PSI - Issue 59

Oleh Yasniy et al. / Procedia Structural Integrity 59 (2024) 271–278 Author name / Structural Integrity Procedia 00 (2023) 000 – 000

276

(2010), Smola and Vishwanathan (2010). The most prevalent model in neural networks is the multilayer neural network, consisting of layers of neurons extending from the input layer to the output layer (Fig. 5) by Haykin (1999).

Fig. 5. Full connected feedforward network with one hidden layer and one output layer by Haykin (1999).

The AMg6 aluminium alloy stress – strain diagram was modelled using the machine learning method according to the experimental data obtained by Fedak (2003). In the process of learning, the dataset was divided into two unequal parts – training (70%) and test (30%) samples. The input parameter was the stress value sp ( ai ), while the strain jump D e ( ai ) was chosen as the output parameter. It is important to note that the models built by ML methods predict the data that was not used in the training sample with high accuracy. It has been investigated that forecasting of jump like deformation correlates well with the experimental data. 4. Results and discussion By using the neural network method of machine learning, there were plotted the dependences of the experimental jump-like strain (   (  i ) true ) on the predicted values of   (  i ) (   (  i ) prediction ) (Fig. 6). The dependences between the value of jump-like increments of strain and the corresponding tensile stress   (  i ) –  p (  i ) are shown in Fig. 7. It was found that the neural networks method gives the lowest prediction error, which is equal to 6.9%.