analysis of the experimental collected strain and acceleration time histories.
Next, the collected strain time histories were used to calculate fatigue life while the
acceleration signals were used to obtain ISO 2631 vertical vibration. Subsequently, a
simple power-law regression was applied to study the fatigue life and ISO 2631 vertical
vibration. The critical results that led to first novelty like generated artificial road
profile, quarter car simulation results, spring designs and estimated fatigue life were
subsequently presented and the evaluation of fatigue multiple linear regression-based
relationship which was related to the first objective was established. In addition, a
vibration-life regression was also developed to provide a prediction for ride related ISO
2631 vertical vibration.
Subsequently, the second objective was to optimise the fatigue and ISO 2631
vertical vibration predictions using artificial neural network (ANN). The selection
results of ANN in terms of different architecture were performed where the ANN-based
durability and vibration architecture with the lowest mean square error (MSE) were
determined. The ANN architecture with the improvised prediction outcome was the
second novelty of this research. Lastly, the third objective was to validate the
regression-based and ANN-based prediction approaches. The regression-based
115
vibration-life approaches were correlated to experimental data using coefficient of
determination (R2) while the vibration-life ANN was validated using a conservative 1:2
or 2:1 fatigue life correlation curve. For ISO 2631 vertical vibration prediction, the
difference between predictions and experimental data were analysed using root mean
square errors (RMSE) and coefficient of determination (R2).
4.2 DETERMINING THE DURABILITY CHARACTERISTICS
4.2.1 Finite Element Analysis of Coil Spring
The preliminary analysis of this work was FEA of a case study vehicle coil spring to
investigate on the stress distribution. This analysis was important because the fatigue
failure area corresponded directly to stress where the crack initiation occurred at the
peak stress region (Meneghetti, Guzella & Atzori 2014). A proper finite modelled
spring possessed details of the geometry captured with finite elements. After the static
analysis was performed, the von Mises stress distribution of the spring was provided in
colour stress contours as shown in Figure 4.1. The region with the highest stress level
is shown in red while the lowest stress region is shown in blue and the highest stress
region was distributed at the inner surface of the spring as similar findings in Del Llano-
Vizcaya, Rubio-Gonzalez and Mesmaque (2006). Although the coil spring is a simple
component, but when the torsional loading was exerted, the stress states of the spring
became complicated.
When a coil spring experiences combined axial and torsional loading, the stress
amplitude of the spring changes (Akiniwa et al. 2002). Stress contour of the case study
coil spring with the combination of axial and twisting load is shown in Figure 4.2. When
compared to the uniaxial loading, the critical point of the spring under combined
torsional and compression remained at the same inner region. Under linear condition,
the stress was proportional to the strain in which the strain was also the highest (Lee et
al. 2005). Even though the spring was loaded under multiaxial forces, the critical point
also remained at the same inner region (Zhu et al. 2014; Abidin et al. 2013). Although
a few high stress regions (FEA contour in red) were observed, the highest stress was
found in the element 41517. With the stress-strain FEA results, the fatigue life contour
of the coil spring was obtained as shown in Figure 4.3. The red contour indicates the
116
region with the lowest fatigue life while the blue contour shows the high fatigue life
regime. The region of spring with low fatigue life was the same with the region of high
stress concentration in uniaxial loading because the stress level has a direct effect on
the fatigue life. The localised fatigue life results were extracted from the highest stress
hot-spot which was 1.26 × 105 blocks to failure. The endurance limit of a steel alloy in
a S-N curve was 2 × 106 cycles under cyclic loadings (Závodská et al. 2016). Although
the fatigue life of the spring was obtained from simulation, the simulated fatigue life
was considered as acceptable because it was within the endurance limit.
Figure 4.1 Stress distribution of the coil spring under axial loading
Figure 4.2 Stress distribution of the coil spring under combination of axial and torsional loading
Stress 989.7 MPa
@ element 41517
Stress (MPa)
Stress 985.9 MPa
@ element 41517
117
Figure 4.3 Fatigue life contour of the coil spring
4.2.2 Observing Signal Characteristics
After determining the critical point of the spring using FEA, signal acquisitions were
performed on a case study passenger car by placing a strain gauge on the coil spring
with a sampling rate of 1000 Hz and same sampling frequency was also applied to the
acceleration measurements (Wang & Zhang 2010). The collected signals for both strain
and acceleration from the testing are obtained as shown in Figures 4.4, 4.5 and 4.6,
respectively. The lengths of strain signals are all at 90 s with 90,000 data points, and
they are sufficient for life assessment because it is repeated based on cycles and are
independent of time. The strain characteristics are tabulated into Table 4.1 and the four-
strain signal amplitude was high because of the extremely uneven road surfaces. The
measured strain amplitude of spring was closely related to the road profile because of
the motion ratio between the wheel and the spring (Patil & Sharma 2015). The
displacement of the wheel was directly linked to the spring response and produced an
equivalent strain of the spring surface (Zhang, Suo, Wang 2008). To characterise the
strain measurements, the statistical properties of the strain time histories were obtained
as shown in Table 4.2. As observed from Table 4.2, all the strain signals had a small
mean value. The positive sign of the strain time histories indicated the mean stress was
in tension while the negative sign was compression.
Fatigue life (blocks to failure)
Fatigue life: 1.264
× 105 @ element
41517
118
Table 4.1 Range of the measured strain time histories
Minimum strain amplitude (µε) Maximum strain amplitude (µε)
Highway -519 537
Campus -1357 1368
Hill -1525 1553
Residential -837 1367
Rural -1550 1770
(a)
(b)
(c)
(d)
(e)
Figure 4.4 Strain time histories of spring collected from various road conditions: (a) highway,
(b) campus, (c) hill, (d) residential, (e) rural
119
Table 4.2 Statistics for the measured strain time histories at different type of roads
Highway Campus Hill Residential Rural
Mean (µε) 58 6.1 4.4 3.0 -4.0
r.m.s (µε) 205.5 373.9 288.3 254.7 375.0
Kurtosis 3.0 3.0 3.4 4.4 5.7
During the data collection process, the rural road was observed to be the
roughest with numerous potholes. Hence, the rural road consisted of the highest SD
value, indicating that the strain amplitudes were distributed far away from the mean
value. Due to the small mean value, the r.m.s value of the strain time histories were
almost the same with the SD value. The r.m.s values indicated the vibration energy in
the strain signal where high energy content led to high fatigue damage that leads to the
fatigue failure (Kihm & Delaux 2013)
Kurtosis value was usually applied to analyse the fatigue damage with
explanation by John & Phillip (2012) higher kurtosis value lead to the greater
randomness of the time histories which contributes to higher fatigue damage. When the
kurtosis value of a time history was greater than a value of 3.0, the time history was
regarded as a non-stationary or non-Gaussian signal. The non-stationary time histories
were time varying spectra where the statistical properties, i.e. mean or SD of the time
histories were varying over time. In this case, the rural road strain time histories had
revealed a heavily non-stationary behaviour with a kurtosis value of 5.7. A few of the
sharp peaks were observed which caused the time histories to become non-stationary.
For the vibration analysis, the collected acceleration time histories from the
vehicle lower arm and top mount are depicted in Figures 4.5 and 4.6, respectively.
Acceleration of lower arm was regarded as the excitation of un-sprung mass while the
top mount of the suspension strut was considered as vibration of the vehicle sprung
mass (Heiβing & Ersoy 2011). The vibrations originated from the wheel of the vehicle
and transmitted through the suspension to the vehicle body. To describe the
characteristic of these vibration time histories, global statistical analysis results of the
lower arm and top mount vibration time histories were performed and tabulated into
Tables 4.3 and 4.4.
120
(a)
(b)
(c)
(d)
(e)
Figure 4.5 Time histories of the measured acceleration from lower arm under various roads:
highway, (b) campus, (c) hill, (d) residential, (e) rural
Table 4.3 Statistics for the measured acceleration time histories at lower arm
Highway Campus Hill Residential Rural
r.m.s 3.6 2.5 3.3 2.6 3.7
Kurtosis 3.4 758 14.3 9.7 4.0
Crest factor 6.7 72.2 12.7 12.3 10.9
121
(a)
(b)
(c)
(d)
(e)
Figure 4.6 Time histories of the measured acceleration from top mount under various roads:
(a) highway, (b) campus, (c) hill, (d) residential, (e) rural
Table 4.4 Statistics for the measured acceleration time histories at top mount
Highway Campus Hill Residential Rural
r.m.s 2.2 1.5 2.1 1.5 2.7
Kurtosis 3.1 47.6 14.3 3.9 3.4
Crest factor 4.5 31.5 21.1 5.6 6.7
122
In Tables 4.3 and 4.4, the SD value of the lower arm was higher than the top
mount which indicates that the vibrations of the un-sprung mass were higher than the
sprung mass. This implies that the suspension system has filtered part of the road
vibrations and the impacts were absorbed by the suspension system. For fatigue
analysis, another statistic parameter to be concerned with is the kurtosis value which is
a measurement of the sensitivity to spikes. The greatest spikes were observed at campus
road measurements with peaks observed at time 61 to 63 s. This happened because of
the pothole strike happened during a curve driving. These peaks led to the high kurtosis
value of the campus road acceleration measurement in both sprung and un-sprung mass.
Based on these observations, all the vibration measurements consisted of a
kurtosis value of higher than three. Thus, it is suggested that the time histories were
possessing non-stationary behaviour. However, the campus road acceleration time
histories were heavily non-stationary because the kurtosis value was much more
deviated from the value of 3.0. Meanwhile, the highway acceleration time histories had
short periods of changed statistics due to the presence of transient behaviour.
Nevertheless, the acceleration time histories from highway road had a kurtosis value
which was closest to 3.0. This meant that the acceleration time history was close to
stationary behaviour. When the kurtosis value above 3.0, the data were expected to
consist of stronger peak with rapid decay and heavy tails (Malik, Wang & Naseem
2017). When the kurtosis level of a time histories increases, the rapid decay and tails of
the histogram distributions were more obvious.
When comparing the acceleration and strain time histories, it was found that the
highway consisted of the lowest kurtosis value which was around 3.0. This implied that
the highway has a better road surface profile which induced less spikes on the
suspension systems. For the r.m.s value, both the strain and acceleration time histories
also revealed that the highest value was during the rural road data collection because
high amplitude was obtained during the testing. The rural road was an unpaved road
with stones and potholes which induced high vibration towards suspension system.
Meanwhile, the acceleration analysis depended on the time while the strain amplitude
was independent of time, which had led to the deviation of statistic parameter for
various road conditions.
123
The collected strain time histories were random and were unable to apply
directly for fatigue life prediction. One of the most famous approaches to analyse the
strain time histories was to decompose the strain time histories into simple stress
reversals which is known as the Rainflow cycle counting method, as reported by
Paraforos et al. (2014). This Rainflow cycle counting was applied to decompose the
collected strain time histories into simple reversal. The fatigue damage histogram of the
strain time histories under all five road conditions were plotted into Figures 4.7 to 4.11.
The x-axis represents the range of the strain, the y-axis represents the mean strain, and
the z-axis represents the fatigue damage to the number of fatigue cycles at the same
range and mean. The observed fatigue damage value was increased proportionally to
the value of strain range and mean value.
The maximum strain range contributed to the maximum fatigue damage, or the
tallest column with red contour, in the damage histogram plots. The fatigue damage
displayed on the histograms was the product of the fatigue damage at one cycle with
the corresponding strain range and mean value. The maximum fatigue damage strain
ranges were observed to have occurred at rural road measurement with fatigue damage
as high as 5.50 × 10-5. It is worth noting that the fatigue damage by calculation was
unitless. During the fatigue life prediction, a single fatigue damage strain range did not
provide significant meaning to the total spring fatigue life. Hence, the Palmgren-Miner
linear fatigue damage rule was applied to summarise all the contributed fatigue damage
because Palmgren-Miner is recognised as a simple but acceptable technique for damage
summation (Fernández-Canteli & Blasón 2014).
When considering these damage histograms, there were a few existing low strain
range cycles which did not contribute much to the fatigue damage. For comparison, the
strain time histories from rural road had obviously shown a few higher fatigue damages
than other road condition measurements. The fatigue damage histogram of highway
road has shown fewer fatigue damage in terms of frequency and amplitude leading to
the cause of the highway road to have a smoother surface and less impact on the
suspension system.
124
(a)
(b)
(c)
Figure 4.7 Damage histograms of spring strain time history under highway road using various
strain life approaches: (a) Coffin-Manson, (b) Morrow, (c) SWT
125
(a)
(b)
(c)
Figure 4.8 Damage histograms of spring strain time history under UKM campus road using
various strain life approaches: (a) Coffin-Manson, (b) Morrow, (c) SWT
126
(a)
(b)
(c)
Figure 4.9 Damage histograms of spring strain time history under hilly road using various strain
life approaches: (a) Coffin-Manson, (b) Morrow, (c) SWT
127
(a)
(b)
(c)
Figure 4.10 Damage histogram of spring strain time history under residential road using various
strain life approaches: (a) Coffin-Manson, (b) Morrow, (c) SWT
128
(a)
(b)
(c)
Figure 4.11 Damage histogram based for spring strain time history under rural road using various
strain life approaches: (a) Coffin-Manson, (b) Morrow, (c) SWT
129
When the measured strain time histories were analysed using the Rainflow cycle
counting technique, a long-time history was usually required to obtain a statistically
representative fatigue damage (Halfpenny & Kihm 2010). Hence, a field of fatigue
based on vibrations was introduced to provide an accelerated fatigue analysis. The
standard procedures to perform a vibration fatigue analysis were by firstly calculating
or estimating the frequency response functions of the structure (Mršnik, Slavi c &
Boltezar 2013). Furthermore, the vibration time histories in the form of PSD were given
as excitation input to obtain the fatigue life of the structures. Through exciting different
vibration modes of the structure, the spread of vibration energy over a frequency range
influences the fatigue damage on the structure. Thus, vibration loading histories play an
important role in fatigue analysis. Nevertheless, automobile vibrations affected not only
structures but also the drivers and passengers (Lakušić, Brčić & Lakušić 2011).
It was worth noting to mention that human perception on vehicle ride dynamics
was also related to the vibration induced by various sources, such as engine and road
profile (Burdzik & Doleček 2012). In some cases, vibration time histories were also
applied to study the structure integrity of automobile components instead of only
vehicle ride quality. The vibration of vehicle sprung mass indicates the level of comfort
provided by the vehicle during interaction between wheel and ground. Meanwhile, the
vibration of the lower arm has shown the excitations of the wheel when the vehicle deals
with the road excitations.
Vibration level of an automobile was analysed using PSD because area under
the curve of a PSD indicates the energy (Vinogradov et al. 2013). The comparison
between the acceleration PSD of lower arm and top mount are plotted into Figure 4.12.
Based on the obtained PSD, it was found that the high amplitude occurred at low
frequency which were below 10 Hz and the energy was decayed across frequency range.
As observed, the energy level in terms of amplitude from PSD of top mount was lower
than lower arm under all five road conditions. This was because of transmissibility and
isolation effects provided by the suspension system (Parekh et al. 2014). To determine
the transmissibility, root mean square error (RMSE) between the top mount and the
lower arm acceleration PSD were performed using Equation 3.14 and listed into Table
4.5. The RMSE value was significant to understand the behaviour of suspension in
130
filtering the vibration where a high RMSE value indicates that more vibrations were
attenuated. The rural road contained the highest RMSE value at 0.0164 because the road
was rough, and the suspension system has worked more for the vibration attenuation.
For automotive suspension components, the interest frequency range for fatigue
analysis was 40 to 60 Hz and frequency above 100 Hz could be neglected (Sener 2012).
Nevertheless, the natural frequency of suspension system and seat was between 1 to 3
Hz (Bouazara, Richard, Rakheja 2006). The resonance frequencies range for engine part
was 20 to 200 Hz with engine speed range of 600 – 6000 rpm (Yu et al. 2001). The
proposed road excitation was below 10 Hz when the vehicle speed was 50 km/hr
(González et al. 2008). Nevertheless, based on the strain PSD under campus and hill
roads, peak was observed at frequency above 100 Hz where this frequency range was
also considered in spring fatigue analysis.
Table 4.5 RMSE between sprung and un-sprung mass acceleration time histories
Road Root mean square error (m/s2)2/Hz
Highway 0.0004
UKM Campus 0.0001
Hilly 0.0010
Residential 0.0007
Rural 0.0164
4.2.3 Regression Analysis for Fatigue Life
The acceleration time histories were used to calculate the vehicle ISO 2631 vertical
vibrations while the spring fatigue life were calculated using strain life approaches. The
results of spring fatigue life and ISO 2631 vertical vibrations were then tabulated into
Table 4.6, indicating that the rural road possessed the lowest fatigue life at 5.71 × 103
blocks to failure due to rough road surface. Meanwhile, the ISO 2631 vertical vibration
of automobile for sprung mass under rural road was the highest at 0.73 m/s2 which
indicates that the ISO 2631 has an inverse relationship towards fatigue life. In addition,
the ISO 2631 vertical vibration for highway road was low (0.57 m/s2) while the fatigue
life was high (3.52 × 104 blocks to failure) due to the smooth road surface. This also
implies that the negative correlation between the fatigue lives and ISO 2631 vertical
vibration were. When ISO 2631 vertical vibration amplitude was high, the vibrational
energy of the suspension was also high and lead to the reduction of fatigue life.
131
(a)
(b)
(c)
(d)
(e)
Figure 4.12 PSD of collected strain and acceleration time histories for various roads: (a) highway,
(b) campus, (c) hill, (d) residential, (e) rural
132
Table 4.6 ISO 2631 vertical accelerations and spring fatigue lives
Weighted
acceleration for
top mount
(m/s2)
Weighted
acceleration
for lower arm
(m/s2)
Coffin-
Manson
(blocks to
failure)
Morrow
(blocks to
failure)
SWT (blocks
to failure)
Highway 0.57 0.98 6.45 × 104 3.64 × 104 3.52 × 104
Campus 0.68 0.86 1.25 × 104 1.18 × 104 1.18 × 104
Hill 0.56 1.87 4.59 × 104 4.61 × 104 5.41 × 104
Residential 0.59 1.06 4.90 × 104 3.52 × 104 3.73 × 104
Rural 0.73 0.97 6.06 × 103 5.71 × 103 5.71 × 103
To investigate the relationship between fatigue life and ISO 2631 vertical
vibrations, a suitable regression approach was needed. Regression analysis is an
extraction of hidden predictive information from large databases with immense
potential to help in extracting the most valuable information in the data. One of the
methods to extract this information is known as linear regression method (Lewis-Beck
2015). This statistical technique is used to find the best-fitting linear relationship
between a dependent variable and its predictors. As prerequisite for the linear
regression, the dependent variable needed to have a linear relationship with the
independent variable.
The correlation study of vibration for durability and vehicle ride was significant
to reduce the number of testing. In nature, fatigue of material possesses a power
relationship with the applied loadings. The power law regression function was applied
to mathematical model of the spring fatigue life and ISO 2631 vertical vibration. The
fitted power curve is shown in Figure 4.13 and the generated power regression
relationships are shown as follows:
NCM_power = 1349Wa-6.523
(4.1)
NMorrow_power = 847Wa-6.859
(4.2)
NSWT_power = 606Wa-7.598
(4.3)
where Ncm_power, NMorrow_power and NSWT_power are the respective Coffin-Manson, Morrow
and SWT approaches with experimental predicted spring fatigue life with units blocks
to failure, Wa is the experimental ISO 2631 vertical vibration with the unit of m/s2. The
133
R2 value for Coffin-Manson vibration-life regression was 0.8419, 0.9433 for Morrow
and 0.9754 for SWT.
Figure 4.13 Correlation of spring fatigue life and ISO 2631 vertical vibration in power form
Equations 4.1 – 4.3 were used to predict the spring fatigue life using the ISO
2631 vertical vibration as input. Based on the R2 value, the relationship between spring
fatigue life and ISO 2631 vertical vibration possessed a strong negative relationship.
This was indicative that the weighted vertical vibration predicted the spring fatigue life
with good accuracy. Hence, these regression relationships were suitable to be proposed
to the automotive industry for spring design assistances. When considering including
additional parameters, the simple power form regression was difficult to handle due to
its non-linearity.
For the ease of use, the regression in power form was derived into linear form
to enable least square calculation (Hanaki et al. 2010). This was through applying
natural logarithm on calculated fatigue life. Subsequently, a linear regression analysis
was conducted on the spring fatigue life and ISO 2631 vertical vibration datasets after
applied natural logarithm and the results of the linear regression analysis are shown in
R2 = 0.8419
R2 = 0.9754
R2 = 0.9344
Fat
igue
life
(b
lock
s to
fai
lure
)
ISO 2631 vertical vibration (m/s2)
134
Figure 4.14. The R2 value for Coffin-Manson predicted that spring fatigue life and ISO
2631 vertical vibration was 0.9613 while the R2 value for Morrow and vertical vibration
was 0.9278. For SWT model, the linear regression fitted with R2 value of 0.9808. Based
on all the obtained R2 value, it could be clarified that the ISO 2631 vertical vibration
has a strong linear relationship with spring fatigue life after the natural logarithm was
applied to the fatigue life unit.
The gradient and intercept of linear regression vibration-life relationship was
obtained and written as follows:
NCM_linear = -5.828Wa + 8.056 (4.4)
NMorrow_linear = -4.699Wa + 7.206 (4.5)
NSWT_linear = -5.359Wa + 7.691 (4.6)
where Ncm_linear, NMorrow_linear, NSWT_linear are the Coffin-Manson, Morrow and SWT
approaches experimental predicted spring fatigue life respectively with the natural
logarithm of units of block to failure, Wa is the ISO 2631 vertical vibration with the unit
of m/s2. All these three approaches provided immediate fatigue life assessment using
ISO 2631 vertical vibration as input.
For Equations 4.4 – 4.6, the vibration-life regressions were linearised and hence,
the coefficients and constants were different from the power law form. The reason to
linearise the power regression approach was because it was easier to interpret the fatigue
life in a linear form than a nonlinear power form (Tomaszewki, Sempruch & Piatkowski
2014). With the high R2 value, it reflected also the quality and repeatability of the
analysis. Moreover, the strong correlation value was also indicative that the feasibility
of linear relationship between spring fatigue life and ISO 2631 vertical vibration. These
proposed new regressions have achieved part of the first objective which was the
establishment of spring vibration-life relationships.
135
Figure 4.14 Correlation of spring fatigue life and ISO 2631 vertical vibration in linear form
To validate the regression predictions, the linear regression predicted fatigue
lives were compared to experimental fatigue lives as shown in Table 4.7 using RMSE
as defined in Equation 3.14. The RMSE for Coffin Manson regression predicted and
experimental fatigue life was 0.0754 blocks to failure in natural logarithm while the
RMSE for predicted Morrow fatigue life was 0.0720 blocks to failure in natural
logarithm. In addition, the RMSE for predicted SWT fatigue life was 0.0507 blocks to
failure in natural logarithm. The suitability of the RMSE was usually compared with
the parameter range where the fatigue lives were ranged from 3.8 to 4.8 blocks to failure
in natural logarithm. Cárdenas et al. (2012) proposed that the suitable RMSE range
when compared with the parameter range was about 10% which indicates that the
proposed regression predictions were acceptable. The correlation studies of prediction
and experimental data for all three strain life models are shown in Figure 4.15 with the
lowest R2 value of 0.8328. According to Sivák and Ostertagová (2012), the R2 value for
fatigue correlation above 0.80 is considered as acceptable. Hence, it suggested that the
regression predicted fatigue life was acceptable.
R2 = 0.91
R2 = 0.95
R2 = 0.9613
R2 = 0.9278
R2 = 0.9808
Fat
igue
life
(b
lock
s to
fai
lure
)
ISO 2631 vertical vibration (m/s2)
136
Table 4.7 Regression predicted and experimental fatigue lives
Linear regression predicted fatigue life
(blocks to failure) Experimental fatigue life (blocks to failure)
Coffin-
Manson Morrow SWT
Coffin-
Manson Morrow SWT
5.42 × 104 3.37 × 104 4.33 × 104 6.45 × 104 3.64 × 104 3.52 × 104
1.24 × 104 1.02 × 104 1.11 × 104 1.25 × 104 1.18 × 104 1.18 × 104
6.20 × 104 3.75 × 104 4.90 × 104 4.59 × 104 4.61 × 104 5.41 × 104
4.14 × 104 2.71 × 104 3.38 × 104 4.90 × 104 3.52 × 104 3.73 × 104
6.33 × 103 6.00 × 103 6.01 × 103 6.06 × 103 5.71 × 103 5.71 × 103
4.3 DATA FOR MODELLING
4.3.1 Artificial Road Profiles
To achieve the first objective, which was establishing multiple input regression
relationships of fatigue life, ISO 2631 vertical vibration and vehicle body frequency,
additional data for analysis were required to expand the analysis. Without experimental
testing, a vehicle quarter car model was proposed and simulated where road inputs were
required to simulate the dynamical behaviour of the suspension. As for the road
acceleration measurements, there were only five acceleration time histories from
different road conditions which were insufficient. To resolve this issue, ISO 8608
artificial road profiles were proposed to generate the additional loading signals where
the fundamental ideas of ISO 8608 were spatial frequency, road displacement and PSD.
Road profile was the variations in height of the road surface which was measured on
single track parallel to the road. This proposed road profile assumed that the given road
was stationary where the statistic properties did not change over time. The generated
road surface was a combination of numerous long or fleeting period bumps with varying
amplitudes.
137
(a)
(b)
(c)
Figure 4.15 Correlation analysis for fatigue life using various fatigue approaches:
(a) Coffin-Manson, (b) Morrow, (c) SWT
0.00E+00
1.00E+04
2.00E+04
3.00E+04
4.00E+04
5.00E+04
6.00E+04
7.00E+04
0.00E+001.00E+042.00E+043.00E+044.00E+045.00E+046.00E+047.00E+04
Lin
ear
regre
ssio
n f
atig
ue
life
(b
lock
s
to f
ailu
re)
Experimental fatigue lives (blocks to failure)
0.00E+00
5.00E+03
1.00E+04
1.50E+04
2.00E+04
2.50E+04
3.00E+04
3.50E+04
4.00E+04
4.50E+04
0.00E+00 1.00E+04 2.00E+04 3.00E+04 4.00E+04 5.00E+04
Lin
ear
regre
ssio
n f
atig
ue
life
(b
lock
s
to f
ailu
re)
Experimental fatigue lives (blocks to failure)
0.00E+00
1.00E+04
2.00E+04
3.00E+04
4.00E+04
5.00E+04
6.00E+04
0.00E+00 1.00E+04 2.00E+04 3.00E+04 4.00E+04 5.00E+04 6.00E+04
Lin
ear
regre
ssio
n f
atig
ue
life
(b
lock
s
to f
ailu
re)
Experimental fatigue lives (blocks to failure)
R2 = 0.8328
R2 = 0.9763
R2 = 0.9312
138
For classification, the PSD between measured road and ISO 8608 road profile
classifications were plotted into a graph as shown in Figure 4.16. The trend of the
measured road profile was different from the ISO 8608 road class because the actual
road excitations was varied, and the road was a combination of various classes due to
different roughness and vehicle speeds. The same trend was also reported by González
et al. (2008). As observed in Figure 4.16, the measured road profiles consisted of the
greatest roughness up to standard road class “D” and the lowest roughness as class “A”
were based on the ISO 8608 classification. The highway measured acceleration PSD
was classified as the class “A” road while the rural area measured acceleration was in
the class “D” road. As support for this classification, many previous researches have
also utilised ISO 8608 class “A” to “D” in quarter car model simulation (Loprencipe &
Zoccali 2017; Joshi et al. 2015). Road profiles beyond road class “D” were unrealistic
for passenger cars because the amplitudes were too high and the signals were considered
as heavily non-stationary with many high peaks from random effects (Chaari et al.
2013). The road class “E” was unrealistic because it simulated off-road conditions
(Balmos et al. 2014). In Figure 4.16, the measured road class was classified below class
“E” and hence the collected acceleration signals were suitable for quarter car model
simulations.
Figure 4.16 Classification of measured road profile according to ISO 8608
139
The generated road profiles in terms of spatial frequency are shown in Figure
4.17. The vehicle speed was determined to be 80 kmph because this was the average
vehicle speed for the objective ride analysis, as proposed by ISO 2631 (Nagarkar et al.
2016; Loprencipe & Zoccali 2017; Panta et al. 2014). When the vehicle speed effect
was not a parameter to be considered, the average vehicle speed of 80 kmph was usually
selected. Meanwhile, the quarter car model simulation required the input of
displacement in time domain so that the behaviour of suspension could be clearly
visualised. Hence, the generated road classes in temporal frequency for simulation were
performed using the generated temporal road profile as shown in Figure 4.18 where the
artificial generated road profiles were in the form of displacement versus time. As
observed, the generated road profiles were stochastic but generally stationary.
Meanwhile, the generated road profiles were also considered as non-deterministic
loading time histories.
To understand the characteristics of these generated road profiles, a statistic
evaluation was performed on the generated road profiles. The mean, r.m.s values and
respective fatigue life of the spring under these road classes were analysed and tabulated
into Table 4.8. The mean value of the road class increased from 4.71 ×10-7 to -1.51 ×
10-5 m across road class “A” to “D”. Nevertheless, the generated road profile contained
a very small mean value. The same trend implied on r.m.s value of the road profile from
road class “A” to “D” where the r.m.s value has been increased from 0.0046 to 0.0369
m. Moreover, the road class “D” contained the highest r.m.s. The r.m.s value indicating
the degree of rough of the generated road profile as well as the energy level. Higher
mean and r.m.s value produced a road class with a higher degree of roughness and
hence, led to lower fatigue life of the coil spring. Muc ka (2016) reported that the road
class “A” to “C” was typical road while class “D” above was unpaved road. After
examining the road characteristics, all four classes of generated road were then used as
input to quarter car simulation model for fatigue loading signal extractions.
Table 4.8 Statistical parameters of artificial generated road classes
Road Mean (m) r.m.s (m) CM Morrow SWT
Class A 4.71 ×10-7 0.0046 6.03 × 106 6.61 × 106 6.76 × 106
Class B -3.77 × 10-6 0.0092 9.33 × 105 1.07 × 106 7.08 × 105
Class C -7.54 × 10-6 0.0185 3.63 × 103 3.63 × 103 3.63 × 103
Class D -1.51 × 10-5 0.0369 2.19 × 102 2.19 × 102 2.19 × 102
140
(a)
(b)
(c)
(d)
Figure 4.17 Generated ISO 8608 road profile in form of spatial frequency for generated road
classes: (a) class A, (b) class B, (c) class C, (d) class D
4.3.2 Quarter Car Model Simulation Results
Quarter car model (QCM) was a vehicle simulation model that could be efficiently used
to study the dynamic interaction between vehicle and road roughness profile, and
therefore applied in this study of suspension dynamic behaviour. Through combining
the masses values, the stiffness and damping of the QCM, it was possible to model any
type of road vehicle (Agostinacchio et al. 2014). Using the QCM, the dynamic
responses of the vehicle were extracted. Two types of time history were extracted for
vibration and fatigue analysis which were force time histories induced on the coil spring
and acceleration time histories exerted on the vehicle mass under influence of different
road roughness.
141
(a)
(b)
(c)
(d)
Figure 4.18 ISO 8608 road profile in form of temporal frequency for various road
classes: (a) class A, (b) class B, (c) class C, (d) class D
An example results of simulated force time histories is shown in Figure 4.19.
This example force time histories were obtained from QCM simulation under road class
“A” with different spring stiffness. As observed, when the spring stiffness increased,
the magnitude of the simulated force time histories was also increased because the
spring force linearly depended on the displacement and spring stiffness according to the
Hooke’s law. Loading a spring with higher stiffness under a fixed displacement,
additional force was required to achieve the same height. Nevertheless, for road class
“A”, the displacement or road surface roughness was not high when compared to other
generated road classes. For other roads, the spring force time histories of other road
classes and measurements were listed in Appendix D. All the simulated spring force
142
time histories were within the range of 200 N and the range of force was drastically
increased after an increment of road class.
(a)
(b)
(c)
(d)
(e)
Continue…
Fo
rce
(N)
20 40 60 80
-200
200
0
20 40 60 80
-200
200
0
20 40 60 80
-200
200
0
20 40 60 80
-200
200
0
20 40 60 80
200
0
-200
143
Continued…
(f)
(g)
(h)
(i)
(j)
Time (s)
Figure 4.19 Simulated force time histories under road class “A” for different spring stiffness:
(a) k14, (b) k16, (c) k18, (d) k20, (e) k22, (f) k24, (g) k26, (h) k28, (i) k30, (j) k32
Fo
rce
(N)
20 40 60 80
-200
200
0
20 40 60 80
-200
200
0
20 40 60 80
-200
200
0
20 40 60 80
-200
200
0
20 40 60 80
-200
200
0
144
Apart from spring force time histories, QCM simulations have also provided the
acceleration of vehicle mass under various road classes. The accelerations of the vehicle
sprung mass under various spring stiffness were plotted into Figure 4.20. The
acceleration of vehicle mass was an indicator for passenger ride perception where high
amplitude reduced the vehicle ride quality. Based on the observations from Figure 4.20,
the acceleration levels were within the range of -1 to 1 m/s2. The acceleration level
under this class “A” road was not high when compared from classes “B” to “D”. The
results of acceleration for road classes “B” to “D” were listed in Appendix E.
The acceleration amplitude was also increased from road classes “A” to “D”
under the influence of road roughness. For measured acceleration simulation input, the
results of spring force time histories are shown in Figure 4.21 while the QCM simulated
acceleration time histories for sprung mass is shown in Figure 4.22. In Figure 4.21, the
spring force time histories were extracted from highway road acceleration input with
quarter car model simulation. The amplitude range for the quarter car model simulation
under highway road excitation was ranged from -200 to 200 N which was a very close
amplitude range to artificial road class “A” due to the similar road roughness
excitations.
In Figure 4.22, the quarter car simulation measured acceleration was collected
from the highway road with the acceleration amplitudes ranging from -0.5 to 0.5 m/s2
which was close to artificial road class “A”. The highway road had similar properties to
artificial road class “A” in terms of amplitude which led to the same extracted data
range. However, the measured road profiles were more dynamic than the artificial road
profile because there were many uncertainties in roads, such as the changing of vehicle
speed due to traffic, lane changes or pothole strikes. When associating the acceleration
amplitude with the ISO 2631 ride standard, the acceleration amplitude was considered
as not high. Hence, these low amplitudes suggest that highway roads provide good ride
quality to the automobile.
145
(a)
(b)
(c)
(d)
(e)
Continue…
Acc
eler
atio
n (
m/s
2)
20 40 60 80
-1
1
0
20 40 60 80 -1
0
1
20 40 60 80
-1
0
1
20 40 60 80 -1
0
1
1
20 40 60 80 -1
0
146
Continued…
(f)
(g)
(h)
(i)
(j)
Time (s)
Figure 4.20 Simulated acceleration time histories under road class “A” for different spring
stiffness: (a) k14, (b) k16, (c) k18, (d) k20, (e) k22, (f) k24, (g) k26, (h) k28, (i) k30,
(j) k32
Acc
eler
atio
n (
m/s
2)
1
20 40 60 80 -1
0
1
20 40 60 80
-1
0
1
20 40 60 80
-1
0
1
20 40 60 80 -1
0
1
20 40 60 80
0
-1
147
(a)
(b)
(c)
(d)
(e)
Continue...
Fo
rce
(N)
200
20 40 60 80
-200
0
200
20 40 60 80 -200
0
200
20 40 60 80 -200
0
20 40 60 80
20 40 60 80
200
-200
0
200
-200
0
148
Continued...
(f)
(g)
(h)
(i)
(j) Time (s)
Figure 4.21 Simulated force time histories under highway road for different spring stiffness:
(a) k14, (b) k16, (c) k18, (d) k20, (e) k22, (f) k24, (g) k26, (h) k28, (i) k30, (j) k32
Fo
rce
(N)
20 40 60 80
20 40 60 80
20 40 60 80
20 40 60 80
20 40 60 80
200
-200
0
200
-200
0
200
-200
0
200
-200
0
200
-200
0
149
(a)
(b)
(c)
(d)
(e)
Continue…
Acc
eler
atio
n (
m/s
2)
20 40 60 80 -1
1
0
20 40 60 80
20 40 60 80
20 40 60 80
20 40 60 80
-1
1
0
-1
1
0
-1
1
0
-1
1
0
150
Continued…
(f)
(g)
(h)
(i)
(j)
Time (s)
Figure 4.22 Simulated acceleration time histories under highway road for different spring
stiffness: (a) k14, (b) k16, (c) k18, (d) k20, (e) k22, (f) k24, (g) k26, (h) k28, (i) k30,
(j) k32
Acc
eler
atio
n (
m/s
2)
20 40 60 80
20 40 60 80
20 40 60 80
20 40 60 80
20 40 60 80
-1
1
0
-1
1
0
-1
1
0
-1
1
0
-1
1
0
151
When considering the various road class in Appendix E, the simulated force and
acceleration time histories of other road conditions were higher than the highway road.
This was because the quarter car simulation has considered the dynamical condition of
hilly, UKM campus, residential and rural road where these road conditions consisted of
coarser road condition and produced higher vibration to the suspension system. The
higher force and acceleration amplitude proposed that these road conditions provide
less ride quality of the automobile.
4.3.3 Spring Stiffness Sensitivity Analysis
The change of a spring design was involving the parameters like spring geometry, bar
diameter, outer diameter or number of active coils (Valsange 2012). To obtain different
spring stiffness, one of the ways was through changing the bar size (SAE 1990). During
the simple regression analysis, the model was analysed with two variables which were
the spring fatigue life and ISO 2631 vertical vibration and, the suspension spring
stiffness was considered as a fixed parameter. In order to include the effects of spring
stiffness as a parameter, spring stiffness sensitivities were performed to obtain the
effects of different spring stiffnesses.
To determine the spring design variants, the adjustment of the bar diameter for
coil spring was performed as the stiffness of the spring increased when the bar diameter
increased (Valsange et al. 2012). However, different bar diameter led to various stress
levels which affected the spring fatigue life. Hence, a series of new spring design
variants were conducted. The examples of spring design variants are shown in Figure
4.23. The stiffest spring design have bar diameter of 13.3 mm which achieved the
maximum stiffness for automobile vehicle body frequency analysis with the
consideration of vehicle ride quality.
The results of all the coil spring diameter adjustment were tabulated into Table
4.9. All the new designed coil springs were obtained through simple calculation and
analysed using FEA. There were some deviations between calculated and FEA stiffness
according to the adjustment of the spring bar diameter. Hence, the difference between
the calculation and FEA simulations were analysed to ensure the validity of the spring
152
design. The maximum difference between the calculated and FEA analysis were
relatively small where the maximum error was only 5.3 %. Based on Zhang et al. (2007),
the acceptance of the FEA results with experimental data was 10 %. Therefore, the FEA
coil spring models were acceptable for fatigue analysis where a more realistic fatigue
life prediction was offered in accordance to spring design.
Figure 4.23 Spring design variants with different bar diameter
Table 4.9 Spring stiffness parameter sensitivity analysis
Bar
diameter
Calculated stiffness
(N/mm)
FEA based stiffness
(N/mm) Difference (%)
11.0 14.3 13.8 1.4
11.3 16.1 15.7 1.9
11.6 18.0 17.6 2.2
12.0 20.3 19.7 1.5
12.2 22.3 21.6 1.8
12.4 24.0 23.9 0.4
12.6 25.7 25.5 1.9
12.8 27.5 27.0 3.5
13.0 29.5 28.4 5.3
13.3 32.4 30.8 3.8
Note: Difference = |Calculated stiffness-FEA based stiffness
FEA based stiffness| ×100% (Equation 3.13)
This spring sensitivity analysis was crucial to determine the ISO 2631 vertical
vibration because the spring stiffness was proportional to the induced vehicle response
153
from ground uncertainties. During the spring sensitivity analysis, the spring stiffness
was adjusted to determine the effects of spring design to fatigue life. Adjusting the
spring stiffness in quarter car model led also to the variation in spring design. The
original nominal spring design did not represent the real fatigue life predictions with the
loading signals from the quarter car model when various spring stiffnesses were applied.
Without the spring sensitivity analysis, the fatigue life of spring could not be accurately
predicted, and the analysis were limited to a fixed spring parameter.
4.3.4 Fatigue Life and Vibration Data
When the spring force time histories and design variants were prepared, the prediction
of spring fatigue life was performed. The equivalent suspension natural frequency was
calculated according to different spring stiffness using Equation 3.10. The results of
spring fatigue life, suspension natural frequency and ISO 2631 vertical vibration under
class “A” were tabulated into Table 4.10 while the results of other road classes were
listed in Appendix F. As observed in Table 4.10, the ISO 2631 vertical acceleration
increased along with increasing the suspension natural frequency. For the other road
conditions, the same increasing trend has been observed. Sekulić and Dedović (2011)
proposed that when the natural frequency was increased, the vertical acceleration
response on the driver side was also increased.
The increasing trends were observed because the spring stiffness changes led to
the increment of suspension natural frequency where the road excitation was amplified.
In addition, the results for highway road simulation predicted fatigue life is shown in
Table 4.11. In Tables 4.10 and 4.11, the vehicle body frequency was ranged from 1.00
to 1.50 Hz because this frequency range was the optimal frequency for passenger cars
(Sekulić & Dedović 2011). The ISO 2631 vertical accelerations were obtained based on
the vehicle sprung mass acceleration through processing the signals with the ISO 2631
weighting values. This value was an indicator of vehicle vibration level and human
perception where the fatigue life of automobile spring was also affected by these
vibrations. When the natural frequencies were increased, the fatigue life of spring was
reduced. This was because the force exerted on the spring was increased according to
154
Hooke’s law. Under the same induced displacement, the force of the spring was higher
when the spring stiffness was increased.
In Tables 4.10 and 4.11, the nominal spring design was 1.20 Hz with spring bar
diameter 12 mm. For this spring stiffness, the obtained fatigue lives were ranged from
1.23 × 107 to 1.46 × 107 blocks to failure while the fatigue lives of spring under highway
road simulation were ranged from 9.80 × 105 to 4.36 × 107 blocks to failure. The
variation of spring during highway excitation was due to the nonlinear road conditions
like pothole striking or waviness which led to varying mean stresses. Kamal and
Rahman (2014) proposed that the fatigue life of a coil spring was about 5.16 × 107
blocks to failure where the predicted fatigue life was reasonable because it fall within
the same range with the proposed fatigue life.
Table 4.10 Predicted fatigue life and ISO 2631 weighted acceleration from Class “A” road
Vertical
frequency (Hz)
Weighted
vertical
acceleration
(m/s2)
Coffin-Manson
fatigue life
(blocks to
failure)
Morrow fatigue
life (blocks to
failure)
SWT fatigue life
(blocks to
failure)
1.00 0.1540 2.02 × 108 2.50 × 108 2.71 × 108
1.06 0.1548 1.69 × 108 1.99 × 108 2.14 × 108
1.13 0.1562 5.60 × 107 6.60 × 107 6.99 × 107
1.20 0.1574 1.23 × 107 1.39 × 107 1.46 × 107
1.25 0.1589 5.96 × 106 6.57 × 106 6.77 × 106
1.30 0.1606 2.83 × 106 3.03 × 106 3.11 × 106
1.36 0.1622 1.51 × 106 1.60 × 106 1.64 × 106
1.41 0.1644 7.69 × 105 7.98 × 105 8.13 × 105
1.45 0.1665 5.39 × 105 5.53 × 105 5.58 × 105
1.50 0.1689 3.93 × 105 3.98 × 105 3.99 × 105
Table 4.11 Predicted fatigue life and ISO 2631 weighted acceleration from highway road
Vertical
frequency (Hz)
Weighted
vertical
acceleration
(m/s2)
Coffin-Manson
fatigue life
(blocks to
failure)
Morrow fatigue
life (blocks to
failure)
SWT fatigue life
(blocks to
failure)
1.00 0.2863 2.68 × 107 6.70 × 109 2.79 × 108
1.06 0.3112 1.75 × 107 1.18 × 108 8.58 × 107
1.13 0.3255 5.26 × 106 2.00 × 107 2.61 × 107
1.20 0.3390 9.80 × 105 4.36 × 107 6.23 × 106
1.25 0.3525 4.22 × 105 7.46 × 106 1.32 × 107
1.30 0.3658 2.11 × 105 7.31 × 105 1.25 × 106
1.36 0.3792 1.22 × 105 2.62 × 105 6.64 × 105
1.41 0.3918 7.58 × 104 7.42 × 104 2.50 × 105
1.45 0.4046 5.52 × 104 5.39 × 104 1.82 × 105
1.50 0.4169 4.15 × 104 4.19 × 104 1.25 × 105
155
As observed from Table 4.11, the ISO 2631 vertical acceleration also increased
with increasing spring stiffness. The same increasing trend for the vertical vibration has
also been proposed by Chen et al. (2017). For fatigue life prediction, a decreasing trend
has been observed where the spring fatigue life was decreased with increased spring
stiffness due to higher amplitude. As reported by Sekulić & Dedović (2011), amplitude
of vertical acceleration increased drastically with spring stiffness of frequency 1.50 Hz
and above. Hence, the vehicle body frequency ranges up to 1.50 Hz was solely
considered in this work. Based on this frequency range, 90 datasets were prepared and
ready for multiple regression analysis which considered all three-spring fatigue life,
vehicle body frequency and ISO 2631 vertical vibrations parameters.
4.4 MULTIPLE INPUT REGRESSION
MLR is a common form of linear regression analysis which explains the relationship
between one dependent variable and two or more independent variables. This method
could be used to explore the relationship between more parameters where the spring
stiffness was included. Hence, this section proposed multiple linear regression for
fatigue life and ISO 2631 vertical vibration including spring stiffness that is related to
the first objective and led to part of the novelty of this research.
4.4.1 Analysis for Life Regression
When there is more than one independent variable (predictors or explanatory variable),
the simple linear method was not able to perform regression fitting because it fitted only
a single variable. Hence, a more complicated regression method known as multiple
linear regression was applied to overcome the limitation of numerous independent
variables. MLR is a type of supervised machine learning method where the target of the
regression approach is needed. This method produces a linear regression relationship
using least square estimate with partial coefficients. A total of 90 datasets were used
where 50 of the datasets were simulated from the measured acceleration while the other
40 datasets were obtained from artificial road profile. The measured acceleration
datasets were collected from various road classes with varying surface conditions.
156
For vibration-life regression approach, the spring fatigue life was applied as the
dependent variable while the ISO 2631 vertical vibration and suspension natural
frequency were set as the independent variables. This relationship is suggested to be
called ‘vibration-life’ analysis where all three strain life fatigue approaches were
applied, i.e. Coffin-Manson, Morrow and SWT. The regression response surfaces are
shown in Figure 4.24 for all three relationships. The yellow contour of the graph
indicates the parameter values with high fatigue life while the purple colour region
shows the area with the lowest fatigue life. The blue colour region shows the region
with middle fatigue life range.
As observed from the response surface plane in Figure 4.24, it is obvious that
the spring fatigue life reduced with increased ISO 2631 vertical acceleration, indicating
linear negative relationship with the spring fatigue life. Apart from that, increments in
suspension natural frequency led to the reduction of spring fatigue life, suggesting a
linear negative relationship between these two parameters. However, the effects of
vehicle body frequency on spring fatigue life were not drastic because the gradient of
the response surface for these two parameters were low. Increasing the spring stiffness
would also cause the spring fatigue life to be reduced.
For the vibration parameters, the increment of vehicle body frequency has led to
higher ISO 2631-1 vertical acceleration. After these relationship surface planes were
generated, the mathematical relationship in the forms of regression were obtained. The
obtained regressions were named as ‘vibration-life’ relationship which were part of the
novelty of this research. This vibration-life regression was created to predict spring
fatigue life with two parameters as input. The application of this regression-based
fatigue relationship to predict spring fatigue life was very straightforward. Conventional
durability analysis needed to measure the spring with strain gauges and performed strain
life calculations using materials cyclic properties. These processes were time
consuming and the data were rarely reusable (Yang & Xu 2012). Vibration-life
regression is intended to provide a solution to this issue through providing spring fatigue
life predictions using solely vibration measurements.
157
(a)
(b)
(c)
Figure 4.24 Multiple linear regression for fatigue life characteristics using various approaches:
(a) Coffin-Manson, (b) Morrow, (c) SWT
158
Using the MLR method, the regression relationships in mathematical form were
also obtained through analysis of variance (ANOVA). For this purpose, ANOVA meant
to relate all three generated regressions for spring strain life predictions according to
Figure 4.24(a) – (c) are listed as follows:
NCM_MLR = 11.61 − 3.65.ωninput − 3.83.Wa_input (4.8)
NMorrow_MLR = 12.62 − 4.32.ωn_input − 4.12.Wa_input (4.9)
NSWT_MLR = 12.78 − 4.40.ωn_input − 4.14.Wa_input (4.10)
where NCM_MLR, NMorrow_MLR, NSWT_MLR are the Coffin-Manson, Morrow and SWT
multiple regression predicted fatigue life with the unit of blocks to failure in natural
logarithm respectively, 𝜔𝑛_𝑖𝑛𝑝𝑢𝑡 is the suspension natural frequency input with unit of
Hz, Wa_input is the ISO 2631 vertical acceleration input with unit of m/s2. In Section
4.2.3, NCM_power, NCM_Morrow, NSWT_power, NCM_linear, NMorrow_linear, NSWT_linear were proposed
to predict fatigue life using ISO 2631 weighted acceleration as input with units of blocks
to failure as output. However, the proposed regressions in Section 4.2.3 were only
applicable to the nominal spring design. In this section, the proposed MLR-based
fatigue life predictions considered the spring design variants which made the regression
more flexible for application.
The proposed coefficients from Equations 4.8 to 4.10 were un-standardised
coefficients from the MLR approach. Un-standardised coefficients indicated the amount
the dependent variable changed if the independent variables changed by one unit while
the other independent variables were kept constant. Un-standardised coefficients do not
eliminate the unit of measurement. Hence, it is not applied to rank the independent
variables (Darlington & Hayes 2016). For example, in Equation 4.8, it implied that for
every increase of a unit of fatigue life, the vehicle body frequency was reduced by 3.65
Hz and ISO 2631 vertical vibration was reduced by 3.83 m/s2 from a constant of 11.61.
The validity of these generated regression relationships was further examined using
analysis of variance (ANOVA), residual histograms, normal probability-probability (P-
P) plot and scatter plot for goodness of fitting, residuals’ normality and
homoscedasticity.
159
To examine the suitability of generated regression-based relationships as shown
in Figure 4.24, the R2 values of each regressions were analysed using ANOVA. The R2
was used to examine the goodness of the fitted data where a high value of R2 indicated
a good correlation (Asadi et al. 2014). With the high R2 value, the regressions were
considered as acceptable for spring fatigue life prediction. The range of acceptable R2
values and its definition is already listed in Table 3.6. In the current analysis, the R2
value show a confidence level of 83% for the developed Coffin-Manson vibration-life
regression with MSE of 0.5285 as shown in Figure 4.24(a). The confidence value is
88% for the Morrow vibration-life regression with MSE of 0.5855 (Figure 4.24(b)) and
83% for the SWT vibration-life regression with MSE of 0.7056 (Figure 4.24(c)). The
low MSE value indicated that the vibration-life regression relationships were acceptable
because the deviations of fatigue life prediction were low. Taghavifar & Mardani (2014)
utilised ANN to predict wheel-soil interaction of off-road vehicles with MSE of up to
0.996. This indicated that the obtained MSEs were within acceptable range. Although
the residuals were determined as normally distributed, it was significant to examine the
residuals using a scatter plot to ensure the homoscedasticity of the vibration-life
regression. The homoscedasticity meant that the variance around the regression line is
the same for all values of the independent variables where the error term was consistent
(Prieto et al. 2016). The homoscedasticity ensured the predictions of fatigue life were
consistent throughout the whole regression range.
The significance of the vibration-life regression in predicting the spring fatigue
life was examined using the F-test. The F-ratio in the ANOVA tested whether the
overall regressions were a good fit of data (Sheridan et al. 2008). The results have shown
that the independent variables were statistically significant in predicting the dependent
variable (Table 4.12). For the Coffin-Manson vibration-life regression, F-test results of
F(2, 87) = 313.5 with p < 0.05 were obtained. The Morrow vibration-life regression was
F(2, 87) = 321.3 with p < 0.05 while the SWT vibration-life regression possessed F(2,
87) = 215.7 with p < 0.05. The regression p-values have indicated that the regression
approaches were significant at an α-level of 0.05, confidence interval of 95 % (Sheridan
et al. 2008). Hence, all three generated regressions have good fit of data which
supported the strong linear relationships (Delijaicov et al. 2010).
160
Table 4.12 F-test for vibration-life regression analysis
Approaches Sample sizes, n Number of independent variables, k F-value P-value
Coffin-Manson 90 2 313.5 <0.05
Morrow 90 2 321.3 <0.05
SWT 90 2 215.7 <0.05
The F-test is used to test the equality of two populations while the t-test is used
to compared to related variables because t-test is a univariate hypothesis test. The t-
values of both predictors of all three regression approaches are listed in Table 4.13. The
higher the absolute t-value, the higher the significance levels of the predictor variable
(Delijaicov et al. 2010). When the p-values of t-test of variables were lower than 0.05,
the variables were considered as significance predictor in the model (Takemoto et al.
2016). Based on Table 4.13, the regression constants have the highest significance level
with the highest t-values while the ISO 2631 vertical vibration were the second. The
third predictors were suspension natural frequency. Nevertheless, all the three
predictors were significant with p-values below 0.05. This indicated that all two
independent variables and constant were significant in predicting spring fatigue life and
the significance level was directly related to the p-value. In this case, all the p-value was
below 0.05 which showed a great significance of the independent variables.
Table 4.13 t-test of vibration-life datasets for various approaches
Approaches
t-value
p-value Constant, α
Vehicle body
frequency, 𝝎𝒏
Weighted
acceleration, Wa
Coffin-Manson 25.85 -10.35 -21.35 <0.05
Morrow 27.27 -11.54 -20.65 <0.05
SWT 21.31 -9.33 -17.26 <0.05
To analyse the relationship between each individual parameter to fatigue life
parameter, standardised coefficients were used to compare the relative importance of
each coefficient in a regression as shown in Table 4.14. The standardised coefficient
was used to explain the relationship between spring fatigue life, vehicle body frequency
and ISO 2631 vertical vibration. For example, for each one standard deviation increase
in vehicle body frequency, predicted Coffin-Manson fatigue life went down by 0.390
of standard deviation. In addition, for each one standard deviation increase in ISO 2631
weighted acceleration, predicted Coffin-Manson fatigue life went down by 0.805 of
standard deviation. These standardised coefficients analysis were also applied to
Morrow and SWT regression approaches. The standardised coefficients explained each
161
parameter in terms of standard deviation and ignored the scale of units which made the
comparisons easy. It suggested that the ISO 2631 vertical vibration has higher influence
on fatigue life when considering the absolute value of standardised coefficients as listed
in Table 4.14.
Table 4.14 Standardised coefficients of independent variables
Approaches Standardised Coefficients
Vehicle body frequency, 𝝎𝒏 ISO 2631 Weighted acceleration, 𝝎𝒂
Coffin-Manson -0.390 -0.805
Morrow -0.432 -0.773
SWT -0.413 -0.764
The vibration-life regression approaches were evaluated using normal P-P plots
and residual histograms to ensure the normality of the datasets. In MLR analysis, there
is more than one explanatory variable making traditional plot impractical to represent
the data. Therefore, the predicted value versus the observed value for these applied
datasets were plotted in the form of normal P-P plots as shown in Figure 4.25. These
data nicely fitted and there was a strong correlation between the regression prediction
and actual results because the points were distributed around the regression line to form
a linear relationship. Meanwhile, the normal P-P plot of regression of these standardised
residuals for the dependent variable also indicated a relatively normal distribution where
the actual values lined up along the diagonal that goes from lower left to upper right
(Sheridan et al. 2008). In this case, the datasets were normally distributed and good
predictions of fatigue life were obtained because the predicted fatigue lives were
distributed between the mean value according to centre limit theorem and identically
distributed.
To examine whether there were existing outliers in the regressions, the residuals
of the regressions were analysed. The residuals were obtained by deducting the
predicted values from the observed values from the regression (Shaman et al. 2015).
The residuals were also significant to determine outliers and how good the regression
was in providing solutions. Therefore, the regression approaches needed to satisfy the
condition that the error terms were normally distributed. If the data followed a normal
distribution with mean µ and variance σ2, then a plot of the theoretical percentiles of the
normal distribution versus the observed sample percentiles were approximately linear
(Sheridan et al. 2008).
162
Apart from normal P-P plot, residual histogram plots of the vibration-life
regression approaches were used to define the error terms. The residual histograms were
plotted into Figure 4.26. As observed in Figure 4.26, the patterns have confirmed the
assumption that the residuals were normally distributed at each level of fatigue life and
constant in variance across all levels of the dependent variable. Since the deviations of
the datasets were small, the datasets were considered as acceptable for spring fatigue
life predictions because all the data were fitted nicely to the curve (Junior & Pires 2014).
(a) (b)
(c)
Figure 4.25 Normal Probability-Probability plot of vibration-life regression standardised
residual for various approaches: (a) Coffin-Manson, (b) Morrow, (c) SWT
The most useful way to plot the residuals was with predicted values from the
regression on the X-axis and standardised residuals on the Y-axis as shown in Figure
4.27. When the data distribution satisfied this random condition, the residuals
distribution in the scatter plot were almost symmetric, tending to cluster towards the
middle of the plot. When the data were homoscedastic, the residuals were
Observed cumulative probability
Exp
ecte
d c
um
ula
tive
pro
bab
ilit
y
Observed cumulative probability
Exp
ecte
d c
um
ula
tive
pro
bab
ilit
y
Observed cumulative probability
Exp
ecte
d c
um
ula
tive
pro
bab
ilit
y
163
approximately equal for all predicted dependent variable scores (Prieto et al. 2016). No
significant pattern was observed on any of the residual scatter plots that violated the
suitability of fit of the regression. The data was distributed evenly among the zero line
and no cluster was observed. To obtain information of the data dispersion, linear
analysis was performed on the regression standardised residual and standardised
predicted values. When a straight linear regression line was obtained, fatigue life
estimation was consistent for all the prediction range (Azadi & Karimi-Jashni 2016).
(a) (b)
(c)
Figure 4.26 Error histogram of vibration-life regression standardised residual for various
approaches: (a) Coffin- Manson, (b) Morrow, (c) SWT
The application of the regression was applied to the vehicle body frequency and
ISO 2631 vertical vibration as inputs with the Coffin-Manson, Morrow and SWT MLR-
based fatigue life predictions shown in Tables 4.15 – 4.17 respectively. The Coffin-
Manson MLR-based predicted fatigue life possessed a RMSE of 0.5165 when compared
with target fatigue life while the Morrow MLR-based predicted fatigue life had an
Regression standardised residual
Fre
qu
ency
Regression standardised residual
Fre
qu
ency
Regression standardised residual
Fre
qu
ency
164
RMSE of 0.5274. Subsequently, the SWT MLR-based predicted fatigue life consisted
of RMSE of 0.3775. For the calculated RMSE of fatigue life, they were defined in
natural logarithm and the RMSE value was acceptable because the fatigue life range
was 2 to 8 (Cárdenas et al. 2012). As observed in Table 4.15, when the vehicle body
frequency and ISO 2631 vertical vibration were 1.00 Hz and 0.15 m/s2 respectively, the
fatigue life was high and the predicted fatigue life deviated a lot from the target value.
This was a limitation of regression approach on data range where its density shape
depended on the parameters that index the distribution (Ospina & Ferrari 2012). The
correlation study of predicted fatigue life was also performed using 1:2 or 2:1
correlation curve as shown in Figure 4.28. However, 10 out of 12 MLR predicted
Morrow fatigue lives distributed out beyond the conservative boundary due to the mean
stress effects. Morrow adjusted the mean stress for fatigue life calculation which led to
the varying results. Nevertheless, the Coffin-Manson and SWT was suitable in
predicting the fatigue life because most of the data points distributed within the
acceptable boundary.
(a) (b)
(c)
Figure 4.27 Scatter plot of vibration-life regression standardised residual for various approaches:
(a) Coffin- Manson, (b) Morrow, (c) SWT
Regression standardised predicted value
Reg
ress
ion
sta
ndar
dis
ed r
esid
ual
Regression standardised predicted value
Reg
ress
ion
sta
ndar
dis
ed r
esid
ual
Regression standardised predicted value
Reg
ress
ion
sta
ndar
dis
ed r
esid
ual
165
Table 4.15 Coffin-Manson MLR-based fatigue life predictions
Vehicle body frequency
(Hz)
ISO 2631 vertical
Acceleration (m/s2)
MLR-based fatigue life
(blocks to failure)
Target fatigue life
(blocks to failure)
1.25 0.35 5.09 × 105 4.27 × 105
1.25 0.67 3.03 × 104 1.20 × 104
1.45 0.75 2.79 × 103 5.75 × 103
1.25 0.66 3.31 × 104 9.55 × 103
1.50 0.73 2.18 × 103 5.75 × 103
1.50 0.40 4.01 × 104 2.82 × 104
1.00 0.15 2.43 × 107 2.04× 108
1.36 0.16 1.08 × 106 1.51 × 106
1.41 0.16 7.09 × 105 7.76 × 105
1.25 0.32 6.64 × 105 9.33 × 105
1.07 0.62 2.14 × 104 3.72 × 104
1.07 1.24 9.02 × 102 7.94 × 102
Table 4.16 Morrow MLR-based fatigue life predictions
Vehicle body frequency
(Hz)
ISO 2631 vertical
Acceleration (m/s2)
MLR-based fatigue life
(blocks to failure)
Target fatigue life
(blocks to failure)
1.40 0.40 8.39 × 104 5.37 × 104
1.24 0.62 5.11 × 104 6.46 × 104
1.35 0.66 1.17 × 104 4.07 × 103
1.40 0.67 6.48 × 103 3.80 × 103
1.06 0.62 3.06 × 105 6.31 × 104
1.45 0.77 1.53 × 103 6.76 × 103
1.06 0.61 3.37 × 105 5.50 × 104
1.35 0.71 7.29 × 103 7.24 × 103
1.30 0.16 2.21 × 106 3.02 × 106
1.00 0.31 1.05 × 107 3.63 × 107
1.45 0.33 9.92 × 104 8.32 × 104
1.50 0.68 2.18 × 104 6.92 × 102
Table 4.17 SWT MLR-based fatigue life predictions
Vehicle body frequency
(Hz)
ISO 2631 vertical
Acceleration (m/s2)
MLR-based fatigue life
(blocks to failure)
Target fatigue life
(blocks to failure)
1.30 0.62 2.82 × 104 1.29 × 104
1.41 0.66 6.45 × 103 5.62 × 103
1.30 0.69 1.45 × 104 8.71 × 103
1.07 0.57 4.45 × 105 1.26 × 105
1.19 0.61 9.25 × 104 1.38 × 104
1.13 0.31 2.89 × 106 4.17 × 106
1.19 0.33 1.32 × 106 1.17 × 106
1.25 0.34 6.59 × 105 5.89 × 105
1.36 0.36 1.83 × 105 1.82 × 105
1.45 0.39 5.61 × 104 9.12 × 104
1.50 0.4 3.10 × 104 7.24 × 104
1.00 0.15 4.81 × 107 2.69 × 108
Note: Difference = |MLR-based fatigue life - target fatigue life
target fatigue life| × 100% (Equation 3.13)
166
(a)
(b)
(c) Figure 4.28 Correlation analysis for MLR fatigue life: (a) Coffin-Manson, (b) Morrow, (c) SWT
1.0E+02
1.0E+03
1.0E+04
1.0E+05
1.0E+06
1.0E+07
1.0E+02 1.0E+04 1.0E+06
1.0E+02
1.0E+03
1.0E+04
1.0E+05
1.0E+06
1.0E+07
1.0E+02 1.0E+04 1.0E+06
1.0E+02
1.0E+03
1.0E+04
1.0E+05
1.0E+06
1.0E+07
1.0E+02 1.0E+04 1.0E+06
Target fatigue life (blocks to failure)
Target fatigue life (blocks to failure)
Target fatigue life (blocks to failure)
ML
R p
red
icte
d f
atig
ue
life
(b
lock
s to
fai
lure
) M
LR
pre
dic
ted
fat
igue
life
(b
lock
s to
fai
lure
) M
LR
pre
dic
ted
fat
igue
life
(b
lock
s to
fai
lure
)
167
As for the vibration-life relationship, the suitability of the regression was
successfully established and evaluated using normal P-P and scatter plots as defined in
the first objective. The main challenges of regression analysis were not only to obtain
the fatigue life prediction relationship but also the consistency of the prediction
outcome. Hence, the normal P-P and scatter plots performed in this analysis were used
to illustrate the consistency of the established regression. The predicted fatigue life was
evaluated according to the target fatigue life. Nevertheless, the regression had a
limitation of wide data range predictions. This suggested vibration-life regression and
fatigue life predictions using ISO 2631 vertical vibration and vehicle body frequency is
the main idea for the novelty of this research which answered the first objective.
4.4.2 Analysis for Vibration Regression
ISO 2631 vertical vibration of vehicle mass associated to coil spring stiffness also
played a significant role in indicating vehicle ride. There was a need to perform
prediction of ISO 2631 vertical vibration using spring fatigue life which is known in
this thesis as vibration prediction. Using the MLR approach, the response surfaces of
regression approaches were obtained and shown in Figure 4.29. The regressions were
defined as follows:
Wa_CM = 2.59 - 0.76.ωninput - 0.22.NCM_MLR (4.11)
Wa_Morrow = 2.58 - 0.83.ωninput - 0.20.NMorrow_MLR (4.12)
Wa_SWT = 2.45 - 0.77.ωninput - 0.19.NSWT_MLR (4.13)
where Ncm_MLR, Nmorrow_MLR, NSWT_MLR are the Coffin-Manson, Morrow SWT fatigue life
input with unit blocks to failure in natural logarithm, 𝜔𝑛_𝑖𝑛𝑝𝑢𝑡 is the suspension natural
frequency with unit Hz having the same range for all three regressionsd, Wa_CM, WMorrow,
WSWT are the regressions estimated ISO 2631 vertical acceleration with unit m/s2 using
various fatigue approaches predicted fatigue life and suspension natural frequency as
input.
168
(a)
(b)
(c)
Figure 4.29 Response surface plot for ISO 2631 vertical vibration prediction using various
approaches: (a) Coffin-Manson, (b) Morrow, (c) SWT
169
The listed regressions were generated using un-standardised coefficients where
the units of the parameters remained. Hence, the unit for dependent variable was m/s2
while the independent variables were vehicle body frequency and fatigue life with units
of Hz and blocks to failure, respectively. Since MLR applied the unstandardized
coefficient, the units of the parameters varied because this regression approach served
to analyse the response change with predictors. When the standardised coefficient was
applied, the units of the parameters were unified with the unit of variance. The
unstandardized coefficients were more commonly used due to its simplicity (Darlington
& Hayes 2016).
The regressions were obtained from ANOVA where the F-test and t-test results
were analysed. For the F-test, the critical F-values were obtained as listed in Table 4.18.
The p-values for all the test were below 0.05 which indicated that the regressions were
significant. The significance levels of each independent variable were determined using
the t-test where the results are tabulated in Table 4.19. The significance of the parameter
depends on the t-test value where a higher t-test value showed greater effects of the
designated variable towards dependent variable. The highest t-value for all regression-
based vibration-life relationships were based on spring fatigue life parameter. This
indicated that the fatigue life played a significant role in defining the prediction
regressions, followed by the constant and vehicle body frequency respectively.
Table 4.18 F-test of fatigue-vibration datasets for various approaches
Approaches Sample sizes, n Number of independent variables, k F-value P-value
Coffin-Manson 90 2 232.1 <0.05
Morrow 90 2 218.8 <0.05
SWT 90 2 151.9 <0.05
Table 4.19 t-test of fatigue-vibration datasets for various approaches
Approaches
t-value
Constant, α Vehicle body
frequency, 𝝎𝒏 Fatigue life
Coffin-Manson 16.91 -7.93 -21.35
Morrow 16.86 -8.43 -20.65
SWT 13.62 -6.65 -17.26
The goodness of fit for the generated regression relationships were analysed and
shown in Figure 4.29. Based on Figure 4.29, there were acceptable correlation between
the regression prediction and actual results and these were supported by the high R2
170
value. The R2 value for vibration prediction Coffin-Manson regression was 0.8420
while the R2 value for Morrow vibration regression was 0.8340. Lastly, the R2 value for
SWT vibration regression was 0.7877 which showed a confidence level of 78 % for ISO
2631 vertical vibration prediction. According to Sivák and Ostertagová (2012), these
models has shown a high correlation value which indicated the data fitted well into the
model as defined in Table 3.6.
Based on the ANOVA analysis, the normality of the datasets was analysed using
Normal P-P plots where the plots of the vibration regression approaches were obtained
as shown in Figure 4.30. When observing the Figure 4.30, the datasets were distributed
along the centre line which indicated the datasets were normally distributed. For the
MLR approach, normality of the datasets was a requirement to perform the analysis
because the assumptions of linearity between parameters were conducted.
Subsequently, examination of residuals was performed for vibration regression
relationship and plotted into error histograms as shown in Figure 4.31. As observed
from Figure 4.31, the residuals were distributed at the centre of the error histograms
which also suggested the normality of the residuals.
In Figure 4.31, the residuals ranged from -3 to 3 for Coffin-Manson vibration
regression with a standard deviation of 0.989. The residuals range of Morrow vibration
relationship predictions were from -3 to 3.2 while the range for SWT relationships were
also from -3 to 3. In this case, the dataset variations were considered as very high and
not suitable for ISO 2631 vertical vibration predictions which ranged from 1.0 to 1.5
m/s2. Indirect causal relationship mentioned that one of the variables exerted a causal
impact on the second variable but only through the impacts on a third variable (Jaccard
& Turrisi 2003). This caused the independent parameters to be less efficient in
predicting the ISO 2631 vertical vibrations.
After analysing the normality of the datasets, the homoscedasticity of the
obtained regressions was analysed using scatter plot as shown in Figure 4.32. As
observed in Figure 4.32, a linear regression with R2 value of zero was obtained at centre
line. When a linear regression lined up with zero and no clear pattern of data was
observed in a scatter plot, the regressions were assumed to be homoscedastic (Casson
171
& Farmer 2014). In this case, the regressions were consistently predicting the ISO 2631
vertical vibrations where the deviations were within the acceptable range.
Observed cumulative probability
Observed cumulative probability
(a) (b)
Observed cumulative probability
(c)
Figure 4.30 Normal Probability-Probability plot of vibration regression standardised residual
for various approaches: (a) Coffin-Manson, (b) Morrow, (c) SWT
The results of ISO 2631 vertical vibration prediction using the MLR approach
and Coffin-Manson fatigue life input were obtained as shown in Table 4.20. In Table
4.20, there were a total of 12 sets of ISO 2631 vertical vibration predictions and targeted
value. The predicted ISO 2631 vertical vibrations were close to the targeted value with
a RMSE of 0.0875. For Morrow, the obtained RMSE was 0.1222 and the data for this
analysis is shown in Table 4.21. The RMSE for SWT predicted ISO 2631 vertical
vibration was 0.1279 with data as shown in Table 4.22. In Table 4.22, the high
difference of prediction up to 75.9% was obtained. This was mainly due to the limitation
of multiple linear regression in predicting the wide range data which induced higher
error range (Ospina & Ferrari 2012). Nevertheless, most of the predicted ISO 2631
vertical vibrations were still within acceptable range.
E
xp
ecte
d c
um
ula
tive
pro
bab
ilit
y
E
xp
ecte
d c
um
ula
tive
pro
bab
ilit
y
E
xp
ecte
d c
um
ula
tive
pro
bab
ilit
y
172
Regression standardised residual
Regression standardised residual
(a) (b)
Regression standardised residual
(c)
Figure 4.31 Error histogram of vibration regression standardised residual for various
approaches: (a) Coffin- Manson, (b) Morrow, (c) SWT
(a) (b)
Regression standardised predicted value
(c)
Figure 4.32 Scatter plot of vibration prediction regression standardised residual for various
approaches: (a) Coffin-Manson, (b) Morrow, (c) SWT
Fre
qu
ency
F
req
uen
cy
Fre
qu
ency
Regression standardised predicted value
Reg
ress
ion
sta
ndar
dis
ed
resi
du
al
Regression standardised predicted value
Reg
ress
ion
sta
ndar
dis
ed
resi
du
al
Reg
ress
ion
sta
ndar
dis
ed
resi
du
al
173
The difference between MLR predicted and target ISO 2631 vertical vibration
values were mostly acceptable. Nevertheless, there were a few points which distributed
beyond the acceptable limit and the highest difference went up to 41.38% which was
far beyond the acceptable limit. The high percentage of difference was due to the tail
value of the distribution which was a limitation of MLR (Ospina & Ferrari 2012). In
this case, the highest percentage difference of 41.38% was predicted from the extreme
value like the smallest frequency value of 1.0 Hz and the highest fatigue life at 8.45
(2.82 × 108 blocks to failure). In most of the cases, the prediction range was within the
acceptable limit of 20% which indicated that the regression vibration predictions were
still acceptable (Manivel & Gandhinathan 2016).
Table 4.20 Coffin-Manson MLR predicted ISO 2631 vertical vibrations
Vehicle
body
frequency
(Hz)
Fatigue life
(blocks to failure in
natural logarithm)
MLR predicted
ISO 2631 vertical
vibration (m/s2)
Target ISO
2631 vertical
vibration
(m/s2)
Difference
(%)
1.13 6.72 0.25 0.33 24.24
1.25 5.63 0.40 0.35 14.28
1.41 4.88 0.44 0.39 12.82
1.19 5.14 0.55 0.59 6.77
1.50 3.72 0.63 0.68 7.35
1.30 3.96 0.73 0.69 5.79
1.50 3.76 0.62 0.73 15.07
1.19 5.68 0.21 0.27 22.22
1.25 6.78 0.15 0.16 6.25
1.30 5.65 0.36 0.32 12.50
1.50 4.68 0.42 0.34 23.5
1.19 3.77 0.86 0.63 36.50
Table 4.21 Morrow MLR predicted ISO 2631 vertical vibrations
Vehicle
body
frequency
(Hz)
Fatigue life
(blocks to failure in
natural logarithm)
MLR predicted
ISO 2631
vertical
vibration (m/s2)
Target ISO
2631 vertical
vibration
(m/s2)
Difference
(%)
1.00 6.00 0.62 0.54 14.81
1.13 4.72 0.70 0.61 14.75
1.30 3.94 0.71 0.70 1.43
1.25 5.53 0.44 0.35 25.71
1.40 4.94 0.43 0.39 10.25
1.13 7.82 0.12 0.16 25.00
1.50 5.60 0.22 0.17 29.40
1.00 7.56 0.24 0.31 22.58
1.30 5.71 0.36 0.32 12.50
1.00 5.41 0.67 0.62 8.06
1.30 2.22 1.06 1.29 17.83
1.36 2.13 1.06 1.30 17.82
174
Table 4.22 SWT MLR predicted ISO 2631 vertical vibrations
Vehicle
body
frequency
(Hz)
Fatigue life
(blocks to failure in
natural logarithm)
MLR predicted
ISO 2631
vertical
vibration (m/s2)
Target ISO
2631 vertical
vibration
(m/s2)
Difference
(%)
1.00 8.45 0.17 0.29 41.38
1.00 5.11 0.71 0.54 31.48
1.40 3.79 0.64 0.73 12.33
1.07 5.10 0.66 0.57 15.78
1.36 3.02 0.83 0.70 18.57
1.07 6.89 0.32 0.30 6.67
1.36 5.26 0.40 0.36 11.11
1.07 8.33 0.10 0.15 33.33
1.07 7.23 0.25 0.31 19.35
1.00 5.38 0.66 0.62 6.45
1.36 3.19 0.80 0.65 23.07
1.40 3.02 0.79 0.66 19.70
Note: Difference = |MLR prediction - target ISO 2631 vertical vibration
target ISO 2631 vertical vibration| × 100% (Equation 3.13)
The novelty of this research was to integrate the spring fatigue life, ISO 2631
vertical vibration and vehicle body frequency into a regression relationship. The spring
fatigue life was processed using the loading signals obtained from simulation of the
quarter car model with material cyclic properties and coil spring stress-strain simulation
from FEA. Meanwhile, the acceleration signals from the vehicle sprung mass were
processed into ISO 2631 ride index. Lastly, the spring stiffness was used to calculate
the natural frequencies for different spring designs. In section 4.4, the regression to link
these three parameters was performed and their performances were evaluated.
In summary, this section has proposed regressions for fatigue life and ISO 2631
vertical vibration predictions using Coffin-Manson, Morrow and SWT strain life
approaches. The proposed regressions have predicted spring fatigue life or ISO 2631
vertical vibration with low RMSE when compared to targeted value. These regressions
were newly established with the purpose of assisting in automotive spring design
through reducing the processes. Hence, the outcome of this section has revealed the
novelty of this research and after an extensive study of literature, no similar proposal
has been found. These established regressions could play a significant role in the field
of fatigue and automotive ride dynamics. Nevertheless, the MLR prediction had some
restriction in predicting extreme values for ISO 2631 vertical vibrations.
175
The first objective was successfully achieved with the determination of these
MLR fatigue life and ISO 2631 vertical vibration predictions. However, MLR is very
limited in capability. Based on a research of bus chassis shear stress prediction, the
capabilities of MLR and ANN approach in predicting the stress results were analysed
(Patel & Bhatt 2013). ANN method has been reported to have higher accuracy than
MLR predictions. Hence, to obtain a more robust prediction model, the same datasets
were proposed to train artificial neural networks (ANN). ANN consist of self-adjusting
capability in adjusting the gradient to obtain optimised fatigue life predictions which
led to the determination of the second objective.
4.5 IMPLEMENTATION OF ARTIFICIAL NEURAL NETWORK FOR
PREDICTION
The second objective of this research is to optimise the fatigue life and ISO 2631 vertical
vibration predictions using ANN. This section presents the results of ANN optimising
procedures, optimised fatigue life and ISO 2631 vertical vibration which was also part
of the novelty of this research because the optimised predictions provided higher
accuracy than the regression predictions. The subsection 4.5.1 depicts the optimised
fatigue life predictions while the subsection 4.5.2 illustrates the optimised ISO 2631
vertical vibration predictions.
4.5.1 Fatigue Life Prediction
To search for a method to optimise the spring fatigue life and automobile ISO 2631
vertical vibration, the same datasets from MLR were used to train ANNs. In this section,
the idea was to optimise the spring fatigue life predictions using ANN approach.
Initially, the feedforward type ANN was trained using Matlab® with Levenberg-
Marquadt training algorithm. The feedforward neural network was selected because the
optimised time and accuracy of the method when compared to other neural networks
(Welch et al. 2009). After determining the type of neural network, the next step was to
train a neural network with a suitable architecture.
176
Random selection of several hidden neurons and layers caused either overfitting
or underfitting (Ghana Sheela & Deepa 2013). The number of neurons and layers were
highly dependent on the characteristics of the data. However, until now, there were no
established rules for ANN architecture design. Previous methods to determine the
suitable number of neurons and layers were based on trial and error (Ghana Sheela &
Deepa 2013). Due to this reason, there was a need to develop a new procedure to
determine the optimum prediction architecture which formed the novelty of this second
objective. For a single hidden layer feedforward neural network, the number of neurons
in the hidden layer was the most important parameter to determine the accuracy of the
trained ANN. The key to assess the goodness of the trained ANN was through analysing
the MSE. Hence, the MSE with one to ten neurons of Coffin-Manson, Morrow and
SWT ANN were selected for these ANN training processes because ten neurons were
sufficient to handle analysis with three parameters (Salim et al. 2015).
When the number of neurons increased or decreased, the MSE of the predictions
were also randomly varying. The results of the MSE across a number of neurons for
Coffin-Manson, Morrow and SWT trained feedforward ANN are plotted into Figure
4.33. The minimum MSE for trained Coffin-Manson datasets was detected at seven
neurons with the value of 0.0885, as shown in Figure 4.34(a). However, different results
were observed for Morrow and SWT ANN-based vibration-life approach. The lowest
MSE for trained Morrow ANN-based vibration-life approach was 0.0827 with nine
neurons in a hidden layer while the lowest MSE for trained SWT ANN-based vibration-
life approach was 0.1769 with only one neuron. The list of the MSE is shown in
Appendix G. When the number of neurons was reduced, the ANN has become simpler
with less weights and biases. Based on Heaton (2015), the minimum number of neurons
was the mean value for the input and output parameters. However, the process to
determine the most suitable number of neurons was an empirical exercise which
consumed a lot of effort. The benefit of ANN with less weights and biases was the
reduced computational power required. Nevertheless, the selection of number of
neurons depended on the MSE which defined the accuracy of predictions.
As observed in Figure 4.33(a), the maximum MSE happened at one neuron for
Coffin-Manson approach while the maximum MSE happened at six neurons for both
177
Morrow and SWT approaches as shown in Figure 4.33(b) and (c), respectively. These
high MSE sets of neuron architecture should be avoided because it yielded great
deviation of predictions. The high MSE was due to the ANN generated random initial
weights that led to varying local minima. The initial weight is difficult to decide (Lee,
Geem & Suh 2016). After considering the MSE, the trained ANN architecture with the
lowest MSE was chosen to be analysed in terms of ‘goodness’ of fit. The total datasets
were divided into three main categories, which was training, validation and ANN test
set. The training set data consisted of 70% of the total data, validation and test set
consisted of 15% each of the total data respectively. This type of data separation was
considered as one of the most optimum when the number of datasets were below 100
(Takahashi et al. 2016). In this analysis, the training dataset was used to adjust the
weights on the ANN. The validation set was used to minimise overfitting while the test
set was used to test the final solution of the actual predictive results of the ANN.
The goodness of fit for the datasets were analysed using indicators known as
Pearson correlation coefficient (r) and the R2 value. The r value was used to present the
strength and direction of a linear relationship while the R2 value was used to show the
proportion of explained variance. The fitted curve for trained Coffin-Manson ANN is
shown in Figure 4.34 where the R2 of each fitting was shown. For Morrow and SWT
curve fitting, the results are listed in Appendix J. The R2 value was determined based
on the Pearson coefficient. All the fitted curves have shown good fit and the datasets
were well distributed along the line. In this case, fitted curve for ‘All’ spent all 90
datasets into weights adjustment have also shown a good agreement between the data.
The R2 value for ANN indicated how good the selected ANN was trained which could
directly affect the fatigue life predictions. A high R2 indicated that the ANN was
predicting with high accuracy and low MSE (Mia, Khan & Dhar 2017).
The summarised R2 value and datasets for all three approaches were tabulated
in Table 4.23. The R2 value for Coffin-Manson training, validation and ANN test dataset
was high and the ANN test was used to assess the performance or generalisation ability
of ANN (Kamp & Savenije 2006). For Coffin-Manson approach, the R2 value for ANN
test was 0.9322 while the R2 value for Morrow approach was 0.9732. This indicated
that the datasets fitted well into the trained ANN. The Morrow training datasets were
178
even closer to the fitted ANN regression. On the other hand, the R2 value for trained
SWT ANN was 0.9502. Although the R2 value of the trained SWT vibration-life ANN
was not high as the Coffin-Manson and Morrow approaches, the datasets were
considered as acceptable. Nevertheless, as observed from all the three-curve fitted
ANN, the data were well distributed around the model line.
(a)
(b)
(c)
Figure 4.33 MSE of trained neural network with single hidden layer for various approaches:
(a) Coffin-Manson, (b) Morrow, (c) SWT
Maximum MSE @ 6 neurons
Number of neurons
Mea
n s
quar
e er
ror
(blo
cks
to f
ailu
re)
Number of neurons
Mea
n s
quar
e er
ror
(blo
cks
to f
ailu
re)
Number of neurons
Mea
n s
quar
e er
ror
(blo
cks
to f
ailu
re)
Maximum MSE @ 1 neuron
Maximum MSE @ 6 neurons
179
In Figure 4.34, the datasets for all training, validation and ANN tests were
trained according to data separation and the ‘ANN Output’ of ANN is labelled on the
vertical axis as the ANN predictions. Meanwhile, the horizontal axis is the ‘Target’
fatigue life which was the supervised learning response. Through this curve fitting, the
difference between ANN predictions and target value for various separated datasets
could be identified and the corresponding correlation between the prediction and target
value identified (Zadeh et al. 2010). These functions could be used to evaluate the
performance of trained ANN.
(a) (b)
(c) (d)
Figure 4.34 Curve fitting of trained Coffin-Manson vibration-life ANN approach with single
hidden layer for various datasets: (a) all, (b) training, (c) validation, (d) ANN test
Target
All: R2 = 0.9448
AN
N o
utp
ut
Target
Training: R2 = 0.9467
AN
N o
utp
ut
Target
Validation: R2 = 0.9520
AN
N o
utp
ut
Target
ANN Test: R2 = 0.9322
AN
N o
utp
ut
180
Table 4.23 R2 value for all approaches with single hidden layer ANN
All Training Validation ANN Test
Coffin-Manson 0.9448 0.9467 0.9520 0.9322
Morrow 0.9700 0.9873 0.9748 0.9732
SWT 0.8885 0.8947 0.7305 0.9504
Table 4.24 summarises the MSE for all three sets of training, validation and
ANN test data where all 90 sets of training data for all three strain life models are listed.
The MSE for trained ANN-based Coffin-Manson vibration-life ANN was below
0.2000. The MSE was relatively low when compared to the datasets. The ANN-based
Morrow vibration-life approach also showed low MSE of below 0.0827. Nevertheless,
the SWT has shown a higher MSE when compared to ANN-based Coffin-Manson and
Morrow vibration-life approach. This was due to the fatigue data characteristics of the
SWT vibration-life approach. Morrow predicted the fatigue life better with compression
mean stress while SWT predicted the fatigue life in tension which led to deviation of
results between Coffin-Manson, Morrow and SWT (Ince & Glinka 2011).
Table 4.24 MSE for various approaches and datasets with single hidden layer ANN
Coffin-Manson Morrow SWT
All 0.1232 0.0839 0.3241
Training 0.1137 0.0753 0.3197
Validation 0.2001 0.1234 0.4911
Testing 0.0885 0.0827 0.1769
A good data fit of R2 value did not indicate that the trained ANN had a low MSE.
To study the error, error histograms of all three trained ANN vibration-life were plotted
into Figure 4.35 with details listed in Appendix I. The error range between the output
and targeted fatigue life of ANN-based Coffin-Manson vibration-life is shown in Figure
4.36(a) (-0.7316 to 0.3082). The error range of Morrow vibration-life ANN ranged from
-0.7621 to 0.6157 while the error range of SWT vibration-life ANN was from -0.7319
to 0.3858 as shown in Figures 4.35(b) & (c) respectively. In this ANN analysis, the
targeted value of fatigue lives was in the range of 2 to 8. This error range was relatively
small when compared to the fatigue life range 2 to 8 due to the optimisation process
loop. When compared to MLR method, the error of single hidden layer ANN has proven
to be lower where the SWT MSE dropped from 0.7056 to 0.1769. This suggests that
ANN provided a better fatigue life prediction because the ANNs were trained and
improved to deal with linear and non-linear dependence datasets.
181
Subsequently, normality test was performed on the residuals of all trained
vibration-life ANN outcome. Examining the normality of residuals is significant in
regression analysis because it implies consistency of predictions (Xie et al. 2017). The
normality of the data was tested using the Lilliefors test because it is one of suitable
methods to determine normality with a sample size of less than 100 (Ayinde et al. 2017).
The results have shown that all the predictions have failed to reject the null hyprothesis
and therefore, the residuals were normally distributed. Based on the Lilliefors test, it
indicated a failure to reject the null hypothesis at the default 5% significance level. The
alternate hypothesis was that the data did not come from a normal distribution. Based
on the Lilliefors test, all the residual datasets failed to reject the null hypothesis which
indicated that all the residuals were normally distributed.
(a) (b)
(c)
Figure 4.35 Error histogram of single hidden layer vibration-life ANN for various approaches:
(a) Coffin-Manson, (b) Morrow, (c) SWT
The analysis of a single hidden layer in optimising the spring fatigue life is
shown in Section 3.4.2. This analysis was a discovery process where the ANN
architecture with the lowest MSE was initially defined. Furthermore, the predicted
182
fatigue lives were evaluated using normality test to ensure the consistency of
predictions. The process was repeated for two and three hidden layers until the most
satisfactory ANN architecture for fatigue life prediction was obtained and the second
objective was achieved.
Other than the number of neurons, the other way to enhance the prediction of
trained ANN was through adding more hidden layers (Hemmat Esfe et al. 2015). Hence,
an additional hidden layer was added to ANN architecture to train a model with
improved accuracy. This modified the ANN from a single hidden layer to become two
hidden layers. Subsequently, MSE analyses of the two hidden layers ANN were also
performed and plotted into Figure 4.36. Coffin-Manson vibration-life ANN with three
neurons in first hidden layer and seven neurons in second layer consisted of the lowest
MSE value of 0.0249. For Morrow vibration-life ANN, the lowest MSE was 0.0333
with eight neurons in first hidden layer and six neurons in second hidden layer. Lastly,
the trained SWT vibration-life ANN consisted of MSE value of 0.0945 with one and
six neurons in first and second hidden layers respectively.
Subsequently, curve fitting of trained Coffin-Manson ANN with two hidden
layers were also plotted into Figure 4.37 but the Morrow and SWT approaches were
plotted into Appendix J. The goodness of the fitted curve was determined using R2 value
and the results were tabulated in Table 4.25. As observed from Table 4.25, all the R2
value for trained Coffin-Manson ANN was above 0.8000 with minimum R2 value of
0.8703 which was considered as good. Based on the R2 value, the trained Morrow ANN
was considered as very good with all the R2 value above 0.9000. The R2 value of the
trained SWT ANN was higher than 0.8000 which was considered as good. All MSE of
these fitted ANNs was tabulated in Table 4.26. The minimum MSE for two-layer
Coffin-Manson ANN was 0.0249 in ANN test datasets while Coffin-Manson and
Morrow were 0.0333, 0.0945 respectively. The obtained MSE was considered as low
when compared to the fatigue life datasets with a range of 2 to 8.
183
(a) (b)
(c)
Figure 4.36 MSE of trained vibration-life ANN with two hidden layers for various approaches:
(a) Coffin-Manson, (b) Morrow, (c) SWT
When compared to single hidden layer, the SWT vibration-life ANN also
possessed the highest MSE value of 0.0945 among all three approaches. Nevertheless,
the overall test datasets have shown some improvements with the reduction of MSE.
However, increasing the hidden layer of ANN did not guarantee with improvement of
the trained model (Ghana Sheela & Deepa 2013). Based on the observation between
single and two hidden layers ANN, adding another hidden layer improved the MSE of
trained ANN-based Coffin-Manson fatigue life from 0.0885 to 0.0249. The MSE of
trained ANN-based Morrow fatigue life dropped from 0.0827 to 0.0333 while the
trained ANN-based SWT fatigue life dropped from 0.1769 to 0.0945. The MSE of
fatigue life for all three trained ANN have shown significant improvements. In this
fatigue data ANN analysis, adding an additional hidden layer has helped to reduce the
test datasets MSE and enhanced fatigue life predictions.
184
(a) (b)
(c) (d)
Figure 4.37 Curve fitting of trained Coffin-Manson vibration-life ANN with two hidden layers for
various datasets: (a) all, (b) training, (c) validation, (d) ANN test
Table 4.25 R2 value for all approaches with two hidden layers ANN
all training validation test
Coffin-Manson 0.9426 0.9432 0.8703 0.9926
Morrow 0.9736 0.9657 0.9775 0.9918
SWT 0.8879 0.8943 0.8122 0.9732
Table 4.26 MSE for various approaches and datasets with two hidden layers ANN
Coffin-Manson Morrow SWT
All 0.1278 0.0738 0.3222
Training 0.1248 0.0823 0.2585
Validation 0.2441 0.0770 0.8316
ANN Test 0.0249 0.0333 0.0945
In Figure 4.37, the ANN output in vertical axis is the ANN predicted fatigue
life. The ‘Target’ is the response for supervised learning where the performance of each
Validation: R2 = 0.8703
Target
All: R2 = 0.9426 A
NN
outp
ut
Target
Training: R2 = 0.9520
AN
N o
utp
ut
Target
AN
N o
utp
ut
Target
ANN test: R2 = 0.9926 A
NN
outp
ut
185
dataset (all, training, validation, test) could be evaluated using the regression. These
fitted curves served to evaluate the performance of ANN training with different datasets
as well as the stability of the ANN. The error range of test datasets for two hidden layers
were plotted into Figure 4.38 with details listed in Appendix I. For Coffin-Manson ANN
test datasets, the errors were distributed in the range of -0.2456 to 0.1578. The error
range for this Coffin-Manson vibration-life ANN predictions was low and acceptable
because the difference was only 10% from the original data range. Meanwhile, the error
range of trained Morrow test datasets were from -0.3460 to 0.2884. This error range
was wider than the trained Coffin-Manson ANN because of the fatigue data
characteristic. The width of error range was also related to the MSE which was an
indicator for accuracy. On the other hand, the trained SWT test datasets showed an error
range of -0.6605 to 0.2620. Based on the MSE, the error of SWT ANN fatigue life
prediction was expected to be higher than Coffin-Manson and Morrow. Nevertheless,
the distribution of the SWT vibration-life ANN error was still within the acceptable
range.
(a) (b)
(c)
Figure 4.38 Error histogram of two hidden layers vibration-life ANN for various approaches:
(a) Coffin-Manson, (b) Morrow, (c) SWT
186
After ANN with single and two hidden layers were investigated, a third hidden
layer was also applied and analysed. The R2 of fitting for all the trained ANN-based
vibration-life is plotted into Table 4.27. All the R2 values were examined to ensure the
ANN-based vibration-life approaches were well fitted. The MSE results of each training
were tabulated into Table 4.28. As observed, the MSE of test datasets were lower than
training and validation datasets due to the back propagation of the training algorithm.
Combined with the varying number of neurons, the MSE of the trained three hidden
layers Coffin-Manson ANN were plotted and shown in Figure 4.39. The minimum MSE
of this trained Coffin-Manson ANN was 0.0356 according to the ANN test datasets.
The MSE of trained three hidden layers Morrow and SWT ANN fatigue life was listed
in Appendix H. Throughout this analysis, the minimum MSE for trained Morrow ANN
was identified as 0.0117 while the lowest MSE of SWT datasets was 0.0824. Three
hidden layers ANN exhibited a significant improvement in MSE when compared to
single or two hidden layers ANN. As observed in Figures 4.39(e), 4.39(i), there was a few
trained ANN with extra high MSE. In Figure 4.39(e), the high MSE architecture was five
neurons in the first hidden layer, ten neurons in the second hidden layer and five neurons in
the third hidden layer while another high MSE point in Figure 4.39(i) was nine neurons in the
first layer, nine neurons in the second layer and ten neurons in the third hidden layer.
This phenomenon demonstrated that the increasing number of neurons did not
guarantee the improvement in performance. These high MSE architecture were
explained via the iterative learning process in which data cases are presented to the
network one at a time, and the weights associated with the input values were adjusted
each time. After all cases were presented, the process often starts over again. During
this learning phase, the network learnt by adjusting the weights to be able to predict the
correct class label of input samples. The random generation on weights and biases have
caused the high MSE. It is significant to know that the initial weights were chosen
randomly. After that the training, or learning, began. The random MSE effects were
also reported by Salim et al. (2011) where they used a loop to determine the ANN
architecture with the lowest MSE and the MSE were fluctuating across different
neurons. The outcome of the MSE as shown in Figure 4.39 was a significant procedure
for exploration of the optimised fatigue life. Utilising the deep learning in fatigue life
prediction also contributed to the novelty of this analysis.
187
(a) (b)
(c) (d)
(e) (f)
(g) (h)
(i) (j)
Figure 4.39 MSE of trained Coffin-Manson vibration-life ANN first hidden layer with various
number of neurons: (a) 1, (b) 2, (c) 3, (d) 4, (e) 5, (f) 6, (g) 7, (h) 8, (i) 9, (j) 10
188
A superior performance of ANN training was mostly dependent on the data
characteristic where a suitable number of layers and neurons were required. To analyse
the trained ANN, a fitted curve of trained three hidden layer Coffin-Manson vibration-
life ANN is shown in Figure 4.40. The r value of all the datasets were considerably high
at above 0.9000. When considering ANN-based Morrow vibration-life ANN, the r
values were also very high, indicating a very good fit of the datasets. The curve fitting
of trained Morrow and SWT ANN using are shown in Appendix J. The trained
vibration-life ANN has also shown a good fitted behaviour of datasets with the lowest
r value of 0.9396. After ensuring the trained ANN fitted nicely, the residuals of the
trained ANN were analysed using error histograms.
In Figure 4.40, the function of ANN ‘Output’ is defined on the vertical axis
while the ‘Target’ is on the horizontal axis for all, training, validation and ANN test
datasets. The significance of the function was to provide a correlation closest to the
target value for ANN performance evaluation. The residuals of the trained ANNs were
plotted into error histograms shown in Figure 4.41 with details listed in Appendix I. As
observed in Figure 4.41(a), the range of residuals for Coffin-Manson vibration-life
ANN was from -0.4041 to 0.2188. The residuals for ANN-based Morrow vibration-life
ANN ranged from -0.0913 to 0.2732 while for the SWT vibration-life ANN, the errors
ranged from -0.4195 to 0.4739. When compared to the two hidden layer vibration-life
ANN, the range of error has shown a significant reduction. To check the normality of
the errors, the Lilliefors test was also applied. The test showed that it failed to reject
null hypothesis which indicated that the residuals were normally distributed for all three
trained three hidden layer ANN residuals.
Table 4.27 R2 value for all approaches with three hidden layers ANN
all training validation test
Coffin-Manson 0.9395 0.9403 0.8530 0.9912
Morrow 0.9811 0.9781 0.9839 0.9944
SWT 0.9017 0.8828 0.9610 0.9559
Table 4.28 MSE for various approaches and datasets with three hidden layers ANN
Coffin-Manson Morrow SWT
All 0.1425 0.0529 0.2831
Training 0.1514 0.0636 0.3636
Validation 0.2102 0.0465 0.1274
ANN Test 0.0356 0.0117 0.0824
189
(a) (b)
(c) (d)
Figure 4.40 Curve fitting of trained Coffin-Manson vibration-life ANN with three hidden layers
for various datasets: (a) all, (b) training, (c) validation, (d) ANN test
Among these three approaches, the trained Morrow ANN has the lowest MSE
and the same trend applies for single and two hidden layers. The three hidden layers
ANN were much more complex than single and two hidden layers because additional
weights and biases as well as layer connectivity. Nevertheless, additional hidden layers
did not promise greater accuracy. The multilayer ANN were compared to single and
two hidden layers and results revealed that the trained ANN Coffin-Manson has fitted
greater with two hidden layers. The results of three hidden layers of Morrow and SWT
have shown better improvement with adding the third layer due to the lower MSE value.
All: R2 = 0.9395
Target
AN
N o
utp
ut
Training: R2 = 0.8703
Target
AN
N o
utp
ut
Validation: R2 = 0.8530
Target
AN
N o
utp
ut
Test: R2 = 0.9912
Target
AN
N o
utp
ut
190
(a) (b)
(c)
Figure 4.41 Error histogram of three hidden layers vibration-life ANN for various approaches:
(a) Coffin-Manson, (b) Morrow, (c) SWT
As a summary for the ANN determination, the best performing ANN
architecture for one, two and three hidden layer vibration-life ANN were identified. The
best architecture among the layers are tabulated in Table 4.29 with the selected ANN
architectures in bold text. Table 4.29 illustrates the ANN architecture with the lowest
MSE for fatigue life prediction which is significant in deciding the most suitable ANN
architecture from one to three hidden layers. The lowest MSE was important to
determine the final ANN architecture among all the designed ANN architecture. This
again contributed to the novelty of this research. In Table 4.29, ANN-based Coffin-
Manson vibration-life ANN with two hidden layers has shown the lowest MSE of
0.0356 and this type of ANN architecture is a hybrid multilayer neural network where
all the hidden layers are interconnected to the output layer. The number of neurons
implied the number of weights in this ANN. For the ANN architecture, the weights and
biases of the ANN were defined in Appendix K. This trained ANN architecture
consisted of two input neurons and three neurons in input layer where inputs and first
hidden layer were connected.
191
Table 4.29 The best performance ANN architecture for vibration-life predictions
Single hidden layer Two hidden layers Three hidden layers
Coffin-Manson 0.0885 0.0356 0.0249
Morrow 0.0827 0.0333 0.0117
SWT 0.1769 0.0945 0.0824
For fatigue life comparisons, the Coffin-Manson, Morrow and SWT ANN-based
predicted fatigue lives using ANN and MLR were listed into Tables 4.30, 4.31 and 4.32
respectively. To compare between ANN and MLR predictions, the RMSE between
ANN and MLR predictions to target fatigue life was identified. The Coffin-Manson
RMSE for MLR-based prediction was 0.5165 blocks to failure in natural logarithm
while the ANN was 0.1578 blocks to failure in natural logarithm. The RMSE has
improved by a significant amount. For Morrow, the RMSE for MLR prediction was
0.5274 blocks to failure in natural logarithm while the ANN prediction was 0.1080
blocks to failure in natural logarithm This also showed a drastic improvement.
Furthermore, the RMSE for MLR SWT prediction was 0.3775 blocks to failure in
natural logarithm while the RMSE for ANN predictions was 0.2870 blocks to failure in
natural logarithm.
Table 4.30 MLR and ANN predicted Coffin-Manson fatigue lives
ANN-based predictions
(blocks to failure)
MLR-based predictions
(blocks to failure)
Target fatigue life
(blocks to failure)
3.27 × 105 5.09 × 105 4.27 × 105
1.07 × 104 3.03 × 104 1.20 × 104
3.13 × 103 2.79 × 103 5.75 × 103
1.17× 104 3.31 × 104 9.55 × 103
3.16 × 103 2.18 × 103 5.75 × 103
3.87 × 104 4.01 × 104 2.82 × 104
1.62 × 108 2.43 × 107 2.04× 108
1.34 × 106 1.08 × 106 1.51 × 106
7.44 × 105 7.09 × 105 7.76 × 105
2.38 × 107 5.93 × 106 3.16 × 107
5.03 × 105 6.64 × 105 9.33 × 105
5.63 × 104 2.14 × 104 3.72 × 104
9.91 × 102 9.02 × 102 7.94 × 102
Although the improvement of RMSE for Morrow was not as high as Coffin-
Manson and SWT, there was still a significant improvement in terms of RMSE. When
the RMSE was low, the predicted value was close to the target value (Kisi 2015). Based
on the obtained RMSE, it suggested that the established ANN has improved the fatigue
life prediction when compared to the MLR approach. This analysis has shown that the
192
ANN approach successfully improved fatigue life predictions when compared to the
MLR approach.
Table 4.31 MLR and ANN predicted Morrow fatigue lives
ANN-based predictions
(blocks to failure)
MLR-based predictions
(blocks to failure)
Target fatigue life
(blocks to failure)
7.06 × 104 8.39 × 104 5.37 × 104
6.73 × 103 5.11 × 104 6.46 × 104
4.66 × 103 2.33 × 104 5.13 × 103
4.25 × 103 1.17 × 104 4.07 × 103
2.94 × 103 6.48 × 103 3.80 × 103
6.70 × 104 3.06 × 105 6.31 × 104
6.15 × 103 1.53 × 103 6.76 × 103
8.23 × 104 3.37 × 105 5.50 × 104
7.66 × 103 7.29 × 103 7.24 × 103
2.54 × 106 2.21 × 106 3.02 × 106
4.11 × 107 1.05 × 107 3.63 × 107
9.88 × 104 9.92 × 104 8.32 × 104
1.36 × 104 2.18 × 104 6.92 × 102
Table 4.32 MLR and ANN predicted SWT fatigue lives
ANN-based predictions
(blocks to failure)
MLR-based predictions
(blocks to failure)
Target fatigue life
(blocks to failure)
6.84 × 103 2.82 × 104 1.29 × 104
2.61 × 103 6.45 × 103 5.62 × 103
2.96 × 103 1.45 × 104 8.71 × 103
4.20 × 105 4.45 × 105 1.26 × 105
7.62 × 104 1.68 × 105 9.33 × 104
2.86 × 104 9.25 × 104 1.38 × 104
9.19 × 106 2.89 × 106 4.17 × 106
2.64 × 106 1.32 × 106 1.17 × 106
8.96 × 105 6.59 × 105 5.89 × 105
1.94 × 105 1.83 × 105 1.82 × 105
7.18 × 104 5.61 × 104 9.12 × 104
5.13 × 104 3.10 × 104 7.24 × 104
1.68 × 108 4.81 × 107 2.69 × 108
For fatigue life analysis, 1:2 or 2:1 correlation curve was used for determining
the accuracy of predictions (Karolczuk 2016). The correlation for ANN-predicted and
target Coffin-Manson fatigue life is shown in Figure 4.42 while the Morrow fatigue life
is shown in Figure 4.43. Based on the Coffin-Manson fatigue life correlation curve, only
three points deviated from the 1:2 or 2:1 boundary while the MLR predicted results
have five deviations. Additionally, the Morrow predictions for ANN approach
possessed only a single point beyond the boundary while the MLR predictions have six
points lying beyond the boundary. Half of the points were outside the acceptability
margin according to the conservative fatigue life analysis. This was due to the mean
stress effects of the Morrow model. The SWT fatigue life correlations are depicted in
193
Figure 4.44 where three points of ANN predictions were beyond the boundary. The
MLR predicted SWT fatigue life has four points lying beyond the boundary. As
observed from all three strain life approaches, the ANN prediction has more points
distributed within the conservative region and the data points were distributed close to
the 1:2 or 2:1 boundary which implied that the predictions were close to the target value.
The MLR predicted fatigue lives were scattered far beyond.
For further analysis, the relationship between prediction and target fatigue lives
were evaluated using linear regression method as shown in Figures 4.45, 4.46 and 4.47.
In Figure 4.45, the R2 value for ANN prediction and target fatigue life was 0.9998 and
the MLR prediction was 0.9954. The R2 value for Morrow fatigue life using ANN was
0.9995 while the MLR prediction with target fatigue life was 0.9805. The R2 value for
ANN predicted SWT with target fatigue life was 0.9991 while the R2 value for MLR
predicted and target fatigue life was 0.9982. The obtained R2 value suggested that the
ANN and MLR prediction was closely related to the target fatigue life (Sivák &
Ostertagová 2012). Nevertheless, based on the fatigue life 1:2 or 2:1 correlation
analysis, the ANN prediction has revealed better accuracy than the MLR predictions
due to the higher number of data points fitted in the conservative boundary region.
In this section, the results of ANN architecture determination and process to find
the optimised spring fatigue life were obtained. The first novel aspect of this research
was the establishment of the prediction regression as discussed in Section 4.4. In this
section, the fatigue life prediction has been improved using ANN which was also the
second objective of this research. The improvised RMSE, R2 and correlation analysis of
the fatigue life predictions suggested that the ANN approach has successfully optimised
the fatigue life predictions. The improvement of fatigue life predictions using ANN
approach was also the novelty of the second objective. Apart from that, the same
procedures were also further analysed to determine the optimised ISO 2631 weighted
acceleration predictions.
194
(a)
(b)
Figure 4.42 Correlation curve for Coffin-Manson target fatigue life using various approaches:
(a) ANN, (b) MLR
1.0E+03
1.0E+04
1.0E+05
1.0E+06
1.0E+07
1.0E+08
1.0E+09
1.0E+03 1.0E+05 1.0E+07 1.0E+09
AN
N p
red
icte
d f
ati
gu
e li
fe (
blo
cks
to f
ail
ure
)
Target fatigue life (blocks to failure)
1:1 correlation
1:2 or 2:1
correlation
1.0E+03
1.0E+04
1.0E+05
1.0E+06
1.0E+07
1.0E+08
1.0E+09
1.0E+03 1.0E+05 1.0E+07 1.0E+09
ML
R p
red
icte
d f
ati
gu
e li
fe (
blo
cks
to f
ail
ure
)
Target fatigue life (blocks to failure)
1:1
correlation
1:2 or 2:1
correlation
195
(a)
(b)
Figure 4.43 Correlation curve for Morrow target fatigue life using various approaches:
(a) ANN, (b) MLR
1.0E+03
1.0E+04
1.0E+05
1.0E+06
1.0E+07
1.0E+08
1.0E+09
1.0E+03 1.0E+05 1.0E+07 1.0E+09
AN
N p
red
icte
d f
ati
gu
e li
fe (
blo
cks
to f
ail
ure
)
Target fatigue life (blocks to failure)
1:1 correlation
1:2 or 2:1
correlation
1.0E+03
1.0E+04
1.0E+05
1.0E+06
1.0E+07
1.0E+08
1.0E+09
1.0E+03 1.0E+05 1.0E+07 1.0E+09
ML
R p
red
icte
d f
ati
gu
e li
fe (
blo
cks
to f
ail
ure
)
Target fatigue life (blocks to failure)
1:1
correlation
1:2 or 2:1
correlation
196
(a)
(b)
Figure 4.44 Correlation curve for SWT target fatigue life using various approaches:
(a) ANN, (b) MLR
1.0E+03
1.0E+04
1.0E+05
1.0E+06
1.0E+07
1.0E+08
1.0E+09
1.0E+03 1.0E+05 1.0E+07 1.0E+09
AN
N p
red
icte
d f
ati
gu
e li
fe (
blo
cks
to f
ail
ure
)
Target fatigue life (blocks to failure)
1:1 correlation
1:2 or 2:1
correlation
1.0E+03
1.0E+04
1.0E+05
1.0E+06
1.0E+07
1.0E+08
1.0E+09
1.0E+03 1.0E+05 1.0E+07 1.0E+09
ML
R p
red
icte
d f
ati
gu
e li
fe (
blo
cks
to f
ail
ure
)
Target fatigue life (blocks to failure)
1:1 correlation
1:2 or 2:1
correlation
197
(a)
(b)
Figure 4.45 Linear regression analysis for prediction and target Coffin-Manson fatigue life using
various approaches: (a) ANN, (b) MLR
1.00E+02
1.00E+03
1.00E+04
1.00E+05
1.00E+06
1.00E+07
1.00E+08
1.00E+09
1.0E+02 1.0E+03 1.0E+04 1.0E+05 1.0E+06 1.0E+07 1.0E+08 1.0E+09
AN
N p
red
icte
d f
atig
ue
life
(b
lock
s to
fail
ure
)
Target fatigue life (blocks to failure)
1.00E+02
1.00E+03
1.00E+04
1.00E+05
1.00E+06
1.00E+07
1.00E+08
1.00E+02 1.00E+03 1.00E+04 1.00E+05 1.00E+06 1.00E+07 1.00E+08 1.00E+09
ML
R p
red
icte
d f
atig
ue
life
(b
lock
s to
fail
ure
)
Target fatigue life (blocks to failure)
R2 = 0.9998
R2 = 0.9917
198
(a)
(b)
Figure 4.46 Linear regression analysis for prediction and target Morrow fatigue life using various
approaches: (a) ANN, (b) MLR
1.00E+02
1.00E+03
1.00E+04
1.00E+05
1.00E+06
1.00E+07
1.00E+08
1.00E+02 1.00E+03 1.00E+04 1.00E+05 1.00E+06 1.00E+07 1.00E+08
AN
N p
red
icte
d f
atig
ue
life
(b
lock
s to
fail
ure
)
Target fatigue life (blocks to failure)
1.00E+02
1.00E+03
1.00E+04
1.00E+05
1.00E+06
1.00E+07
1.00E+08
1.00E+02 1.00E+03 1.00E+04 1.00E+05 1.00E+06 1.00E+07 1.00E+08
ML
R p
red
icte
d f
atig
ue
life
(b
lock
s to
fail
ure
)
Target fatigue life (blocks to failure)
R2 = 0.9995
R2 = 0.9805
199
(a)
(b)
Figure 4.47 Linear regression analysis for prediction and target SWT fatigue life using various
approaches: (a) ANN, (b) MLR
4.5.2 ISO 2631 Vertical Vibration Prediction
Prediction of fatigue life using ANN has been successfully performed using vehicle
body frequency and ISO 2631 vertical vibration as input. This section presents the
improvement of ISO 2631 vertical vibration prediction for vehicle sprung mass using
trained ANN using vehicle body frequency and spring fatigue life as input. All three
Coffin-Manson, Morrow and SWT predicted spring fatigue lives were used as the input
to train ANN for optimised ISO 2631 vertical vibration prediction.
1.00E+02
1.00E+03
1.00E+04
1.00E+05
1.00E+06
1.00E+07
1.00E+08
1.00E+09
1.00E+02 1.00E+03 1.00E+04 1.00E+05 1.00E+06 1.00E+07 1.00E+08 1.00E+09
AN
N p
red
icte
d f
atig
ue
life
(b
lock
s to
fail
ure
)
Target fatigue life (blocks to failure)
1.00E+02
1.00E+03
1.00E+04
1.00E+05
1.00E+06
1.00E+07
1.00E+08
1.00E+02 1.00E+03 1.00E+04 1.00E+05 1.00E+06 1.00E+07 1.00E+08 1.00E+09
ML
R p
red
icte
d f
atig
ue
life
(b
lock
s to
fail
ure
)
Target fatigue life (blocks to failure)
R2 = 0.9991
R2 = 0.9982
200
To determine a suitable ANN architecture for ISO 2631 vertical vibration
predictions, the MSE for using all three strain life approaches were obtained through
using different numbers of neurons and hidden layers. In this case, ANN with
architecture from one to ten neurons and one to three hidden layers were constructed
and analysed. The results of the lowest MSE across varied number of neurons and
hidden layers were tabulated in Table 4.33. As observed from Table 4.33, the
architecture with three hidden layers has shown the lowest MSE for all three strain life
fatigue approaches. The lowest MSE for Coffin-Manson vibration prediction ANN
predicted ISO 2631 vertical vibration was 0.0007 m/s2 while the lowest MSE for
Morrow ANN predictions was 0.0004 m/s2. For SWT predicted ISO 2631 vertical
vibration the MSE was determined to be 0.0016 m/s2. The determination of suitable
ANN architecture highly depended on the lowest MSE because the accuracy of
predictions was directly influenced by the MSE. Hence, the proposed MSE was the
lowest obtained value which indicated the suitable architecture for this analysis.
Table 4.33 MSE for different trained vibration prediction ANN
Hidden layer Coffin-Manson Morrow SWT
1 0.0885 0.0025 0.0053
2 0.0016 0.0010 0.0031
3 0.0007 0.0004 0.0016
The detailed 3D scatter plot of MSE for all neurons and hidden layers are shown
in Appendix H. The same datasets from Section 4.3.4 were used to train these vibration-
life ANN. For data separation, a total of 72 datasets were used to train the ANN, 14
datasets for validation and 14 data were selected for ANN ‘test’. Curve fitting of the
trained Coffin-Manson ANN-based vibration-life prediction is shown in Figure 4.48
where the r value of 0.9912 for ANN ‘test’ was obtained. When the r value was squared,
R2 value of 0.9825 was obtained. For Morrow vibration prediction ANN, the r value of
ANN ‘test’ was 0.9983 and R2 as high of 0.9966 was obtained. Furthermore, the trained
SWT vibration prediction ANN consisted of r value 0.9787 which was equivalent to R2
value of 0.9579. In Figure 4.48, the obtained R2 value for Coffin-Manson, Morrow and
SWT were considered as very high which indicating the weights and biases of the ANN
were well trained and optimised for the prediction outcome (Sivák & Ostertagová
2012). The ordinate of Figure 4.48 revealed the ANN predicted ISO 2631 with respect
to the target trained value. When the obtained R2 value of ANN ‘test’ was high, the
201
designed ANN performed better (Atici 2011). Hence, the prediction of the ISO 2631
vertical vibration was expected to be greater with high R2 value.
The residuals of the trained ANN were analysed using error histograms as shown
in Figure 4.49. The residuals range of trained Coffin-Manson vibration prediction ANN
was from -0.0494 to 0.0315 m/s2. For Morrow vibration prediction ANN, the error range
was from -0.0385 to 0.0410 m/s2 while the SWT vibration prediction ANN possessed a
range of -0.0929 to 0.0420 m/s2. The original range of the ISO 2631 vertical vibration
was from 0.2900 to 1.3600 m/s2. Based on the obtained residuals, the range was
relatively small. The residuals were also normally distributed. The Lilliefors test
performed to verify the normality of the residuals indicated that the critical value was
greater than the statistic k value, pointing to a normal distribution. This also proved that
the outputs of the trained ANN were well-behaved and suitable to use for ISO 2631
vertical vibration prediction.
In this analysis, a three hidden layer architecture of the Coffin-Manson vibration
prediction ANN has shown the lowest MSE when compared to two hidden layer ANN.
When observing the t-test value for vibration prediction MLR, the t-value of the
parameters were varied when the dependent variables or independent variable were
interchanged. The t-test value expresses the functionality of the independent variable to
dependent variable (Prieto et al. 2017). When the spring fatigue life and ISO 2631
vertical vibration were exchanged, the R2 value and un-standardised coefficients were
also varied. Therefore, three hidden layers Coffin-Manson ANN has shown better
capability in ISO 2631 vertical vibration predictions instead of ANN with two hidden
layers as proposed in vibration-life analysis because the data characteristics were varied.
The selected ANN architectures are listed in Section 3.4.2. All the selected and
trained feedforward network was a hybrid ANN and the weights and biases for this
trained ANN are listed in Appendix K. With the weights and biases, the ANN could be
reproduced and used for optimised ISO 2631 vertical vibration prediction. ISO 2631
vertical vibration is an essential element in assessing vehicle ride (Sezgin & Arslan
2012). Excessive vertical vibration leads to discomfort for drivers and passengers
because human body parts resonate in low frequency range, and the cause of human
202
fatigue. Hence, the ISO 2631 vertical vibration is a main consideration in automotive
suspension design which is relevant to the durability analysis of a vehicle.
(a) (b)
(c)
Figure 4.48 Curve fitting of trained vibration prediction ANN for various approaches:
(a) Coffin-Manson, (b) Morrow, (c) SWT
Utilising the designed ANN architecture as proposed in Section 3.4.2, the
predicted ISO 2631 vertical vibration using Coffin-Manson fatigue lives are shown in
Table 4.34. In addition, the predicted ISO 2631 vertical vibrations for Morrow and SWT
input are shown in Tables 4.35 and 4.36 respectively. In Table 4.34, the comparisons
between MLR and ANN predicted ISO 2631 vertical vibrations were performed. As
observed, the ANN predicted ISO 2631 vertical vibrations were closer in range to the
target value. This was proved by the RMSE between the predictions and target value
using Equation 3.14 where the Coffin-Manson ANN predictions RMSE was 0.0264
Target
AN
N o
utp
ut
AN
N o
utp
ut
AN
N o
utp
ut
Target
Target
203
m/s2. The MLR predicted ISO 2631 vertical vibration was 0.0875 m/s2 which was
higher than the ANN predictions.
(a) (b)
(c)
Figure 4.49 Error histogram of three hidden layers vibration prediction ANN for various
approaches: (a) Coffin-Manson, (b) Morrow, (c) SWT
Table 4.34 MLR and ANN Coffin-Manson predicted ISO 2631 vertical vibration
ANN-based
predictions
(m/s2)
MLR-based
predictions
(m/s2)
Target ISO 2631
vertical
vibration (m/s2)
ANN
difference
(%)
MLR
difference
(%)
0.32 0.25 0.33 3.03 24.24
0.34 0.40 0.35 2.86 14.28
0.42 0.44 0.39 7.69 12.82
0.60 0.55 0.59 1.69 6.77
0.71 0.63 0.68 4.41 7.35
0.65 0.73 0.69 5.79 5.80
0.70 0.62 0.73 4.11 15.06
0.27 0.21 0.27 0.00 22.22
0.38 0.44 0.33 15.15 33.33
0.18 0.15 0.16 12.50 6.25
0.15 0.20 0.16 6.25 25.00
0.32 0.36 0.32 0.00 12.50
0.39 0.42 0.34 14.71 23.52
Errors
Inst
ance
Errors
Inst
ance
Errors
Inst
ance
204
Table 4.35 MLR and ANN Morrow predicted ISO 2631 vertical vibration
ANN-based
predictions
(m/s2)
MLR-based
predictions
(m/s2)
Target ISO 2631
vertical
vibration (m/s2)
ANN
difference
(%)
MLR
difference
(%)
0.58 0.62 0.54 7.41 14.81
0.61 0.75 0.61 0.00 22.95
0.62 0.70 0.61 1.64 14.75
0.65 0.71 0.70 7.14 1.43
0.32 0.17 0.30 6.67 43.33
0.36 0.44 0.35 2.86 25.71
0.38 0.43 0.39 2.56 10.26
0.16 0.12 0.16 0.00 25.00
0.15 0.21 0.17 11.76 23.53
0.34 0.24 0.31 9.68 22.58
0.34 0.36 0.32 6.25 12.50
0.61 0.67 0.62 1.61 8.06
1.30 1.06 1.29 0.78 17.83
Table 4.36 MLR and ANN SWT predicted ISO 2631 vertical vibration
ANN-based
predictions
(m/s2)
MLR-based
predictions
(m/s2)
Target ISO 2631
vertical
vibration (m/s2)
ANN
difference
(%)
MLR
difference
(%)
0.27 0.07 0.29 6.89 41.38
0.57 0.71 0.54 5.56 31.48
0.69 0.64 0.73 5.48 12.33
0.63 0.66 0.57 10.50 15.79
0.68 0.83 0.70 2.86 18.57
0.33 0.32 0.30 10.00 6.67
0.39 0.40 0.36 8.33 11.11
0.17 0.04 0.15 13.33 33.33
0.30 0.25 0.31 3.23 19.35
0.57 0.66 0.62 8.06 6.45
0.61 0.82 0.63 3.17 30.16
0.59 0.81 0.64 7.81 26.56
0.62 0.80 0.65 4.62 23.08
Note: Difference = |Predicted ISO 2631 vertical vibration - target ISO 2631 vertical vibration
target ISO 2631 vertical vibration| × 100% (Equation 3.13)
For Morrow predicted ISO 2631 vertical vibrations, the RMSE for ANN
predictions was 0.0207 m/s2 while the MLR predicted ISO 2631 vertical vibration was
0.1222 m/s2. The difference between RMSE also showed that the ANN predictions were
superior than the MLR predicted ISO 2631 vertical vibration. RMSE using SWT input
were 0.0423 and 0.1279 m/s2 with respect to ANN and MLR approaches where an
obvious improvement of RMSE has been observed. The RMSE is a method to validate
the predictions towards experimental data where a smaller RMSE indicate the
predictions are closer to the actual value (Beltramo et al. 2016). When considering the
difference in predictions, the maximum difference of ANN was 15.15% while the
maximum difference of MLR prediction was 41.38%. However, the acceptable limit for
deviation is only within 20% (Manivel & Gandhinathan 2016). Approximately, half of
205
the MLR predictions were beyond the proposed acceptable limit due to the regression
range error (Ospina & Ferrari 2012). Both the RMSE and difference results proposed
that the ANN have improved the ISO 2631 vertical vibration predictions.
The correlation between ANN prediction and target ISO 2631 vertical vibration
using Coffin-Manson fatigue life input was analysed as shown in Figure 4.50(a) while
the correlation for MLR prediction is shown in Figure 4.50(b). In this analysis, the
obtained R2 value for ANN and MLR predictions were 0.9767 and 0.8596 respectively.
The R2 value for Morrow ANN predictions and target vertical vibrations was 0.9938
while the MLR predictions R2 value was 0.8649 as depicted in Figure 4.51. On the other
hand, in Figure 4.52, the R2 value for ANN predictions using SWT input was 0.9539
and the R2 value MLR predictions was 0.8346. In overall terms, the ISO 2631 vertical
vibration prediction using ANN has shown significant improvement when compared
with the MLR approach according to the R2 value. Nevertheless, the ANN prediction
could be classified as “very good” while the MLR predictions was still considered as
good (Sivák and Ostertagová 2012).
In this subsection, multilayer ANN was successfully used to optimise the ISO
2631 vertical vibrations according to the reduced RMSE and R2 value which indicated
a greater correlation to the target value. This analysis aimed to determine the optimised
ISO 2631 vertical vibrations which was relevant to the second novel aspect of this
research. The results suggested that the second objective for prediction improvement
has been successfully achieved. Nevertheless, although most of the ANN predictions
nicely fitted, the accuracy of the predictions remained controversial because the data
used to generate the MLR and trained ANN were artificially generated. The capability
of the quarter-sized car model to predict actual conditions using strain or acceleration
data was unknown. Hence, there was a need to validate the approaches using actual
vehicle experimental data. Experimental validation using collected strain and
acceleration data were performed to verify the generated MLR and ANN based
regression, leading to the third objective.
206
(a)
(b)
Figure 4.50 Correlation between Coffin-Manson predicted and target vertical vibration using
various approaches: (a) ANN, (b) MLR
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
AN
N p
red
icte
d v
erti
cal
vib
rati
on (
m/s
2)
Target vertical vibration (m/s2)
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
ML
R p
red
icte
d v
erti
cal
vib
rati
on (
m/s
2)
Target vertical vibration (m/s2)
R2 = 0.9767
R2 = 0.8596
207
(a)
(b)
Figure 4.51 Correlation between Morrow predicted and target vertical vibration using various
approaches: (a) ANN, (b) MLR
0
0.2
0.4
0.6
0.8
1
1.2
1.4
0 0.2 0.4 0.6 0.8 1 1.2 1.4
AN
N p
red
icte
d v
erti
cal
vib
rati
on (
m/s
2)
Target vertical vibration (m/s2)
0
0.2
0.4
0.6
0.8
1
1.2
1.4
0 0.2 0.4 0.6 0.8 1 1.2 1.4
ML
R p
red
icte
d v
erti
cal
vib
rati
on (
m/s
2)
Target vertical vibration (m/s2)
R2 = 0.9938
R2 = 0.8649
208
(a)
(b)
Figure 4.52 Correlation between SWT predicted and target vertical vibration using various
approaches: (a) ANN, (b) MLR
4.6 PREDICTION VALIDATION
Prediction of automotive component fatigue life using measured strain time histories is
considered as one of the most acceptable methods due to its acceptable accuracy and
real time monitoring capability (Prakash, Nandi & Sivakumar 2016). Nowadays, many
researches are focusing on developing an algorithm with prediction outcome close to
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
ML
R p
red
icte
d v
erti
cal
vib
rati
on (
m/s
2)
Target vertical vibration (m/s2)
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
ML
R p
red
icte
d v
erti
cal
vib
rati
on (
m/s
2)
Target vertical vibration (m/s2)
R2 = 0.9539
R2 = 0.8346
209
the strain time history predicted fatigue life because strain life fatigue approaches
provided good accuracy (Slavi c et al. 2012; Ferreira et al. 2016). Hence, for this
regression validation, five sets of experimental measured data collected from the actual
vehicles were applied to the MLR and ANN approach. The validation datasets were
divided into two groups, which were strain and acceleration signals. Initially, the
measured strain time histories were used to calculate the spring fatigue life while the
acceleration time histories were used to obtain ISO 2631 vertical vibration. For
vibration-life regression validation, the experimental ISO 2631 vertical vibration and
vehicle body frequency were used as input for MLR fatigue lives prediction where the
outcome were plotted against experimental fatigue lives as shown in Figure 4.53. It is
noteworthy to mention that the proposed MLR-based regressions were obtained from
simulated data but more applicable than the NCM_linear and NCM_power and, hence only the
MLR regressions were validated.
As observed from Figure 4.53(a), the Coffin-Manson vibration-life regression
predicted fatigue life correlated well with experimental fatigue life and an R2 value of
0.9643 was obtained. The RMSE between prediction and experimental results was
0.4857 blocks to failure in natural logarithm. For Morrow vibration-life regression, an
R2 value of 0.9275 was obtained with the RMSE of 0.6401 blocks to failure in natural
logarithm. Meanwhile, SWT vibration-life regression has yielded a better correlation
compared to other generated regression with an R2 value of 0.9806. For SWT regression
prediction, the obtained RMSE value was 0.6240 blocks to failure in natural logarithm.
Although the generated SWT vibration-life regression consisted of higher RMSE, the
regression predictions have correlated well to the experimental strain time history
predicted fatigue life. Nevertheless, all the three vibration-life regressions have shown
a good correlation with high R2 value which was above 0.9000. Therefore, it is
suggested that the fatigue life predictions using vibration-life regression is closely
related to the experimental predicted spring fatigue life which is considered as
acceptable. Thus, these generated regressions are suitable for automotive applications.
210
(a)
(b)
(c)
Figure 4.53 Correlation curve for three vibration-life regressions: (a) Coffin-Manson, (b) Morrow,
(a) SWT
3.6
3.8
4
4.2
4.4
4.6
4.8
5
5.2
3.4 3.6 3.8 4 4.2 4.4 4.6 4.8 5Vib
rati
on
-lif
e r
eg
ress
ion
pre
dic
ted
fati
gu
e li
fe (
blo
cks
to f
ail
ure
)
Experimental strain predicted fatigue life (blocks to failure)
3.6
3.8
4
4.2
4.4
4.6
4.8
5
5.2
5.4
3.4 3.6 3.8 4 4.2 4.4 4.6 4.8
Vib
rati
on
-lif
e r
eg
ress
ion
pre
dic
ted
fa
tig
ue
life
(b
lock
s to
fail
ure
)
Experimental strain predicted fatigue life (blocks to failure)
3.6
3.8
4
4.2
4.4
4.6
4.8
5
5.2
5.4
3.4 3.6 3.8 4 4.2 4.4 4.6 4.8Vib
rati
on
-lif
e r
eg
ress
ion
pre
dic
ted
fati
gu
e li
fe (
blo
cks
to f
ail
ure
)
Experimental strain predicted fatigue life (blocks to failure)
R2 = 0.9642
R2 = 0.9275
R2 = 0.9806
211
The regressions for ISO 2631 vertical vibration predictions were validated using
the same experimental datasets. The results of the Coffin-Manson vibration prediction
regressions were plotted into Table 4.37. As seen from Table 4.37, the highest
difference between MLR predicted and experimental ISO 2631 vertical vibration was
17.86% with a RMSE value of 0.0933 m/s2. The percentage of difference between
Morrow regression-based predicted and experimental ISO 2631 vertical vibration is
illustrated in Table 4.38. The highest difference was 19.31% while the minimum error
was 14.40% with the RMSE of 0.1026 m/s2. Subsequently, the SWT predicted and
experimental ISO 2631 vertical vibration was tabulated in Table 4.39 where the highest
error was 17.52%. with RMSE of 0.0844 m/s2. In this analysis, all the RMSE was
calculated using Equation 3.14 where the difference between the MLR predicted and
experimental ISO 2631 vertical vibration was calculated. A high R2 value and low
RMSE indicated the correlation between prediction and experimental data was good
(Beltramo et al. 2016).
Table 4.37 Difference between experimental and MLR vibration-life prediction for
Coffin-Manson datasets
MLR predicted vibration (m/s2) Experimental vibration (m/s2) Difference (%)
0.63 0.57 10.09
0.78 0.68 15.34
0.66 0.56 17.86
0.65 0.59 10.81
0.85 0.73 16.92
Table 4.38 Difference between experimental and MLR vibration-life prediction for Morrow
datasets
MLR predicted vibration (m/s2) Experimental vibration (m/s2) Difference (%)
0.68 0.57 19.31
0.78 0.68 14.40
0.66 0.56 17.78
0.68 0.59 15.76
0.84 0.73 15.20
Table 4.39 Difference between experimental and MLR vibration-life prediction for SWT
datasets
MLR predicted vibration (m/s2) Experimental vibration (m/s2) Difference (%)
0.67 0.57 17.52
0.76 0.68 11.77
0.63 0.56 13.28
0.67 0.59 12.72
0.82 0.73 12.32
Note: Difference = |MLR predicted vertical vibration - Experimental vibration
Experimental vibration| × 100% (Equation 3.13)
212
The percentage of difference for regression predictions were obtained but this
gave rise to the question of how much difference between the predicted and
experimental value was acceptable in automotive applications. Manivel and
Gandhinathan (2016) proposed that the acceptable reliable statistical analysis of error
for regression analysis was below 20%. Additionally, Zhang et al. (2007) proposed that
the 20% deviation between experimental and simulation results was considered as
acceptable and 10% deviation was considered as good. In addition, the RMSE for all
the vibration prediction regression predictions were relatively low. To analyse the error,
the differences between all three vibration prediction regression predictions were
obtained and plotted into Figure 4.54. As observed, most of the outcomes were below
20%. Within the proposed acceptable range of 20%, the predictions of all generated
vibration prediction regressions were considered as acceptable. This demonstrated that
the vibration prediction regression has predicted the automobile ISO 2631 vertical
vibration with acceptable accuracy. This analysis corresponded to objective three of this
research which was the validation of the generated regression to integrate the all three-
fatigue life, vehicle body frequency and ISO 2631 vertical vibration parameters.
Figure 4.54 Difference between prediction and experiment ISO 2631 vertical vibration for various
approaches
0
5
10
15
20
25
Highway Campus Hill Residential Rural
Dif
fere
nce
(%
)
Coffin-Manson
Morrow
SWT
213
After the validation of the MLR approach, the validation analysis was performed
for the trained ANN. The following validation of trained ANN was performed using
strain-based experimental fatigue life data which were collected from different road
conditions. The correlation plots between ANN predicted and experimental fatigue lives
are shown in Figure 4.55. As observed from Figure 4.55(a), the ANN predicted Coffin-
Manson fatigue lives all fitted into the boundary of 1:2 or 2:1 life correlation graph.
This implied that the trained ANN has provided very good prediction of Coffin-Manson
fatigue life (Kim et al. 2002). For Morrow ANN predicted fatigue life (Figure 4.55(b)),
there were three fatigue life data points located beyond the boundary of life factor of 2.
These three data points were strain time histories from hill, residential and rural area.
Based on the observation, these three points have very low mean value of 4.4, 3.0 and
-4.0 µε when compared to the other two strain measurements. Morrow strain life model
is very sensitive to mean stress (Manson & Halford 2006). Hence, the predictions of
this fatigue life deviated from the experimental results.
In Figure 4.55(c), all the fatigue life predictions fitted into the 1:2 or 2:1 region
of fatigue life correlation graph. SWT vibration-life ANN has provided a better
prediction results than Morrow ANN even though both fatigue strain life model
considered the mean stress effects. Ince & Glinka (2011) has provided research results
through comparing Morrow and SWT strain life models where the results have shown
that the SWT model was superior in predicting fatigue life than Morrow model for steel
components. This was the reason that SWT vibration-life ANN provided nicely fitted
and accurate fatigue life predictions inside the fatigue life scatter band. When compared
to the MLR prediction, vibration-life ANN have shown a good improvement in fatigue
life prediction where the MSE dropped from 0.7056 to 0.1769. The reduction of MSE
suggested that the fatigue life prediction was improved. Although vibration-life ANN
was more accurate than MLR vibration-life method, the structure of ANN was much
more complex than MLR. Implementation of MLR regression was easier and the results
could be obtained through simple calculations.
214
(a)
(b)
(c)
Figure 4.55 Correlation curve of experimental and vibration-life ANN fatigue life predictions
using various approaches: (a) Coffin-Manson, (b) Morrow, (c) SWT
1.0E+03
1.0E+04
1.0E+05
1.0E+06
1.0E+03 1.0E+04 1.0E+05 1.0E+06
AN
N p
red
icu
ted
fa
tig
ue
life
(b
lock
s to
fail
ure
)
Experimental fatigue life (blocks to failure)
1:1 correlation
1:2 or 2:1
correlation
1.0E+03
1.0E+04
1.0E+05
1.0E+06
1.0E+03 1.0E+04 1.0E+05 1.0E+06
AN
N p
red
icu
ted
fa
tig
ue
life
(b
lock
s to
fail
ure
)
Experimental fatigue life (blocks to failure)
1:1 correlation
1:2 or 2:1
correlation
1.0E+03
1.0E+04
1.0E+05
1.0E+06
1.0E+03 1.0E+04 1.0E+05 1.0E+06
AN
N p
red
icu
ted
fa
tig
ue
life
(b
lock
s
to f
ail
ure
)
Experimental fatigue life (blocks to failure)
1:1 correlation
1:2 or 2:1
correlation
215
For vibration-life ANN training results, the validation was performed using the
same experimental data. The results of validation for ISO 2631 vertical vibration
prediction using Coffin-Manson fatigue life as input was tabulated in Table 4.40. For
Coffin-Manson ANN predicted ISO 2631 vertical vibration, the highest difference
between prediction and experimental ISO 2631 vertical vibration results was 10.96%.
For Morrow vibration prediction ANN, the maximum difference was also 10.96% while
the SWT ANN for ISO 2631 vertical vibration was 11.69%. The results for Morrow
and SWT life vibration ANN are illustrated in Tables 4.41 and 4.42 respectively while
the percentage difference was plotted into a chart as shown in Figure 4.56. The
difference between ANN-based prediction and experimental ISO 2631 vertical
vibration was below 15% while the MLR approaches were below 20%. Based on
Manivel and Gandhinathan (2016), the prediction of vibration for experimental and
simulation within 20% were acceptable and below 10% was considered as good. Hence,
the ANN prediction was considered as good and the prediction results were also
improved from MLR approaches.
When compared to MLR, the percentage difference of trained ANN was lower
because the ANN has deeper learning with Levenberg-Marquadt algorithm.
Nevertheless, these functions also made the trained vibration-life ANN more complex
to handle than MLR method. ANN method has shown superior capability in
development and widely applied in recent automotive researches. For example, Patel
and Bhatt (2016) applied the ANN method for chassis identical stress predictions. ANN
was also introduced to predict the spine acceleration from road excitation signal of an
automobile (Gohari et al. 2014). The vehicle suspension system was then optimised
using metaheuristic algorithms. In this research, the proposed MLR relationship and
trained ANN were also applicable to automotive suspension system design. MLR
provided a less accurate solution but was easier to use. The trained ANN provided more
accurate predictions but knowledge of ANN was required when it was applied.
Table 4.40 Difference between experimental and ANN vibration-life prediction for
Coffin-Manson approach
ANN predicted vibration (m/s2) Experimental vibration (m/s2) Difference (%)
0.61 0.57 6.20
0.62 0.68 8.70
0.61 0.56 8.69
0.61 0.59 3.07
0.65 0.73 10.96
216
Table 4.41 Difference between experimental and ANN vibration-life prediction for Morrow
approach
ANN predicted vibration (m/s2) Experimental vibration (m/s2) Difference (%)
0.63 0.57 10.53
0.63 0.68 7.35
0.62 0.56 10.71
0.63 0.59 6.78
0.65 0.73 10.96
Table 4.42 Difference between experimental and ANN vibration-life prediction for SWT
approach
ANN predicted vibration (m/s2) Experimental vibration (m/s2) Difference (%)
0.63 0.57 11.35
0.60 0.68 11.69
0.62 0.56 10.11
0.63 0.59 7.36
0.65 0.73 10.96
Note: Difference = |ANN Predicted vertical vibration - Experimental vibration
Experimental vibration| × 100% (Equation 3.13)
Figure 4.56 Difference between prediction and experiment ISO 2631 vertical vibration for various
strain life approaches
In this section, validation of both MLR and ANN vibration-life approaches for
spring fatigue life and ISO 2631 vertical vibration predictions were performed. The
MLR regression predicted fatigue life has correlated to the experimental strain predicted
fatigue life while the ANN predicted fatigue life has correlated well with the
experimental fatigue life under a conservative fatigue scatter band approach. On the
other hand, the vibration prediction ANN has shown a significant improvement of ISO
2631 vertical vibration with percentage of difference below 11.69%. All this validation
processes suggested that the third objective has been successfully achieved where the
0
5
10
15
20
Highway Campus Hill Residential Rural
Dif
fere
nce
(%
)
Coffin-Manson
Morrow
SWT
217
proposed approaches were acceptable for spring fatigue life or ISO 2631 vertical
vibration predictions.
4.7 SUMMARY
This chapter has presented all the analysis results and discussions according to the
sequence of three objectives which led to the achievement of research novelty. There
were two main findings that contributed to the research novelty where the first was
regression-based prediction approaches for spring fatigue life, vehicle body frequency
and ISO 2631 vertical vibration. The integrated fatigue life regressions were analysed
using R2, normal P-P, F-test, t-test to show normality. The second novelty was
described by the second objective which was predictions of the fatigue life or ISO 2631
vertical vibration using ANN. The ANN approach with Levenberg-Marquadt algorithm
in training where the ANN method has the capability of adjusting weighting and
improvised the prediction results. The designed ANN were analysed using R2, error
histogram and the Lilliefors test to ensure that it had satisfactory performance which
served as the second novelty of this research.
Based on the validation results, the vibration prediction ANN have shown
significant improvement when compared with regression-based approaches according
to experimental data. The vibration prediction ANN have also shown a lower percentage
of MSE. Nevertheless, both regression and ANN-based vibration predictions were
acceptable due to the low percentage of differences. Overall, both proposed prediction
approaches contribute to the automotive industry through reducing spring or suspension
design process in field of fatigue and automobile ride. In the automotive industry,
suspension design process involved many steps including the spring fatigue life and
automobile ride analysis associated to the spring stiffness. In this thesis, the solutions
to assist these tedious processes have been proposed which are known as the vibration-
life approaches. This proposed MLR and ANN vibration-life approaches for optimised
spring fatigue life predictions have led to the novelty of this research.
218
CHAPTER V
CONCLUSION AND RECOMMENDATIONS
5.1 CONCLUSION
This thesis focuses on determining and establishing fatigue life prediction
relationships using vehicle body frequency and automotive ride as inputs. Multiple
linear regression relationships and trained ANNs were proposed for modelling these
parameters to simplify the automotive suspension design process through predictions.
This study was divided into the following three main stages:
1. Determination of multiple input regression-based fatigue relationships that
predict vehicle ISO 2631 weighted vibrations using quarter car model
simulations with different spring stiffness variants;
2. Establishment of fatigue life predictions using a multilayer perceptron ANN by
determining the architecture with the lowest mean squared error; and
3. Validation of the regression and ANN predictions using various experimental
data.
The execution of these stages produced the optimised durability relationships for
predicting the spring fatigue life or ISO 2631 vertical vibration with good and
acceptable accuracy, as well as repeatability.
5.1.1 Determination of Durability Relationships
The first objective of the current study is to determine a relationship that predicts the
spring fatigue life or vertical vibration with consideration of suspension parameter.
219
The characterisation of road measurement data successfully verified the feasibility of
the regression method. With the combination of the generated artificial road profile
and the acceleration inputs to the quarter car model simulation, spring force and
vehicle spring mass acceleration responses were obtained and processed into fatigue
life and ISO 2631 weighted acceleration, respectively. For the nominal spring design,
the simulated fatigue life was 6.23 × 106 blocks to failure while the weighted
acceleration was 0.3390 m/s2. Then, MLR approaches were used to model these
parameters with R2 value of 0.8320 and an MSE of 0.7056. Based on these
parameters, the proposed durability relationship was suitable for fatigue life and ISO
2631 vertical vibration predictions.
5.1.2 Establishment of Fatigue Life Prediction
The second objective is to establish the fatigue life prediction relationship using ANN.
ANNs performed better than the MLR predictions for vibration–life analyses. For the
Morrow vibration–life ANN, the R2 and MSE obtained were 0.9944 and 0.0117,
respectively. For the SWT vibration–life ANN, R2 and MSE values of 0.9926 and
0.0824 m/s2 were respectively obtained. For the life–vibration Morrow ANN, the
trained ANN possessed an R2 value of 0.9966 with a low MSE of 0.0004 m/s2. The
SWT fatigue–vibration ANN obtained an R2 value of 0.9580 with an MSE value of
0.0016 m/s2. The ANN exhibited greater improvement in terms of MSE compared
with the MLR method. From these findings, the life estimations can be established
using ANN through the analysis of vibration and strain data.
5.1.3 Validation of MLR- and ANN-Predicted Fatigue Life
The third objective of this study is to validate the regression and ANN predictions
using various experimental data. To achieve this goal, the predicted life of the
vibration–life MLR approach was correlated to the experimental strain results. The
correlations showed high R2 values of 0.9275 and 0.9806 for the Morrow and SWT
predictions. The fatigue life was considered following the trend of experimental
predictions. Meanwhile, the maximum difference of the predicted ISO 2631 vertical
vibrations using MLR was 19.31%. For vibration–life ANN, the fatigue lives were
220
correlated to the experimental data using a 1:2 or 2:1 correlation curve. Most of the
data were fitted within the boundary, and the outcome of the predictions was
considered acceptable. For fatigue life to ISO 2631 vertical vibration ANN, the
maximum deviation of the prediction was 11.69% which indicating that it is
acceptable. Finally, it suggested that the ANN predictions could be suitable for the
cases of ride-related vibration and life prediction in automotive suspension
applications.
5.2 RESEARCH CONTRIBUTIONS
The automotive suspension design process requires considerable and lengthy efforts.
Thus, this study established prediction relationships to provide a quick solution to
reduce the time and effort required. To the best of the author’s knowledge, the
established prediction relationship has yet to be found in previous works. With the
vehicle body frequency and ISO 2631 vertical vibration as inputs, spring fatigue lives
were predicted with acceptable accuracy. Conversely, the ISO 2631 vertical vibrations
were estimated using vehicle body frequency and spring fatigue life as inputs. This
study involved three academic disciplines of mechanical engineering, namely,
automotive suspension design, structure durability and vehicle ride analysis. The gap
of these three disciplines could be bridged using an advanced supervised learning
method. Considerable attention has been devoted to the improvement of predictions
using ANN architectures in the field of machine learning in terms of mechanical
engineering applications. The outcomes of this study could offer solutions to
automotive industries for addressing issues in prototyping costs and time. In addition,
academicians from the field of fatigue could better understand vehicle vibration to
analyse spring fatigue life.
221
5.3 RECOMMENDATIONS
There are some future works that can be furthered to improve this current work, which
are:
1. In real applications, the springs are subjected to multiaxial loadings because
they possess round and helical geometry where the stress states are complex.
In future, the fatigue life of the spring could be assessed using a multiaxial
fatigue model to obtain fatigue life. The proposed analysis was based on ISO
2631 vertical vibration as a ride indicator. This indicator could be expanded
into three axes where the lateral and longitudinal vibrations are also
considered.
2. This current strain life approaches do not take load sequence effects into
consideration. When the loading time histories are random and stochastic, the
fatigue life could deviate from the actual conditions. Effective strain damage
models could be used to enhance the predictions.
3. The quarter car simulation model was applied in this analysis to extract the
vehicle and spring responses. A half car or full car model could provide greater
accuracy of simulation because of the consideration of time-delayed input.
4. The developed prediction relationship could be further expanded with
additional parameters, such as variable damping coefficient of dampers, hard
points of the lower arm, tyre stiffness to perform a more feasible suspension
design. Additional variables could be added to train the ANN for a more
advanced application.
5. Recurrent neural network is a type of neural network where the output is
feedback to the input for weights and biases adjustment. This kind of neural
network is more complicated but provides greater accuracy. This type of
neural network is suggested for further analyses.
222
REFERENCES
Abdullah, S. 2005. Wavelet Bump Extraction (WBE) for Editing Variable Amplitude Fatigue
Loadings. Ph.D. Thesis. The University of Sheffield.
Abidin, M.I.Z., Mahmud, J., Abd Latif, M.J., Jumahat, S. 2013. Experimental and numerical
investigation of SUP12 steel coil spring. Procedia Engineering 68: 251 – 257.
Aboud, W.S., Haris, S.M., Yaacob, Y. 2014. Advances in the control of mechatronic
suspension system. Journal of Zhejiang University-Science C 15(10): 848 – 860.
Abulizi, N., Kawamura, A., Tomiyama, K., Fujita, S. 2016. Measuring and evaluating road
roughness conditions with a compact road profiler and ArcGis. Journal of Traffic and
Transportation Engineering 3(5): 398 – 411.
Abyaneh, Z. 2014. Evaluation of multivariate linear regression and artificial neural networks
in prediction of water quality parameters. Journal of Environmental Health Science and
Engineering 12: 40.
Agostinacchio, M., Chiampa, D., Olita, S. 2014. The vibrations induced by surface
irregularities in road pavements – A matlab® approach. European Transport Research
Review 6(3), 267 – 275.
Ahlin, K., Granlund, J., Lundström, R. 2000. Whole-body vibration when riding on rough
roads – a shocking study, Swedish National Road Administration, Vol. 31E, 2008-02-
24.
Ahmed, A.E.S., Ali, A.S., Ghazaly, N.M., Jaber, G.T.A. 2015. PID controller of active
suspension system for a quarter car model. International Journal of Advances in
Engineering and Technology 8(6), 899 – 909.
Akiniwa, Y., Stanzl-Tschegg, S., Mayer, H., Wakita, M., Tanaka, K. 2008. Fatigue strength of
spring steel under axial and torsional loading in the very high cycle regime.
International Journal of Fatigue 30: 2057 – 2063.
Al-assadi, M., Kadi, H.E., Deiab, I.M. 2010. Predicting the fatigue life of different composite
materials using artificial neural network. Applied Composite Materials 17: 1 – 14.
Amin Yahari, S., Wauthle, R., van der Stok, J., Riemslag, A.C., Janssen, M., Mulier, M., Kruth,
J.P., Schrooten, J., Weinans, H., Zadpoor, A.A. 2013. Fatigue behaviour of porous
biomaterials manufactured using selective laser melting. Materials Science and
Engineering C 33: 4849 – 4858.
Amsterdam, E. & Grooteman, F. 2016. The influence of stress rate on the exponent in the
power law equation of fatigue crack growth. International Journal of Fatigue 82: 572
– 578.
Anes, V., Caxias, J., Freitas, M., Reis, L. 2017. Fatigue damage assessment under random and
variable amplitude multiaxial loading conditions in structural steels. International
Journal of Fatigue 100: 591 – 601.
223
Anes, V., Reis, L., Li, B., Freitas, M. 2014. New cycle counting method for multiaxial fatigue.
International Journal of Fatigue 67: 78 – 94.
Arjun, K.S. & Peter, P.T. 2014. Design of hydraulic spring loaded rail clamp of ship unloader.
International Journal of Engineering Research and Science and Technology 3(2): 324
– 332.
Asadi, S., Amiri, S.S., Mottahedi, M. 2014. On the development of multi-linear regression
analysis to access energy consumption in the early stages of building design. Energy
and Buildings 85: 246 – 255.
ASTM E1049-85. 2005. Standard Practice for Cycle Counting in Fatigue Analysis. West
Conshohocken: ASTM International.
ASTM E112-96. 2004. Standard Test Methods for Determining Average Grain Size. West
Conshohocken: ASTM International.
Atici, U. 2011. Prediction of the strength of mineral admixture concrete using multivariable
regression analysis and an artificial neural network. Expert Systems with Applications
38: 9609 – 9618.
Aygül, M., Al-Emrani, M. & Urushadze, S. 2012. Modelling and Fatigue Life Assessment of
Orthotropic Bridge Deck Details using FEM. International Journal of Fatigue 40: 129
– 142.
Ayinde, K., Kuranga, K.O., Solomon, G.S. 2016. Empirical investigation of type 1 error rate
of univariate tests of normality. International Journal of Computer Applications
148(8): 24 – 31.
Azadi, S. & Karimi-Jashni, A. 2015. Verifying the performance of artificial neural network
and multiple linear regression in predicting the mean seasonal municipal solid waste
generation rate: a case study of Fars province, Iran. Waste Management 18: 40 – 23.
Bae, D.H., Sohn, I.S., Jung, W.S., Kim, N.S., Jung, W.W., Park, S.C. 2005. Effect of residual
stress by shot peening on fatigue strength of LCV leaf spring. International Journal of
Automotive Technology 6(6): 671 – 676.
Balike, K.P., Rakheja, S. & Stiharu, I. 2013. Kineto-dynamic performance analysis of vehicle
with an asymmetric suspension damper using a roll-plane model. International Journal
of Vehicle Perfomance 1(1): 1 – 23.
Balmos, A.D., Layton, A.D., Aaron, A., Krogmeier, J.V., Buckmaster, D.R. 2014. Toward
understanding the errors in online air-ride suspension based weight estimation, 2014
ASABE and CSBE/SCG AB Annual International Meeting, Canada, 141897468.
Barbosa, R.S. 2011. Vehicle dynamic response due to pavement roughness. Journal of
Brazillian Society of Mechanical Sciences and Engineering 33(3): 302 – 307.
224
Barulich, N.D, Godoy, L.A. & Dardati, P.M. 2017. A simple procedure to evaluate Cauchy
stress tensor at the macro level based on computational micromechanics under general
finite strain states. Mechanics of Materials 117: 73 – 80.
Baykal, H. & Yildirim, H.K. 2013. Application of artificial neural networks (ANN) in wine
technology. Critical Reviews in Food Science and Nutrition 53(5): 415 – 421.
Beltramo, T., Ranzan, C., Hinrichs, J., Hitzmann, B. 2016. Artificial neural network
prediction of the biogas flow rate optimised with an ant colony algorithm.
Biosystems Engineering 143: 68 – 78.
Beretta, S. & Regazzi, D. 2016. Probabilistic fatigue assessment for railway axles and
derivation of a simple format for damage calculations. International Journal of
Fatigue 86: 13 – 23.
Bhanage, A & Padmanabhan, K. 2015. Static and fatigue simulation of automotive anti roll bar
before DBTT. International Journal of Applied Engineering Research 10(71): 472 –
476.
Bižal, A., Klemenc, J. & Fajdiga, M. 2015. Evaluating the statistical significance of a fatigue
life reduction due to macro porosity. Strojniški vestnik - Journal of Mechanical
Engineering 60: 407 – 416.
Blundell, M. & Harty, D. 2004. Vehicle Dynamics Theory and Application. U.S.A.: Elsevier
limited.
Bogdan, M., Wilamowski, J., Irwin, D. 2011. The Industrial Electronics Handbook, Industrial
Communication Systems. U.S.A.: CRC Press.
Borah, T.R., Sarma, K.K. & Talukdar, P. 2015. Handbook of Research on Emerging
Perspectives in Intelligent Pattern Recognition, Analysis, and Image Processing,
Edition: 1st. USA: IGI Global.
Bouzara, M., Richard, M.J., Rakheja, S. 2006. Safety and comfort analysis of a 3-D vehicle
model with optimal non-linear active seat suspension. Journal of Terramechanics 43:
97 – 118.
Bozorg-Haddad, O., Zarezadeh-Mehrizi, M., Abdi-Dehkordi, M., Loaiciga, H.A., Marino,
M.A. 2016. A self-tuning ANN model for simulation and forecasting of surface flows.
Water Resources Management 30: 2907 – 2929.
Brabazon, A., O’Neill, M., McGarraghy, S. 2015. Natural Computing Algorithm. U.K:
Springer-Verlag Berlin Heidelberg.
Bruchhausen, M., Fischer, B., Ruiz, A., González, S., Hähner, P., Soller, S. 2015. Impact of
hydrogen on the high cycle fatigue behaviour of Inconel 718 in asymmetric push-pull
mode at room temperature. International Journal of Fatigue 70: 137 – 145.
Burdzik, R. 2014. Identification of structure and directional distribution of vibration transferred
to car body from road roughness. Journal of Vibroengineering 16(1): 324 – 333.
225
Burdzik, R. & Dolecek, R. 2012. Research of vibration distribution in vehicle constructive.
Perner’s Contacts 14(2): 16 – 25.
Burdzik, R. & Konieczny, L. 2013. Research on structure, propagation and exposure to general
vibration in passenger car for different damping parameters. Journal of
Vibroengineering 15(4): 1680 – 1688.
Cantisani, G. & Loprencipe, G. 2010. Road roughness and whole body vibration: evaluation
tools and comfort limits. Journal of Transportation Engineering 136(9): 818 – 826.
Cao, D., Rakheja, S., Su, C.Y. 2008. Heavy vehicle pitch dynamics and suspension tuning. Part
1: unconnected suspension. Vehicle System Dynamics 46(10): 931 – 953.
Capponi, L., Česnik, M., Slavič, J., Cianetti, F., Boltežar, M. 2017. Non-stationarity index in
vibration fatigue: theoretical and experimental research. Fatigue and Fracture of
Engineering Materials and Structures 34: 868 – 876.
Cárdenas, J.J., Romeral, L., Garcia, A., Andrade, F. 2012. Load forecasting framework of
electricity consumptions for an intelligent energy management system in the user-side.
Expert Systems with Applications 39: 5557 – 5565.
Carpinteri, A. & Paggi, M. 2009. A unified interpretation of the power laws in fatigue and the
analytical correlations between cyclic properties of engineering materials. International
Journal of Fatigue 31: 1524 – 1531.
Carrillo, C., Obando Montano, A.F., Cidras, J., Diaz-Dorado, J. 2013. Review of power curve
modelling for wind turbines. Renewable and Sustainable Energy Reviews 21: 572 –
581.
Casson, R.J. & Farmer, L.D. 2014. Understanding and checking the assumptions of linear
regression: a primer for medical researchers. Clinical and Experimental Ophthalmology
42(6): 590 – 596.
Castro, D.B.V., Ventura, J.M., Terra Ruckert, C.O.F., Spinelli, D., Bose Filho, W.W. 2010.
Influence of Phosporus content and quenching/tempering temperatures on fracture
toughness and fatigue life of SAE 5160 steel. Materials Research 13(4): 445 – 455.
Chaari, F., Abbes, M.S., Rueda, F.V., Rincon, A.F., Haddar, M. 2013. Analysis of planetary
gear transmission in non-stationary operations. Frontiers of Mechanical Engineering
8(1): 88 – 94.
Chandwani, V., Agrawal, V., Nagar, R. 2015. Modeling slump of ready mix concrete using
genetic algorithms assisted training of artificial neural networks. Expert Systems with
Applications 42: 885 – 893.
Chen, H.C., Chen, W.C., Liu, Y.P., Chen, C.Y., Pan, Y.T. 2009. Whole-body vibration
exposure experienced by motorcycle riders – An evaluation according to ISO 2631-1
and ISO 2631-5 standards. International Journal of Industrial Ergonomics 39: 708 –
718.
226
Chen, L., Liu, C., Liu, W., Nie, J., Shen, Y., Chen, G. 2017. Network synthesis and parameter
optimization for vehicle suspension with inerter. Advances in Mechanical Engineering
9(1): 1 – 7.
Chen, X. 2014. Analysis of Crosswind Fatigue of Wind-Excited Structures with Nonlinear
Aerodynamic Damping. Engineering Structures 74: 145 – 156.
Chetan, J., Khushbu, P., Nauman, M. 2012. The fatigue analysis of a vehicle suspension system
– a review article. International Journal of Advanced Computer Research 2(6): 386 –
390.
Chindamo, D., Gadola, M. & Marchesin, F.P. 2017. Reproduction of real-world road profiles
on a four poster rig for indoor vehicle chassis and suspension durability testing.
International Journal of Vehicle Design 73(1/2/3): 54 – 62.
Chiou, Y. & Yang, J.K. 2012. The effects of Pre-deformation on the subsequent fatigue
behaviours of SUS 430 strainless steel in load-control. International Journal of Solids
and Structures 49L: 3263 – 3268.
Chiu, M.C., Chang, Y.C., Yeh, L.J., Chung, C.H. 2016. Numerical assessment of a one-mass
spring based electromagnetic energy harvester on a vibrating object. Archives of
Acoustics 41(1): 119 – 131.
Choube, A.M. 2016. Finite element analysis of helical suspension in mono suspension system.
International Journal of Engineering Science and Computing 6(5): 4947 – 4950.
Coppola, T., Cortese, L. & Folgarait, P. 2009. The effect of stress invariants on ductile fracture
limit in steels. Engineering Facture Mechanics 76: 1288 – 1302.
Cossalter, V., Doria, A., Pegoraro, R., Trombetta, L. 2010. On the non-linear behaviour of
motorcycle shock absorbers. Proceeding of the Institute of Mechanical Engineers, Part
D: Journal of Automobile Engineering 224(1): 15 – 27.
Cossalter, V., Roberto, L., Rota, S. 2010. Objective and subjective evaluation of an advanced
motorcycle riding simulator. European Transport Research Review 2(4): 223 – 233.
Coutinho, I., Landre, J. & Magalhaes, P.A. 2016. Rear suspension damper support virtual
durability and load reconstruction. Proceedings of the XXXVII Iberian Latin-American
Congress on Computational Methods in Engineering, Brazil.
Cui, W. 2002. A State-of-the-Art Review on Fatigue Life Prediction Methods for Metal
Structures. Journal of Marine Science and Technology 7: 43 – 46.
Cui, Y., Rollin, S.B., Germano, G. 2017. Full and fast calibration of the Heston stochastic
volatility model. European Journal of Operational Research 263: 625 – 638.
Dabell, B. 1985. Development of automotive components through integrated
design/manufacture. International Journal of Vehicle Design 6(4/5): 514 – 518.
227
Dang, J., Chen, H., Gao, B., Li, Q., Li, M., Watanabe, T., Hayama, R., Lou, L., Nakano, S.
2014. Optimal design of on-center steering force characteristic based on correlations
between subjective and objective evaluations. SAE Technical Paper 2014-01-0137.
Darban, H., Nosrati, M. & Djavanroodi, M. 2014. Multiaxial fatigue analysis of stranded-wire
helical springs. International Journal of Damage Mechanics 0(0): 1 – 13.
Darlington, R.B. & Hayes, A.F. 2016. Regression Analysis and Linear Models: Concepts,
Applications and Implementation. U.K.: The Guildford Press.
de Jesus, A.M.P. & da Silva, A.L.L. 2010. Identification of the net effective strain life Model
for a puddle iron from the Portuguese Fão riveted road bridge. Procedia Engineering
2: 1181 – 1190.
Del Llano-Vizcaya, L., Rubio-Gonzalez, C., Mesmaque, G. 2006. Multiaxial fatigue and
failure analysis of helical compression springs. Engineering Failure Analysis 13: 1303
– 1313.
Delijaicov, A.T., Fleury, A.T., Martins, F.P.R. 2010. Application of multiple regression and
neural networks to synthesize a model for peen forming process planning. Journal of
Achievements of Materials and Manufacturing Engineering 43: 651 – 656.
Dimitris, S.P. 2016. Country road and field surface profile acquisition, modelling and synthetic
realisation for evaluating fatigue life of agricultural machinery. Journal of
Terramechanics 63: 1 – 12.
Dixon, J.C. 2007. The Shock Absorber Handbook. U.S.A.: John Wiley & Sons, Inc.
Dodge, Y. 2010. The Concise Encyclopedia of Statistics. New York: Springer-Verlag.
Dor, F. & Vielhaber, M. 2014. Simulation-based concept generation for mechatronic systems.
International Design Conference – Design 2014: Croatia.
Dore, M.J. & Maddox, S.J. 2013. Accelerated fatigue crack growth in 6082 T651 Aluminium
alloy subjected to periodic underloads. Procedia Engineering 66: 313 – 322.
Dorofki, M., Elshafir, A.H., Jaafar, O., Karim, O.A., Mastura, F. 2012. Comparison of artificial
neural network transfer function abilities to simulate extreme runoff data. 2012
International Conference on Environment, Energy and Biotechnology 33: 39 – 44.
dos Santos, H.A.F.A., Auricchio, F. & Conti, M. 2012. Fatigue life assessment of
cardiovascular balloon-expandable stents: a two-scale plasticity-damage model
approach. Journal of the Mechanical Behavior of Biomedical Materials 15: 78 – 92.
Dowling, N. E. 2004. Mean stress effects in stress-life and strain-life fatigue. SAE Technical
Paper 2004-01-2227.
Dowling, N.E. 1999. Mechanical Behaviour of Materials: Engineering Methods for
Deformation, Fracture and Fatigue 2nd ed. NJ: Prentice hall.
228
Du, F. 2016. Frequency weighting filter design for automotive ride comfort evaluation. Chinese
Journal of Mechanical Engineering 29(4): 727 – 738.
Du, Y., Liu, C., Wu, D. 2016. Application of vehicle mounted accelerometers to measure
pavement roughness. International Journal of Distributed Sensor Networks 8413146.
Du, Y., Liu, C., Wu, D., Jiang, C. 2014. Measurement of road roughness index by using Z-axis
accelerometers and GPS. Mathematical Problems in Engineering 928980.
Eger, T., Stevenson, J., Boileau, P.E., Salmoni, A., Vibrg. 2008. Predictions of health risk
associated with the operation of load-haul-dump mining vehicles: Part 1- analysis of
whole body vibration exposure using ISO 2631-1 and ISO 2631-5 standards.
International Journal of Industrial Ergonomics 38: 726 – 738.
Elbayoumi, M., Md Yusof, N.F., Ramli, N.A. 2015. Development and comparison of
regression models and feedforward backpropagation neural network models to predict
seasonal indoor PM2.5_10 and PM2.5 concentrations in naturally ventilated schools.
Atmospheric Pollution Research 6: 1013 – 1023.
Els, P.S. 2005. The applicability of ride comfort standards to off-road vehicles. Journal of
Terramechanics 42: 47 – 64.
Els, P.S., Theron, N.J., Uys, P.E., Thoresson, M.J. 2007. The ride comfort vs. handling
compromise for off-road vehicles. Journal of Terramechanics 44(4): 303 – 317.
Ertas, A.H. & Sonmez, F.O. 2008. A parametric study on fatigue strength of spot-welded joints.
Fatigue and Fracture of Engineering Materials and Structures 31(9): 766 – 776.
Evin, E., Tomas, M., Kmec, J., Nemeth, S., Katalinic, B., Wessely, E. 2014. The deformation
properties of high strength steel sheets for auto-body components. Procedia
Engineering 69: 758 – 767.
Facchinetti, M.L. 2018. Fatigue damage of materials and structures assessed by Wöhler and
Gassner frameworks: recent insights about load spectra for the automotive. Procedia
Engineering 213: 117 – 125.
Fajdiga, G. & Sraml, M. 2009. Fatigue crack initiation and propagation under cyclic contact
loading. Engineering Fracture Mechanics 76: 1320 – 1335.
Fang, J., Gao, Y., Sun, G., Xu, C., Li, Q. 2015. Multiobjective robust design optimization of
fatigue life for a truck cab. Reliability Engineering and System Safety 135: 1 – 8.
Ferreira, W.G., Meehan, T.S., Bishop, N., Cardoso, V.N. 2016. A comparative study of
automotive system fatigue models processed in the time and frequency domain. SAE
Technical Paper 2016-01-0377.
Fernández-Canteli, A., Blasón, S. 2014. A probabilistic interpretation of the Miner number for
fatigue life prediction. Frattura ed Integrità Strutturale 30: 327 – 339.
229
Feng, J., Zhang, X., Guo, K., Ma, F., Karimi, H.R. 2013. A frequency compensation algorithm
of four-wheel coherence random road. Mathematical Problems in Engineering 986584.
Fielder, M. & Vormwald, M. 2016. Considering fatigue load sequence effects by applying the
local strain approach and a fracture mechanics based damage parameter. Theoretical
and Applied Fracture Mechanics 83: 31 – 41.
Figueira Pujol, J.C. & Andrade Pinto, J.M. 2011. A neural network approach to fatigue life
prediction. International Journal of Fatigue 33: 313 – 322.
Fragoudakis, R., Karditsas, S., Savaidis, G., Michailidis, N. 2014. The effect of heat and
surface treatment on the fatigue behaviour of 56SiCr7 Spring steel. Procedia
Engineering 73: 309 – 312.
Francisco, E.R., Aranha, F., Zambaldi, F., Goldszmidt, R. 2006. Electricity consumption as a
predictor of household income: an spatial statistic approach. VIII Brazilian Symposium
on GeoInformatics, November, Brazil, 175 – 192.
Fu, T.T. & Cebon, D. 2000. Predicting fatigue lives for bi-modal stress spectral densities.
International Journal of Fatigue 22: 11 – 21.
Fu, W., Ling, F., Peng, M., Zhang, Y. 2013. Comparison of wheelbase filtering effect and
suspension tuning between two axle and tri-axle vehicle with tandem suspension. The
2nd International Conference Computer Application and System Modeling: 1124 – 1127.
Gandomi, A.H. 2015. Structural optimization using interior search algorithm. Iranian Journal
of Structural Engineering 2(1): 19 – 27.
Ganjian, E., Khorami, M. & Maghsoudi, A.A. 2009. Scrap-tyre-rubber replacement for
aggregate and filler in concrete. Construction and Building Materials 23(5): 1828 –
1836.
Gao, H., Huang, H.Z., Zhu, S.P., Li, Y.F., Yuan, R. 2014. A modified nonlinear damage
accumulation model for fatigue life considering load interaction effects. The Scientific
World Journal 164378.
Gao, Y. 2017. Fatigue stress concentration sensitivity and stress ratio effect of a
40CrNi2Si2MoVA steel. Materials Letters 186: 235 – 238.
Gazi, O. 2017. Understanding Digital Signal Processing. Singapore: Springer Nature.
Ghafoori, E., Motavalli, M., Nussbaumer, A., Herwig, A., Prinz, G.S. & Fontana, M. 2015.
Determination of minimum CFRP pre-stress levels for fatigue crack prevention in
retrofitted metallic beams. Engineering Structures 84: 29 – 41.
Ghana Sheela, K. & Deepa, S.N. 2013. Review on methods to fix number of hidden neurons
in neural networks. Mathematical Problems in Engineering 425740.
Ghasemi, A. & Zahediasl, S. 2012. Normality tests for statistical analysis: a guide for non-
statisticians. International Journal of Endocrinology Metabolism 10(2): 486 – 489.
230
Giannakis, E., Malikoutsakis, M. & Savaidis, G. 2016. Fatigue design of leaf springs for new
generation trucks. Materials Science and Engineering 161: 012065.
Gohari, M., Rahman, R.A., Tahmasebi, M., Nejat, P. 2014. Off-road vehicle seat suspension
optimisation part 1: derivation of an artificial neural network model to predict seated
human spine acceleration in vertical vibration. Journal of Low Frequency Noise,
Vibration and Active Control 33(4): 429 – 442.
Gokhale, N.S., Deshpande, S.S., Bedekar, S.V., Thite, A.N. 2008. Practical Finite Element
Analysis. India: Finite to Infinite.
Gonzalez, A., Obrien, E.J., Li, Y.Y., Cashell, K. 2008. The use of vehicle acceleration
measurements to estimate road roughness. Vehicle System Dynamics 46(6): 483 –
499.
Graves, A. 2011. Practical Variational Inference for Neural Networks. NIPS: 2348 - 2356.
Graves, A., Mohamed, A., Hinton, G. 2013. Speech recognition with deep recurrent neural
networks. 2013 IEEE International Conference on Acoustics, Speech and signal
Processing, Canada.
Gribbin, S., Bicknell, J., Jorgensen, L., Tsukrov, I., Knezevic, M. 2016. Low cycle fatigue
behaviour of direct metal laser sintered Inconel alloy 718. International Journal of
Fatigue 93: 156 – 167.
Grujicic, M., Arakere, G., Bell, W.C., Marvi, H., Yalavarthy, H.V., Pandurangan, B., Haque,
I., Fadel, G.M. 2010. Reliability-based design optimization for durability of ground
vehicles suspension system components. Journal of Materials Engineering and
Performance 9482.
Gu, S. 2017. Application of finite element method in mechanical design of automotive parts.
Materials Science and Engineering 231, 012180.
Gupta, C. & Kumawat, B. 2016. Analysis of flow behaviour of iron and steel using power law
models. Materials Science and Technology 33(8): 971 – 983.
Hackl, A., Hirschberg, W., Lex, C., Magele, C. 2017. Parameterisation of a Maxwell model for
transient tyre forces by means of an extended firefly algorithm. Advances in Mechanical
Engineering 9(1): 1 – 11.
Halfpenny, A. & Kihm, F. 2010. Rainflow cycle counting and acoustic fatigue analysis
technique for random loading. 10th International Conference RASD 2010,
Southampton, UK.
Hassaan, G.A. 2015. Car dynamics using quarter car model and passive suspension, part II: a
novel simple harmonic bump. IOSR Journal of Mechanical and Civil Engineering
12(1): 93 – 100.
Hassaan, G.A. 2015. Car dynamics using quarter model and passive suspension, Part VI:
Sprung-mass step response. IOSR Journal of Computer Engineering 17(2): 65 – 74.
231
He, B.Y., Wang, S.X. & Gao, F. 2010. Failure Analysis of an Automobile Damper Spring
Tower. Engineering Failure Analysis 17: 498 – 505.
He, H., Yi, G. & Sun, J. 2016. Measurement of foil bonded strain gauges with 3 wire quarter
bridge. 2016 International Conference on Applied Mechanics, Electronics and
Mechatronics Engineering Engineering, May, China.
Heidari, E., Sobati, M.A. & Movahedirad, S. 2016. Accurate prediction of nanofluid viscosity
using a multilayer perceptron artificial neural network (MLP-ANN). Chemometrics and
Intelligent Laboratory Systems 155: 73 – 85.
Heiβing, B. & Ersoy, M. 2011. Chassis Handbook Fundamentals, Driving Dynamics,
Components, Mechatronics, Perspectives. Germany: Vieweg Teubner Verlag.
Hemmat Esfe, H., Yan, W., Akbari, M., Karimipour, A., Hassani, M. 2015. Evaluation of
thermal conductivity of COOH-functionalized MWCNTs/water via temperature and
solid volum fraction by using experimental data and ANN methods. Journal of Thermal
Analysis and Calorimetry 121(3): 1273 – 1278.
Hsu, W.Y. 2012. Application of competitive hopfield neural network to brain computer
interface systems. International Journal of Neural Systems 22(1): 51 – 62.
Hu, J., Chen, B., Smith, D.J., Flewitt, P.E.J, Cocks, A.C.F. 2016. On the evaluation of
Bauschinger effect in an austenitic steel - The role of multi-scale residual stresses.
International Journal of Plasticity 84: 203 – 223.
Huang, Y.S., Huang, C.D. & Ho, J.W. 2017. A customized two-dimensional extended warranty
with preventive maintenance. European Journal of Operational Research 257(3): 971
– 978.
Huang, M., Luo, H. & Yi, S. 2013. Determination analysi of the torsion beam suspension FEA
model and fatigue strain based on road test. Advanced Materials Research 779-780:
921 – 924.
Huang, Y. & Griffin, M.J. 2014. A model predicting discomfort caused by combined noise and
vibration. The 21st International Congress on Sound and Vibration, July, China.
Hui, Y., Li, Bo., Kawai, H., Yang, Q. 2017. Non-stationary and non-Gaussian characteristics
of wind speeds. Wind and Structures 24(1): 59 – 78.
Ibrahim, D. 2013. An overview of soft computing. Procedia Computer Science 102: 34 – 38.
Ibrahimy, M.I., Ahsan, M.R., Khalifa, O.O. 2013. Design and optimization of Levenberg-
Marquadt based neural network classifier for EMG signals to identify hand motions.
Measurement Science Review 13(1): 202 – 209.
Ince, A. & Glinka, G. 2011. A modification of Morrow and Smith-Watson-Topper mean stress
correction models. Fatigue and Fracture of Engineering Materials and Structures
34(11): 854 – 867.
232
ISO 8608. 2016. Mechanical vibration – road surface profile – reporting of measured data,
Switzerland, Geneva: International Standardization Organisation.
Jaccard, J. & Turrisi, R. 2003. Interaction effects in multiple regression. Materials and
Structures 49(11): 4749 – 4763.
Janežič, M., Klemenc, J., Fajdiga, M. 2010. A neural network approach to describe the scatter
of cyclic stress-strain curves. Materials and Design 31: 438 – 448.
Jazar, R.N. 2014. Vehicle Dynamics Theory and Application. U.S.A.: Springer Verlag.
Jiang, R. & Murthy, D.N.P. 2011. A Study of Weibull Shape Parameter: Properties and
Significance. Reliability Engineering and System Safety 96: 1619 – 1626.
Jin, L., Yu, Y. & Fu, Y. 2016. Study on the ride comfort of vehicles driven by in-wheel motors.
Advances in Mechanical Engineering 8(3): 1 – 9.
John, V.B. & Phillip, V.B. 2012. The Fatigue Damage Spectrum and Kurtosis Control.
Sound & Vibration Magazine: 10 – 12.
Joshi, O.P., Jadhav, T.A., Pawar, P.R., Saraf, M.R. 2015. Investigating effect of road
roughness and vehicle speed on the dynamic response of the seven degrees-of-
freedom vehicle model by using artificial road profile. International Journal of
Current Engineering and Technology 5(4): 2596 – 2602.
Junior, A.B.B. & Pores, M.S.M. 2014. An approach to outlier detection and smoothing
applied to a trajectography radar data. Journal of Aerospace Technology and
Management 6(3): 237 – 248.
Juvinall, R.C. & Marshek, K.M. 2017. Fundamentals of Machine Component Design, 6th ed.
U.S.A.: John Wiley & Sons Inc.
Kamal, M. & Rahman, M.M. 2014. Finite element-based fatigue behaviour of springs in
automobile suspension. International Journal of Automotive and Mechanical
Engineering 10: 1910 – 1919.
Kamer-Ainur, A., & Marioara, M. 2007. Errors and limitations associated with regression and
correlation analysis. Statistic and Economic Informatics, 709.
Kamp, R.G., & Savenjie, H.H.G. 2006. Optimising training data for ANNs with genetic
algorithm. Hydrology and Earth System 10: 603 – 608.
Kang, J.Y., Choi, B.I., Lee, H.J., Kim, J.S., Kim, K.J. 2006. Neural network application in
fatigue damage analysis under multiaxial random loadings. International Journal of
Fatigue 28:132 – 140.
Kanoje, N.K., Sharma, S.C. & Harsha, S.P. 2013. Response surface method based railway
wheel optimisation for natural frequency. International Journal of Vehicle Structures
and Systems 5(3-4): 90 – 94.
233
Kaoua, S.A., Taibi, K., Benghanem, N., Azouaoui, K., Azzaz, M. 2011. Numerical modelling
of twin helical spring under tensile loading. Applied Mathematical Modelling 35: 1378
– 1387.
Karakas, Ö. & Tomasella, M. 2013. Fatigue life estimation of non-penetrated butt weldments
in light metals by artificial neural network approach. Materialwissenschaft und
Werkstofftechnik 44(10): 847 – 855.
Karlik, B. & Vehbi Olgac, A. 2011. Performance analysis of various activation functions in
generalized MLP architecture of neural networks. International Journal of Artificial
Intelligence and Expert System 1(4): 111 – 122.
Karlsson, M. 2007. Load Modelling for Fatigue Assessment of Vehicles – a Statistical
Approach, PhD thesis, Göteborg university.
Karolczuk, A. 2016. Analysis of revised fatigue life calculation algorithm under proportional
and non-proportional loading with constant amplitude. International Journal of Fatigue
88: 111 – 120.
Karthik, J.P., Chaitanya, K.L. & Sasanka, C.T. 2012. Fatigue Life Prediction of a Parabolic
Spring under non-Constant Amplitude Proportional Loading using Finite Element
Method. International Journal of Advanced Science and Technology 46: 143 – 156.
Kavita, C., Abinav Shankar, S., Ashok, B., Denis Ashok, S., Ahmed, H., Kaisan, M.U. 2018.
Adaptive suspension strategy for a double wishbone suspension through camber and
toe optimization. Engineering Science and Technology 21: 149 – 158.
Khademi, F., Jamal, S.M., Deshpande, N., Londhe, S. 2016. Predicting strength of recycled
aggregate concrete using artificial neural network, adaptive neuro-fuzzy inference
system and multiple linear regression. International Journal of Sustainable Built
Environment 5: 355 – 369.
Khan, S.M.A., Benyahia, F., Bouiadjra, B.B., Albedah, A. 2014. Analysis and repair of crack
growth emanating from V-notch under stepped variable fatigue loading. Procedia
Engineering 74: 151 – 156.
Khotimah, S.N., Viridi, S., Widayani, Khairurrijal. 2011. The dependence of the spring
constant in the linear range on spring parameters. Physical Education 46(5): 540 – 543.
Kihm, F. & Delaux, D. 2013. Vibration Fatigue and Simulation of Damage on Shaker Table
Tests: the Influence of Clipping the Random Drive Signal. Procedia Engineering
66: 549 – 564.
Kihm, F., Ferguson, N.S., Antoni, J. 2015. Fatigue life from kurtosis controlled excitations.
Procedia Engineering 133: 698 – 713.
Kim, D.H., Choi, D.H., Kim, H.S. 2014. Design optimization of a carbon fiber reinforced
composite automotive lower arm. Composites: Part B 58: 400 – 407.
234
Kim, K.S., Chen, X., Han, C., Lee, H.W. 2002. Estimation methods for fatigue properties of
steels under axial and torsional loading. International Journal of Fatigue 24: 783 – 793.
Kisi, O. 2015. Streamflow forecasting and estimation using least square support vector
regression and adaptive neuro-fuzzy embedded Fuzzy c-means clustering. Water
Resource Management 29(14): 5109 – 5127.
Klemenc, J., Janezic, M. & Fajdiga, M. 2012. Modelling the dependency of the Smith-Watson-
Topper parameter on the cycles to failure using serial hybrid neural networks. Fatigue
and Fracture of Engineering Materials and Structures 35: 809 – 825.
Kluger, K. & Lagoda, T. 2016. Fatigue life estimation for selected materials in multiaxial stress
states with mean stress. Journal of Theoretical and Applied Mechanics 54(2): 385 –
396.
Knapczyk, J. & Maniowski, M. 2010. Optimization of 5-rod car suspension for elastokinematic
and dynamic characteristics. The Archive of Mechanical Engineering LVII(2): 133 –
147.
Koch, G., Pellegrini, E., Spirk, S., Lohmann, B. 2010. Design and modeling of a quarter-
vehicle test rig for active suspension, Technical Reports on Automatic Control 5.
Kosobudzki, M. 2014. The use of acceleration signal in modeling process of loading an element
of underframe of high mobility wheeled vehicle. Maintenance and Reliability 16(4):
595 – 599.
Kovaci, H., Yetim, A.F., Baran, O., Celik, A. 2016a. Fatigue crack growth behaviour of DLC
coated AISI 4140 steel under constant and variable amplitude loading conditions.
Surface and Coatings Technology 304: 316 – 324.
Krauze, P. & Kasprzyk, J. 2016. Mixed skyhook and FxLMS control of a half-car model with
magnetorheological dampers. Advances in Acoustics and Vibration 7428616.
Kreyszig, E. 2011. Advanced Engineering Mathematics, 10th Ed. U.S.A.: John Wiley & Sons,
Inc.
Kropac, O. & Mucka, P. 2007. Indicators of longitudinal road unevenness and their mutual
relationships. Road Materials and Pavement Design 8(3): 523 – 549.
Krzanowski, W.J. 2010. An Introduction to Statistical Modelling. U.K.: Wiley.
Kubo, P., Paiva, C., Ferreira, A., Larocca, A. 2015. Influence of shock absorber condition on
pavement fatigue using relative damage concept . Journal of Traffic and Transportation
Engineering 2(6): 406 – 413.
Kulkarni, A., Ranjha, S.A. & Kapoor, A. 2016. Fatigue analysis of a suspension for an in-wheel
electric vehicle. Engineering Failure Analysis 68: 150 – 158.
Kun, F., Carmona, H.A., Andrade, J.S., Herrmann, H. J. 2008. Universality behind Basquin’s
law of fatigue. Physical Review Letters 100, 094301.
235
Kuruvilla, J. & Gunavathi, K. 2014. Lung cancer classification using neural networks for CT
images. Computers Methods and Programs in Biomedicine 113(1): 202 – 209.
Kwofie, S. & Rahbar, N. 2012. A fatigue driving stress approach to damage and life prediction
under variable amplitude loading. International Journal of Damage Mechanics 0(0): 1
– 12.
LabVIEW. 2010. Overview of Curve Fitting Models and Methods in Labview, Labview
technical manual.
Lakušić, S., Brčić, D. & Lakušić, V.T. 2011. Analysis of vehicle vibration – new approach to
rating pavement condition of urban roads. Traffic and Transportation 23(6): 485 – 494.
Lavanya, N., Rao, P.S. & Pramod Reddy, M. 2014. Design and analysis of a suspension coil
spring for automotive vehicle. International Journal of Engineering Research and
Applications 4(9): 151 – 157.
Ledesma, R., Jenaway, L., Wang, Y., Shih, S. 2005. Development of accelerated durability
tests for commercial vehicle suspension components. SAE Technical Paper 2005-01-
3565.
Lee, A., Geem, Z.W. & Suh, K.D. 2016. Determination of optimal initial weights of an artificial
neural network by using the Harmony search algorithm: application to breakwater
armor stones. Applied Sciences 6, 164.
Lee, Y.L., Barkey, M.E. & Kang, H.T. 2012. Metal Fatigue Analysis Handbook. U.S.: Elsevier
Butterworth-Heinemann.
Lee, Y.L., Pan, J., Hathaway, R. & Barkey, M. 2005. Fatigue Testing and Analysis: Theory
and Practice. Oxford: Elsevier Butterworth-Heinemann.
Lennie, A., Abdullah, S., Z.M. Nopiah, M.N. Baharin. 2010. Behavioural investigation of
fatigue time series using the statistical approach. Journal of applied Sciences 10(16):
171 – 1722.
Lewis-Beck, M.S. 2015. Applied Regression: An Introduction. U.K.: Sage publication.
Ligaj, B. 2011. An analysis of the influence of cycle counting methods on fatigue life
calculations of steel. Scientific Problems on Machines Operation and Maintenance
4(168): 25 – 43.
Lin, G, Pang, H., Zhang, W., Wang, D., Feng, L. 2014. A self-decoupled three-axis force sensor
for measuring the wheel force. Proceeding of Institute of Mechanical Engineers, Part
D: Journal of Automobile Engineering 228(3): 319 – 334.
Liu, F., Zhou, S., Xia, C., Zeng, D., Shi, T. 2016. Optimization of fatigue life distribution model
and establishment of probabilistic S-N curves for a 165 ksi grade super high strength
drill pipe. Journal of Petroleum Science and Engineering 145: 527 – 532.
236
Lopez-Garcia, R.D., Garcia-Pastor, F.A., Castro-Roman, M.J., Alfaro-Lopez, E., Acosta-
Gonzalez, F.A. 2015. Effect of immersion routes on the quenching distortion of a long
steel component using a finite element model. Transactions of the Indian Institute of
Metals 69(9): 1645 – 1656.
Lopez-Moreno, I., Gonzalez-Dominiguez, J., Plchot, J., Martinez, D., Gonzalez-Rodriguez, J.,
Moreno, P. 2014. Automatic language identification using deep neural networks, 2014
IEEE International Conference on Acoustics, Speech and Signal Processing, 5337 –
5341.
Loprencipe, G. & Zocalli, P. 2017. Ride quality due to road surface irregularities: comparison
of different methods applied on a set of real road profiles. Coatings 7(59): 1 – 16.
Loprencipe, G. & Zoccali, P. 2017. Use of generated artificial road profiles in road
roughness evaluation. Journal of Modern Transportation 25(1): 24 – 33.
Loutridis, S., Gialamas, T., Gravalos, I., Moshou, D., Kateri, D., Xyradakis, P., Tsiropoulos,
Z. 2011. A study on the effect of electronic engine speed regulator on agricultural
tractor ride vibration behavior. Journal of Terramechanics 48(2): 139 – 147.
Lu, Z., & Liu, Y. 2011. Experimental investigation of random loading sequence effect on
fatigue crack growth. Materials and Design 32: 4773 – 4785.
Luo, H., Huang, M., Zhou, Z. 2018. Integration of multi-Gaussian fitting and LSTM neural
networks for health monitoring of an automotive suspension component. Journal of
Sound and Vibration 428: 87 – 103.
Mucka, P. 2018. Simulated road profiles according to ISO 8608 in vibration analysis.
Journal of Testing and Evaluation 46(1): 405 – 418.
Mahala, M.K., Gadkari, P., Anindya, D. 2009. Mathematical models for designing vehicles for
ride comfort. Proceedings of the 2nd International Conference on Research into Design,
168 – 175.
Mahmoodi-Kaleibar, M., Javanshir, I., Asadi, K., Afkar, A., Paykani, A. 2013. Optimization
of suspension system of off road vehicle for vehicle performance improvement. Journal
of Central South University 20: 902 – 910.
Malik, F., Wang, F., Naseem, M.A. 2017. Econometric estimation of banking stocks. The
Journal of Developing Areas 51(4): 207 – 237.
Manivel, D. & Gandhinathan, R. 2016. Prediction of cutting tool vibration in hard turning using
response surface methodology. International Journal of Advanced Engineering
Technology 13(1): 1 – 5.
Manson, S.S. & Halford G.R. 2006. Fatigue and Durability of Structural Materials. U.S.A.:
ASTM International.
237
Marsh, G., Wignall, C., Thies, P.R., Barltrop, N., Incecik, A., Venugopal, V., Johanning, l.
2016. Review and application of Rainflow residue processing techniques for accurate
fatigue damage estimation. International Journal of Fatigue 82: 757 – 765.
Mas, D.M.L & Ahlfeld, D.P. 2007. Comparing artificial neural networks and regression models
for predicting faecal coliform concentrations. Hydrological Sciences Journal 52(4):
713 – 731.
Mattetti, M., Molari, G., Vertua, A. 2015. New methodology for accelerating the four-post
testing of tractors using wheel-hub displacements. Biosystems Engineering 129: 307 –
314.
Mayén, J., Abúndez, A., Pereyra, I., Colín, J., Blanco, A., Serna, S. 2017. Comparative analysis
of the fatigue short crack growth on Al 6061-T6 alloy by the exponential crack growth
equation and a proposed empirical model. Engineering Faracture Mechanics 177: 203
– 217.
Meneghetti, G., Guzzella, C., Atzori, B. 2014. The peak stress method combined with 3D finite
element models for fatigue assessment of toe and root cracking in steel welded joints
subjected to axial or bending loading. Fatigue and Fracture of Engineering Materials
and Structures 37: 722 – 739.
Meneghetti, G., Ricotta, M., Atzori, B. 2016. A two-parameter, heat energy-based approach to
analyse the mean stress influence on axial fatigue behaviour of plain steel specimens.
International Journal of Fatigue 82: 60 – 70.
Mia, M., Khan, M.A., Dhar, N.R. 2017. Performance prediction of high-pressure coolant
assisted turning of Ti-6Al-4V. International Journal of Advanced Manufacturing
Technology 90: 1433 – 1445.
Mitra, A.C. & Banerjee, N. 2015. A LabVIEW-based data acquisition system in a quarter car
test rig to optimize vehicle suspension system. Intelligent Computing and Application
343: 593 – 601.
Mohammadi, M. 2012. Road classification and condition determination using hyperspectral
imagery. International Archives of the Photogrammetry, Remote Sensing and Spatial
Information Science, Australia.
Mokhtari, M.H., Busu, I., Mokhtari, H., Zahedi, G., Sheikhattar, L., Mohaved, M.A. 2013.
Neural network and multiple linear regression for estimating surface Albedo from
ASTER visible and near infrared spectral bands. Earth Interactions 17(3): 1 – 20.
Moon, S.I., Cho, I.J. & Yoon, D. 2011. Fatigue life evaluation of mechanical components using
vibration fatigue analysis technique. Journal of Mechanical Science and Technology
25(3): 631 – 637.
Mottitschka, T., Pusch, G., Biermann, H., Zybell, L., Kuna, M. 2010. Influence of overloads on
the fatigue crack growth in nodular cast iron: experiments and numerical simulation.
Procedia Engineering 2: 1557 – 1567.
238
Mrsnik, M., Slavic, J., Boltezar, M. 2013. Frequency-domain methods for a vibration-
fatigue-life estimation – application to real data. International Journal of Fatigue
47: 8 – 17.
Mrad, H., Erchiqui, F., Bouazara, M., Hammoud, R. 2016. Effect of rainflow cycle number on
fatigue lifetime of an arm of a vehicle suspension system. International Journal of
Advanced Research in Engineering 2(3): 6 – 10.
Naeem, S.A.A. & Jagtap, P.V. 2017. Analysis of upper control arm for finding optimised
model using FEA and experimentation. Imperial Journal of Interdisciplinary
Research 3(3): 1814 – 1818.
Nagarkar, M.P., Vikhe Patil, G.J., Zaware Patil, R.N. 2016. Optimization of nonlinear
quarter car suspension-seat-driver model. Journal of Advanced Research 7: 991 –
1007.
Naik, R.D. & Singru, P.M. 2010. Resonance, stability and chaotic vibration of a quarter-car
vehicle model with time-delay feedback. Communications in Nonlinear Science and
Numerical Simulation 16: 3397 – 3410.
Nakano, T., Sakakibara, T., Wakita, M., Sugimoto, A. 2000. Development of high-strength
suspension coil springs in which corrosion fatigue strength is important. Journal of
Materials and Manufacturing 109(5): 203 – 209.
Nam, J.S., Shin, H.W. & Choi, G.J. 2014. Durability prediction for automobile aluminium front
subframe using nonlinear models in virtual test simulations. International Journal of
Automotive Technology 15(4): 593 – 601.
Ngai, E.W.T., Hu, Y., Wong, Y.H., Chen, Y., Sun, X. 2011. The application of data mining
techniques in financial fraud detection: A classification framework and an academic
review of literature. Design Support Systems 50: 559 – 569.
Ngwanga, H.M., Heyns, P.S., Breytenbach, H.G.A., Els, P.S. 2014. Reconstruction of road
defects and road roughness classification using artificial neural network simulation and
vehicle dynamic responses: Application to experimental data. Journal of
Terramechanics 53: 1 – 8.
Ning, X., Zhao, C., Shen, J. 2013. Dynamic analysis of car suspension using ADAMS/car for
development of a software interface for optimisation. Procedia Engineering 16: 333 –
341.
Norton, M.P. & Karczub, D.G. 2007. Fundamentals of Noise and Vibration Analysis for
Engineers, 2nd Ed. UK: Cambridge University Press.
Ogunoiki, A.O. & Olatunbosun, O. 2015. Characterisation of the effects of vehicle parameter
variations on vehicle road load data. SAE Technical Paper 2015-01-0634.
Ogunoiki, A.O. 2015. Artificial Road Input Data Generation Tool for Vehicle Durability
Assessment using Artificial Intelligence. PhD thesis, University of Birmingham.
239
Oh, C.S. 2001. Application of Wavelet Transform in Fatigue History Editing. International
Journal of Fatigue 23: 241 – 250.
Ospina, R. & Ferrari, S.L.P. 2012. A general class of zero-or-one inflated beta regression
models. Computational Statistics and Data Analysis 56(6): 1609 – 1623.
Ozmen, B., Altiok, B., Guzel, A., Kocyigit, I., Atamer, S. 2015. A novel methodology with
testing and simulation of leaf springs based on measured load collectives. Procedia
Engineering 101: 363 – 371.
Paddan, G.S., Mansfield, N.J., Arrowsmith, C.J., Rimell, A.N., King, S.K., Holmes, S.R. 2012.
The influence of seat backrest angle on perceived discomfort during exposure to
vertical whole-body vibration. Ergonomics 55(8): 923 – 936.
Padzi, M.M., Abdullah, S., Nuawi, M.Z. 2014. Fatigue damage assessment correlating with I-
Kaz coefficient. Applied Mechanics and Materials 471: 235 – 240.
Palowitch, J., Shabalin, A., Zhou, Y.H., Nobel, A.B., Wright, F.A. 2016. Estimation of csi-
eQTL effect sizes using a log linear model. Biometrics, 12810.
Pandra, K.C. 2016. Dealing with noise and vibration in automotive industry. Procedia
Engineering 144: 1167 – 1174.
Pang, H., Li, H.Y., Fang, Z.D., Shan, N. 2010. Ride comfort optimization and test research on
an 8x4 heavy trucky. Noise and Vibration Worldwide 41(10): 65 – 71.
Panta, S.R., Avasarala, R., Koona, R. 2014. Ride dynamic behaviour of coupled human-
vehicle vibratory model. International Journal of Human Factors Modelling and
Simulation 4(3/4): 250 – 265.
Pantazopoulos, G., Vazdirvanidis, A., Rikos, A. & Toulfatzis, A. 2014. Analysis of Abnormal
Fatigue Failure of Forklift Forks. Case Studies in Engineering Failure Analysis 2: 9 –
14.
Paraforos, D.S., Griepentrog, H.W. & Vougioukas, S.G. 2016. Country road and field surface
profiles acquisition, modeling and synthetic realisation for evaluating fatigue life of
agricultural machinery. Journal of Terramechanics 63: 1 – 12.
Paraforos, D.S., Griepentrog, H.W., Vougioukas, S.G., Kortenbruck, D. 2014. Fatigue life
assessment of a four-rotor swather based on rainflow cycle counting. Biosystems
Engineering 127: 1 – 10.
Parekh, A., Khumbar, S.B., Joshi, S.G. 2014. Transmissibility analysis of a car driver’s seat
suspension system with an air bellow type damper. International Journal on Recent
Technologies in Mechanical and Electrical Engineering 1(3): 12 – 19.
Patel, T.M. & Bhatt, N.M. 2013. ANN and MLR model for shear stress prediction of eicher
11.10 chassis frame: A comparative study. International Journal of Mechanical
Engineering and Technology 4(5): 216 – 223.
240
Patel, T.M. & Bhatt, N.M. 2016. Development of FEA-ANN hybrid model for equivalent stress
prediction of automobile structural member. Automatic Control and Computer Sciences
50(5): 293 – 305.
Patil, M. & Sharma, V. 2015. Optimization, design and analysis of push rod actuated double
wishbone suspension system. Journal of Material Science and Mechanical Engineering
2(13): 16 – 20.
Pawar Prashant, R. & Saraf, M.R. 2009. Measurement of road profile and study it’s effect on
vehicle durability and ride. SAE Technical Paper 2009-26-0070.
Petaś, M., Mróz, K., Doliński, K. 2013. Damage modelling in graded layer system. Acta
Mechanica et Automatica 7(2): 99 – 106.
Pevec, M., Oder, G., Potrc, I., Sraml, M. 2014. Elevated temperature low cycle fatigue of grey
cast iron used for automotive disc brakes. Engineering Failure Analysis 42: 221 – 230.
Phalke, T.P. & Mitra, A.C. 2017. Analysis of ride comfort and road holding of quarter car
model by Simulink. Materials Today: Proceedings 4: 2425 – 2430.
Pionke, R., Bocik, J.G. 2011. Semi-Active Suspension System. US Patent 8,075,002.
Popovic, V., Vasic, B., Petrovic, M., Mitic, S. 2011. System approach to vehicle suspension
system control in CAE environment. Journal of Mechanical Engineering 57(2): 100 –
109.
Pospíchal, J. & Kvasnička, V. 2015. 70th Anniversary of publication: Warren McCulloh &
Walter Pitts – A Logical Calculus of the Ideas Immanent in Nervous Activity.
U.S.A.: Springer cham.
Prakash, J., Nandi, S. & Sivakumar, K. 2006. Dynamic stress and durability analyses: time vs
frequency domain approaches, MSC Software Corporation’s 2006 Americas Virtual
Produce Development Conference, U.S.A.
Prawoto, Y., Ikeda, M., Manville, S.K., Nishikawa, A. 2008. Design and failure modes of
automotive suspension springs. Engineering Failure Analysis 15: 1155 – 1174.
Prieto, A.J., Silva, A., Brito, J., Macias-Bernal, J.M., Alejandre, F.J. 2017. Multiple linear
regression and fuzzy logic models applied to the functional service life prediction of
cultural heritage. Journal of Cultural Heritage 27: 20 – 35.
Prieto, M., Tanner, P., Andrade, C. 2016. Multiple linear regression model for the assessment
of bond strength in corroded and non-corroded steel bars in structural concrete.
Materials and Structures 49(11): 4749 – 4763.
Purusthotham, A. 2013. Comparative simulation studies on MacPherson suspension system.
International Journal of Modern Engineering Research 3(3): 1377 – 1381.
Putra, T.E. 2016. Improving the Fatigue Life Prediction of Automotive Components using
Simulated Strain Signals Method, PhD thesis, Universiti Kebangsaan Malaysia.
241
Pyttel, B., Brunner, I., Kaiser, B., Berger, C., Mahendran, M. 2014. Fatigue behaviour of helical
compression springs at a very high cycle number of cycles – investigation of various
influences. International Journal of Fatigue 60: 101 – 109.
Qian, J. & Fatemi, A. 1995. Cyclic deformation and fatigue behaviour of ion-nitrided steel.
International Journal of Fatigue 17(1): 15 – 24.
Radhakrishna, D.V. & Gurmukhdas, M.A. 2013. Experimental investigation of inimitable
platform for heavy vehicle chassis. International Journal of Automobile Engineering
Research & Development 3(3): 7 – 12.
Rafael, J.M.S. 2007. Development of An On-line Ride Comfort Evaluation Tool. M. Sc. Thesis,
Royal Institute of Technology.
Ramamurti, V. & Sujatha, C. 1990. Bus vibration study – theoretical response analysis and
experimental verification. International Journal of Vehicle Design 11(4/5): 401 – 409.
Rao, S.J. 2009. Vehicle Modeling and ADAMS-simulink Co-simulation with Integrated
Continuously Controlled Electronic Suspension (CES) and Electronic Stability Control
(ESC) Models. M. Sc. Thesis, The Ohio State University.
Rasyidi, R., Pratiwi, D.K. 2015. Mechanical behavior of medium carbon steel welded with
SMAW and corroded with sea water. Journal of Mechanical Science and Engineering
2(1): 17 – 23.
Rathore, G.S., & Joshi, U.K. 2013. Fatigue stress analysis of helical spring, a review.
International Journal of Emerging Trends in Engineering and Development 3(2): 512
– 520.
Razali, N.M. & Yap, B.W. 2011. Power comparisons of Shapiro-Wilk, Kolmogorov-Smirnov,
Lilliefors and Anderson-Darling tests. Journal of Statistical Modeling and Analytics
2(1): 21 – 33.
Reddy, K.V., Kodati, M., Chatra, K., Bandyopadhyay, S. 2016. A comprehensive kinematic
analysis of the double wishbone and Macpherson strut suspension systems. Mechanism
and Machine Theory 105: 441 – 470.
Repetto, M.P. & Torrielli, A. 2017. Long term simulation of wind-induced fatigue loadings.
Engineering Structures 132: 551 – 561.
Reza-Kashyzadeh, K, Ostad-Ahmad-Ghorabi, M.J. & Arghavan, A. 2014. Investigating the
effect of road roughness on automotive component. Engineering Failure Analysis 41:
96 – 107.
Ricotta, M. 2015. Simple regressions to estimate the Coffin-Manson curves of ductile cast
irons. International Journal of Fatigue 78: 38 – 45.
Rizvi, S.M.H., Abid, M., Khan, A.Q., Satti, S.G., Latif, J. 2016., H∞ control of 8 degrees of
freedom vehicle active suspension system. Journal of King Saud University, article in
press: https://doi.org/10.1016/j.jksues.2016.02.004.
242
Rizzi, S.A., Behnke, M.N., Przekop, A. 2010. The effect of a non-Gaussian random loading on
high-cycle fatigue of a thermally post-buckled structure. 10th International Conference
RASD, Southampton.
Roman, L., Florea, A., Cofaru, I. 2014. Mathematical model and software simulation of
suspension’s system from Opel cars. Annals of the Oradea University Fascicle of
Management and Technological Engineering 23(3): 94 – 99.
Roshanfar, M. & Salimi, M.H., 2014. Comparing of methods of cycle calculating and counting
to the rain flow method. Extensive Journal of Applied Sciences 3(7): 291 – 296.
Rothleutner, L., Cyderman, R., Van Tyne, C.J., Burke, J. 2015. Torsional fatigue performance
of induction hardened 1045 and 10V45 steels. Proceedings of the 28th ASM Heat
Treating Society Conference, U.S.A.
Runciman, A., Xu, D., Pelton, A.R. & Ritchie, R.O. 2011. An Equivalent Strain / Coffin-
Manson Approach to Multiaxial Fatigue and Life Prediction in Superelastic Nitinol
Medical Devices. Biomaterials 32: 4987 – 4993.
Ryu, Y.I., Kang, D.O., Heo, S.J., Yim, H.J., Jeon, J.I. 2010. Development of analytical process
to reduce side load in strur-type suspension. Journal of Mechanical Science and
Technology 24: 351 – 356.
Salim, M.I.M., Ahmad, A.H., Ariffin, I., Rosidi, B., Supriyanto, E. 2015. Diagnosis of breast
cancer using hybrid magnetoacoustic method and artificial neural network,
Proceedings ACS’11 of the 11th WSEAS International Conference on Applied Computer
Science: 64 – 69, Penang.
Santecchia, E., Hamouda, A.M.S., Musharavati, F., Zalnezhad, E., Cabibbo, M., El Mehtedi,
M., Spigarelli, S. 2016. A review on fatigue life prediction for metals. Advances in
Materials Science and Engineering 9573524.
Saurabh, Y.S., Kumar, S., Jain, K.K., Behera, S.K., Gandhi, D., Raghavendra, S., Kalita, K.
2016. Design of suspension system for Formula student race car. Procedia Engineering
144: 1138 – 1149.
Saoudi, A., Bouazara, M. & Marceau, D. 2011. Fatigue failure study of the lower suspension
vehicle arm using a multiaxial criterion of the strain energy density. Strojniški vestnik
- Journal of Mechanical Engineering 57(4): 345 – 356.
Sashikumar, P., Sujatha, C., Chinaraj, K. 2017. Transient fatigue analysis of exhaust system
mounting brackets for commercial vehicle – correlation. SAE Technical Paper 2017-
01-1333.
Satishkumar, P., Jancirani, J., John, D., Manikandan, S. 2014. Mathematical modelling and
simulation quarter car vehicle suspension. International Journal of Innovative Research
in Science, Engineering and Technology 3(1): 1280 – 1283.
243
Scheiblegger, C, Roy, N., Pfeffer, P.E., Hillis, A.J. 2016. Modeling hydro mounts in vehicles
for durability load analyses, ride comfort and vehicle dynamics simulation. In: 13th
International Symposium on Advance Vehicle Control, 703 – 712.
Schmidhuber, J. 2014. Deep Learning in Neural Networks: An Overview. Technical report
IDSIA-03-13.
Schneider, C.R.A. & Maddox, S.J. 2003. Best Practice Guide on Statistical Analysis of Fatigue
Data. TWI statistic report, Cambridge UK.
Schramm, D., Hiller, M., Bardini, R. 2014. Vehicle Dynamics Modeling and Simulation.
Springer-Verlag Berlin Heidelberg, London UK.
Sedlák, P., Frost, M., Kruisová, A., Hiřmanová, K., Heller, L., Šittner, P. 2014. Simulations of
mechanical response surface of superelastic NiTi helical spring and its relation to
fatigue resistance. Journal of Materials Engineering and Performance 23(7): 2591 –
2598.
Seifi, A., Hassannejad, R. & Hamed, M.A. 2015. Optimum design for passive suspension
system of a vehicle to prevent rollover and improve ride comfort under random road
excitations. Proceedings of Institute of Mechanical Engineers, Part K: Journal of
Multibody Dynamics 0(0): 1 – 16.
Sekulic, D. & Dedovic, V. 2011. The effect of stiffness and damping of the suspension system
elements on the optimization of the vibrational behaviour of a bus. International
Journal for Transport and Traffic Engineering 4(1): 231 – 244.
Sekulic, D., Dedovic, V., Rusov, S., Obradovic, A., Salinic, S. 2016. Definition and
determination of the bus oscillatory comfort zones. International Journal of Industrial
Ergonomics 53: 328 – 339.
Sener, A.S. 2011. Determination of vehicle components fatigue life based on FEA method and
experimental analysis. International Journal of Electronics, Mechanical and
Mechatronics Engineering 2(1): 133 – 145.
Senthil Kumar, M. & Vijayarangan, S. 2007. Analytical and experimental studies on active
suspension system of light passenger vehicle to improve ride comfort. Mechanika
3(65): 34 – 41.
Sequeira, A.A, Singh, R.K., Shetti, G.K. 2016. Comparative analysis of helical steel springs
with composite springs using finite element method. Journal of Mechanical
Engineering and Automation 6(5A): 63 – 70.
Sezgin, A. & Arslan, Y.Z. 2012. Analysis of the vertical vibration effects on ride comfort of
vehicle driver. Journal of Vibroengineering 14(2): 559 – 571.
Shafiqur Rahman, M. & Golam Kibria, K.M. 2014. Investigation of vibration and ride
characteristics of a five degrees of freedom vehicle suspension system. Procedia
Engineering 90: 96 – 102.
244
Shaman, N.A., Shukor, M.S., Masdor, N.A., Sabullah, M.K., Shuor, M.Y. 2015. Testing the
normality of residuals on regression model for the growth of Moraxella sp.B on
Monobromoacetic acid. Bulletin of Environment Science and Management 3(1): 16 –
18.
Shamsaei, N. & McKelvey, S.A. 2014. Multiaxial life predictions in absence of any fatigue
properties. International Journal of Fatigue 67: 62 – 72.
Shamsaei, N., Gladskyi, M., Panasovskyi, K., Shukaev, S., Fatemi, A. 2010. Multiaxial fatigue
of titanium including step loading and load path alteration and sequence effects.
International Journal of Fatigue 32: 1862 – 1874.
Shao, X., Xu, N. & Liu, X. 2010. Investigation and application on the vertical vibration models
of the seated human body. Automotive Innovation 1(3): 263 – 271.
Sharma, P., Saluja, N., Saini, D., Parveen, S. 2013. Analysis of automotive passive suspension
system with Matlab program generation. International Journal of Advancement in
Technology 4(2): 115 – 119.
Sheridan, J.C., Steed, L., Ong., C. 2008. Analysis Without Anguish using SPSS Version 16.0
for Windows. U.K.: John Wiley & Sons.
Shevale, D.V. & Khaire, N.D. 2016. Review on failure analysis of helical compression spring.
International Journal of Science, Engineering and Technology Research 5(4): 892 –
898.
Shim, T.& Velusamy, P.C. 2007. Suspension design and dynamic analysis of a lightweight
vehicle. International Journal of Vehicle Design 43(1–4): 258 – 280.
Shirahatti, A., Prasad, P.S.S., Panzade, P., Kulkarni, M.M. 2008. Optimal design of passenger
car suspension for ride and road holding. Journal of the Brazilian Society of Mechanical
Sciences and Engineering 30(1): 66 – 76.
Sim, K., Lee, H., Yoon, J.W., Choi, C., Hwang, S.H. 2017. Effectiveness evaluation of hydro-
pneumatic and semi-active cab suspension for the improvement of ride comfort of
agricultural tractors. Journal of Terramechanics 69: 23 – 32.
Sivák, P. & Ostertagová, E. 2012. Evaluation of Fatigue Tests by Means of Mathematical
Statistics. Procedia Engineering 48: 636 – 642.
Skagerstrand, H. 2014. Shock Absorber Modelling, Master thesis. Chalmers University of
Technology.
Skrobachi, Z. 2007. Selected methods for the estimation of the logistic function parameters.
Maintenance and Reliability 3(35): 52 – 56.
Slavic, J., Mrsnik, M., Boltezar, M. 2013. Frequency-domain methods for a vibration-fatigue-
life estimation: application to real data. International Journal of Fatigue 47: 8 – 17.
Smyth, G. K. 2002. Nonlinear Regression. Encyclopedia of Environmetrics.
245
SoMat. 2002. eDAQ Manual. Champaign: SoMat Corporation.
Song, L.K., Fei, C.W., Wen, J., Bai, G.C. 2017. Multi-objective reliability-based design
optimization approach of complex structure with multi-mode failures. Aerospace
Science and Technology 64: 52 – 62.
Sonsino, C. 2007. Course of SN-curves especially in the high-cycle fatigue regime with regard
to component design and safety. International Journal of Fatigue 29(12): 2246 – 2258.
Stanzl-Tschegg, S. 2014. Very high cycle fatigue measuring techniques. International Journal
of Fatigue 60: 2 – 17.
Stojanovic, B., Milivojevic, M., Ivanovic, M., Milivojevic, N., Divac, D. 2013. Adaptive
system for dam behavior modeling based on linear regression and genetic algorithms.
Advances in Engineering Software 65: 182 – 190.
Sun, Q., Dui, H.N. & Fan, X.L. 2014. A statstically consistent fatigue damage model based on
Miner’s rule. International Journal of Fatigue 69: 16 – 21.
Tafti, M.A. 2013. Experimental High Cycle Fatigue Testing and Shape Optimization of
Turbine Blades, Master Thesis, University of Toronto.
Taghavifar, H. & Mardani, A. 2014. Application of artificial neural networks for the prediction
of traction performance parameters. Journal of the Saudi Society of Agricultural
Sciences 13: 35 – 43.
Taheri, S., Vincent, L. & Le-roux, J.C. 2013. A new model for fatigue damage accumulation
of austenitic stainless steel under variable amplitude loading. Procedia Engineering 66:
575 – 586.
Takahashi, K., Amano, T., Ando, K., Takahashi, F. 2011. Improvement of fatigue limit by shot
peening for high strength steel containing a crack-like surface defect. International
Journal of Structural Integrity 2(3): 281 – 292.
Takahashi, M.B., Roca, J.C., o Fernández Núnez, E.G. 2016. Optimization of artificial neural
network by genetic algorithm for describing viral production from uniform design data.
Process Biochemistry 51(3): 422 – 430.
Takechi, H. 1990. Recent developments in steel products for automotive application.
International Journal of Vehicle Design 11(1): 3 – 17.
Takemoto, M., Boissiere, L., Vital, J.M., Pellise, F., Perez-Grueso, F. Acaroglu, E.A., Alanya,
A., Obeid, I. 2016. Are sagittal spinopelvic radiographic parameters significantly
associated with quality of life of adult spinal deformity patients? Multivariate linear
regression analysed for pre-operative and short-term-post-operative health-related
quality of life. European Spine Journal 26(8): 1 – 11.
Tanaka, K., Mutoh, Y., Sakoda, S., Leadbeater, G. 1985. Fretting fatigue in 0.55c spring steel
and 0.45c carbon steel. Fatigue and Fracture of Engineering Materials and Structures
8(2): 129 – 142.
246
Tatak, M., Omheni, K., Aloui, A., Dammak, F., Haddar, M. 2014. Dynamic optimization
design of a cylindrical helical spring. Applied Acoustics 77: 178 – 183.
Tayyebi, A., Pijanowski, B. 2014. Modeling multiple land use changes using ANN, CART and
MARS: Comparing tradeoffs in goodness of fit and explanatory power of data mining
tools. International Journal of Applied Earth Observation and Geoinformation 28: 102
– 116.
Thies, P.R., Johanning, L., Harnois, V., Smith, H.C.M., Parish, D.N. 2014. Mooring line
fatigue damage evaluation for floating marine energy converters: field
measurements and prediction. Renewable Energy 63: 133 – 144.
Thite, A.N. 2012. Development of a refined quarter car model for the analysis of discomfort
due to vibration. Advances in Acoustics and Vibration 863061.
Thomas, D.J. 2012. Characterisation of cut-edges for improved automotive chassis and
suspension fatigue CAE life predictions. Journal of Failure Analysis and Prevention
12(4): 408 – 418.
Thomas, J.J., Bignonnet, A. & Perroud, G. 2005. Fatigue design and experimentations with
variable amplitude loadings in the automotive industry. Symposium on Fatigue testing
and analysis under variable amplitude loading conditions, ASTM STP 1439.
Tomaszewski, T., Sempruch, J., Piatkowski, T. 2014. Verification of selected models of
the size effect on high-cycle fatigue testing on mini specimens made of EN AW-
6063 aluminium alloy. Journal of Theoretical and Applied Mechanics 52(4): 888 –
894.
Tosun, E., Aydin, K., Bilgili, M. 2016. Comparison of linear regression and artificial neural
network model of a diesel engine fueled with biodiesel-alcohol mixtures. Alexandria
Engineering Journal 55: 3081 – 3089.
Uberti, S., Gadola, M., Chindamo, D., Romano, M., Galli, F. 2015. Design of a double
wishbone front suspension for an orchard-vineyard tractor: kinematic analysis. Journal
of Terramechanics 57: 23 – 39.
Unaune, D.R., Pawar, M.J., Mohite, S.S. 2011. Ride analysis of quarter vehicle model.
Proceedings of the 1st International Conference on Modern Trends in Industrial
Engineering, India.
Upadhyaya, Y.S. & Sridhara, B.K. 2012. Fatigue Life Prediction: a Continuum Damage
Mechanics and Fracture Mechanics Approach. Materials and Design 35: 220 – 224.
Valipour, M., Banihabib, M.E., Reza Behbahani, S.M. Comparison of the ARMA, ARIMA,
and the autoregressive artificial neural network models in forecasting the monthly
inflow of Dez dam reservoir. Journal of Hydrology 476: 433 – 441.
Valsange, P.S. 2012. Design of a helical coil spring, a review. International Journal of
Engineering Research and Applications 2(6): 513 – 522.
247
Vassilopoulos, A.P., Georgopoulos, E.F., Dionysopoulos, V. 2007. Artificial neural networks
in spectrum fatigue life prediction of composite materials. International Journal of
Fatigue 29: 20 – 29.
Velmurugan, P., Kumaraswamidhas, L.A. & Sankaranarayanasamy, K. 2012. Optimisation of
whole body vibration analysis for suspended cabin tractor semitrailer. Journal of
Vehicle Noise and Vibration 8(2): 152 – 165.
Vinogradov, A., Lazarev, A., Linderov, M., Weidner, A., Biermann, H. 2013. Kinematics of
deformation processes in high-alloy cast transformation-induced plasticity/twinning-
induced plasticity steels determined by acoustic emission and scanning electron
microscopy: Influence austenite stability on deformation mechanicsms. Acta
Materialia 61: 2434 – 2449.
Vishay Inc. 2005a. Strain Gauge Selection - Criteria, Procedures, Recommendations. Vishay
technical notes, TN-505-4.
Vishay Inc. 2005b. Errors due to Misalignment of Strain Gages. Vishay technical notes, TN-
511.
Vukelic, G. & Brcic, M. 2016. Failure analysis of a motor vehicle coil spring. Procedia
Structural Integrity 2: 2944 – 2950.
Wang, B., Jalil, N.A.A. & Voon, W.S. 2012a. Prediction of Rotary Machinery Degradation
Status based on Vibration Data using Back Propagation Neural Network. Scientific
Research and Essays 7: 1393 – 1404.
Wang, Y., Chen, H. & Zhou, H. 2016. A fatigue life estimation algorithm based on statistical
energy analysis in high-frequency random processes. International Journal of Fatigue
83: 221 – 229.
Welch, R.L., Ruffing, S.M., Venayagamoorthy, G.K. 2009. Comparison of feedforward and
feedback neural network architectures for short term wind speed prediction,
Proceedings of International Joint Conference on Neural Networks, U.S.A.
Williams, C. R., Lee, Y.-L. & Rilly, J. T. 2003. A practical method for statistical analysis of
strain–life fatigue data. International Journal of Fatigue 25(5): 427-436.
Wong, J.Y. 2001. Theory of Ground Vehicles, 3rd edi. U.S.A.: Wiley-Interscience.
Woo, S. 2017. Reliability Design of Mechanical System. U.S.A.: Springer Cham.
Woods, D.E. & Jawad, B.A. 1999. Numerical design of race car suspension parameters. SAE
Technical Series 1999-01-2257.
Wu, J., Luo, Z., Zhang, N., Zhang, Y. 2015. A new uncertainty analysis method and its
application in vehicle dynamics. Mechanical Systems and Signal Processing 50 – 51:
659 – 675.
248
Xia, Q. & Quail, F. 2016. Principles and validation of strain gauge shunt design for large
dynamic strain measurement. Sensors and Actuators A: Physical 241: 124 – 134.
Xie, J., Hong, T., Laing, T., Kang, C. 2015. On normality assumption in residual simulation
for probability load forecasting. IEEE Transactions on Smart Grid 8(3): 1046 – 1053.
Yan, X.L. & Wang, X.L., Zhang, Y.Y. 2014. Influence of roughness parameters skewness and
kurtosis on fatigue life under mixed elastohydrodynamic lubrication point contacts.
Journal of Tribology 136: 031503.
Yang, W.S., Doedhar, C., Streit, D.A., Kulakowski, B.T. 1995. Measurement of heavy vehicle
dynamic wheel forces using a bolt-on transducer. Road Transport Technology 4: 317 –
321.
Yang, X. & Xu, P. 2012. Road Load Analysis Techniques in Automotive Engineering, Metal
Fatigue Analysis Handbook: 1 – 60.
Yarmohamadi, H. & Berbyuk, V. 2013. Kinematic and dynamic analysis of a heavy truck with
individual front suspension. Vehicle System Dynamics 51(6): 877 – 905.
Ye, M., Yin, P., Lee, W.C. 2010. Location recommendation for location-based social
networks. Proceedings of the 18th SIGSPATIAL International Conference on
Advances in Geographic Information Systems, U.S.A: 458 – 461.
Ye, X.W., Su, Y.H. & Han, J.P. 2014. A state-of-the-art review on fatigue life assessment of
steel bridges. Mathematical Problems in Engineering 956473.
Yildiz, B.S & Lekesiz, H. 2017. Fatigue-based structural optimisation of vehicle components.
International Journal of Vehicle Design 73(1/2/3): 54 – 62.
Yu, W., Zhang, X., Guo, K., Karimi, H.R., Ma, F., Zheng, F. 2013. Adaptive real-time
estimation on road disturbances properties considering load variation via vehicle
vertical dynamics. Mathematical Problems in Engineering 283528.
Yu, Y., Peelamedu, S.M., Naganathan, N.G., Dukkipati, R.V. 2001. Automotive vehicle engine
mounting systems: a survey. Transactions of the ASME 123: 186 – 194.
Yunoh, M.F.M, Abdullah, S., Saad, M.H.M., Nopiah, Z.M., Nuawi, M.Z., Ariffin, A. 2018.
Classification of fatigue damaging segments using artificial neural network. Journal of
Mechanical Engineering SI 5(3): 61 – 72.
Zadeh, M.R., Amin, S., Khalili, D., Singh, V.P. 2010. Daily outflow prediction by multi layer
perceptron with logistic sigmoid and tangent sigmoid activation functions. Water
Resource Management 24: 2673 – 2688.
Zakaria, K.A., Abdullah, S., Ghazali, M.J., Nuawi M.Z., Beden, S.M., Nopiah, S.M. 2014.
Fatigue strain signal behaviour and damage assessment of an engine mount bracket.
Journal of Scientific and Industrial Research 73: 112 – 116.
249
Závodská, D., Guagliano, M., Bokuvka, O., Trško, L. 2016. Fatigue resistance of low alloy
steel after shot peening. Materialstoday:Proceedings 3: 1220 – 1225.
Zhang, J., Li, W., Wang, H., Song, Q., Lu, L., Wang, W., Liu, Z. 2016. A comparison of the
effects of traditional shot peening and micro-shot peening on the scuffing resistance of
carburized and quenched gear steel. Wear 368 – 369: 253 – 257.
Zhang, W., Suo, C., Wang, Q. 2008. A novel sensor system for measuring wheel loads of
vehicles on highways. Sensor 8: 7671 – 7689.
Zhang, Z., Zhang, W., Zhai, Z., Chen, Q. 2007. Evaluation of various turbulence models in
predicting airflow and turbulence in enclosed environments by CFD: Part-2:
comparison experimental data from literature. HVAC & R Research 13(6): 1 – 25.
Zhao, L.H., Zheng, S.L., Feng, J.Z., Zhou, H.F., Xing, Y.F. 2014. Fatigue assessment of rear
axle under service loading histories considering the strengthening and damaging effects
of loads below fatigue limit. International Journal of Automotive Technology 15(5):
843 – 852.
Zhao, X. & Schindler, C. 2014. Evaluation of whole-body vibration exposure experienced by
operators of a compact wheel loader according to ISO 2631-1:1997 and ISO 2631-5:
2004. International Journal of Industrial Ergonomics 44: 840 – 850.
Zheng, E. & Zhou, X. 2014. Modeling and simulation of flexible slider-crank mechanism with
clearance for a closed high speed press system. Mechanism and Machine Theory 74: 10
– 30.
Zhu, J.H., Zhang, W.H. & Xia, L. 2015. Topology optimization in aircraft and aerospace
structures design. Archives of Computational Methods in Engineering 23(4): 595 – 622.
Zhu, S.P., Lei, Q., Huang, H.Z., Yang, Y.J., Peng, W. 2016. Mean stress effect correction in
strain-energy based fatigue life prediction of metals. International Journal of Damage
Mechanics 0(0): 1 – 23.
Zhu, Y., Wang, Y. & Huang, Y. 2014. Failure analysis of a helical compression spring for a
heavy vehicle’s suspension system. Case Studies in Engineering Failure Analysis 2:
169 – 173.
Zielesny, A. 2014. From Curve Fitting to Machine Learning: An Illustrative Guide to Scientific
Data Analysis and Computational Intelligence. Germany: Springer-Verlag.
Zonfrillo, G. & Rossi, A. 2012. A fuzzy approach for drawing S-N curves in the finite life
region. Fatigue and Fracture of Engineering Materials and Structures 34: 868 – 876.
250
APPENDIX A
SPRING FAILURE REPORTED BY THEAA
A.1 Reported coil spring breakdowns
Source: https://www.theaa.com/driving-advice/service-repair/coil-springs-breaking
251
APPENDIX B
MESH SENSITIVITY
(a) (b)
(c) (d)
(e)
B.1 Element sizes: (a) 3.0, (b) 2.5, (c) 2.0, (d) 1.5, (e) 1.0 mm
252
(a) (b)
(c) (d)
(e)
B.2 Von Mises stress for various element sizes: (a) 3.0, (b) 2.5, (c) 2.0, (d) 1.5, (e) 1.0 mm
B.3 Mesh sensitivity results
Mesh Size (mm) Von Mises Stress (MPa) Required Time (minute)
3.0 1010 < 1
2.5 1002 < 1
2.0 990 2
1.5 988 4
1.0 987 25
253
APPENDIX C
OPTIMISED MULTILAYER ANN ARCHITECTURE
C.1 ANN architecture for optimised Morrow vibration-life predictions
C.2 ANN architecture for SWT vibration-life predictions
Spring fatigue
life
ISO 2631-1
weighted
acceleration
Characteristic
frequency
Input layer Second hidden layer First hidden layer Third hidden layer
Second hidden
layer Input First hidden
layer
Spring
fatigue life
Characteristic
frequency
ISO 2631-1
weighted
acceleration
Third hidden
layer Output
254
C.3 ANN architecture for Coffin-Manson ISO 2631 vertical vibration prediction
C.4 ANN architecture for Morrow ISO 2631 vertical vibration prediction
ISO 2631-1
weighted
acceleration
Second hidden
layer Input First hidden
layer
Spring
fatigue life
Characteristic
frequency
Third hidden
layer Output
ISO 2631
weighted
acceleration
Second hidden
layer
Input First hidden
layer
Spring
fatigue life
Suspension
frequency
Third hidden
layer Output
255
C.5 ANN architecture for SWT ISO 2631 vertical vibration prediction
ISO 2631
weighted
acceleration
Spring
fatigue life
Suspension
frequency
Second hidden
layer Input
First hidden
layer Third hidden
layer
Output
256
APPENDIX D
ARTIFICIAL ROAD SIMULATED TIME HISTORIES
D.1 Spring force time histories of road class “B” for various spring stiffness: (a) k14, (b) k16,
(c) k18, (d) k20, (e) k22, (f) k24, (g) k26, (h) k28, (i) k30, (j) k32
257
D.2 Spring force time histories of road class “C” for various spring stiffness: (a) k14, (b) k16,
(c) k18, (d) k20, (e) k22, (f) k24, (g) k26, (h) k28, (i) k30, (j) k32
258
D.3 Spring force time histories of road class “D” for (a) k14, (b) k16, (c) k18, (d) k20, (e) k22,
(f) k24, (g) k26, (h) k28, (i) k30, (j) k32
259
D.4 Vehicle acceleration time histories of road class “A” for (a) k14, (b) k16, (c) k18, (d) k20,
(e) k22, (f) k24, (g) k26, (h) k28, (i) k30, (j) k32
260
D.5 Vehicle acceleration time histories of road class “B” for (a) k14, (b) k16, (c) k18, (d) k20,
(e) k22, (f) k24, (g) k26, (h) k28, (i) k30, (j) k32
261
D.6 Vehicle acceleration time histories of road class “C” for (a) k14, (b) k16, (c) k18, (d) k20,
(e) k22, (f) k24, (g) k26, (h) k28, (i) k30, (j) k32
262
D.7 Vehicle acceleration time histories of road class “D” for (a) k14, (b) k16, (c) k18, (d) k20,
(e) k22, (f) k24, (g) k26, (h) k28, (i) k30, (j) k32
263
APPENDIX E
MEASURED ROAD SIMULATED TIME HISTORY
E.1 Spring force time histories of highway road for (a) k14, (b) k16, (c) k18, (d) k20, (e) k22,
(f) k24, (g) k26, (h) k28, (i) k30, (j) k32
264
E.2 Spring force time histories of campus road for (a) k14, (b) k16, (c) k18, (d) k20, (e) k22,
(f) k24, (g) k26, (h) k28, (i) k30, (j) k32
265
E.3 Spring force time histories of hill road for (a) k14, (b) k16, (c) k18, (d) k20, (e) k22, (f) k24,
(g) k26, (h) k28, (i) k30, (j) k32
266
E.4 Spring force time histories of residential road for (a) k14, (b) k16, (c) k18, (d) k20, (e) k22,
(f) k24, (g) k26, (h) k28, (i) k30, (j) k32
267
E.5 Simulated acceleration time histories of road class “D” for (a) k14, (b) k16, (c) k18, (d) k20,
(e) k22, (f) k24, (g) k26, (h) k28, (i) k30, (j) k32
268
E.6 Simulated acceleration time histories of road class “D” for (a) k14, (b) k16, (c) k18, (d) k20,
(e) k22, (f) k24, (g) k26, (h) k28, (i) k30, (j) k32
269
E.7 Simulated acceleration time histories of road class “D” for (a) k14, (b) k16, (c) k18, (d) k20,
(e) k22, (f) k24, (g) k26, (h) k28, (i) k30, (j) k32
270
E.8 Simulated acceleration time histories of road class “D” for (a) k14, (b) k16, (c) k18, (d) k20,
(e) k22, (f) k24, (g) k26, (h) k28, (i) k30, (j) k32
271
APPENDIX F
FATIGUE LIFE AND VIBRATION DATA
F.1 Predicted fatigue life and ISO 2631 weighted acceleration from road Class “B”
Suspension
frequency (Hz)
Weighted
Acceleration
(m/s2)
Coffin-Manson
fatigue life
(blocks to
failure)
Morrow fatigue
life (blocks to
failure)
SWT fatigue life
(blocks to
failure)
1.00 0.3085 3.14 × 108 3.65 × 108 2.23 × 108
1.06 0.3109 2.44 × 107 2.79 × 107 1.69 × 107
1.13 0.3133 8.20 × 106 9.20 × 106 5.88 × 106
1.20 0.3160 1.87 × 106 2.12 × 106 1.40 × 106
1.25 0.3191 9.35 × 105 1.07 × 106 5.85 × 105
1.30 0.3224 4.42 × 105 5.14 × 105 3.47 × 105
1.36 0.3262 2.34 × 105 2.80 × 105 1.93 × 105
1.41 0.3303 1.00 × 105 1.25 × 105 9.17 × 104
1.45 0.3346 6.45 × 104 8.33 × 104 5.78 × 104
1.50 0.3392 4.78 × 104 5.00 × 104 3.79 × 104
F.2 Predicted fatigue life and ISO 2631 weighted acceleration from road Class “C”
Suspension
frequency (Hz)
Weighted
Acceleration
(m/s2)
Coffin-Manson
fatigue life
(blocks to
failure)
Morrow fatigue
life (blocks to
failure)
SWT fatigue life
(blocks to
failure)
1.00 0.6158 2.97 × 105 2.57 × 105 2.38 × 105
1.06 0.6201 3.72 × 104 3.41 × 104 3.26 × 104
1.13 0.6244 1.64 × 104 1.55 × 104 1.50 × 104
1.20 0.6298 5.84 × 103 5.75 × 103 5.67 × 103
1.25 0.6360 3.66 × 103 3.64 × 103 3.60 × 103
1.30 0.6433 2.30 × 103 2.30 × 103 2.28 × 103
1.36 0.6507 1.62 × 103 1.60 × 103 1.55 × 103
1.41 0.6587 1.08 × 103 1.06 × 103 1.05 × 103
1.45 0.6672 8.70 × 102 8.54 × 102 8.43 × 102
1.50 0.6771 6.89 × 102 6.98 × 102 6.89 × 102
F.3 Predicted fatigue life and ISO 2631 weighted acceleration from road Class “D”
Suspension
frequency (Hz)
Weighted
Acceleration
(m/s2)
Coffin-Manson
fatigue life
(blocks to
failure)
Morrow fatigue
life (blocks to
failure)
SWT fatigue life
(blocks to
failure)
1.00 1.234 2.48 × 103 2.48 × 103 2.47 × 103
1.06 1.241 7.95 × 102 8.05 × 102 8.05 × 102
1.13 1.251 5.11 × 102 5.15 × 102 5.17 × 102
1.20 1.262 2.86 × 102 2.86 × 102 2.87 × 102
1.25 1.274 2.19 × 102 2.18 × 102 2.18 × 102
1.30 1.288 1.68 × 102 1.67 × 102 1.67 × 102
1.36 1.303 1.37 × 102 1.36 × 102 1.35 × 102
1.41 1.319 1.28 × 102 1.06 × 102 1.06 × 102
1.45 1.335 9.31 × 101 9.26 × 101 9.21 × 101
1.50 1.356 8.15 × 101 8.11 × 101 8.11 × 101
272
F.4 Predicted fatigue life and ISO 2631 weighted acceleration from campus road
Suspension
frequency (Hz)
Weighted
vertical
acceleration
(m/s2)
Coffin-Manson
fatigue life
(blocks to
failure)
Morrow fatigue
life (blocks to
failure)
SWT fatigue life
(blocks to
failure)
1.00 0.4816 9.74 × 105 2.17 × 107 8.33 × 107
1.06 0.5350 8.77 × 105 8.62 × 105 1.02 × 107
1.13 0.5631 3.72 × 105 1.04 × 105 4.13 × 106
1.20 0.5861 1.38 × 105 2.26 × 105 5.41 × 104
1.25 0.6068 1.03 × 105 8.87 × 103 8.20 × 104
1.30 0.6249 7.85 × 104 6.49 × 103 1.30 × 104
1.36 0.6413 6.52 × 104 5.18 × 103 9.43 × 103
1.41 0.6563 5.30 × 104 4.03 × 103 5.62 × 103
1.45 0.6701 5.23 × 104 3.80 × 103 5.21 × 103
1.50 0.6819 5.26 × 103 3.66 × 103 4.93 × 103
F.5 Predicted fatigue life and ISO 2631 weighted acceleration from hilly road
Suspension
frequency (Hz)
Weighted
vertical
acceleration
(m/s2)
Coffin-Manson
fatigue life
(blocks to
failure)
Morrow fatigue
life (blocks to
failure)
SWT fatigue life
(blocks to
failure)
1.00 0.5391 2.92 × 105 9.92 × 105 1.29 × 105
1.06 0.5916 5.17 × 104 2.65 × 105 3.49 × 104
1.13 0.6202 3.01 × 104 6.03 × 104 2.28 × 104
1.20 0.6451 1.58 × 104 2.44 × 104 1.35 × 104
1.25 0.6678 1.19 × 104 1.59 × 104 1.07 × 104
1.30 0.6906 9.19 × 103 1.24 × 104 8.63 × 103
1.36 0.7115 7.75 × 103 1.01 × 104 7.40 × 103
1.41 0.7300 6.24 × 103 8.07 × 103 6.13 × 103
1.45 0.7489 5.75 × 103 7.35 × 103 5.71 × 103
1.50 0.7657 5.36 × 103 6.77 × 103 5.38 × 103
F.6 Predicted fatigue life and ISO 2631 weighted acceleration from residential road
Suspension
frequency (Hz)
Weighted
vertical
acceleration
(m/s2)
Coffin-Manson
fatigue life
(blocks to
failure)
Morrow fatigue
life (blocks to
failure)
SWT fatigue life
(blocks to
failure)
1.00 0.2740 2.93 × 107 3.50 × 107 4.63 × 107
1.06 0.2981 3.93 × 106 7.99 × 107 7.75 × 106
1.13 0.3122 1.72 × 106 3.30 × 107 4.18 × 106
1.20 0.3250 4.82 × 105 1.05 × 107 1.17 × 106
1.25 0.3374 2.37 × 105 9.82 × 105 5.92 × 105
1.30 0.3503 1.17 × 105 3.40 × 105 2.93 × 105
1.36 0.3630 7.35 × 104 1.87 × 105 1.81 × 105
1.41 0.3759 4.56 × 104 1.14 × 105 1.17 × 105
1.45 0.3888 3.56 × 104 8.77 × 105 9.17 × 104
1.50 0.4013 2.82 × 104 6.95 × 104 7.30 × 104
273
F.7 Predicted fatigue life and ISO 2631 weighted acceleration from rural road
Suspension
frequency (Hz)
Weighted
vertical
acceleration
(m/s2)
Coffin-Manson
fatigue life
(blocks to
failure)
Morrow fatigue
life (blocks to
failure)
SWT fatigue life
(blocks to
failure)
1.00 0.5040 2.98 × 105 1.04 × 106 8.47 × 105
1.06 0.5740 5.73 × 104 2.71 × 105 1.25 × 105
1.13 0.5919 2.56 × 104 5.55 × 104 9.43 × 104
1.20 0.6098 2.37 × 104 5.29 × 104 1.37 × 104
1.25 0.6619 9.55 × 103 1.24 × 104 4.12 × 103
1.30 0.6818 8.18 × 103 1.02 × 104 1.28 × 103
1.36 0.6990 7.05 × 103 8.71 × 103 1.05 × 103
1.41 0.7130 6.03 × 103 7.29 × 103 7.37 × 102
1.45 0.7236 5.86 × 103 6.98 × 103 6.97 × 102
1.50 0.7310 5.78 × 103 6.80 × 103 4.92 × 103
274
APPENDIX G
MSE FOR SINGLE LAYER ANN
G.1 MSE of trained neural network with single hidden layer for Coffin-Manson
Number of Neurons MSE
1 0.4490
2 0.1461
3 0.2691
4 0.1132
5 0.1418
6 0.1715
7 0.0885
8 0.1537
9 0.2114
10 0.2551
G.2 MSE of trained neural network with single hidden layer for Morrow
Number of Neurons MSE
1 0.1430
2 0.2127
3 0.1230
4 0.1430
5 0.1372
6 0.3389
7 0.0905
8 0.1175
9 0.0827
10 0.1315
G.3 MSE of trained neural network with single hidden layer for SWT
Number of Neurons MSE
1 0.1769
2 0.5395
3 0.1818
4 0.2106
5 0.3576
6 1.2619
7 0.6108
8 0.2437
9 0.2241
10 0.2404
276
Continued…
(i) (j)
H.1 MSE of trained Morrow vibration-durability first hidden layer ANN with various number of neurons:
(a) 1, (b) 2, (c) 3, (d) 4, (e) 5, (f) 6, (g) 7, (h) 8, (i) 9, (j) 10
(a) (b)
(c) (d)
(e) (f)
Continue…
277
Continued…
(g) (h)
(i) (j)
H.2 MSE of trained SWT vibration-durability first hidden layer ANN with various number of neurons: (a) 1,
(b) 2, (c) 3, (d) 4, (e) 5, (f) 6, (g) 7, (h) 8, (i) 9, (j) 10
278
APPENDIX I
DATA FOR ANN ERROR HISTOGRAM
I.1 Error histogram data for single
layer Coffin-Manson ANN
I.2 Error histogram data for single
layer Morrow ANN
Data
points
Residuals Data
points Residuals
1 0.0192 1 0.1201
2 -0.0388 2 -0.1902
3 -0.1407 3 -0.0824
4 -0.7894 4 -0.1845
5 0.0201 5 -0.7051
6 0.0595 6 -0.2369
7 0.3660 7 -0.2758
8 0.3417 8 -0.3243
9 0.2170 9 -0.0489
10 0.0957 10 0.3004
11 -0.4284 11 0.4710
12 -0.2765 12 0.0951
13 -0.1489 13 -0.0838
14 -0.0388 14 -0.0340
I.3 Error histogram data for single layer SWT ANN
Data points Residuals
1 -0.7940
2 -0.6770
3 -0.0884
4 0.3144
5 -0.6802
6 0.0617
7 0.3868
8 0.3802
9 0.4321
10 0.2567
11 -0.1106
12 0.1764
13 0.4479
14 0.1560
279
I.4 Error histogram data for two
layers Coffin-Manson ANN
I.5 Error histogram data for two
layers Morrow ANN
Data
points
Residuals Data
points Residuals
1 -0.1149 1 -0.3812
2 -0.0495 2 0.0582
3 -0.2646 3 0.1013
4 0.0872 4 0.0028
5 -0.2610 5 -0.2983
6 0.1381 6 -0.0192
7 -0.0997 7 0.0901
8 -0.0533 8 0.1498
9 -0.0183 9 0.0040
10 -0.1237 10 0.1288
11 -0.2680 11 0.3236
12 0.1802 12 -0.1451
13 0.0961 13 -0.2092
14 0.1621 14 0.0345
I.6 Error histogram data for two layers SWT ANN
Data points Residuals
1 -0.7118
2 -0.4118
3 0.1854
4 0.2901
5 0.3072
6 0.2381
7 0.3133
8 -0.0183
9 0.0016
10 0.2048
11 0.2160
12 0.3070
13 0.0825
14 0.2992
280
I.7 Error histogram data for single
layer Coffin-Manson ANN
I.8 Error histogram data for single
layer Morrow ANN
Data
points
Residuals Data
points Residuals
1 -0.4387 1 0.1189
2 -0.1059 2 0.0182
3 0.0275 3 -0.0416
4 -0.0249 4 0.0188
5 0.0068 5 -0.1115
6 0.1061 6 0.0260
7 0.1081 7 -0.0408
8 0.0461 8 0.1753
9 -0.1317 9 0.0243
10 -0.0556 10 -0.0747
11 -0.3277 11 0.0539
12 -0.2755 12 0.0748
13 0.2534 13 0.2934
14 0.0243 14 0.0213
I.9 Error histogram data for single layer SWT ANN
Data points Residuals
1 -0.2752
2 -0.3334
3 -0.4692
4 0.5236
5 -0.0881
6 0.3157
7 0.3435
8 0.3515
9 0.1823
10 0.0268
11 -0.1038
12 -0.1502
13 -0.2055
14 0.1169
281
APPENDIX J
CURVE FITTING FOR ANN
(a) (b)
(c) (d)
J.1 Curve fitting of trained Morrow vibration- fatigue ANN model with single hidden layer for
various datasets: (a) all, (b) training, (c) validation, (d) ANN test
(a) (b)
Continue…
282
Continued…
(c) (d)
J.2 Curve fitting of trained SWT vibration- fatigue ANN model with single hidden layer for
various datasets: (a) all, (b) training, (c) validation, (d) ANN test
(a) (b)
(c) (d)
J.3 Curve fitting of trained Morrow vibration- fatigue ANN model with two hidden layers for
various datasets: (a) all, (b) training, (c) validation, (d) ANN test
283
(a) (b)
(c) (d)
J.4 Curve fitting of trained SWT vibration- fatigue ANN model with two hidden layers for
various datasets: (a) all, (b) training, (c) validation, (d) ANN test
(a) (b)
Continue…
284
Continued…
(c) (d)
J.5 Curve fitting of trained Morrow vibration- fatigue ANN model with three hidden layers for
various datasets: (a) all, (b) training, (c) validation, (d) ANN test
(a) (b)
(c) (d)
J.6 Curve fitting of trained SWT vibration- fatigue ANN model with three hidden layers for
various datasets: (a) all, (b) training, (c) validation, (d) ANN test
285
APPENDIX K
WEIGHTS AND BIASES FOR ANN
K.1 Weights of input layer for trained Coffin-Manson ANN
WIi,j WIi1 WIi2
WI1j -2.2843 0.6346
WI2j 4.0773 -1.1071
WI3j 0.0954 -2.1144
K.2 Weights of first hidden layer for trained Coffin-Manson ANN
W2i,j W2i1 W2i2 W2i3
W21j -0.4810 0.4841 0.5051
W22j 0.8562 1.1494 0.7742
W23j -0.4734 -0.5309 -0.6807
W24j 0.02823 -0.4950 0.3696
W25j -0.7406 0.8840 -0.4637
W26j 0.02554 -0.1258 -0.4063
W27j 1.0695 -0.7864 0.7782
K.3 Weights of second hidden layer for trained Coffin-Manson ANN
WOi,j WOi1 WOi2 WOi3 WOi4 WOi5 WOi6 WOi7
WO1j 0.0593 -1.1444 -0.1317
WO2j 1.4178 -1.3051 -0.5066 -1.0216 -0.9601 -0.5815 1.4264
K.4 Biases for trained Coffin-Manson ANN
Bi,j Bi1 Bi2
B1j -5.335 1.0025
B2j 0.2126 -1.1008
B3j -2.1111 -0.2645
B4j -0.6274
B5j -0.7800
B6j -0.4187
B7j 1.2197
K.5 Weights of input layer for trained Morrow ANN
WIij WIi1 WIi2
WI1j -1.6465 3.8170
WI2j -1.6633 -3.3086
WI3j -4.0989 -0.1864
WI4j -2.3028 4.1623
WI5j 2.2595 -3.3822
WI6j 1.942 1.9043
WI7j -0.9415 -4.3768
WI8j -2.2554 6.5517
WI9j 2.5697 -3.3467
286
K.6 Weights of first hidden layer for trained Morrow ANN
W1ij WIi1 W1i2 W1i3 W1i4 W1i5 W1i6 W1i7 W1i8 W1i9
W11j 0.9527 0.3362 0.5431 0.3380 0.0772 0.5009 0.6537 0.7318 0.2898
W12j 0.9181 0.3032 0.2514 0.5188 0.7691 0.8235 0.1485 0.5031 0.5786
W13j 0.6756 0.6931 0.4256 0.4646 0.7967 0.7466 0.8653 0.2610 0.4493
W14j 1.0211 1.0168 0.7203 0.2622 0.6877 0.7144 0.5856 0.5634 0.4557
W15j 0.5575 0.2048 0.4135 0.4912 0.6442 0.6290 0.4135 0.9610 0.0955
W16j 0.4864 0.7796 0.0204 0.7247 0.6901 0.5001 0.6143 0.2616 0.8713
W17j 0.9278 0.4753 0.9148 0.5578 0.4243 0.2143 0.0206 0.0110 0.2143
K.7 Weights of second hidden layer for trained Morrow ANN
W2ij W2i1 W2i2 W2i3 W2i4 W2i5 W2i6 W2i7 W2i8 W2i9
W21j 0.9332 0.3810 0.3960 0.9501 0.8211 0.3950 0.3167 0.0042 0.1904
W22j 0.4814 0.4829 0.7698 0.9784 0.1981 0.8531 0.6161 0.1351 0.9220
W23j 0.1289 0.6535 0.4215 0.5788 0.9862 0.5374 0.4559 0.2884 0.8550
W24j 0.4440 0.4556 0.1042 0.8450 0.6675 0.7242 0.5206
W25j 0.4704 0.5964 0.8376 0.6293 0.5061 0.3583 0.4967
W26j 0.9775 0.5690 0.0173 0.1312 0.1987 0.3375 0.0442
K.8 Weights of third hidden layer for trained Morrow ANN
W3ij W3i1 W3i2 W3i3 W3i4 W3i5 W3i6 W3i7 W3i8 W3i9
W31j 0.0939 0.6367 0.3584 0.2294 0.3612 0.6167 0.7527 0.4798 0.9156
W32j 0.9825 0.0487 1.2130 0.2337 0.0096 0.1908 1.0180
W33j 0.1490 0.1000 0.5146
K.9 Bias for trained Morrow ANN
Bij Bi1 Bi2
B1j 1.9779 1.4289
B2j 3.7479 -0.6773
B3j 1.5133 2.1153
B4j -1.6643 -0.2913
B5j 0.6234 0.2567
B6j 1.4991 0.1413
B7j -2.521 -0.9489
B8j 3.6138
B9j 4.2704
K.10 Weights of input layer for trained SWT ANN
WIij WIi1 WIi2
WI1j 2.6847 2.2466
WI2j -0.3294 -2.5249
WI3j -0.2808 0.9913
WI4j 1.1835 0.4237
287
K.11 Weights of first hidden layer for trained SWT ANN
W1ij W1i1 W1i2 W1i3 W1i4
W11j 0.6588 -0.2927 -0.7191 0.1382
W12j 0.6696 0.2984 0.753 -0.7009
W13j 0.8316 0.3809 -0.8153 0.0307
W14j 0.6308 -0.0236 0.3668 0.5656
W15j -0.127 -0.0644 0.7814 -0.7755
W16j -0.198 -0.2983 -0.2633 -0.8177
W17j 0.4486 0.2628 0.4799 0.3082
W18j 1.0334 0.9514 0.6678 -0.6896
W19j 0.1513 -0.2279 -0.1566 0.4566
W110j -0.817 -0.827 -0.0604 -0.2201
K.12 Weights of second hidden layer for trained SWT ANN: Part 1
W2ij W2i1 W2i2 W2i3 W2i4
W21j 0.0857 0.2385 0.6906 -0.7706
W22j 0.1564 -0.6227 0.3521 -0.3331
W23j 0.1215 -0.7911 0.5255 0.8570
W24j -0.4073 0.8366 -0.5305 -0.1828
W25j -0.4545 -0.8891 -0.841 0.1596
W26j -0.191 -0.5893 -0.5577 0.8493
W27j 0.2866 -0.7959 0.0447 0.6888
W28j -0.8247 -0.6177 0.2179 -0.5031
K.13 Weights of second hidden layer for trained SWT ANN: Part 2
W2ij W2i1 W2i2 W2i3 W2i4 W2i5 W2i6 W2i7
W21j 0.3651 -0.3194 0.2035 0.7496 -0.1942 0.2132 0.0255
W22j -0.2167 -0.0419 -0.1911 0.2413 0.2265 -0.9164 -0.0828
W23j 0.0575 -0.3303 0.8003 -0.3519 -0.2428 0.1589 0.0624
W24j -0.0277 -0.7334 0.4025 0.0536 0.0498 -0.8695 0.0366
W25j -0.5373 -0.1226 -0.1859 -0.5005 -0.3357 0.942 -0.0289
W26j 0.3062 -0.714 0.5921 -0.4777 0.9051 -0.8714 0.5242
W27j 0.529 -0.3133 -0.3739 0.237 0.425 -0.716 0.8697
W28j 0.8158 0.0935 0.1499 0.7025 -0.2701 0.0045 -0.8567
W2i8 W2i9 W2i10 W2i8 W2i9 W2i10
W21j -0.822 0.6717 0.2898 W28j -0.097 -0.7101 0.7193
W22j -1.0831 0.2174 -0.3881
W23j 0.0995 0.7386 -0.3866
W24j 0.7115 -0.0015 -0.2078
W25j 0.1601 -0.2957 -0.2848
W26j 0.534 -0.3738 0.8251
W27j 0.7373 -0.2374 -0.5994
K.14 Weights of third hidden layer for trained SWT ANN
Wij Wi1 Wi2 Wi3 Wi4 Wi5 Wi6 Wi7 Wi8
W1j -0.8055 0.5143 0.277 -0.6102
W2j 0.866 0.9021 -0.2945 0.8391 -0.4476 0.2799 -0.064 -0.4499
W3j -0.0019 -0.5066 -0.4311 0.5389 -0.5407 0.1726 -0.2061 0.8255
Wi9 Wi10
W2j -0.9228 -0.3769
288
K.15 Biases for trained SWT ANN
Bij Bi1 Bi2
B1j -2.9837 1.1637
B2j -1.1411 0.2212
B3j 0.1448 -0.2593
B4j 0.7207 0.9473
B5j -0.5979
B6j -0.7493
B7j -0.5716
B8j 0.3258
B9j -0.1556
B10j 0.4510
289
APPENDIX L
LIST OF PUBLICATIONS
1. Kong, Y.S., Abdullah, S., Omar, M.Z., Haris, S.M. 2016. Failure assessment of a
leaf spring eye design under various load cases. Engineering Failure Analysis 63:
146 – 159. (ISI Q2, Published)
2. Kong, Y.S., Abdullah, S., Omar, M.Z., Haris, S.M. 2016. Topological and
topographical optimization of automotive spring lower seat. Latin America Journal
of Solids and Structures 13(7): 1388 – 1405. (ISI Q2, Published)
3. Kong, Y.S., Abdullah, S., Omar, M.Z., Haris, S.M. 2016. Side force analysis of
suspension strut under various load cases. Jurnal Teknologi 78(6-10): 85 – 90.
(Scopus, Published)
4. Kong, Y.S., D. Schramm, Abdullah, S., Omar, M.Z., Haris, S.M. 2017. The
significance to establish a durability model for automotive ride. SAE Technical
Paper 2017-01-0347. (Scopus, Published)
5. Kong, Y.S., D. Schramm, Abdullah, S., Omar, M.Z., Haris, S.M., Bruckmann, T.
2017. Mission profiling of road data measurement for coil spring fatigue life.
Measurement 107: 99 – 110. (ISI Q2, Published)
6. Kong, Y.S., D. Schramm, Abdullah, S., Omar, M.Z., Haris, S.M. 2018. Vibration
fatigue analysis of carbon steel coil spring under various road excitations. Metals
8(8): 617 (ISI Q1, Published)
7. Kong, Y.S., D. Schramm, Abdullah, S., Omar, M.Z., Haris, S.M., Bruckmann, T.
2017. The need to generate a force time history towards life assessment of a coil
spring. Journal of Mechanical Engineering SI4 (5): 11 – 26. (Scopus, Published)
8. Kong, Y.S., D. Schramm, Abdullah, S., Omar, M.Z., Haris, S.M., Bruckmann, T.,
Kracht, F. 2018. Observing the durability effects of a Formula Student Electric Car
using acceleration and strain signals. Mobilität und digitale Transformation,
Springer Gabler. (Book Chapter, Published)
9. Kong, Y.S., D. Schramm, Abdullah, S., Omar, M.Z., Haris, S.M., Kracht, F. 2018.
Characterising spring durability for automotive ride using artificial neural network.
International Journal of Engineering and Technology 7: 47 - 53. (Scopus, Published)
10. Kong, Y.S., D. Schramm, Abdullah, S., Omar, M.Z., Haris, S.M. 2019.
Development of multi linear regression-based models for fatigue life evaluation of
automotive coil spring. Mechanical Systems and Signal Processing 118: 675 - 695
(ISI Q1, Published)
11. Kong, Y.S., D. Schramm, Abdullah, S., Omar, M.Z., Haris, S.M. 2018. Generation
of artificial road profile for spring durability analysis. Jurnal Kejuruteraan 30(2).
(ECSI, Published)
290
12. Kong, Y.S., D. Schramm, Abdullah, S., Omar, M.Z., Haris, S.M. 2018.
Optimisation of spring fatigue life prediction model for vehicle ride using hybrid
multi-layer perceptron artificial neural networks. Mechanical Systems and Signal
Processing 122: 597- 621 (ISI Q1, Published)
13. Kong, Y.S., D. Schramm, Abdullah, S., Omar, M.Z., Haris, S.M. 2018. Ride and
durability characterisation under road excitation in frequency domain.
Measurement. (ISI Q2, Minor corrections)
14. Kong, Y.S., D. Schramm, Abdullah, S., Omar, M.Z., Haris, S.M. 2018. Correlation
of uniaxial and multiaxial fatigue models for spring fatigue life assessment.
Experimental Techniques (ISI Q3, Major corrections)
15. Kong, Y.S., D. Schramm, Abdullah, S., Omar, M.Z., Haris, S.M. 2018. Design of
artificial neural network using particle swarm optimisation for spring durability.
Journal of Mechanical Science and Technology (ISI Q3, Under Review)
16. Kong, Y.S., D. Schramm, Abdullah, S., Omar, M.Z., Haris, S.M. 2018. Validation
of established predictive models for automotive ride and spring durability. Journal
of Mechanical Science and Technology (ISI Q3, Under Review)
17. Kong, Y.S., D. Schramm, Abdullah, S., Omar, M.Z., Haris, S.M. 2018. Evaluation
of regression-tree based durability models for spring fatigue life assessment.
ICMFM XI Porto (Scopus, Under Review)
18. Kong, Y.S., D. Schramm, Abdullah, S., Omar, M.Z., Haris, S.M. 2018. Determining
optimal suspension system parameters for spring fatigue life using design of
experiment. Mechanics and Industry (ISI Q4, Under Review)
19. Kong, Y.S., D. Schramm, Abdullah, S., Omar, M.Z., Haris, S.M. 2018. Evaluation
of automobile spring energy-based model for strrain signals generation. Metals (ISI
Q2, Submission)
20. Kong, Y.S., D. Schramm, Abdullah, S., Omar, M.Z., Haris, S.M. 2018. A
comparison of multiple linear regression and artificial neural network for
automobile vertical vibration predictions. Mechanical Systems and Signal
Processing (ISI Q1, Submission)
21. Kong, Y.S., Abdullah, S., Omar, M.Z., Haris, S.M. 2018. Kebolehtahanan terpecut
bagi komponen sistem ampaian dengan pelbagai ujaan jalan. (Book Chapter,
Submission)
22. Kong, Y.S., Abdullah, S., Omar, M.Z., Haris, S.M. 2018. Kaedah ujian jahitan bagi
kebolehtahanan komponen sistem ampaian kenderaan. (Book Chapter, Submission)
23. Kong, Y.S., D. Schramm, Abdullah, S., Omar, M.Z., Haris, S.M. 2018. K-Nearest
Neighbors Classification of Spring Strain Signal according to ISO 8608 using
Hilbert Huang Transform. Mechanical Systems and Signal Processing (ISI Q1,
Submission)
291
24. Kong, Y.S., Abdullah, S., Omar, M.Z., Haris, S.M. 2018. Prediction of automobile
spring fatigue life using support vector machine approach. (Scopus, Submission)
Diese Dissertation wird über DuEPublico, dem Dokumenten- und Publikationsserver derUniversität Duisburg-Essen, zur Verfügung gestellt und liegt auch als Print-Version vor.
DOI:URN:
10.17185/duepublico/70111urn:nbn:de:hbz:464-20190523-091034-1
Dieses Werk kann unter einer Creative Commons Namensnennung -Nicht kommerziell - Keine Bearbeitungen 4.0 Lizenz (CC BY-NC-ND 4.0) genutzt werden.