International Journal of Fatigue 158 (2022) 106730

Contents lists available at ScienceDirect

International Journal of Fatigue

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

Gaussian process regression based remaining fatigue life prediction for

metallic materials under two-step loading
Jingjing Gao a, Cunjun Wang a, Zili Xu a, *, Jun Wang b, Song Yan b, Zhen Wang b
a
State Key Laboratory for Strength and Vibration of Mechanical Structures, Xi’an Jiaotong University, Xi’an 710049, China
b
Science and Technology on Liquid Rocket Engine Laboratory, Xi’an 710100, China

A R T I C L E I N F O A B S T R A C T

Keywords: Remaining fatigue life prediction is vital for engineering structures to ensure safety and reliability. It can be more
Gaussian Process Regression challenging when the structures suffer variable amplitude loadings because of the complex, non-uniform of the
Fatigue life prediction fatigue damage accumulation and inherent noise, uncertainty in the data. To further tackle the problem, the
Machine learning
Gaussian process regression (GPR) is introduced, which can simultaneously estimate the output value and
Two-step loading
Remaining fatigue life
quantify the associated uncertainty. Therefore, a GPR-based remaining fatigue life prediction method is proposed
to predict the remaining fatigue life for metallic materials under two-step loading in this paper. The proposed
method is comprehensively evaluated on the dataset containing 12 materials, 328 samples in total. The proposed
method achieves the lowest mean square error (MSE), mean absolute percentage error (MAPE), residual standard
deviation (RSD) values and the highest correlation coefficient (CC) values among the six machine learning
methods and the two model-driven methods. Those results indicate that the proposed method can achieve greater
accuracy and reliability in remaining life prediction under two-step loading, which illustrate the effectiveness of
the proposed method as a data-driven method in the field of remaining life prediction.

n1 n2
1. Introduction + =1 (1)
N1 N2
Fatigue, as one of the major issues among in engineering structures,
where n1 and n2 are experienced cycles under the two-step loadings,
can lead to fracture and failure in vulnerable component of structures.
respectively. N1 and N2 are the fatigue life corresponding to the stress
Structural components are usually subjected to variable amplitude
level.
loading in practice. The fatigue damage prediction of variable amplitude
However, the PM model takes no account of the load sequence and
loading is still a challenging puzzle, which remains to be resolved. Thus,
the load interaction effects due to its empirical nature. In order to
the remaining fatigue life prediction plays a vital role in design and
overcome the drawbacks of the PM model, a wide range of nonlinear
condition monitoring of the components.
damage accumulation models are established. Marco and Starkey pro­
In practice, fatigue damage accumulation of structural components is
posed the first nonlinear damage accumulation model [3]. Then lots of
a process of which the various material performance degenerates with
accumulation models have been done based on the Marco and Starkey
the number of loading cycles in essence related with environmental
model by taking the damage curve, the S-N curve, isodamage lines,
factors, material defects. To model the fatigue damage accumulation
continuum damage mechanics, material degradation and energy into
process and predict the remaining fatigue life, amount of theories and
account [4]. In this process, additional material parameters have been
methods have been proposed. In this respect, Palmgren-Miner (PM)
added into the models, like σ u ultimate tensile strength, Ne the minimum
model [1] based on the linear accumulation assumption used for two-
fatigue life at the fatigue limit on the S-N curve.
step loading is shown in Eq. (1). Due to simplicity and convenience,
Among various nonlinear models, the Ye-Wang (YW) model [5] is
the PM model, as a standard for steel structures design [2], has been
usually regard as a useful alternative model to the PM model. The YW
widely utilized in engineering practice.
model is built with the natural logarithm on the basis of the exhaustion
of static toughness of the materials to predict the remaining fatigue life

* Corresponding author.
E-mail address: [email protected] (Z. Xu).

https://doi.org/10.1016/j.ijfatigue.2022.106730
Received 7 October 2021; Received in revised form 16 December 2021; Accepted 7 January 2022
Available online 11 January 2022
0142-1123/© 2022 Elsevier Ltd. All rights reserved.
J. Gao et al. International Journal of Fatigue 158 (2022) 106730

under the two-step loading as in Eq. (2). any finite subset of the data generated by the Gaussian process
throughout some domain can follow a multivariate Gaussian distribu­
n1 n2 ln(N1 ) { }
+ ( )ln(N2 ) = 1 (2) tion [32]. A set of n observations y = y1 , ..., yn is imagined as a sample
N1 N2
point from some multivariate (n-variate) Gaussian distribution. Thus, it
Conventional life prediction damage accumulation models are semi- can be partnered with a GP by working backwards. Generally, the mean
empirical based on the understanding of the phenomenon of the fatigue of this partner GP is considered to be zero. A variety of covariance
experiments. Although many fatigue damage accumulation models were functions are one of the most important features of GP, which relate one
established, there hasn’t been a nonlinear accumulation model which observation to another.
has been widely accepted and employed by the industry and academia Any observation y can be considered as connected with an under­
researchers. So more effective model should be explored in the field of lying function f(x) by a Gaussian noise model [32,33] in Eq.(3). In other
remaining life prediction. words, the relationship between input vector and the target of a GP is
Apart from the semi-empirical conventional fatigue life prediction expressed in Eq. (3), meanwhile noise is considered:
models, machine learning-based methods have been demonstrated their
potential in the fatigue field [6–9]: fatigue characteristics and material y = f (x) + ε (3)
properties estimation [10–12], fatigue life prediction [13,14], random where x is the input data, f(x) is the arbitrary regression function and
loading fatigue analysis [15–18], remaining life prediction [19,20] and y is the observed target observation, ε is the noise, following an inde­
fatigue crack growth prediction [21–23]. Recently, Chen et al. [24] pendent, identically distributed Gaussian distribution with zero mean
incorporated physics/knowledge constraints in artificial neural network and variance σ 2n , in other words ε ∼ N(0, σ2n ).
(ANN) for probabilistic fatigue S-N estimation. He et al. [11] employed The Gaussian process of f(x) is defined by the covariance and mean
ANN, support vector regression (SVR) and random forest machine of functions in Eqs. (4) and (5), where E denotes expectation.
learning approaches to predict the fatigue life for oblique hyperbola-
and bilinear-mode S-N curves of metallic materials. Barbosa et al. [25] m(x) = E(f (x)) (4)
generated a new constant life diagram for metallic materials on the basis
of ANN. Moreover, due to various sources of uncertainties in the mate­ (5)
′ ′ ′
k(x, x ) = E(f (x) − m(x))(f (x ) − m(x ))
rial data and parameters, a probabilities quantity should be considered The mean function m(x) of the prior GP embodied a prior knowledge
in fatigue life prediction. Therefore, it is desirable to find an efficient
about the latent function f(x). The covariance function k(x, x ) is used to
′

data-driven approach which can consider uncertainty quantification.

describe the relationship between one observation to another, also
Among machine learning methods, the GPR has good prediction
called kernel. So, the Gaussian process f(x) follows the relationship as in
ability and has been used in various fields [26–29]. The GPR, as a
Eq. (6)
Bayesian nonlinear regression method, infers a probability distribution
over all possible values with a solid theoretical foundation. It should f (x) ∼ GP(m(x), k(x, x ))
′
(6)
note that GPR has good performance for nonlinear regression problems
with small datasets [30]. And GPR can quantify uncertainty from both Fundamentally, regression problem is to search the latent function
intrinsic noise of the dataset and the parameter estimation errors [15]. f(x) utilizing a GP, then the objective y* can be obtained.
GPR gets the simultaneous estimation of the output value and the The covariance matrix in Eqs. (7) and (8) is calculated to prepare for
associated error/uncertainty. Therefore, the GPR is gradually employed the GPR. K aims to describe the relationship between each pair of
to tackle regression problem in various engineering fields, like fatigue training data x. And K* is calculated of the covariance between training
reliability of offshore wind turbines [31], machining-induced surface data x with test data x* .
⎡ ⎤
residual stress [28] and the degradation trend in bearings [29]. Hence in k(x1 , x2 ) k(x1 , x2 ) ⋯ k(x1 , xn )
this paper, Gaussian process regression (GPR) is introduced to quantify ⎢ k(x1 , x2 ) k(x1 , x2 ) ⋯ k(x2 , xn ) ⎥
K=⎢ ⎥ (7)
the uncertainties and improve the accuracy in remaining fatigue life ⎣ ⋮ ⋮ ⋱ ⋮ ⎦
prediction. k(xn , x1 ) k(xn , x2 ) ⋯ k(xn , xn )
In summary, this paper proposes a novel intelligent GPR-based
remaining fatigue prediction method to predict remaining fatigue life K∗ = [ k(x∗ , x1 ) k(x∗ , x2 ) ⋯ k(x∗ , xn ) ] (8)
under two-step loading. A dataset containing 12 materials, 328 samples
under two-step loading fatigue experimental results, is established to 2.1.2. The GPR-based remaining fatigue prediction method
train and test the GPR-based remaining fatigue prediction method and The GPR uses the Bayesian framework that interprets the probability
other methods used in this paper. In addition, the comparison with other as a degree of belief and employs Bayes’ theorem to perform supervised
five kinds of machine learning-based methods, the ordinary least machine learning to establish nonparametric kernel-based probabilistic
squares (OLS) regression, the ridge regression (RR), the support vector models. Therefore, the GPR-based remaining fatigue prediction method
regression (SVR), the decision tree (DT) and back propagation (BP) predicts the remaining fatigue life y* under two-step loading value with
neural network and the two conventional methods (PM and YW) is new input x* , the selected input variables, using Eq. (9).
conducted to verify the performance of the proposed method.
p(y|y* )p(y* )
The framework of this work is structured as follows: Section 2 gives a p(y* |y) = (9)
p(y)
detailed description of the GPR-based remaining fatigue prediction
method. Section 3 elaborates the two-step loading fatigue experiment where p(y|y* ) represents the prior distribution.p(y* ) indicates the like­
dataset. Section 4 validates the performance of the GPR-based remaining lihood. p(y* |y) is the posterior distribution.
fatigue prediction method. Conclusions are drawn in Section 5. The joint distribution of the target remaining fatigue life y* and the
training remaining fatigue life y still meets a Gaussian distribution, as
2. The GPR-based remaining fatigue prediction method shown in Eq. (10). K* is the transpose of K* .
′

[ ] ( [ ])
2.1. Methodology
′
y K + σ 2n I K*
∼ N 0, (10)
y* K* K**
2.1.1. Gaussian process
A Gaussian process (GP) is a probabilistic nonparametric model, The posterior predictive distribution p(y* |y) is obtained after condi­
which can be used for regression and classification problems. Formally, tioning the prior of the training data n observations y, x and x* . Hence,

2
J. Gao et al. International Journal of Fatigue 158 (2022) 106730

the key predictive equation for the GPR-based remaining fatigue pre­ regression model fits a dataset. And CC describes the correlation be­
diction method is in Eq. (11): tween prediction lives and experimental lives.
p(y* |y) ∼ N(y* , cov(y* )) (11)
3. Dataset establishment

y* = K* (K + σ 2 − 1
n I) y (12)
Two-step loading fatigue experiments involve Low-High (L-H) and
High-Low (H-L) loadings spectrums as shown in Fig. 1, which are per­
(13)
− 1 ′
cov(y* ) = K** − K* (K + σ2n I) K* formed to verify the life prediction accuracy and capability of the
Consequently, the GPR-based remaining fatigue prediction method damage accumulation models. Here N1 and N2 are the fatigue life cycles
can predict the remaining fatigue life by using the mean function y* , the at the corresponding loading stress Δσ1 and Δσ2 obtained from the S-N
covariance function cov(y* ) and the training data. y* is the prediction curve as shown in Fig. 2. So N1 and N2 are able to represent the fatigue
result given in Eq. (12). cov(y* ) represents the uncertainty of the pre­ load level. In the first step of H-L, the higher stress or strain is firstly
diction in Eq.(13). It can be seen that the advantage of employing GPR is imposed n1 cycles while inverse loading mode is imposed n1 cycles in L-
the simultaneous estimation of the output value and the associated H.
uncertainty. The dataset in this paper is composed of 12 materials, 328 fatigue
As mentioned above, the covariance function performs a crucial role experimental results under two-step loading which are to train and test
in GPR. Therefore, the crucial step in GPR process is to select a covari­ the GPR-based remaining fatigue prediction method. The dataset dis­
ance function which is symmetric and positive semi-definite by defini­ tribution is listed in Table 1. In addition, except the 316 stainless steel
tion. There are a variety of covariance functions. Each of these functions and P355NL1(A) are conducted under strain-controlled mode, all ma­
has a number of parameters called hyperparameters, which specify the terials are tested in stress-controlled mode. Moreover, normalized 45
property of the covariance function and should be tuned. According to steel, normalized 16Mn steel, 300CVM steel, SAE 4130, hot rolled 16Mn
the characteristics of the experimental data and evaluating results of steel and some of ductile iron GS61 are under fully-reversed rotary-
squared exponential kernel, exponential kernel, Matern 5/2 kernel, the bending loading. The fatigue experiments of Al-2024-T42 aluminum
exponential kernel is selected in this paper which is shown in Eq. (14). alloy are carried out under completely reverse bending loading.
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅ P355NL1(B) are performed under tension–compression loading ac­
′
(xi − xj ) (xi − xj ) cording to the stress ratios R = 0, R = 0.15 and R = 0.3. More details of
2
k(xi , xj ) = σf exp(− ) (14) the experimental data can be found out in the corresponding references.
l
The input variables of machine learning methods should be carefully
where l is a characteristic length scale, indicating the smoothness of the selected, which have to represent the relevant physical properties of the
candidate functions generated by the GP. σf is the standard deviation of mechanical component of fatigue experiments. As shown the two-step
the data. fatigue experiment design in Fig. 1, n1, N1 and N2 are direct acting
{ }
These optimal hyperparameters θ = l, σ f , σn are estimated factors to the fatigue life n2 without regard to material defects and other
through maximizing the logarithm likelihood function logp(y|x, θ) in Eq material factors. Especially n1, N1 and N2 play a crucial role in the or­
(15). using the gradient descent algorithm. ganization of the conventional empirical formulas so their physical ra­
tionality has been demonstrated. As shown in Fig. 2, N1 and N2 are
logp(y|x, θ) = −
1 ′ −1 1 n
y K y − log|K| − log(2π) (15) obtained from the S-N curve, which indicate the material properties and
2 2 2 behavior. Consequently, n1, N1 and N2 are selected as inputs to train the
machine learning methods and the remaining fatigue life n2 is selected as
2.2. Performance analysis output variable in this study. Then the adoption of logarithm trans­
formation with base 10 is in application with all the inputs and outputs
To assess the performance of the proposed method given above, the as data preprocessing to improve the prediction capability of the pro­
mean square error (MSE), mean absolute percentage error (MAPE), re­ posed method. As the span of fatigue life is so wide, the logarithm
sidual standard deviation(RSD) and correlation coefficient (CC) are used transformation should be adopted in case that some data points are
respectively, given in Eqs. (16)–(19) respectively. submerged.

1∑ n
MSE = y − yi )2
(̂ (16) 4. Results analysis and comparison
n i=1 i
4.1. Comparison methods
n ⃒ ⃒
100% ∑ ⃒̂y i − yi ⃒
MAPE = ⃒ ⃒ (17)
n i=1 ⃒ yi ⃒ In order to validate GPR-based remaining fatigue prediction method,
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅ there are five machine learning methods including OLS, RR, SVR, DT
√∑
√n ⌢ 2 and BP and two conventional models, the YW and the PM, are employed
√ ( y − yi ) in this paper, which are introduced in this part.
√
(18)
i
RSD = i=1
df
(1) The ordinary least squares
∑n ⌢ ⌢
i=1 (yi − mean(y))( − mean( ))
y y
i
CC = √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
̅ (19) OLS regression is finding the best function matching of dataset by
∑ n ∑
(yi − mean(y))2 ( y − mean( y ))2
n ⌢ ⌢
minimizing the sum of squares of errors for linear regression, trying to
i=1 i=1 i find a straight line to minimize the sum of Euclidean distances from all
samples to the straight line. The general linear regression formula is
⌢
where y is the predicted remaining fatigue life and y is the actual shown in Eq. (20) and the OLS regression solution is shown in Eq. (21).
experimental fatigue life. df is the degrees of freedom of the model.
y = wT X + b (20)
mean(y) is the mean of y. The value of MSE is in the range of 0 to 1. The
MSE value is close to 0 when the model prediction is accuracy. The range
(21)
⌢
w = (X T X)− 1 X T y
of MAPE is 0% to + ∞. When the value of MAPE is 0%, the model is
perfect. As for the RSD, when the RSD value is closer to 0, the better a

3
J. Gao et al. International Journal of Fatigue 158 (2022) 106730

Fig. 1. The illustrations for the H-L and L-H loading spectrums.

Fig. 2. The graphical representation of N1 and N2.

penalty function to the sum of the squared residuals to reduce the

Table 1
variability of the prediction. RR solution is shown in Eq. (22).
The distribution of two-step fatigue experiment results according to material
types. ⌢
w = (X T X + λI)− 1 X T y (22)
Serial number Material Number

1 Al-2024-T42 [34,35] 6
where I is unit matrix and λ is positive constant lying in the range of (0,
2 300CVM steel [35,36] 62 1).
3 SAE 4130 [36] 55
4 316L stainless steel [37] 9 (3) Support Vector Regression
5 Hot rolled 16Mn steel [38] 19
6 P355NL1(A) [39] 16
7 C-35 steel [40] 22 SVR is extended from the support vector machine, which is proposed
8 30NiCrMoV12 steel [39,41–43] 47 by former Soviet scientist Vapnik [48]. SVR uses structural risk mini­
9 Ductile iron GS61 [44] 6 mization principle to solve the over-fitting problems and thus to improve
10 Normalized 45 steel [45] 7 generalization ability. SVR maps input samples from low dimensional
11 Normalized 16Mn steel [45] 8
12 P355NL1(B) [46] 71
space to high dimensional space by adopting kernel functions using
insensitive loss, which can effectively tackle nonlinear problems.
Total 328

(4) Decision tree

where X and y are the input variables and output variables separately.
DT as a nonparametric machine learning algorithm builds regression
or classification models in the form of tree structure. When dealing with
(2) Ridge regression
nonlinear data, DT does not require any transformation of the features.
DT is not affected by high correlations among independent variables.
RR, proposed by Hoerl and Kennard [47], is introducing a small
The most important variables are included while the insignificant ones
amount of bias on the basis of linear regression. The RR is a regulari­
are excluded. It is also discovered could provide accurate estimation for
zation method, also known as Tikhonov regularization by adding

4
J. Gao et al. International Journal of Fatigue 158 (2022) 106730

small dataset. predicted value and a corresponding level of uncertainty [15]. There­
fore, the employed method can almost predict the remaining fatigue life
(5) Back propagation neural network under two-step loading accurately.
The remaining life prediction performance of the GPR-based
ANN is a machine learning algorithm which is inspired by neuron remaining fatigue prediction method, other five machine learning
nodes interconnection of human brains. It has good self-learning ability, methods and two conventional models (shaded area) is shown in Fig. 4.
which can be employed to simulate the nonlinear relation with high As shown in Fig. 4(a), the MSE values of the six machine learning
accuracy. Currently, the feedforward neural network optimized by BP methods are in the range of 0–0.1, which means that the machine
algorithm has been widely used because of its excellent ability in arbi­ learning methods used in this work perform very well in predicting the
trary complex pattern classification and multidimensional function remaining life under two-step loading. Particularly, the proposed
mapping, as well as its mature training method. In theory, a BP neural method has the lowest MSE value, which is less than 0.04 while it of the
network with a single hidden layer can approximate any nonlinear other five machine learning method is greater than 0.04. As for the
mapping for most forecasting problems [49]. In this paper, a three-layer conventional models, the MSE values are higher than all the six machine
network with 10 neurons in a hidden layer is established. The dataset is learning methods and over 0.1, of which the MSE value of the nonlinear
divided into three parts, 70% for training, 15% for validation and 15% accumulation model YW is lower than the linear model PM. In addition,
for testing. the six machine learning methods have better performance than the two
In general, no mandatory requirements need to be complied with conventional models. To sum up, according to the MSE results presented
when dividing data set. Considering the reliability and of machine in Fig. 4(a), the GPR-based remaining fatigue prediction method pre­
learning method, for all the data except the AL-2024-T24 material, 80% diction performance is the best among all the methods in this paper.
of the data is random selected as training data and the rest is classified as Also, to support the MSE results, MAPE results are obtained for each
test data. For the GPR-based remaining fatigue prediction method, the prediction method shown in Fig. 4(b). According to the MAPE values,
OLS, RR, SVR, DT and BP methods, the ten-fold cross-validation is used the GPR-based remaining fatigue prediction method prediction results
to obtain the best parameters in the training process. are better than the OLS, RR, SVR, DT and BP prediction results. More­
over, the MAPE results of the machine learning methods are better than
the those of the two conventional models. The GPR-based remaining
4.2. Prediction results and method comparisons
fatigue prediction method is able to successfully model and predict the
remaining fatigue life among these methods.
The remaining life prediction results of the proposed method for test
In Fig. 4(c), the RSD values of all regression methods in this paper are
data are shown in Fig. 3. The X-axis stands for the number of test data,
displayed. The proposed method has the minimum RSD value indicating
and Y-axis stands for the corresponding experimental fatigue life. As
that the proposed method has better remaining fatigue prediction per­
stated in section 2, the prediction results of GPR-based remaining fatigue
formance and higher prediction accuracy compared with other methods
prediction method obey Gaussian distribution, including the mean and
used in this paper.
the covariance. The mean shows the prediction results of the GPR-based
Again, comparing the CC values in Fig. 4(d), it can be concluded that
remaining fatigue prediction method. And the covariance reflects the
the GPR-based remaining fatigue prediction method has the highest CC
uncertainty. Comparison with the prediction fatigue lives of the GPR-
value. The CC value of 0.965 indicates better linear relationship and
based remaining fatigue prediction method, all the experiment fatigue
high strength of correlation between actual experimental fatigue lives
lives of test data locate in the 95% prediction interval except one close to
and predicted fatigue lives. In this premise, the GPR-based remaining
the positive variance as shown in Fig. 3. There are various sources of
fatigue prediction method has very good performance in remaining fa­
uncertainty due to the material defects, environmental factors and
tigue life prediction under two-step loading.
experimental design [50], which lead to the dispersing of fatigue life
In order to verify the life prediction accuracy and capability of the
prediction. The uncertainty should be probabilities quantified by a

Fig. 3. The predicted lives of the GPR-based remaining fatigue prediction method.

5
J. Gao et al. International Journal of Fatigue 158 (2022) 106730

Fig. 4. The remaining life prediction performance of the GPR-based remaining fatigue prediction method, other five machine learning methods, OLS, RR, SVR, DT,
BP, and two conventional models, YW and PM.

GPR-based remaining fatigue prediction method, comparisons between experimental data of the Al-2024-T42 alloy are used, which are excluded
the experimental fatigue lives and the predicted fatigue lives of the test from the training data. As can be seen in Fig. 6, all the six points of the
dataset are presented in Fig. 5. The blue dash line and the red dash line Al-2024-T42 alloy fall within the ±2 scatter band. And five of the six
represent the ±2 and ±3 scatter band separately. All the predicted lives points are close to the black solid line which represents the perfect
by the GPR-based remaining fatigue prediction method locate within the prediction. Consequently, the trained GPR-based remaining fatigue life
±3 scatter band to the experimental lives under the two-step loading and prediction method can get good prediction performance even for two-
most of the predicted lives lie in the ±2 scatter band in Fig. 5(a). While step loading fatigue experiments of untrained material to some extent.
the predication performance of other five machine learning methods are
worse according to Fig. 5(b)–(f). Although most of the predicted points 4.4. Latent variable analysis
falling with in the ±3 scatter band, there are some prediction points
outside of the ±3 scatter band. Moreover, there are more prediction In this part we attempt to verify if the prediction accuracy of the
points of two conventional prediction models lying outside of ±3 scatter proposed method can be increased with more information given. On the
band than the six machine learning methods shown in Fig. 5(g)–(h). On basis of n1 , N1 , N2 , σu and Ne are took into account. σ u is the ultimate
the whole, the GPR-based remaining fatigue prediction method gets best tensile strength. Ne is the minimum number of cycles that initiates
prediction performance among all the methods in this paper. The overall damage, namely the fatigue life at the knee point of S-N curve. Adopting
results confirm that using the proposed GPR method could successfully the logarithm transformation to scale down value of σ u and Ne is to
model and predict the remaining fatigue life of the two-step loading ensure they are in similar magnitude with the fatigue life.
condition. In Fig. 7, there are four kinds of inputs as shown in Table 2, set A
represents n1 , N1 , N2 ; set B represents n1 , N1 , N2 and Ne ; set C represents
4.3. Applied to untrained materials n1 , N1 , N2 and σu ; set D represents n1 , N1 , N2 Ne and σ u . As shown in
Fig. 7, from set A to set D, the values of the MSE, the MAPE and the RSD
To validate the performance of the proposed method in predicting gradually become smaller. On the contrary, the CC values of the four sets
the remaining fatigue life of a new material, the input variables n1 , N1 , gradually become larger with the increase of parameters.
N2 are needed or the S-N curve of the new material should be known. In On the whole, despite of the more information can increase the
this part, in order to verify the reliability of the proposed method in performance of the remaining life performance, the increase is little. As
prediction the remaining fatigue life of the new material, the shown in Fig. 7, the MSE value of the best set D minus that of the worst

6
J. Gao et al. International Journal of Fatigue 158 (2022) 106730

Fig. 5. Comparison between experimental lives and predicted lives by the GPR-based remaining fatigue prediction method and other five machine learning methods,
OLS, RR, SVR, DT, BP, and two conventional models, YW and PM.

7
J. Gao et al. International Journal of Fatigue 158 (2022) 106730

set A is just closed to 0.01. Similarly, the difference value of MAPE is just
approximates 0.6. And the difference value of the RSD is near 0.03,
while the difference value of CC is closed to 0.01. Consequently, the
extra input variables cannot significantly improve the remaining life
prediction accuracy of the GPR-based remaining fatigue life prediction
method. A possible reason is discussed here. For giving σu and Ne , there
is more material information added to the proposed method. So, the
performance gets improvement gradually. However, the extra input
variables cannot significantly improve the remaining life prediction
performance is that the added information is not enough for the per­
formance gets significantly improvement.

5. Conclusion

This research work is aimed to propose a machine learning method

for the remaining life prediction of metallic materials under two-step
loading. For this purpose, the GPR-based remaining fatigue prediction
method is employed to map the relationship between the selected inputs
and the corresponding remaining fatigue life. To validate the method,
the results of the proposed method are compared with other five ma­
chine learning methods, OLS, RR, SVR, DT, BP, and two conventional
models, YW and PM. Moreover, 328 two-step experimental results of 12
Fig. 6. Comparison between experimental lives and predicted lives of the AL- materials are used in this research. The main conclusions of this work
2024-T42 alloy by using the trained GPR-based remaining fatigue life predic­ can be drawn as follows:
tion method.
(1) The GPR-based remaining fatigue life prediction method can es­
timate the output life value and the associated uncertainty at the
same time in remaining fatigue life prediction under two-step
loading condition. The prediction results can provide partial
basis for fatigue reliability design of structures and operational
instructions and insights for structural health monitoring.
(2) Comparison results show that the GPR-based remaining fatigue
life prediction method has a higher prediction accuracy than the
OLS, RR, SVR, DT, BP methods and YW and PM conventional
models under the two-step loading. Specifically, the six machine
learning methods have better prediction performance than the
two conventional models, which can be inferred that the machine
learning methods are effective in remaining fatigue life predic­
tion to some extent. Such result illustrates the possibility of taking
the proposed method as a useful substitute to conventional
damage models in the field of remaining life prediction.
(3) The trained GPR-based remaining fatigue life prediction method
performs very well in the estimation of the untrained Al-2024-
T42 material, which shows the proposed method has effective
application in new material.
(4) Given more information, the prediction accuracy of the GPR-
based remaining fatigue life prediction method only has little
improvement. Therefore, the input variables, n1 , N1 , N2 , of the
GPR-based remaining fatigue life prediction method can main­
tain the basic physical information.

Declaration of Competing Interest

Fig. 7. The GPR-based remaining fatigue life prediction method prediction
performance with different input variables. The authors declare that they have no known competing financial
interests or personal relationships that could have appeared to influence
the work reported in this paper.
Table 2
List of the representative input variables. Acknowledgements
Input variables Output variable
This work is supported by the Science and Technology on Liquid
Set A n1 , N1 , N2 n2
Rocket Engine Laboratory (NO. 6142704190404) and the National
Set B n1 , N1 , N2 , Ne
Natural Science Foundation of China(NO. 51675406).
Set C n1 , N1 , N2 , σu
Set D n1 , N1 , N2 , Ne , σu
References

[1] Miner MA. Cumulative damage in fatigue. J Appl Mech 1945;67:A159–64.

8
J. Gao et al. International Journal of Fatigue 158 (2022) 106730

[2] B.S. Institution. BS 7608:2014 Guide to fatigue design and assessment of steel [25] Barbosa JF, Correia JAFO, Júnior RCSF, Jesus AMPD. Fatigue life prediction of
products. 389 Chiswick High Road, London; 2014. metallic materials considering mean stress effects by means of an artificial neural
[3] Marco SM, Starkey WL. A concept of fatigue damage. Trans ASME 1954;76: network. Int J Fatigue 2020;135:105527. https://doi.org/10.1016/j.
627–32. ijfatigue.2020.105527.
[4] Hectors K, De Waele W. Cumulative damage and life prediction models for high- [26] Richardson RR, Osborne MA, Howey DA. Gaussian process regression for
cycle fatigue of metals: A review. Metals 2021;11. forecasting battery state of health. J Power Sources 2017;357:209–19.
[5] Duyi Y, Zhenlin W. A new approach to low-cycle fatigue damage based on [27] Satria Palar P, Rizki Zuhal L, Shimoyama K. Gaussian process surrogate model with
exhaustion of static toughness and dissipation of cyclic plastic strain energy during composite kernel learning for engineering design. AIAA J 2020;58:1864–80.
fatigue. Int J Fatigue 2001;23(8):679–87. [28] Cheng M, Jiao Li, Yan P, Feng L, Qiu T, Wang X, et al. Prediction of surface residual
[6] Ulasa M, Aydura O, Gurgencb T, Ozel C. Surface roughness prediction of machined stress in end milling with Gaussian process regression. Measurement 2021;178:
aluminum alloy with wire electrical discharge machining by different machine 109333. https://doi.org/10.1016/j.measurement.2021.109333.
learning algorithms. J Mater Res Technol 2020;9:12512–24. [29] shankar kumar P, Kumaraswamidhas LA, Laha SK. Bearing degradation assessment
[7] Kalombo RB, Pestana MS, Freire Júnior RCS, Ferreira JLA, Silva CRM, and remaining useful life estimation based on Kullback-Leibler divergence and
Veloso LACM, et al. Fatigue life estimation of an all aluminium alloy 1055 MCM Gaussian processes regression, Measurement. Measurement 2021;174:108948.
conductor for different mean stresses using an artificial neural network. Int J https://doi.org/10.1016/j.measurement.2020.108948.
Fatigue 2020;140:105814. https://doi.org/10.1016/j.ijfatigue.2020.105814. [30] Sheng H, Liu X, Bai L, Dong H, Cheng Y. Small sample state of health estimation
[8] Eleftheroglou N, Zarouchas D, Benedictus R. An adaptive probabilistic data-driven based on weighted Gaussian process regression. J Storage Mater 2021;41:102816.
methodology for prognosis of the fatigue life of composite structures. Compos https://doi.org/10.1016/j.est.2021.102816.
Struct 2020;245:112386. https://doi.org/10.1016/j.compstruct.2020.112386. [31] Wilkie D, Galasso C. Gaussian process regression for fatigue reliability analysis of
[9] Pestana MS, Kalombo RB, Freire Júnior RCS, Ferreira JLA, da Silva CRM, offshore wind turbines. Struct Saf 2021;88:102020. https://doi.org/10.1016/j.
Araújo JA. Use of artificial neural network to assess the effect of mean stress on strusafe.2020.102020.
fatigue of overhead conductors. Fatigue Fract Eng Mater Struct 2018;41:2577–86. [32] Ebden M. Gaussian Processes for Regression: A Quick Introduction. [Online]
[10] Yan F, Song K, Liu Y, Chen S, Chen J. Predictions and mechanism analyses of the Available at: http://www.robots.ox.ac.uk/~mebden/reports, GPtutorial; 2008.
fatigue strength of steel based on machine learning. J Mater Sci 2020;55: [33] Rasmussen CE, Williams CKI. Gaussian processes for machine learning. MIT Press;
15334–49. 2006.
[11] He L, Wang Z, Akebono H, Sugeta A. Machine learning-based predictions of fatigue [34] DG P. A phenomenological fatigue damage accumulation rule based on hardness
life and fatigue limit for steels. J Mater Sci Technol 2021;90:9–19. increasing, for the 2024-T42 aluminum. Eng Struct 2002; 24(11): 1363–8.
[12] Agrawal A, Deshpande PD, Cecen A, Basavarsu GP, Choudhary AN, Kalidindi SR. [35] Mesmacque G, Garcia S, Amrouche A, Rubiogonzalez C. Sequential law in
Exploration of data science techniques to predict fatigue strength of steel from multiaxial fatigue, a new damage indicator. Int J Fatigue 2005;27:461–7.
composition and processing parameters. Integr Mater Manuf Innov 2014;3(1): [36] Ensign CR, Freche J, Manson SS. Application of a double linear damage rule to
90–108. cumulative fatigue. 1967.
[13] Zhang M, Sun C-N, Zhang X, Goh PC, Wei J, Hardacre D, et al. High cycle fatigue [37] Kamaya M, Kawakubo M. Loading sequence effect on fatigue life of Type 316
life prediction of laser additive manufactured stainless steel: A machine learning stainless steel. Int J Fatigue 2015;81:10–20.
approach. Int J Fatigue 2019;128:105–94. [38] Xie L, Lu W, Shi Z. Experimental Study on Fatigue Damage under two Level
[14] Yang J, Kang G, Liu Y, Chen K, Kan Q. Life prediction for rate-dependent low-cycle Loading. J Mech Strength 1994;16:52–4.
fatigue of PA6 polymer considering ratchetting: Semi-empirical model and neural [39] Zhu S, Liao D, Liu Q, Correia JAFO, De Jesus AMP. Nonlinear fatigue damage
network based approach. Int J Fatigue 2020;136:105619. accumulation: Isodamage curve-based model and life prediction aspects. Int J
[15] Farid M. Data-driven method for real-time prediction and uncertainty Fatigue 2019;128.
quantification of fatigue failure under stochastic loading using artificial neural [40] Subramanyan S. A cumulative damage rule based on the knee point of the S-N
networks and Gaussian process regression. Int J Fatigue 2022;155:106415. https:// curve. J Eng Mater Technol 1976;98:316.
doi.org/10.1016/j.ijfatigue.2021.106415. [41] Xia F-L, Zhu S-P, Liao D, Dantas R, Correia JAFO, De Jesus AMP. Isodamage curve-
[16] Ramachandra S, Durodola JF, Fellows NA, Gerguri S, Thite A. Experimental based fatigue damage accumulation model considering the exhaustion of static
validation of an ANN model for random loading fatigue analysis. Int J Fatigue toughness. Eng Fail Anal 2020;115:104575. https://doi.org/10.1016/j.
2019;126:112–21. engfailanal.2020.104575.
[17] Jang D-W, Lee S, Park J-W, Baek D-C. Failure detection technique under random [42] Dattoma V, Giancane S, Nobile R, Panella F. Fatigue life prediction under variable
fatigue loading by machine learning and dual sensing on symmetric structure. Int J loading based on a new non-linear continuum damage mechanics model. Int J
Fatigue 2018;114:57–64. Fatigue 2006;28:89–95.
[18] Durodola JF, Ramachandra S, Gerguri S, Fellows NA. Artificial neural network for [43] Fang Y, Hu M, Luo Y. New continuous fatigue damage model based on whole
random fatigue loading analysis including the effect of mean stress. Int J Fatigue damage field mearurement. J Mech Strength 2006;28:582–6.
2018;111:321–32. [44] Aid A, Amrouche A, Bouiadjra BB, Benguediab M, Mesmacque G. Fatigue life
[19] Gan L, Zhao X, Wu H, Zhong Z. Estimation of remaining fatigue life under two-step prediction under variable loading based on a new damage model. Mater Des 2011;
loading based on kernel-extreme learning machine. Int J Fatigue 2021;148: 32:183–91.
106190. https://doi.org/10.1016/j.ijfatigue.2021.106190. [45] Shang DG, Yao WX. A nonlinear damage cumulative model for uniaxial fatigue. Int
[20] Hou M, Pi D, Li B. Similarity-based deep learning approach for remaining useful J Fatigue 1999;21:187–94.
life prediction. Measurement 2020;159:107788. [46] Pereira HFSG, De Jesus AMP, Ribeiro AS, Fernandes AA. Fatigue Damage Behavior
[21] Strohmann T, Starostin-Penner D, Breitbarth E, Requena G. Automatic detection of of a Structural Component Made of P355NL1 Steel Under Block Loading. J Pressure
fatigue crack paths using digital image correlation and convolutional neural Vessel Technol 2009;131.
networks. Fatigue Fract Eng Mater Struct 2021;44:1336–48. [47] Hoerl A, Kennard R. Ridge Regression: Biased Estimation for Nonorthogonal
[22] Ma X, He X, Tu ZC. Prediction of fatigue–crack growth with neural network-based Problems. Technometrics 1970;12(1):55–67.
increment learning scheme. Eng Fract Mech 2021;241:107402. https://doi.org/ [48] Drucker H, Burges CJC, Kaufman L, C CJ, Kaufman BL, Smola A, Vapnik V. Support
10.1016/j.engfracmech.2020.107402. vector regression machines. Adv Neural Inf Process Syst 1997;28:779–84.
[23] Mortazavi SNS, Ince A. An artificial neural network modeling approach for short [49] V CG. Approximation by superpositions of a sigmoidal function. Math Control
and long fatigue crack propagation. Comput Mater Sci 2020;185:109962. https:// Signals Syst 1989;2(4):303–14.
doi.org/10.1016/j.commatsci.2020.109962. [50] Chowdhurya S, Deeba M, Zabelb V. Effects of parameter estimation techniques and
[24] Chen J, Liu Y. Probabilistic physics-guided machine learning for fatigue data uncertainty on the selection of fatigue crack growth model. Structures 2019;19:
analysis. Expert Syst Appl 2021;168:114316. https://doi.org/10.1016/j. 128–42.
eswa.2020.114316.