PSI - Issue 72

Oleh Yasniy et al. / Procedia Structural Integrity 72 (2025) 181–187

185

Fig. 3. The predicted and experimental values of the crack length, obtained by method of neural networks

Fig. 4. The predicted and experimental dependences of the crack length on the number of loading cycles for R = 0.03, 0.1, 0.3 obtained by method of neural networks

Fig. 5 shows the dependence of crack length on the number of loading cycles , using three different data sets, for example, training data marked with yellow dots. This is the data on which the neural network was trained. It clearly shows the basic pattern of increasing crack length with increasing load cycles. Green dots indicated unseen data. This data was not used in the training process but to validate the model. They demonstrate the consistency between the actual crack length relationship and the model prediction. In addition, the NN Prediction is the blue line that shows the modeling results of the neural network. This line represents the network's prediction for the entire cycle region, including the unreleased data. The neural network generally demonstrates high accuracy in predicting the crack length . Predicted values (blue line) almost coincide with the real values from both sets (yellow and green dots). The absence of significant deviations between the trained and unseen data confirms that the model has not overfitted and generalizes well. In addition, Fig. 6 shows that the dependence of the crack length on the number of cycles is nonlinear. In the early stages (low values of ), the crack growth is slow, but with increasing cycles, the process accelerates, which is typical for fatigue crack development.

Fig. 5. The predicted and experimental values of the crack length at R = 0.03, obtained by the method of neural networks

Fig. 6. The predicted and experimental dependences of the crack length on the number of loading cycles for R = 0.03, obtained by the method of neural networks