International Journal of Fatigue 33 (2011) 313–322

Contents lists available at ScienceDirect

International Journal of Fatigue

journal homepage: www.elsevier.com/locate/ijfatigue

A neural network approach to fatigue life prediction

João Carlos Figueira Pujol a,⇑, João Mário Andrade Pinto b
a
Cento de Desenvolvimento da Tecnologia Nuclear, Brazil
b
Universidade FUMEC, Belo Horizonte, Brazil

a r t i c l e i n f o a b s t r a c t

Article history: In this work, a novel approach to fatigue life prediction under step-stress conditions is introduced, where
Received 8 January 2010 the cumulative distribution function for the failure of components was implemented by means of a neu-
Received in revised form 31 August 2010 ral network. The model was fit to experimental data on the fatigue life of steel under step-stress condi-
Accepted 6 September 2010
tions. For comparison, a standard approach based on the lognormal distribution function was also
Available online 15 September 2010
implemented and fit to the same experimental data. Both models were optimized by evolutionary com-
putation, using a maximum likelihood estimator. The Kolmogorov–Smirnov test was applied to compare
Keywords:
the results of the new approach to those obtained with the lognormal distribution function.
Fatigue
Lognormal distribution
Ó 2010 Elsevier Ltd. All rights reserved.
Step-stress
Neural networks
Evolutionary computation

1. Introduction real-world situations where significant deviations from these dis-

tributions may occur. In this context, a novel approach is proposed
To predict failure by mechanical fatigue, the response of mate- where a statistical distribution is built by means of an artificial
rials to the several loads expected to occur during the lifetime of neural network.
components must be considered. As fatigue properties cannot be Artificial neural networks, or neural networks (NNs) for short,
deduced from static mechanical properties only, they must be di- are a class of computational tools inspired by the biological ner-
rectly measured by specifically designed mechanical experiments, vous system [7]. They consist of partially or fully interconnected
where a constant alternate stress S is cyclically applied to the simple processing units called neurons. The neuron is a nonlinear
material until failure. These experiments allow the construction unit that receives input signals from other units or from the envi-
of the well-known S–N curves, which express the lifetime of the ronment (the space of input data relevant to the problem at hand),
material as measured by the logarithm of the number of cycles yielding an output. The signals received by a neuron are modulated
N, plotted against the logarithm of S. However, in the majority of by real numbers called synaptic weights (or simply weights). By
the situations, mechanical components are submitted to more than adjusting the weights through a training process, NNs can learn
one level of stress during their lifetime. Such load variations make underlying relations from a given set of representative examples
it inadequate to use a single S–N curve for fatigue life prediction, to solve a particular problem instead of following a predefined
since these curves are built under constant stress conditions. set of rules. In feedforward networks, the neurons are usually ar-
Cumulative damage models [1], such as the cumulative exposure ranged in layers: input and output layers, which interact with
model [2–4], are commonly used in this situation. the environment, and one or more layers of hidden neurons, which
To complicate matters, random factors lead to great variability do not have contact with the environment. There is a definite order
in the results of fatigue tests, and a probabilistic approach is neces- of evaluation of the neurons, the network receives the input sig-
sary to account for the uncertainties present in the lifetime of nals, propagates them through all the layers, and returns signals
structures [1,5,6]. Usually, these probabilistic approaches are built to the environment through the output neurons (see Fig. 1). Feed-
over standard statistical distributions. However, there are many forward networks with a single hidden layer and a continuous non-
linear activation function are universal approximators, in that any
continuous function can be approximated to any degree of accu-
⇑ Corresponding author. Address: Rua Itapitocai 70 apt, 702-90820120 Porto racy if a sufficient number of hidden neurons is provided [8].
Alegre-RS, Brazil. Tel.: +55 31 30693237; fax: +55 31 3069 3257. Neural networks have been successfully applied to a number of
E-mail addresses: [email protected], [email protected] (J.C. Figueira Pujol), problems in many fields of science and engineering [9].
[email protected] (J.M. Andrade Pinto).
URLs: http://www.cdtn.br (J.C. Figueira Pujol), http://www.fumec.br
Neural networks have also been used for fatigue life prediction.
(J.M. Andrade Pinto). The most direct approach has been to use material properties and

0142-1123/$ - see front matter Ó 2010 Elsevier Ltd. All rights reserved.
doi:10.1016/j.ijfatigue.2010.09.003
314 J.C. Figueira Pujol, J.M. Andrade Pinto / International Journal of Fatigue 33 (2011) 313–322

Table 1
Chemical composition of the SAE 8620 steel [45].

Element % Min. % Max.

C 0.18 0.23
Si 0.15 0.30
Mn 0.70 0.90
Cr 0.40 0.60
Ni 0.40 0.70
Mo 0.15 0.25
P <0.03
S <0.04
Fig. 1. The neuron is a nonlinear unit that receives input signals from neurons in
Co <0.1
the previous layer or from the environment, and transfers output signals to neurons
Pb 0 0.15
in the next layer or back to the environment.
Cu <0.3
Al <0.1
U <0.1
experimental conditions as input data to a neural network, whose W 0 0.1
output is the number of cycles until failure. For example, Mathur
et al. [10] demonstrated the ability of neural networks to predict
fatigue life. In this case, the input to the neural net include: volume
fraction, tensile modulus, tensile strength, applied load parame- Table 2
Mechanical properties of the SAE 8620 steel with 95%
ters, probability of failure, statistical parameters of fatigue life, confidence level [45].
etc. The output is the logarithm of the number of fatigue life cycles.
Troudt and Merril [11] implemented a feedforward neural network Property Value

for diagnosis and prognosis purposes. The neural network is sup- Yield strength – r0.2 (MPa) 370 ± 10
posed to evaluate degradation on components under mechanical Tensile strength (MPa) 602 ± 24
Elongation (%) 21 ± 2
stress in real time to predict when they will eventually fail. Alter-
Reduction of area (%) 39 ± 1
natively, the neural network can be combined with other models to Breaking strength (MPa) 432 ± 19
estimate fatigue life. For example, Artymiak et. al [12] trained four Endurance limit (MPa) 194 ± 5
multilayered feedforward neural neural networks to build S–N
curves. Vassilopoulos et al. [13] proposed to predict fatigue life
by using a feedforward neural network to build constant life dia-
grams. Kang et al. [14] combined a feedforward neural network specimens were designed and manufactured according to sugges-
with the critical plane method to predict fatigue life. Another strat- tions by Cazaud [24] and the ASTM standards [25], to produce
egy has been to use neural networks to estimate the fatigue crack specimens whose geometry and smoothness do not interfere with
growth rate [15–17]. Neural networks have also been applied to the test results (see Fig. 2). Moreover, the experiments were carried
address the stochastic aspects of the fatigue phenomenon. For out under refrigeration provided by natural water. Each specimen
example, Janezic et al. [18] implemented a feedforward neural net- was submitted to three levels of stress, viz.: S1, S2 and S3. S1 was
work to estimate the parameters of the Weibull distribution. Sim- set to 258 MPa during 35,000 cycles and S2 = 238 MPa was applied
ilarly, Buc̆ar et al. [19] designed a probability density function for 65,000 cycles. S3 was set to 20 values selected at regular inter-
using a weighted sum of Weibull density functions modulated by vals of 10 MPa to cover a broad load range below the yield strength
real-valued coefficients, whose values are determined by a feedfor- of the material (see Table 2). S3 was applied until failure or until
ward neural network. The list of such applications is endless (see the machine reached 5  106 cycles, a condition that qualifies as
[20–22] for additional material on the subject). a type I censoring mechanism [26].
The next logical step would be to replace the standard statistical The test suite was planned according to the S–N–P curves of the
distributions with a probability distribution built from scratch. material (see Fig. 3). In preliminary experiments (not reported in
That is the proposal of the present work, to use a feedforward neu- the paper), five specimens were submitted to a step-stress condi-
ral network to compute the probability that a component submit- tion with S1 set to 258 MPa for 50,000 cycles, and S2 set to
ted to a specified level of stress fails until a specified number of 238 MPa for 100,000 cycles. A third level of stress was then applied
cycles. The numerical model was tested on fatigue data collected to each of the five specimens by setting S3 to 218 MPa, 198 MPa,
from step-stress experiments carried out on steel specimens. The 178 MPa, 158 MPa and 148 MPa. The specimens submitted to
new paradigm was compared to a standard statistical approach S3 = 158 MPa and 158 MPa failed around 1.0  106 cycles. Even
based on the lognormal distribution. with the great dispersion expected in fatigue test results, such load
This work is organized as follows: in Section 2 the experimental configuration would probably not allow censoring (an undesirable
setup is described, Section 3 describes the step-stress model, in simplification of the model). Therefore, the number of cycles for
Section 4 the maximum likelihood is discussed, in Section 5 the the application of the first and the second levels of stress were re-
optimization procedure is introduced, in Section 6 the approach duced to 35,000 cycles and 65,000 cycles, respectively.
using a standard statistical distribution function is described, in
Section 7 the approach based on the neural network is presented
and, in Sections 8 and 9, results comparing both methods are re-
ported and discussed. Finally, Section 10 concludes and provides
suggestions for further research.

2. Description of the fatigue experiments

Flex-rotational fatigue tests were carried out on SAE 8620 steel

specimens [23] (chemical composition and mechanical properties
of the material are shown in Tables 1 and 2, respectively). The test Fig. 2. Geometry of the specimens used in the fatigue tests [45].
 J.C. Figueira Pujol, J.M. Andrade Pinto / International Journal of Fatigue 33 (2011) 313–322 315

SNP curves
undamaged
350

330
1% 50% 99%
310

Stress amplitude (Mpa)

290

270 258

250 238

230 218

210
198

190 178

170 158

150

130
1000 10000 100000 1000000 10000000 100000000
Number of cycles (N)

Fig. 3. S–N–P curves for the SAE 8620 steel, with probability of failure of 1%, 50% and 99% [48].

The results of the fatigue tests are summarized in Table 3. Two where the equivalent time ni1 is the solution to:
hundred and thirty two specimens were used in the experiments,
FðDt i1 þ ni2 ; Si1 Þ ¼ Fðni1 ; Si Þ ð2Þ
resulting in 225 experiments with failure by fatigue, and 7 exper-
iments where censoring was introduced at 5  106. with
Dti1 ¼ t i1  ti2 ; t 0 ¼ n0 ¼ 0 ð3Þ

3. Step-stress model From Eq. (1), the corresponding probability density f0(t) of the life-
time of the specimens can be computed as:
To predict fatigue life under step-stress conditions, it is neces- @F 0 ðtÞ
sary to model the accumulated damage experienced by the mate- f0 ðtÞ ¼ ð4Þ
@t
rial. In other words, a cumulative damage model is required [1]. In
this work, the cumulative distribution function (CDF) of the num- In other words, F0(t) is conceptually defined by the cumulative
ber of cycles until failure was built by means of the widely used exposure model as the combination of the segments of the three
linear cumulative exposure model [2–4,27–37]. This model was distribution functions for constant stress. F0(t) is analytically differ-
adopted for its simplicity of implementation, and its basic assump- entiated in each segment to produce the corresponding probability
tions are: density f0(t).
To apply the cumulative damage exposure model to a specific
 The fatigue mechanisms are the same for any level of stress. case, the distribution function F(t, S) still has to be specified. In
 The cumulative damage is linear, i.e., the remaining life of a the case of metal components, the lifetime distribution is usually
component depends only on the current cumulative failed frac- modeled using standard distribution functions such as the lognor-
tion and on the level of stress, regardless of how the fraction mal, for example. These functions actually represent a family of
was accumulated. functions characterized by free parameters which must be ad-
 Should the level of stress be kept constant, the survivors will fail justed to fit the experimental data. In the current work, these
according to the CDF corresponding to that level of stress, but parameters were estimated by means of the maximum likelihood
starting at the previously accumulated failed fraction. method.

In this model, the CDF for the lifetime of a specimen submitted 4. Maximum likelihood method
to a step-stress load regime is built from the one under constant
stress. For example, Fig. 4 shows the three distribution functions The maximum likelihood method (ML) has the advantage of
for stress S1, S2 and S3, applied at times t0 = 0, t1 and t2, respectively. producing estimators whose distributions, under certain condi-
Initially, the specimens follow the CDF for S1 until time t1. Those tions of regularity, converge to the normal distribution as the size
specimens that have not failed yet, will follow the CDF for S2 until of the sample increases. Also, it allows censoring information to be
time t2. And, similarly, after time t2 the survivors will follow the incorporated [38]. From the sample of experimental points, the fol-
CDF for S3 until failure or censoring. This procedure makes it pos- lowing likelihood function is built:
sible to build a CDF, F0(t), for the lifetime of the specimens, as a " # " #
Ynf Ync
combination of the segments of the three distribution functions L¼ f0 ðt k Þ  ð1  F 0 ðt k ÞÞ ð5Þ
for constant stress. k¼1 k¼1
Mathematically, this model means that if the CDFs for S1, S2, and
where
S3 are F(t, S1), F(t, S2) and F(t, S3), respectively, then the CDF for the
step-stress pattern can be written as:
 nf is the number of failures observed in the sample;
F 0 ðtÞ ¼ Fðt  ti1 þ ni1 ; Si Þ; ti1 6 t 6 t i ; i ¼ 1; 2; 3 ð1Þ  nc is the number of censored observations in the sample;
316 J.C. Figueira Pujol, J.M. Andrade Pinto / International Journal of Fatigue 33 (2011) 313–322

Table 3
Experimental data. Columns labeled m indicate the trial number, columns labeled S indicate the stress applied to the specimen in the third step of the experiment (see description
in Section 2), and t is the number of cycles until failure or censoring. Experimental points with t = 5  106 indicate experiments that have been censored at this number of cycles.
Each row corresponds to a different test specimen.

m S t m S t m S t m S t
1 148 5  106 61 198 445,791 121 238 142,653 181 288 206,477
2 148 5  106 62 198 240,103 122 238 166,208 182 288 163,047
3 148 5  106 63 198 268,343 123 238 152,377 183 298 152,988
4 148 5  106 64 198 260,625 124 238 215,350 184 298 164,159
5 148 5  106 65 198 272,910 125 238 228,643 185 298 121,481
6 148 5  106 66 198 564,735 126 238 163,040 186 298 149,512
7 148 5  106 67 198 433,231 127 238 467,993 187 298 129,177
8 148 645,666 68 198 366,461 128 238 258,595 188 298 120,788
9 148 3,696,070 69 198 375,056 129 238 487,602 189 298 107,364
10 148 2,895,030 70 198 644,623 130 238 478,663 190 298 125,676
11 148 3,448,194 71 198 925,586 131 238 229,460 191 298 129,993
12 158 1,107,437 72 198 1,009,238 132 248 244,031 192 298 126,412
13 158 1,761,868 73 198 972,100 133 248 239,407 193 298 156,552
14 158 1,498,143 74 198 1,218,589 134 248 154,292 194 298 142,482
15 158 1,745,894 75 198 1,436,466 135 248 288,401 195 298 146,422
16 158 1,298,980 76 198 966,459 136 248 229,549 196 298 128,181
17 158 1,635,304 77 198 916,017 137 248 335,877 197 308 130,182
18 158 1,999,236 78 198 933,758 138 258 139,769 198 318 162,707
19 158 1,692,219 79 198 1,048,277 139 258 164,322 199 318 117,808
20 158 1,307,811 80 198 1,092,700 140 258 184,018 200 318 130,086
21 158 1,243,014 81 208 560,982 141 258 152,176 201 318 189,829
22 158 710,201 82 208 1,029,295 142 258 128,712 202 318 114,203
23 158 1,821,045 83 208 845,353 143 258 151,535 203 318 148,712
24 158 1,545,562 84 208 462,717 144 258 232,154 204 318 107,238
25 158 1,691,022 85 208 691,987 145 258 231,113 205 318 122,397
26 158 1,902,363 86 208 453,310 146 258 166,334 206 318 181,694
27 168 3,742,946 87 218 374,035 147 258 174,258 207 318 123,566
28 168 976,895 88 218 326,242 148 258 192,930 208 318 121,719
29 168 1,699,593 89 218 204,764 149 258 192,930 209 318 124,857
30 168 708,623 90 218 243,306 150 258 486,218 210 318 117,750
31 168 1,151,776 91 218 222,596 151 258 235,209 211 318 127,576
32 168 2,356,311 92 218 277,006 152 258 260,596 212 318 113,965
33 168 779,788 93 218 222,144 153 258 247,523 213 318 119,675
34 178 673,479 94 218 380,956 154 258 157,640 214 318 118,574
35 178 1,500,728 95 218 503,890 155 268 352,117 215 328 138,044
36 178 253,841 96 218 238,335 156 268 226,310 216 338 107,677
37 178 1,126,587 97 218 318,598 157 268 231,301 217 338 133,770
38 178 280,493 98 218 334,537 158 268 170,774 218 338 140,230
39 178 848,011 99 218 781,316 159 268 236,910 219 338 119,029
40 178 551,128 100 218 593,163 160 268 199,035 220 338 129,288
41 178 254,679 101 218 802,863 161 278 140,406 221 338 110,424
42 178 361,823 102 218 1,052,459 162 278 128,431 222 338 116,634
43 178 456,656 103 218 1,016,447 163 278 127,216 223 338 121,211
44 178 1,271,267 104 218 808,388 164 278 136,492 224 338 108,553
45 178 702,117 105 218 458,517 165 278 151,323 225 338 119,207
46 178 1,706,990 106 218 873,946 166 278 133,177 226 338 132,840
47 178 1,696,530 107 218 787,923 167 278 113,253 227 338 112,781
48 178 1,716,303 108 218 879,073 168 278 169,013 228 338 123,291
49 178 1,634,609 109 228 541,674 169 278 167,633 229 338 122,761
50 178 1,724,898 110 228 308,427 170 278 152,740 230 338 113,062
51 178 1,687,304 111 238 334,986 171 278 150,285 231 338 118,773
52 178 1,413,503 112 238 339,404 172 278 114,803 232 338 119,388
53 188 1,282,886 113 238 187,427 173 278 151,806 – – –
54 188 874,291 114 238 263,557 174 278 195,735 – – –
55 188 1,211,409 115 238 172,913 175 278 222,941 – – –
56 188 1,414,851 116 238 145,354 176 278 284,093 – – –
57 188 1,544,262 117 238 170,160 177 278 153,727 – – –
58 188 570,450 118 238 187,501 178 288 171,066 – – –
59 198 328,536 119 238 162,415 179 288 143,103 – – –
60 198 288,377 120 238 303,807 180 288 160,953 – – –

 k is the index of the specimen; function with singular points. To address this task a procedure
 tk is the time of failure or censoring of the kth specimen mea- based on evolutionary computation was used.
sured in cycles;
 F0 and f0 are given by Eqs. (1) and (4), respectively. 5. Evolutionary computation

The idea is to find the values of the free parameters of the dis- Evolutionary computation is a class of global optimization tech-
tribution function F(t, S) that maximize L. Unfortunately, it is not niques also inspired by the natural evolution of species [39]. Evo-
that simple. Although easy to implement, L is rarely a simple con- lutionary algorithms are computational tools based on the
vex function. On the contrary, it is often a complex, multimodal collective learning process of a population of individuals where
 J.C. Figueira Pujol, J.M. Andrade Pinto / International Journal of Fatigue 33 (2011) 313–322 317

i1 
X  x
Sj
ni1 ¼ ðt j  t j1 Þ  ; i ¼ 2; 3 ð9Þ
j¼1
Si

Introducing Eqs. (7) and (8) in Eq. (1) results in:

   x 
1 Si
F 0 ðtÞ ¼ 1  U   log ðt  t i1 þ ni1 Þ  ð10Þ
r A
where ti1 6 t 6 ti, i = 1, 2, 3 (t3 = 1).
Introducing Eqs. (10) and (4) in Eq. (5) results in the likelihood
function of the free parameters A, x and r, whose optimal values
are estimated by maximizing L.

7. CDF for the step-stress experiment based on the neural

network
Fig. 4. Distribution functions for three levels of constant stress, S1, S2, S3, and the
equivalent times ni.
The idea was to build a probability distribution function based
on a neural network, in order to replace the lognormal distribution
function discussed in the previous section, and then use the max-
imum likelihood method to adjust the weights of the neural net-
each individual is a trial solution to a given problem. In the present work by maximizing the likelihood function L.
task, each individual represents a distribution function with initial
random values assigned to the free parameters. This population is 7.1. The neural network and the probability distribution function
then evolved by means of the repeated application of evolutionary
operators. These operators are biased to favor individuals with bet- A feedforward neural network with the following architecture
ter performance in the task at hand (in the current case, the max- was implemented:
imization of the likelihood function). As a result, a new population
with better performance is created at each generation to explore  Two input neurons: one to represent the number of cycles t and
the space of trial solutions. Ideally, this process will eventually another to represent the stress S applied to the specimen.
converge to a solution. The ability of evolutionary algorithms to  A hidden layer with nh neurons.
search large and complex spaces has been demonstrated in a wide  An output layer with a single neuron.
range of tasks [40]. As a consequence, they seem to be natural can-
didates to optimize the likelihood function. 1
The logistic function gðuÞ ¼ 1þexpðuÞ was used as an activation
For that purpose, differential evolution [41], a very popular evo- function where u is the total input to the neuron. This results in
lutionary algorithm, was implemented. The idea was to carry out the following neural network output:
an initial global search with this procedure and then to perform !! !
X
nh X
2
an additional local search to refine the result. The global search netðx1 ; x2 Þ ¼ g vj  g ðwij  xi Þ þ bhj þ bo ð11Þ
was carried out with a population of 20 individuals. Subsequently, j¼1 i¼1
the global search was complemented by a local search carried out
with the Newton–Raphson method. where

 wij is the weight of the connection between the ith input neuron
6. CDF for the step-stress experiment based on the lognormal and the jth hidden neuron;
distribution  vj is the weight of the connection between the jth hidden neu-
ron and the output neuron;
In the current work, the lognormal distribution was used to  bhj is the bias of the jth hidden neuron, and bo is the bias of the
implement the distribution function F(t, S) as follows: output neuron. The bias terms work as activation thresholds,
  and are also adjusted similarly to the connection weights;
1  xi is the ith input data.
Fðt; SÞ ¼ 1  U   ðlogðtÞ  lðSÞÞ ð6Þ
r
By making x1R = t and x2 = S in net(t, S), a probability function can
The deterministic part of the model was introduced via the t
netðt; SÞdt
  l(S) given by the widely used inverse power
location parameter be defined as R 1
0
. However, it is common practice in the
netðt; SÞdt
x 0
law: lðSÞ ¼ log AS [3,42,26]. Resulting in: implementation of neural networks, to normalize the input data
   x  to avoid saturation of the activation function and to deal with
1 S the input data uniformly. As a consequence, S was divided by SMAX
Fðt; SÞ ¼ 1  U   log t  ð7Þ
r A t
and the superior limit of integration was replaced with ðtþTMAXÞ

where U is the standard normal distribution function [43], and A, x (values for SMAX and TMAX are given in the discussion of the
and r are positive real numbers whose values depend on the spec- experimental results). In the case of the number of cycles, the nor-
imen and the experimental procedure. malization procedure had the additional benefit of avoiding
Applying Eq. (2) recursively leads to: numerical overflow in the evaluation of the improper integral. As
a consequence, the distribution function and the probability den-
 x
Si1 sity were defined as follows:
ni1 ¼ ðDt i1 þ ni2 Þ  ; i ¼ 2; 3 ð8Þ
Si R ðtþTMAXÞ
t  
S
0 net t; SMAX dt
Fðt; SÞ ¼ R1   ð12Þ
It is straightforward to demonstrate that this expression is S
net t; SMAX dt
0
equivalent to the following non-recursive expression:
318 J.C. Figueira Pujol, J.M. Andrade Pinto / International Journal of Fatigue 33 (2011) 313–322

@Fðt; SÞ
f ðt; SÞ ¼ ð13Þ
@t

7.2. Building the neural network

Similarly to Section 6, one can build F0, f0 by using Eqs. (1)–(4).

To do that, it is necessary to compute the equivalent times of the
step-stress experiments. In the case of the lognormal distribution
function, this can be performed analytically by Eq. (9). With the
neural network, this must be carried out numerically.
In many applications of neural networks, a set of input patterns
is presented to the network and, for each pattern, the output of the
network is computed and compared to an expected response or
target value yielding an error. An iterative procedure is then exe-
cuted to minimize this error. However, in the current application, Fig. 6. Lognormal probability density for stress S = 198–238.
there are no target values. Moreover, the architecture of the neural
network is not fixed in advance. Consequently, there are no opti-
mal values to be found. To cope with this situation, the value of
the maximum likelihood function computed with the optimized
lognormal distribution function was taken as the optimization cri-
terion. In other words, a neural network was built to produce a va-
lue of the maximum likelihood function better than that obtained
with the lognormal distribution function (this is further discussed
in the experimental results).

8. Experimental results

8.1. Lognormal-based CDF

The first step was to maximize the likelihood function based on

the lognormal distribution function. The optimization procedure
described in Section 5 converged to the optimal values A = Fig. 7. Lognormal probability density for stress S = 248–288.

1.751  103, x = 6.3 and r = 0.358, corresponding to log(L) =

2.945  103. This was the target value to be beaten with the neu-
ral-network-based approach. Plots of the corresponding optimal
f(t, S) and F(t, S) are shown in Figs. 5–12.

8.2. Neural-network-based CDF

As described in Section 7, the neural network used to imple-

ment the new model has two input neurons and an output neuron.
But the number of neurons in the hidden layer still has to be spec-
ified. Unfortunately, there is no simple, general, reliable and auto-
matic method to do that. The hidden layer was built by trial and
error in the course of the training process. The weights and biases
were initialized within the range [1, +1]. Initially, nh was set to 5.
It was soon realized that such a small architecture was not rich en-
ough to learn the task. The hidden layer was then increased to 10
Fig. 8. Lognormal probability density for stress S = 298–338.

Fig. 5. Lognormal probability density for stress S = 148–188. Fig. 9. Lognormal probability function for stress S = 148–188.
 J.C. Figueira Pujol, J.M. Andrade Pinto / International Journal of Fatigue 33 (2011) 313–322 319

and eventually to 15 neurons. The search was interrupted when

the value of the likelihood function log(L) = 2.922  103 was
achieved, a result better than that obtained with the lognormal dis-
tribution function.
The solution obtained can be summarized by introducing the
optimized weights and biases in Eq. (11), resulting in the following
function:

netðx1 ; x2 Þ ¼ gð356 þ 1:08  gð135  69:5  x2 þ 2:23  x1 Þ

þ 1080  gð29:8  4:21  x2  2:43  x1 Þ
þ 68  gð145  0:504  x2 þ 14:8  x1 Þ
þ 8:61  gð0:342 þ 3:54  x2  299  x1 Þ
 109  gð33:9  35:3  x2  466  x1 Þ
Fig. 10. Lognormal probability function for stress S = 198–238.
 9:76  gð7:62  9:69  x2  20  x1 Þ
 37:3  gð24:5 þ 68:1  x2  4:14  x1 Þ
 2590  gð37:1  147  x2  57:2  x1 Þ
þ 46:7  gð202 þ 6:45  x2 þ 1:36  x1 Þ
 698  gð6:26  5:94  x2 þ 1:33  x1 Þ
þ 1:63  gð25:4  0:803  x2 þ 16:3  x1 Þ
þ 5:11  gð5:64  1:42  x2  4:96  x1 Þ
þ 2:74  gð25:5  10:8  x2  4:03  x1 Þ
þ 1:20  gð0:929 þ 2:90  x2  7:37  x1 Þ
 2:06  gð177 þ 321  x2  218  x1 ÞÞ ð14Þ
To build the probability distribution function as described by
Eq. (12), TMAX = 5  106 cycles and SMAX = 338 MPa were used.
They represent the censoring value of the number of cycles and
the maximum stress applied to the specimens, respectively. Plots
of the corresponding optimal f(t, S) and F(t, S) are shown in Figs.
Fig. 11. Lognormal probability function for stress S = 248–288. 13–20.

Fig. 12. Lognormal probability function for stress S = 298–338. Fig. 14. Neural-network-based probability density for stress S = 198–238.

Fig. 13. Neural-network-based probability density for stress S = 148–188. Fig. 15. Neural-network-based probability density for stress S = 248–288.
320 J.C. Figueira Pujol, J.M. Andrade Pinto / International Journal of Fatigue 33 (2011) 313–322

Fig. 16. Neural-network-based probability density for stress S = 298–338. Fig. 20. Neural-network-based probability function for stress S = 298–338.

Table 4
Equivalent time n2 as recursively defined
by Eq. (2), and used to build F0 and f0. S is
the third stress level applied to each
specimen. n1 = 66,500.

S Equivalent time
148 249,850
158 241,252
168 225,478
178 210,666
188 196,504
198 187,995
208 172,941
218 157,482
228 144,064
238 131,500
Fig. 17. Neural-network-based probability function for stress S = 148–188.
248 119,052
258 106,637
268 94,162
278 81,482
288 68,437
298 55,135
308 42,304
318 30,618
328 19,833
338 13,150

Similarly to the lognormal distribution, Eqs. (1)–(4) were used

to build F0 and f0. However, in the current case, the equivalent
times as defined by Eq. (2) had to be computed numerically. That
means for each different step-stress pattern Eq. (2) had to be iter-
atively solved for the equivalent times. The results are summarized
in Table 4.
Fig. 18. Neural-network-based probability function for stress S = 198–238.

9. Discussion of results

By comparing Figs. 5–14 one can see that the neural-network-

based probability densities show a steep increase followed by a
slight deecrease at t  0.2  106 cycles, as compared to the lognor-
mal ones. These features are more noticeable for S = 148, 158 and
168 MPa. Also, the neural-network-based probability densities
are more skewed to the left than the lognormal ones. This behavior
is a consequence of the fact that the neural-network-based ap-
proach has more degrees of freedom to fit the experimental data.
In other words, it is a consequence of the sample of experimental
points. However, it is important to avoid overfitting. For this rea-
son, the optimization of the neural network was interrupted as
soon as the value of L obtained with this model was superior to
Fig. 19. Neural-network-based probability function for stress S = 248–288. that achieved with the lognormal distribution.
 J.C. Figueira Pujol, J.M. Andrade Pinto / International Journal of Fatigue 33 (2011) 313–322 321

Table 5 that, on the current experimental data, the new model was supe-
Summary of the Kolmogorov–Smirnov test with significance level a = 5%. S is the last rior to the standard one.
level of stress applied to the specimens. N is the number of experimental points
obtained with each value of S. Column 3 gives the critical values of the K–S test.
Moreover, it must be pointed out that the neural-based CDF
Columns 4 and 5 represent the maximum distance of the cumulative histogram of the introduced in this work is not limited to or by the linear cumulative
experimental data to the hypothesized distribution functions. LND is the distance to exposure model. In addition, in the new approach, the deterministic
the lognormal distribution function, and NND is the distance to the distribution dependence of fatigue life with the stress is not hypothesized, it is
function given by the neural network.
built in the neural network.
S N CV LND NND Experimental results of fatigue tests on 232 SAE 8620 steel spec-
148 11 0.624 0.230 0.106 imens were also reported. In the experiments, each specimen was
158 15 0.338 0.593 0.341 submitted to three levels of stress on a flex-rotational machine
168 7 0.483 0.272 0.228 with a type I censoring mechanism.
178 19 0.301 0.217 0.196
The results of the new paradigm are promising, but to estimate
188 6 0.519 0.262 0.291
198 22 0.281 0.190 0.175 fatigue life is a difficult task, and the new approach should be eval-
208 6 0.519 0.271 0.149 uated more thoroughly by performing new experiments. Also, it
218 22 0.281 0.264 0.139 should be compared with other standard statistical distributions,
228 2 0.842 0.099 0.136 e.g. the Weibull distribution.
238 21 0.287 0.291 0.092
248 6 0.519 0.170 0.330
258 17 0.318 0.196 0.119 Acknowledgments
268 6 0.519 0.315 0.379
278 17 0.318 0.301 0.253
288 5 0.563 0.227 0.215
The authors would like to thank the Fundação de Amparo à
298 14 0.349 0.246 0.278 Pesquisa do Estado de Minas Gerais (http://www.fapemig.br) for
308 1 0.975 0.554 0.523 the financial support to carry out this research.
318 17 0.318 0.196 0.158
328 1 0.975 0.166 0.100
338 17 0.318 0.107 0.159 References

[1] Collins J. Failure of materials in mechanical design. 2nd ed. John Wiley; 1993.
[2] Nelson W. Accelerated life testing – step-stress models and data analysis. IEEE
To quantify the adequacy of the models to the experimental Trans Reliab 1980;R-29(2):103–8.
[3] Nelson W. Accelerated life testing, statistical models, test plans and data
data, both distribution functions were submitted to the Kolmogo- analysis. New York: J. Wiley; 1990.
rov–Smirnov (K–S) test [43]. The test was carried out to each final [4] Tang L, Sun Y, Goh T, Ong H. Analysis of step-stress accelerated-life-test data: a
stress S applied to the specimens. That means, the experimental new approach. IEEE Trans Reliab 1996;45(1):69–74.
[5] Sobczyk K, Spencer jr B. Random fatigue: from data to theory. Academic Press;
points were grouped according to S, and the test was then applied
1992.
to each distribution function F0 built from each of the hypothesized [6] Johnson W, Hillberry B, editors. Probabilistic aspects of life prediction. ASTM
distribution functions. The results are summarized in Table 5. Int; 2004.
The neural-network-based distribution function can be rejected [7] Haykin S. Neural networks a comprehensive foundation. 2nd ed. Prentice Hall;
1999.
in just one case of the K–S test, whereas the lognormal distribution [8] Hornik K. Multilayer feedforward networks are universal approximators.
function can be rejected in two of them. So it seems that, in the Neural Networks 1989;2:359–66.
current experiments, the results support the conclusion that the [9] Fiesler E, Beale R, editors. Handbook of neural computation. Oxford University
Press; 1997.
neural-network-based approach is superior to the lognormal- [10] Mathur S, Gope P, Sharma J. Prediction of fatigue lives of composites material
based one. It may be objected that the new approach has more de- by artificial neural network. In Proceedings of the SEM 2007 annual conference
grees of freedom to fit the experimental data, and is expected to be and exposition, Springfield, Massachusetts, USA; June 2007 [Paper 260].
[11] Troudt T, Merrill W. A real time neural net estimator of fatigue life. In
computationally more expensive than the lognormal-based one. Proceedings international joint conference on neural networks (SPIE), vol. 2,
That is true. On the other hand, with the new approach it is not pages 59–64, San Diego, CA, USA, June 1990.
necessary to assume a life-stress relationship such as the semi- [12] Artymiak P, Bukowsky L, Feliks J, MarberHaus S, Henner Z. Determination of S–
N curves with the application of neural networks. J Fatigue Fract Eng Mater
empirical inverse power law. The relationship is embedded in the Struct 1999(22):23–728.
neural network. Although frequently used, the inverse power law [13] Vassilopoulos A, Georgopoulos E, Dionysopoulos V. Artificial neural networks
is only one of many options [3,44,45]. in spectrum fatigue life prediction of composite materials. Int J Fatigue
2007:20–9.
It must be pointed out that, instead of trial and error, a more
[14] Kang J, Choi B, Lee H, Kim J, Kim K. Neural network application in fatigue
elaborate approach based on constructive or destructive methods damage analysis under multiaxial random loadings. Int J Fatigue
can be used to build the neural network [46,47]. This strategy is 2006:132–40.
worth investigating, as it may result in a robust and fully automatic [15] Mohantya J, Vermaa B, Parhib D, Rayb P. Prediction of residual fatigue life
under interspersed mixed-mode (I and II) overloads by artificial neural
procedure for fatigue life prediction based on neural networks. network. J Fatigue Fract Eng Mater Struct 2009;32(12):1020–31.
[16] Lee D, Hong S, Cho S, Joo W. A study on fatigue damage modeling using neural
networks. J Mech Sci Technol 2005;19(7):1393–404.
[17] Fathi A, Aghakouchak A. Prediction of fatigue crack growth rate in welded
10. Conclusion tubular joints using neural network. Int J Fatigue 2007(29):261–75.
[18] Janezĭc̆ MM, Klemenc J, Fajdiga M. A neural-network approach to describe the
In this work, a novel approach to fatigue life prediction based on scatter of cyclic stressstrain curves. J Mater Des 2007(31):38–448.
[19] Buc̆ar T, Nagode M, Fajdiga M. A neural network approach to describing the
neural networks has been presented. A feedforward neural net- scatter of sn curves. Int J Fatigue 2006(28):311–23.
work was used to build a probability distribution. This distribution [20] Bhadeshia H. Neural networks in materials science. ISI Int J
was then combined with the linear cumulative exposure model to 1999;39(10):66–979.
[21] Malinov S, Sha W. Software products for modelling and simulation in materials
estimate fatigue life under step-stress conditions.
science. J Comput Materi Sci 2003(28):179–98.
The new model was compared to a standard approach based on [22] Sha W, Edwards K. The use of artificial neural networks in materials science
the lognormal distribution. Both models were fit to experimental based research. J Mater Des 2007(28):1747–52.
data by means of the maximum likelihood method and evolution- [23] SAE International, SAE handbook. Number J404 in chemical compositions of
SAE alloy steels, 1994, p. 1.12–3.
ary computation. A Kolmogorov–Smirnov test was then performed [24] Cazaud R. La fadiga de los metales. 3rd French ed. Madrid, Espanha: Aguilar;
to check the goodness of fit of both distributions. It has been shown 1957.
322 J.C. Figueira Pujol, J.M. Andrade Pinto / International Journal of Fatigue 33 (2011) 313–322

[25] American Society for Testing and Materials – ASTM. Standard terminology [37] Abdel-Hamid A, AL-Hussaini E. Estimation in step-stress accelerated life tests
relating to fatigue and fracture testing, vol. 03.01. Annual Book of ASTM for the exponentiated exponential distribution with type-I censoring. J
Standards; 2001 [ASTM E 1823–96]. Comput Statist Data Anal 2009(53):1328–38.
[26] Yang G. Life cycle reliability engineering. John Wiley & Sons; 2007. [38] Cordeiro G. Introdução à teoria de verossimilhança. In: Course during the 10o
[27] Nelson W. Residuals and their analyses for accelerated life tests with step and Simpósio Nacional de Probabilidade e Estatı́stica, Associaç ao Brasileira de
varying stress. IEEE Trans Reliab 2008;57(2):360–8. Estatı́stica, Rio de Janeiro; 1992.
[28] Miller R, Nelson W. Optimum simple step-stress plans for accelerated life [39] Jong K. Evolutionary computation a unified approach. MIT Press; 2006.
testing. IEEE Trans Reliab 1983;R-2(1):59–65. [40] Baeck T, Fogel D, Michalewicz Z. Handbook of evolutionary. Oxford University
[29] Xiong C, Milliken G. Step-stress life-testing with random stress-change times Press; 1997.
for exponential data. IEEE Trans Reliab 1999;48(2):141–8. [41] Price K, Storn R, Lampinen J. Differential evolution a practical approach to
[30] Yin X, Seng B. Some aspects of accelerated life testing by progressive stress. global optimization. Springer-Verlag; 2005.
IEEE Trans Reliab 1987;R-36(1):150–5. [42] Allegri G, Zhang F. On the inverse power laws for accelerated random fatigue
[31] Bai D, Kim M, Lee S. Optimum simple step-stress accelerated life tests with testing. Int J Fatigue 2008(30):967–77.
censoring. IEEE Trans Reliab 1989;38(5):528–32. [43] Kreyszig E. Introductory mathematical statistics. John Wiley & Sons; 1970.
[32] McSorley E, Lu J, Li C. Performance of parameter-estimates in step-stress [44] Meeker W, Escobar L. Statistical methods for reliability data. John Wiley &
accelerated life-tests with various sample-sizes. IEEE Trans Reliab Sons; 1998.
2002;51(3):271–7. [45] Pinto J. Modelagem de testes acelerados com esforço aplicado em nı´veis em
[33] Mettas A, Vassiliou P. Modeling and analysis of time-dependent stress um estudo de fadiga mecânica. PhD thesis, Instituto de Pesquisas Energéticas e
accelerated life data. In: Proceedings of the annual reliability and Nucleares, São Paulo, Brazil; 2004.
maintainability symposium, Seattle WA, USA; January 2002. p. 343–8. [46] Śmieja F. Neural network constructive algorithms: trading generalization for
[34] Xiong C. Inferences on a simple step stress model with type-II censored learning efficiency? Circ Syst Signal Process 1993;12(2):331–74.
exponential data. IEEE Trans Reliab 1998;47(2):142–6. [47] Reed R. Pruning algorithms: a survey. IEEE Trans Neural Networks
[35] Teng S, Yeo K. A least-squares approach to analyzing life-stress relationship in 1993;4(5):40–747.
step-stress accelerated life tests. IEEE Trans Reliab 2002;51(2):177–82. [48] Alvarenga Jr A. Acúmulo de danos por fadiga no aço SAE 8620. MSc
[36] Dharmadhikari A, Rahman M. A model for step-stress accelerated life testing. dissertation, Departamento de Engenharia Mecânica da Universidade
Naval Res Logis 2003;50:841–68. Católica de Minas Gerais, Belo Horizonte, Brazil; 2001.