Engineering Fracture Mechanics 282 (2023) 109161

Contents lists available at ScienceDirect

Engineering Fracture Mechanics

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

Machine learning method for predicting the fatigue life of

sucker rods
Wu Yuandeng, Liu Shaohu *, Ma Weiguo, Ran Xiaofeng, Qu Baolong
School of Mechanical Engineering, Yangtze University, Jingzhou 434023, PR China

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

Keywords: As an important part of the rod pump production system, the safety of sucker rods directly affects
Sucker rod the normal production of the oil production system. In order to predict the fatigue life of defective
Fatigue life sucker rods, a novel method based on data-driven machine learning (ML) is proposed. Firstly, the
Machine learning
mechanical properties of sucker rods with defects are tested and the damage mechanism of sucker
S–N curve
Nominal stress method
rods is analyzed from the fracture morphology. Then the finite element method (FEM) is used to
analyze the stress distribution of the defective sucker rod. Based on the S–N curve of sucker rods
obtained from the experiment, the fatigue life of defective sucker rods is calculated by the
nominal stress method. A large sample dataset consisting of rod diameter, defect diameter, defect
depth, axial load and fatigue life is established to construct the training set and test set of ML
models. Different from the traditional elastic–plastic mechanics or FE methods, this method can
effectively calculate the fatigue life of sucker rods by inputting working parameters. The results
show that BPNN has good generalization ability by comparing the prediction results of three ML
models, and can effectively predict the fatigue life of defective sucker rods under different
working conditions, which provides a feasible method for rod safety monitoring.

1. Introduction

As a strategic resource, petroleum plays a key role in the industrial field and manufacturing industry, and accounts for a relatively
large proportion of the energy consumption in the world. With the continuous exploitation of oil and gas resources, the proportion of
exploitation of deep oil and gas resources gradually increases, making oil exploitation more difficult [1–2]. Depending on statistics,
more than 50 % of oil wells use rod pumps. Sucker rod is one of the most important devices in rod pump production system, and its
safety plays a crucial role in oil and gas exploitation. In the process of oil production, sucker rods bear complex alternating loads for a
long time, which can easily lead to fatigue fracture and other failure problems of sucker rods, thus affecting oil well production and
causing huge economic losses [3–6].
In order to ensure the normal operation of sucker rods within the safe service life, scholars have done a lot of research on the rod
failure problem. In the early stage, the maximum allowable stress of steel sucker rod was utilized according to the Goodman diagram.
Later, Hein and Hermanson et al. [7] made improvements on this basis and proposed the allowable stress curve for sucker rods. Pons
[8] proposed a rod stress analysis method based on the modified Everitt-Jennings algorithm, which could calculate stress values at any
depth. Woldesenbet [9] studied the failure location and cause of composite sucker rod by FEM and experimental tests. Zhou et al. [10]
applied damage mechanics to fatigue damage research of sucker rods, providing a theoretical basis for fatigue life prediction of sucker

* Corresponding author.
E-mail address: [email protected] (L. Shaohu).

https://doi.org/10.1016/j.engfracmech.2023.109161
Received 7 September 2022; Received in revised form 22 February 2023; Accepted 23 February 2023
Available online 2 March 2023
0013-7944/© 2023 Elsevier Ltd. All rights reserved.
W. Yuandeng et al. Engineering Fracture Mechanics 282 (2023) 109161

rods. Zhang and Wei [11–12] established a model of sucker rods with defects using magnetic flux leakage (MFL) detection technology,
and the results showed that the size and shape of transverse defects on sucker rods could be effectively determined by this method.
Zhang et al. [13] analyzed the reasons of sucker rod fracture in corrosive environment according to the sucker rod fracture in corrosive
environment. Cai et al. [14] proposed a three-parameter Weibull fitting method, which could estimate the fatigue limit of sucker rods
more accurately. Zhang et al. [15] established a cyclic stress model of sucker rod string, and proposed a fatigue life reliability pre­
diction model based on the randomness of material properties, defect size and cyclic load. The above researches mainly focused on the
failure mechanism of sucker rods with defects and the calculation of its bearing limit range, but did not involve specific fatigue life
prediction of sucker rods. However, if the service life of sucker rod can be predicted in practical engineering, it will effectively avoid
the occurrence of operational accidents.
It is difficult to establish a model to predict the fatigue life of sucker rods due to the complex alternating loads and the randomness
of defects. Fracture mechanics and experimental methods are generally utilized. ML shows excellent performance in processing
complex data and has great application prospects in evaluating fatigue parameters of material structures [16–18]. ML can be divided
into unsupervised learning algorithms and supervised learning algorithms. Regression algorithm can describe the continuous rela­
tionship between variables and is usually used for life prediction problems. Among them, artificial neural network (ANN) and support
vector machine (SVM) have been widely concerned as promising ML methods [19–20]. Compared with traditional statistical methods,
ML approach has higher computational accuracy and efficiency in prediction and nonlinear regression analysis [21–25]. For fatigue
problems which are difficult to establish prediction models, ML can be a good solution. Therefore, an increasing number of studies
have applied ML methods to the prediction of structural fatigue life [26–28]. At present, few scholars have used ML to study the fatigue
life of sucker rods. G et al. [29] used an optimized support vector machine (SVM) model to predict the fatigue damage of remanu­
factured sucker rods, providing a new feasible method for predicting the remaining life of sucker rods. Lv et al. [30] established a fault
diagnosis model for rod production system by using incremental support vector machine. The prediction results showed that the
diagnosis model could achieve satisfactory diagnosis accuracy, but the feature extraction method of the model needed to be further
improved. Chen et al. [31] used Extreme Gradient Boosting (XGBoost) algorithm to diagnose the working state of the rod pump. In
order to accurately predict polished rod load and improve service life of the sucker rod, ZHOU et al. [32] established a prediction
model of polished rod load by using improved artificial neural network based on the data of polished rod load collected in the field.
Peng [33] utilized a deep learning algorithm to analyze the relationship between electrical data and corresponding dynamometer cards
under different working conditions, which could be used to monitor downhole operation.
There are few researches on the application of ML approach to predict the fatigue life of sucker rods. The main objective of this
paper is to establish a data-driven ML method to predict the fatigue life of rods by combining experiment, finite element model and ML

Fig. 1. Schematic diagram of pumping unit operation.

2
W. Yuandeng et al. Engineering Fracture Mechanics 282 (2023) 109161

algorithm. Firstly, the fatigue performance of sucker rods is obtained according to the experimental results, and a data set containing
91 sets of data is established by using FEM to provide sufficient data for the ML model. Finally, three kinds of ML models are established
to predict the fatigue life of sucker rods, and the performance of ML models is evaluated.

2. Fatigue test of sucker rod

2.1. Experiments

During service, the pumping rods are subjected to long-term alternating loads. The load on the sucker rod changes periodically with
the change of up-down stroke of the oil pumping unit. Fig. 1 is the schematic diagram of pumping unit operation. The fatigue load of
the upper rod string is tension -tension alternating load, while the lower rod string is subjected to tension -compression fatigue load
once it is unstable [34]. The existence of defects can significantly reduce the fatigue life of sucker rods. In order to accurately study the
influence of defects on the fatigue life of sucker rods, the fatigue performance of sucker rods with defects was investigated by fatigue
experiments.
The PLG-300C electromagnetic resonance high frequency fatigue testing machine was used to conduct tensile fatigue test on the
sucker rods, as shown in Fig. 2. During experiments, both ends of the sucker rods were connected with the clamp of the testing ma­
chine. The loading mode is sinusoidal loading, operating at a frequency of 100 Hz, and the load ratio is 0.1(R = Fmin /Fmax ), the average
load is 84.6 kN, the alternating load is 69.2 kN, that is, the maximum axial tensile loadFmax = 153.8 kN, the minimum axial tensile
loadFmin = 15.38 kN.
In the experiments, 24 HL grade sucker rods were selected, with a diameter of 19 mm and a length of 500 mm. There were 8 groups
of experiments with 3 samples in each group, including 1 group of complete rods and 7 groups of sucker rods with different defect
depths. Fig. 3 shows sucker rod samples with defects. The spherical defects were processed in the middle position of the rod body, the
diameter of the defects was 10 mm, and the depth of the defects was 0.75 mm, 1.5 mm, 1.75 mm, 2 mm, 3 mm, 4 mm, and 6 mm,
respectively. The prepared samples were tested by fatigue testing machine, and the fatigue cycle times were measured. The sucker rods
were made of 30CrMo alloy structural steel. The specific material parameters and properties are shown in Table 1.

2.2. Analysis of experimental results

As shown in Figs. 4, 8 groups of experiments were carried out under the condition of maximum load of 153.8 kN and load ratio of
0.1. With the increase of defect depth, the fatigue life of sucker rods decreased successively, and the fatigue life of sucker rods
decreased sharply when the defect depth reached 2 mm. Through the analysis of the experimental results, it is concluded that the first

Fig. 2. High frequency fatigue testing machine.

3
W. Yuandeng et al. Engineering Fracture Mechanics 282 (2023) 109161

Fig. 3. Sucker rod samples with defects.

Table 1
Material properties of 30CrMo.
Sample Young’s modulus/MPa Poisson’s ratio Yield strength /MPa Tensile strength/MPa Elongation/% Reduction of area/%

Rod body 210,000 0.278 1034 1073 18 63

SY/T 5029-2013 / / ≥793 965–1195 ≥10 ≥45

group (no defect) and the third group (h = 1.5 mm) each have a bad point, and these two data should be discarded to avoid affecting
the subsequent analysis results.
Fig. 5(a) and (b) show the macroscopic fracture morphology of the sucker rod without defect and with defect depth h = 4 mm,
respectively. From the fracture morphology, three regions can be clearly seen, namely, the crack source zone, the fatigue crack
propagation zone and the transient fracture zone. Along the direction of crack propagation, the crack source and shear lip can be found,
indicating a typical fatigue fracture [35]. The fracture originates from the surface of the rod, and the macroscopic morphology of the
crack source area is relatively smooth. With the continuous expansion of the crack, the expansion area gradually becomes rough. When
the stress level in the remaining material reaches the tensile strength limit of the material, the sucker rod fractures instantaneously.
This area is the transient fracture zone, and obvious shear lip can be observed.

Fig. 4. Fatigue test results of sucker rods.

4
W. Yuandeng et al. Engineering Fracture Mechanics 282 (2023) 109161

Fig. 5. Macroscopic appearance of sucker rod fracture.

3. Numerical analysis of sucker rod

3.1. Establishment of FEM

In order to determine the stress distribution and the possible failure location of sucker rod under axial load, FEM of sucker rod was
established in ABAQUS, as shown in Fig. 6. Both ends of the sucker rod were simplified, one end of sucker rod was fixed, and axial
tensile load was applied at the other end. A variety of FEM for sucker rods with defects were established respectively. The mesh type of
the sucker rod is hexahedral, the mesh size is 2 mm, and the element type of the model is C3D8R. The mesh at the defect is refined.

Fig. 6. FEM of sucker rod.

5
W. Yuandeng et al. Engineering Fracture Mechanics 282 (2023) 109161

3.2. Stress analysis

Based on FE analysis, the stress distribution of defective and non-defective sucker rods under axial tensile load were calculated. As
shown in Table 2, the maximum von Mises stress and its distribution location of sucker rods with different defect depths were obtained
at a loading ratio of 0.1. The stress distribution of sucker rods is illustrated in Fig. 7. Under tensile load, the maximum stress of non-
defective sucker rod is distributed at the transition between rod body and upsetting, which is consistent with the actual fracture
position of the non-defective sucker rod. The maximum stress of sucker rod with defect is distributed at the defect, which is consistent
with the fracture position of the sample. It can be concluded that the stress concentration is the main cause of sucker rod failure.

3.3. Fatigue life calculation

Under the given stress ratio, if fatigue failure occurs after N cycles under the maximum stress, N is referred to as the fatigue life
under the maximum stress condition. In order to evaluate and estimate the fatigue life, the relationship between the load and the
fatigue life of material has to be established. The curve of the relationship between applied stress S and fatigue life N, namely S–N
curve, tend to evaluate the fatigue performance of structures [36]. In this paper, S–N curve of sucker rod combined with nominal
stress method is used to predict the fatigue life of sucker rods. Fig. 8 shows the process of estimating the fatigue life of structures by the
nominal stress method:
In order to accurately estimate the fatigue life of components, it is necessary to obtain the relationship curve that can reflect the
fatigue performance of materials. For high cycle fatigue, the S–N curve is desirable. The S–N curve is usually obtained by regression
analysis of fatigue experimental results. Commonly used empirical models include power function or an exponential function, etc. In
order to facilitate calculation, power function formula [37] is adopted, as shown in Eq. (1). The S–N curve of sucker rod is shown in
Fig. 9:
Sα N = C (1)
where α and C are the material constant. Take the logarithm of both sides of this equation, then Eq. (2) is obtained:
lgN = a + blgS (2)
where S is the stress level of the experiment. N is the cycle number of fatigue failure, a and b are constant.
The S–N curve obtained from the fatigue test applies only to the fatigue performance at a specific load ratio. The components bear
different load ratios in the actual loading process. Therefore, it is necessary to modify the S–N curve so that it can be applied to the
cyclic loading process under different load ratios [38]. Considering the effect of load ratio is actually considering the effect of average
stress, for which there are many modified models to quantify the effect of average stress on fatigue performance, such as Soderberg,
Goodman, Gerber and Morrow, et al. [39]. The most commonly used models are as following:

(1) Soderberg model

Sa Sm
+ =1 (3)
S− 1 Sy

(2) Goodman model

Sa Sm
+ =1 (4)
S− 1 Su

(3) Gerber model

( )2
Sa Sm
+ =1 (5)
S− 1 Su

Table 2
Maximum von Mises stress and distribution position of sucker rods.
Defect depth/mm Von Mises stress/MPa Distribution position

0 588.6 Transition
0.75 647.5 Defect
1.5 730.3 Defect
1.75 745.8 Defect
2 801.9 Defect
3 961.2 Defect
4 1197 Defect
6 1439 Defect

6
W. Yuandeng et al. Engineering Fracture Mechanics 282 (2023) 109161

Fig. 7. Stress diagram of sucker rods.

Fig. 8. Process of fatigue life estimation by nominal stress method.

Fig. 9. S–N curve of sucker rod.

(4) Morrow model

Sa Sm
+ =1 (6)
S− 1 Sf

where Sa is the stress amplitude, Sm is the average stress, MPa, S− 1 is the fatigue limit of the component when the stress ratio is − 1,

7
W. Yuandeng et al. Engineering Fracture Mechanics 282 (2023) 109161

Sy is yield strength, Sf is the true fracture stress, MPa. For tensile mean stress, the results of Soderberg model are conservative and
rarely used; Goodman model is suitable for brittle materials and conservative for ductile metals; Gerber model is suitable for ductile
metals, but its application is limited by its nonlinear function; Morrow model predicts that the sensitivity of ductile steel to average
stress is low [40]. So the S–N curve of sucker rod is modified by Morrow model.
Eq. (6) is transformed to obtain the Morrow model when the stress ratio isR0 , as shown in Eq. (7):
Sa,R0 Sm,R0
+ =1 (7)
S− 1 Sf

where Sa,R0 and Sm,R0 are stress amplitude and average stress respectively when the stress ratio isR0 . Morrow model was used to
obtain the modified S–N curve, and combined with the FE results, the modified S–N curve was modified, as shown in Fig. 10.
The stress results during the cycle and S–N fatigue performance curve of sucker rods calculated by ABAQUS were imported into Fe-
Safe to calculate the fatigue life of sucker rods. Fig. 11 Fig. 12 Fig. 13 Fig. 14 Fig. 15 Fig. 16 Fig. 17 Fig. 18 are the fatigue life dis­
tribution of the whole sucker rod. It can be concluded from the figure that the minimum fatigue life of the sucker rod without defect is
located in the arc transition zone, which is consistent with the fracture position of the non-defective sucker rod in the experiment. The
minimum fatigue life of the rod with defects is located at the defect, which is consistent with the fracture position of the rod with
defects in the experiment, which verifies the accuracy of the FE results.
Fig. 19 shows the comparison between the FE results and the experimental results of fatigue life of sucker rods. It can be observed in
the figure that the fatigue life results calculated by FEM are distributed within the range of the experimental results, with a small
difference from the experimental values. The errors between the FE results and the experimental results are relatively large in the two
groups of sucker rods without defects and with defect depth h = 1.5 mm. The main reason is that the two groups of fatigue test results
are relatively scattered, which has a certain impact on the results. Fig. 20 shows the comparison of errors between FE results and
experimental average values. It can be seen from the figure that the absolute errors of non-defective sucker rod and defect depth h =
1.5 mm are 20.67 % and 26.44 % respectively, and the errors of other data groups are less than 10 %, which further verifies the
feasibility of FEM to calculate the fatigue life of sucker rods.

4. Machine learning methods

Conventional machine learning algorithms used for regression include BP neural network (BPNN), support vector machine (SVM),
Elamn algorithm, etc., which show good potential in the field of prediction. In this section, three typical ML models (BP, Elman and
SVR) are selected for fatigue life prediction of sucker rods.

4.1. BP neural network

As a new branch of computing, artificial neural network has a powerful performance in dealing with complex linear and nonlinear
problems. BPNN is a widely used multilayer feedforward neural network based on error back propagation and gradient descent al­
gorithm [41,42]. Under the condition of reasonable structure and proper weight of network, BPNN can approximate any continuous
nonlinear function with maximum accuracy [43–45]. Fig. 21 is the schematic structural diagram of the neural network, which consists
of an input layer, an output layer and several hidden layers. Neurons in the same layer are not connected to each other, while neurons
in adjacent layers are fully connected.
After receiving the input signal Xi , each neuron in the input layer transmits the signal to all neurons in the hidden layer. Neurons Yj
in the hidden layer perform a weighted sum of the received input signals and calculate the output signal through the activation

Fig. 10. Corrected S–N curve.

8
W. Yuandeng et al. Engineering Fracture Mechanics 282 (2023) 109161

Fig. 11. Sucker rod without defect.

Fig. 12. Defect depth h = 0.75 mm.

Fig. 13. Defect depth h = 1.5 mm.

9
W. Yuandeng et al. Engineering Fracture Mechanics 282 (2023) 109161

Fig. 14. Defect depth h = 1.75 mm.

Fig. 15. Defect depth h = 2 mm.

Fig. 16. Defect depth h = 3 mm.

10
W. Yuandeng et al. Engineering Fracture Mechanics 282 (2023) 109161

Fig. 17. Defect depth h = 4 mm.

Fig. 18. Defect depth h = 6 mm.

Fig. 19. Comparison between FEM and experimental results.

11
W. Yuandeng et al. Engineering Fracture Mechanics 282 (2023) 109161

Fig. 20. Error between FEM and experimental results.

Fig. 21. General structure of BP neural network.

function, as shown in Eq. (8):

∑
Yj = fϕ ( Wij ⋅ Xi − θj ) (8)
i=1

where Wij is the weight between input layer unit Xi and hidden layer unitYj , θj is the threshold of hidden layer neuronYj , and fϕ is the
activation function of hidden layer neuron. fϕ is the activation function of neurons at the output layer, and sigmoid function is usually
selected, as defined in Eq. (9):
1
f (t) = t
(9)
1 + e−
Since BPNN adopts the fastest descent method and adjusts the weight according to the direction of the negative gradient of the error
function, there are usually-two problems: (1) low learning efficiency and slow convergence; (2) Easy to fall into local error. Therefore,
levenberg-Marquardt training algorithm (LM) is used to train the neural network to solve the above two problems [46,47], as shown in
Eq. (11):
[ ]
Jk ∗ JkT + λdiag(Jk ∗ JkT ) ΔW k = Jk ∗ Ek (11)

where Jk is the gradient of loss function, Ek is the vector of error at pointWk , λ is the non-negative damping factor. LM algorithm
makes the neural network update parameters along the direction of small gradient, so as to avoid falling into local optimum.

12
W. Yuandeng et al. Engineering Fracture Mechanics 282 (2023) 109161

The data obtained from the experiments have different dimensions and the values vary greatly, which will affect the results of data
analysis. Therefore, in order to eliminate the influence of dimensions, the input and output variables should be normalized within the
range (0,1) before applying the data to the neural network [48,49], as shown in Eq. (12):
xi − xmin
(12)
′
xi = a +b
xmax − xmin

where xi is the normalized data, xmin and xmax are the minimum and maximum values of all training and test data respectively, a and
′

b are scaling factors. Here a and b are equal to 0.

In addition to the influence of the raw data on the prediction accuracy, the structure of neural network also has great impact on the
prediction accuracy. The learning ability of a neural network can be improved by increasing the number of neurons in the hidden layer
or the number of hidden layers. However, excessive increaseing the number of neurons in the hidden layer or the number of hidden
layers can also lead to over-fitting, which results in larger errors [28,50]. After several calculations, the selected optimal network
structure is given in Table 3.

4.2. Elman neural network

Elman neural network is a kind of recurrent neural network with local memory unit and local feedback connection. An undertaking
layer is added to the hidden layer of the feedforward network structure, which receives feedback signals from the hidden layer [51,52].
Each node of the hidden layer has its corresponding associated node, which acts as a delay operator to memorize and store the
memorize of the hidden layer, so that the network has time-varying characteristics, and the network structure as shown in Fig. 22.
The expression used to describe the nonlinear state space of Elman neural network is as follows:
y(k) = g[wj,k H(k) + b2 ] (13)

H(k) = f [αF(k) + wi,j (x(k − 1)) + b1 ] (14)

F(k) = H(k − 1) (15)

where wi,j is the weight between input layer and hidden layer, wj,k is the weight between the output layers of the hidden layer, α is
the weight of the hidden layer and the undertaking layer, b1 is the threshold between the input layer and the hidden layer, b2 Is the
threshold between hidden layer and output layer, y(k) is the output of the output layer, H(k) is the output of the hidden layer, F(k) is
the output of the receiving layer, x(k − 1) is the (k-1)th input data, f(*) is the activation function of hidden layer, and sigmoid function
is adopted, g(*) is the activation function of output layer.

4.3. SVR model

The core idea of SVR is to improve the efficiency of nonlinear classification and regression analysis by mapping the input infor­
mation to a high-dimensional feature space using a kernel function and finding an interval hyperplane in the high-dimensional
mapping space [53,54]. The learning strategy of SVR is to transform its interval maximization into a convex quadratic program­
ming problem, and the objective function can be expressed as:

1 ∑N
( )
min ‖ω‖2 + C ξn + ξ*n
2 n=1
⎧
⎪
⎪
⎪
⎪ yn − ωφ(xn ) − b⩽ε + ξn
⎪
⎪
⎪
⎪ (16)
⎨ ωφ(xn ) − b − yn ⩽ε + ξ*n
⎪
s.t.
⎪
⎪
⎪ ξn , ξ*n ⩾0
⎪
⎪
⎪
⎪
⎪
⎪ n = 1, 2, ⋯, N
⎩

where ω is the weight vector, b is the threshold, C is the penalty factor, ξn and ξ*n is the slack variable, φ(x) is a nonlinear function, ε

Table 3
Optimal architecture of neural networks.
Network Number

Neurons of input layer 4

Hidden layer 2
Neurons of the 1st hidden layer 20
Neurons of the 2st hidden layer 30
Neurons of output layer 1

13
W. Yuandeng et al. Engineering Fracture Mechanics 282 (2023) 109161

Fig. 22. Structure of Elman neural network.

Fig. 23. Dataset of ML models: (a)Training set; (b)Test set.

14
W. Yuandeng et al. Engineering Fracture Mechanics 282 (2023) 109161

Fig. 24. The prediction results of ML models: (a)BPNN; (b)Elman; (c)SVR.

15
W. Yuandeng et al. Engineering Fracture Mechanics 282 (2023) 109161

is the error accuracy. Eq. (16) can be transformed into a dual optimization problem according to Lagrange function [55].
∑
N
[ ] 1∑ N
min αn (ε − yn ) − α*n (ε + yn ) + (αi − α*i )(αj − α*j )K(xi , xj )
n=1
2 i,j=1
⎧∑
⎪
n
(17)
⎪
⎨ (αi − α*i ) = 0
s.t. n=1
⎪
⎪
⎩
αi , α*i ∈ [0, C]

where K(xi , xj ) = φ(xi )φ(xj ) is kernel function, αi and α*i are Lagrange multipliers. The relationship between input vector and output
vector established by SVR is as follows:
∑
N
f (x) = (αi − α*i ) ⋅ K(x, xi ) + b (18)
n=1

Since the radial basis function has good nonlinear characteristics in high dimensional space, it is adopted as the kernel function:
( ⃦ ⃦ )
⃦xi − xj ⃦2
K(xi , xj ) = exp - (19)
2σ2

where σ is the kernel parameter of the radial kernel function, and the value of this parameter has an impact on the effect of the
model. If the parameter is too small, the generalization ability of the model will be weakened, and if the parameter is too large, the
model will be overfitted.

4.4. Evaluation and validation

Considering the influence of various factors on the fatigue life of sucker rods, the rod diameter, defect diameter, defect depth and
axial load are taken as the input data, and the fatigue life of the sucker rod is taken as the output data. 91 sets of samples are established
by FEM, 70 % of the total samples are selected as the training set, and the remaining 30 % of the data are selected as the test set. The
data of the training set and test set are chosen by random sampling method. Fig. 23 shows the prediction results of training set and test
set respectively. As can be seen from the figure, the prediction results of training set and test set are evenly distributed within the range
of 1.5 times.
In order to evaluate the accuracy and effectiveness of BPNN, Elman and SVR models for the prediction of fatigue life, mean absolute
percentage error (MAPE) and coefficient of determination (R2) are used to evaluate the predicted performance of the ML models,
where R2 represents the fitting degree of prediction results. The specific expression formula is as follows:
N ⃒
⃒ ⃒
1 ∑ Opred − Treal ⃒
MAPE = × 100% (20)
N n=1 |Treal |

∑N ( )2
Opred − Treal
R2 = 1 − ∑n=1
N ( ) 2
(21)
n=1 Treal − Oavg

where Treal is the actual value, Opred is the predicted value, Oavg is the average value of the predicted result.
To verify the generalization of the ML models, the fatigue life under eight experimental conditions in Section III is predicted, and
the predicted results are compared with the FEM results (these results are new to the dataset of ML and can therefore be used to verify
the generalization of ML models). The prediction results of the three models are presented in Fig. 24. Fig. 24(a) shows the prediction
results of BPNN. It can be seen from the figure that all the prediction results are within 1.5 times error band. Fig. 24(b) shows the
prediction results of Elman neural network. It can be seen from the figure that there are 6 prediction results within 1.5 times error
band, 2 prediction results outside 1.5 times error band. Fig. 24(c) shows the prediction results of SVR. It can be seen from the figure
that 7 prediction results are within 1.5 times error band, and 1 prediction result is outside 1.5 times error band. It can be concluded that

Table 4
Comparison results of FE and BPNN model.
Sample number FE results/cycles BPNN results/cycles Difference/cycles Diference/% R2 MAPE

1 247,642 255,305 7663 3.09 0.955 0.084

2 190,898 226,501 35,603 18.65
3 135,653 124,512 11,141 8.2
4 127,933 109,987 17,946 14.02
5 105,915 98,568 7347 6.93
6 65,537 71,984 6447 9.84
7 36,373 37,748 1375 3.78
8 21,064 20,503 561 2.66

16
W. Yuandeng et al. Engineering Fracture Mechanics 282 (2023) 109161

Table 5
Comparison results of FE and Elman model.
Sample number FE results/cycles Elman results/cycles Difference/cycles Diference/% R2 MAPE

1 247,642 230,844 16,798 6.78 0.874 0.375

2 190,898 235,316 44,418 23.28
3 135,653 165,375 29,722 21.91
4 127,933 146,897 18,964 14.82
5 105,915 130,698 24,783 23.4
6 65,537 85,198 19,661 30
7 36,373 61,400 25,027 68.8
8 21,064 44,515 23,451 111.33

Table 6
Comparison results of FE and SVR model.
Sample number FE results/cycles SVR results/cycles Difference/cycles Diference/% R2 MAPE

1 247,642 260,597 12,955 5.23 0.975 0.165

2 190,898 205,360 14,462 7.58
3 135,653 129,969 5684 4.19
4 127,933 114,596 13,337 10.43
5 105,915 102,765 3150 2.97
6 65,537 75,402 9865 15.05
7 36,373 52,353 15,985 43.93
8 21,064 12,147 8917 42.33

BP neural network has better performance in extracting multi-feature input.

As the data of the training and test set of the neural network are based on the prediction of existing data, in order to further verify
the accuracy of the neural network and test its generalization ability, the error between the prediction results and the FEM results, the
coefficient of determination R2 and the mean absolute percentage error (MAPE) are calculated. The specific results are shown in
Table 4 Table 5. Table 6. It can be seen that the maximum and minimum error of BPNN are 18.65 % and 2.66 % respectively; the
maximum and minimum error of Elman are 111.33 % and 6.78 % respectively; the maximum and minimum error of SVR are 43.93 %
and 2.97 % respectively. According to the calculation results of R2 and MAPE, BPNN and SVR have better prediction performance than
Elman. By further comparison, the R2 of SVR is slightly higher than that of BPNN, indicating that the difference of total prediction
effect between the two algorithms is small. However, according to the results of MAPE, the MAPE of SVR is 0.165 and that of BPNN is
0.084, indicating that the prediction results of SVR are more discrete. It can be concluded that the BPNN has better generalization
ability and high prediction performance, this model has higher accuracy in predicting the fatigue life affected by multi-characteristic
factors.

5. Conclusion

The fatigue life of sucker rods with defects is studied by combining experimental research, FE simulation and theoretical approach.
A ML method is proposed to predict the fatigue life of sucker rods by establishing a data set through FEM. The following conclusions
can be drawn from the results:

(1) The fatigue life of sucker rod is calculated by S–N curve and nominal stress method. The error between the fatigue life
calculation results and the experimental results is less than 10 %, which verifies the feasibility of the fatigue calculation method.
(2) Three machine learning models are used to predict the fatigue life of sucker rods, including the influencing factors of rod
diameter, defect diameter, defect depth, axial load. The results show that the prediction results of BPNN and SVR have better
correlation with the FE results, indicating that these two network models have good training performance.
(3) Taking the experimental conditions as input parameters, the prediction errors of BPNN are all less than 20 %, and the prediction
performance of BPNN is better than that of SVR, which indicates that BPNN neural network has good generalization ability and
provides a feasible method for rod fatigue life prediction.

It should be noted that the method proposed is only for the specific defect shape and load, and does not consider factors such as the
damage degree, actual load, environmental corrosion and other factors during the actual operation of the rod. Therefore, these
influencing factors should be taken into account in the subsequent study.

CRediT authorship contribution statement

Wu Yuandeng: Conceptualization, Methodology, Software, Writing – original draft, Writing – reviewing and editing. Liu Shaohu:
Investigation, Supervision, Funding acquistion, Formal analysis. Ma Weiguo: Project administration, Resources, Investigation,

17
W. Yuandeng et al. Engineering Fracture Mechanics 282 (2023) 109161

Supervision. Ran Xiaofeng: Data curation, Validation, Software. Qu Baolong: Data curation, Software.

Declaration of Competing Interest

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.

Data availability

All the data are in the article

Acknowledgements

This research was supported by the National Natural Science Foundation of China (No.51974036, No.51604039), the Hubei
Provincial Outstanding Young and middle-aged Science and Technology Innovation Team Project (No.T2021035) and the Young Top-
notch Talent Cultivation Program of Hubei Province.

References

[1] Ding H, Xie JF, Bai ZQ, et al. Fracture analysis of a connecting rod for oil pumping unit in China western oilfield. Engng. Fail. Anal. 2019;105:313–20. https://
doi.org/10.1016/j. engfailanal.2019.07. 021.
[2] Liang H, Li XM. Analysis on failure mechanism of sucker rod pumping system. Adv. Mat. Res. 2014;875:1219–24. https://doi.org/10.4028/www.scientific.net/
AMR.875-877.1219.
[3] Guo Hua L, Shun Li H, Zhi Y, et al. A prediction model for a new deep-rod pumping system. J. Pet. Sci. Engng. 2011;80(1):75–80. https://doi.org/10.1016/j.
petrol.2011. 10.011.
[4] Ding H, Zhang AB, Qi DT, et al. Failure analysis of a sucker rod fracture in an oilfield. Engng. Fail. Anal. 2020;109:104300. https://doi.org/10.1016/j.
engfailanal.2019.104300.
[5] Xu J, Wu X, Kang Y. Defect detection in transition zones of sucker rods using magnetostrictive guided waves. Int. J. Appl. Electromagn. Mech. 2012;39(1–4):
229–35. https://doi.org/10.3233/JAE-2012-1465.
[6] Duan DL, Geng Z, Jiang SL, et al. Failure mechanism of sucker rod coupling. Engng. Fail. Anal. 2014;36:166–72. https://doi.org/10.1016/j.
engfailanal.2013.10.003.
[7] Hein NW, Hermanson DE. A new look at sucker rod fatigue life. SPE 1993;26558. https://doi.org/10.2118/26558-MS.
[8] Pons V. Optimal stress calculations for sucker rod pumping systems. SPE 2014;171346. https://doi.org/10.2118/171346-MS.
[9] Woldesenbet E. Finite element stress analysis of composite sucker rods. J. Energy Res. Technol. 2003;125(4):299–303. https://doi.org/10.1115/1.1618264.
[10] Zhou RF, Feng MX, Jiang MZ, et al. Research on fatigue damage of sucker rod based on damage mechanics. Appl. Mech. Mater. 2014;633:1117–23. https://doi.
org/10.4028/www.scientific.net/AMM.633-634.1117.
[11] Zhang O, Wei X, Yan S. Numerical analysis of magnetic flux leakage of transverse defects of sucker rod. J. Test. Eval. 2018;46(5):2265–73. https://doi.org/
10.1520/JTE20160543.
[12] Zhang O, Wei X. Analysis of MFL model for sucker rod defects and its MFL signal processing. J. Test. Eval. 2018;47(5):3765–80. https://doi.org/10.1520/
JTE20170687.
[13] Zhang J, Zeng W, Guo Y, et al. Fracture failure analysis of type HL sucker rod in H2S-CO2 environment. In: NACE International Corrosion Conference
Proceedings. NACE International; 2020. p. 1–11.
[14] Cai W, Li W, Xu J. Study on the P-S-N curve of sucker rod based on three-parameter Weibull distribution. Materials 2022;15(2):560. https://doi.org/10.3390/
ma15020560.
[15] Zhang H, Dong S, Chen Z, et al. Reliability prediction method of fatigue life for rod string. In: 2012 International Conference on Quality, Reliability, Risk,
Maintenance, and Safety Engineering. IEEE; 2012. p. 895–9. doi: 10.1109/ICQR2MSE.2012.6246370.
[16] Durodola JF, Ramachandra S, Gerguri S, et al. Artificial neural network for random fatigue loading analysis including the effect of mean stress. Int. J. Fatigue
2018;111:321–32. https://doi.org/10.1016/j.ijfatigue.2018.02.007.
[17] Naik DL, Kiran R. Identification and characterization of fracture in metals using machine learning based texture recognition algorithms. Engng. Fract. Mech.
2019;219:106618. https://doi.org/10.1016/j.engfracmech.2019.106618.
[18] Ma XR, He XF, Tu ZC. Prediction of fatigue–crack growth with neural network-based increment learning scheme. Engng. Fract. Mech. 2020;242:107402.
https://doi.org/10.1016/j.engfracmech.2020.107402.
[19] Agrawal A, Choudhary A. An online tool for predicting fatigue strength of steelalloys based on ensemble data mining. Int. J. Fatigue 2018;113:389–400. https://
doi.org/10.1016/j.ijfatigue.2018.04.017.
[20] Artero-Guerrero JA, Pernas-Sánchez J, Martín-Montal J, et al. The influence of laminate stacking sequence on ballistic limit using a combined Experimental/
FEM/Artificial Neural Networks (ANN) methodology. Compos. Struct. 2018;183:299–308. https://doi.org/10.1016/j.compstruct.2017.03.068.
[21] Horňas J, Běhal J, Homola P, et al. Modelling fatigue life prediction of additively manufactured Ti-6Al-4V samples using machine learning approach. Int. J.
Fatigue 2023;169:107483. https://doi.org/10.1016/j.ijfatigue.2022.107483.
[22] He R, Yang H, Sun S, et al. A machine learning-based fatigue loads and power prediction method for wind turbines under yaw control. Appl. Energy 2022;326:
120013. https://doi.org/10.1016/j.apenergy.2022.120013.
[23] Xu BW, Ye S, Li M, et al. Deep learning method for predicting the strengths of microcracked brittle materials. Engng. Fract. Mech. 2022;271:108600. https://doi.
org/10.1016/j.engfracmech.2022.108600.
[24] Tan L, Yang XG, Shi DQ, et al. Unified fatigue life modelling and uncertainty estimation of Ni-based superalloy family with a supervised machine learning
approach. Engng. Fract. Mech. 2022;275:108813. https://doi.org/10.1016/j.engfracmech.2022.108813.
[25] Zhan Z, Ao N, Hu Y, et al. Defect-induced fatigue scattering and assessment of additively manufactured 300M-AerMet100 steel: An investigation based on
experiments and machine learning. Engng. Fract. Mech. 2022;264:108352. https://doi.org/10.1016/j.engfracmech.2022.108352.
[26] Zhang M, Sun C, Zhang X, et al. High cycle fatigue life prediction of laser additive manufactured stainless steel: A machine learning approach. Int. J. Fatigue
2019;128:105194. https://doi.org/10.1016/j.ijfatigue.2019.105194.
[27] Voet V, Van Loock F, De Fruytier C, et al. Machine learning aided modelling of thermomechanical fatigue of solder joints in electronic component assemblies.
Int. J. Fatigue 2023;167:107298. https://doi.org/10.1016/j.ijfatigue.2022.107298.
[28] Yang J, Kang G, Liu Y, et al. Life prediction for rate-dependent low-cycle fatigue of PA6 polymer considering ratchetting: Semi-empirical model and neural
network based approach. Int. J. Fatigue 2020;136:105619. https://doi.org/10.1016/j.ijfatigue.2020.105619.
[29] Yatian G, Jiancheng L, Siqi L. Residual life prediction method for remanufacturing sucker rods based on magnetic memory testing and a support vector machine
model. Insight-Non-Destructive Testing and Condition Monitoring 2019;61(1):44–50. https://doi.org/10.1784/insi.2.019.61.1.44.

18
W. Yuandeng et al. Engineering Fracture Mechanics 282 (2023) 109161

[30] Lv X, Wang H, Zhang X, et al. An evolutional SVM method based on incremental algorithm and simulated indicator diagrams for fault diagnosis in sucker rod
pumping systems. J. Pet. Sci. Engng. 2021;203:108806. https://doi.org/10.1016/j.petrol. 2021.108806.
[31] Chen L, Gao X, Li X. Using the motor power and XGBoost to diagnose working states of a sucker rod pump. J. Pet. Sci. Engng. 2021;199:108329. https://doi.org/
10.1016/j.petrol.2020.108329.
[32] Zhou RF, Bai L, Dong KX, et al. Suspended load prediction on sucker rod suspension load based on artificial neural network. Adv. Mat. Res. 2011;217:1040–3.
https://doi.org/10.4028/www.scientific.net/AMR. 217-218.1040.
[33] Peng Y. Artificial intelligence applied in sucker rod pumping wells: Intelligent dynamometer card generation, diagnosis, and failure detection using deep neural
networks. SPE 2019;196159. https://doi.org/10.2118/196159-MS.
[34] Wang Y, Wang S, Yang L, et al. A new model to evaluate polished rod load of sucker rod pumping system. SPE 2018;191803. https://doi.org/10.2118/191803-
MS.
[35] Zhang W, Zeng L. Experimental investigation and low-cycle fatigue life prediction of welded Q355B steel. J. Constr. Steel Res. 2021;178:106497. https://doi.
org/10.1016/j.jcsr. 2020.106497.
[36] Murakami Y, Takagi T, Wada K, et al. Essential structure of SN curve: Prediction of fatigue life and fatigue limit of defective materials and nature of scatter. Int.
J. Fatigue 2021;146:106138. https://doi.org/10.1016/j.ijfatigue. 2020.106138.
[37] Alencar G, Hong JK, de Jesus A, et al. The Master S-N curve approach for fatigue assessment of welded bridge structural details. Int. J. Fatigue 2021;152:
106432. https://doi.org/10.1016/j.ijfatigue.2021.106432.
[38] He N, Feng PF, Li ZW, et al. Fatigue life prediction of centrifugal fan blades in the ventilation cooling system of the high-speed-train. Engng. Fail. Anal. 2021;
124:105373. https://doi.org/10.1016/j.engfailanal.2021.105373.
[39] Korba P, Huňady R, Hovanec M, et al. Fatigue life analysis of an aircraft brake component to prevent damage and ensure operational safety. Engng. Fail. Anal.
2021;129:105653. https://doi.org/10.1016/j.engfailanal.2021.105653.
[40] Shittu AA, Mehmanparast A, Hart P, et al. Comparative study between S-N and fracture mechanics approach on reliability assessment of offshore wind turbine
jacket foundations. Reliab. Engng. Syst. Saf. 2021;215:107838. https://doi.org/10.1016/j.ress.2021.107838.
[41] Nasiri S, Khosravani MR, Weinberg K. Fracture mechanics and mechanical fault detection by artificial intelligence methods: A review. Engng. Fail. Anal. 2017;
81:270–93. https://doi.org/10.1016/j.engfailanal.2017.07.011.
[42] Han YL. Artificial neural network technology as a method to evaluate the fatigue life of weldments with welding defects. Int. J. Press. Vessel. Pip. 1995;63(2):
205–9. https://doi.org/10.1016/0308-0161(94)00055-N.
[43] Luo H, Li Z, Xiong Q. Study on wind-induced fatigue of heliostat based on artificial neural network. J. Wind Engng. Ind. Aerodyn. 2021;217:104750. https://doi.
org/10.1016/j.jweia.2021.104750.
[44] Mortazavi SNS, Ince A. An artificial neural network modeling approach for short and long fatigue crack propagation. Comput. Mater. Sci 2020;185:109962.
https://doi.org/10.1016/j.commatsci.2020.109962.
[45] Wu L, Yang Y, Maheshwari M. Strain prediction for critical positions of FPSO under different loading of stored oil using GAIFOA-BP neural network. Mar. Struct.
2020;72:102762. https://doi.org/10.1016/j.marstruc.2020.102762.
[46] D’Agostino L, De Santis A, Di Cocco V, et al. Fatigue crack propagation in ductile cast irons: An artificial neural networks based model. Procedia Struct. Integrity
2017;3:291–8. https://doi.org/10.1016/j.prostr.2017.04.048.
[47] Kong YS, Abdullah S, Schramm D, et al. Optimization of spring fatigue life prediction model for vehicle ride using hybrid multi-layer perceptron artificial neural
networks. Mech. Syst. Sig. Process. 2019;122:597–621. https://doi.org/10.1016/j.ymssp. 2018.12.046.
[48] Gan L, Zhao X, Wu H, et al. Estimation of remaining fatigue life under two-step loading based on kernel-extreme learning machine. Int. J. Fatigue 2021;148:
106190. https://doi.org/10.1016/j.ijfatigue. 2021.106190.
[49] Genel K. Application of artificial neural network for predicting strain-life fatigue properties of steels on the basis of tensile tests. Int. J. Fatigue 2004;26(10):
1027–35. https://doi.org/10.1016/j.ijfatigue.2004.03.009.
[50] Zhan Z, Li H. Machine learning based fatigue life prediction with effects of additive manufacturing process parameters for printed SS 316L. Int. J. Fatigue 2021;
142:105941. https://doi.org/10.1016/j.ijfatigue.2020.105941.
[51] Yang L, Wang F, Zhang J, et al. Remaining useful life prediction of ultrasonic motor based on Elman neural network with improved particle swarm optimization.
Measurement 2019;143:27–38. https://doi.org/10.1016/j.measurement.2019.05.013.
[52] Merainani B, Laddada S, Bechhoefer E, et al. An integrated methodology for estimating the remaining useful life of high-speed wind turbine shaft bearings with
limited samples. Renew. Energy 2022;182:1141–51. https://doi.org/10.1016/j.renene.2021.10.062.
[53] Bartošák M. Using machine learning to predict lifetime under isothermal low-cycle fatigue and thermo-mechanical fatigue loading. Int. J. Fatigue 2022;163:
107067. https://doi.org/10.1016/j.ijfatigue.2022.107067.
[54] Zhang XC, Gong JG, Xuan FZ. A deep learning based life prediction method for components under creep, fatigue and creep-fatigue conditions. Int. J. Fatigue
2021;148:106236. https://doi.org/10.1016/j.ijfatigue.2021.106236.
[55] Bao H, Wu S, Wu Z, et al. A machine-learning fatigue life prediction approach of additively manufactured metals. Engng. Fract. Mech. 2021;242:107508.
https://doi.org/10.1016/j.engfracmech.2020.107508.

19