In the current study, the fatigue life of QSTE340TM steel was modelled using a machine learning method, namely, a neural network. This problem was solved by a Multi-Layer Perceptron (MLP) neural network with a 3-75-1 architecture, which allows the prediction of the crack length based on the number of load cycles N, the stress ratio R, and the overload ratio Rol. The proposed model showed high accuracy, with mean absolute percentage error (MAPE) ranging from 0.02% to 4.59% for different R and Rol. The neural network effectively reveals the nonlinear relationships between input parameters and fatigue crack growth, providing reliable predictions for different loading conditions.
QSTE340TM steel is a thermomechanically hardened low-alloy steel used in the automotive
and mechanical engineering industries. Due to its high strength, QSTE340TM steel can reduce
the weight of structures, which is important for automotive parts such as chassis, suspensions,
and body components. It has good fatigue resistance, which ensures durability in harsh
environments. The chemical composition of the steel includes manganese, silicon, phosphorus,
sulfur, and other alloying elements that give it the required mechanical properties [
Machine learning methods allow us to model the fatigue life of QSTE340TM steel and effectively predict the material's durability under cyclic loading. By applying machine learning algorithms, a large amount of experimental data can be analyzed automatically and the relationship between various parameters affecting material properties can be determined
Neural networks allow us to model the fatigue life of QSTE340TM steel and effectively
predict crack growth in the material under cyclic loading. Functional dependencies were
modelled for experimental data obtained in [
The first 80% of the load cycles were used for the training, testing, and validation process. The accuracy of crack length prediction as a function of N, R, and Rol was tested on the data of the next 20% loading cycles. During the training process, the dataset was divided into 3 parts: training, testing, and verification. The training, testing, and validation samples contained 1791 items, 80% of which were randomly selected for the training sample, 10% for the validation sample, and 10% for testing and evaluating the model's prediction quality. The forecasting error was calculated using the formula for the mean absolute percentage error (MAPE): 1 n |attreuste (i )−atpersetd . (i )| MAPE=100 % ∙ ∑ n i=1 |attreuste (i )| where n is the size of the test dataset, attreuste (i ) is the true value of the crack length in the test dataset, atpersetd (i ) is the predicted value of the crack length in the test dataset.
The Multi-Layer Perceptron (MLP) neural network [
As can be seen from Figure 1, the model is highly accurate, as almost all points lie along the bisector of the first coordinate angle, which means that the predicted values are almost identical to the experimental ones. The prediction error calculated by MAPE is only 0.34%. This indicates that the neural network effectively models fatigue crack growth in QSTE340TM steel, providing accurate predictions with minimal deviation from the actual values.
To test the accuracy of crack length prediction as a function of N, R, and Rol, data from the next 20% of loading cycles were used. This data was removed at the initial stage and was not used to test the model accuracy. Figure 2 shows the relationship between the experimental crack length values atrue and the predicted values apred, obtained for a stress ratio R = 0.1 at constant amplitude and with a single overload with overload factors Rol = 1.5, 2.0.
As can be seen from Figure 2, the predicted crack lengths are very close to the experimental values, which is also confirmed by the low value of the MAPE prediction error (Table 1). c) Rol = 2.0
With a load factor of R = 0.3, the model showed high forecasting accuracy, which is also confirmed by the low value of the MAPE forecasting error (Table 2). c) Rol = 2.0
At stress ratio R = 0.5, the predicted crack lengths are also very close to the experimental ones, which is confirmed by the low value of the MAPE prediction error (Table 3).
Figure 5 shows the dependence between the experimental values of the crack length a and the predicted values obtained for a load factor of R = 0.7 at CA and Rol = 1.5, 2.0. a) CA
Similarly to the previous cases, at a load factor of R = 0.7, the predicted crack lengths are quite close to the experimental ones. The values of the MAPE prediction error are given in Table 4.
The obtained results demonstrate the high generalization capability of the model and its effectiveness in reflecting the real fatigue crack growth under cyclic loading.
The crack length a was predicted as a function of the number of load cycles N for four stress ratio R = 0.1, 0.3, 0.5, and 0.7 at a constant amplitude and overload factors Rol = 1.5, 2.0 by a neural network. The neural network, trained on experimental data, is able to predict the crack length a based on the input parameters, thus providing sufficiently accurate predictions for assessing the fatigue life of QSTE340TM steel.