PSI - Issue 48

3

Oleh Yasniy et al. / Procedia Structural Integrity 48 (2023) 183–189 Author name / Structural Integrity Procedia 00 (2019) 000 – 000

185

One of the powerful methods of machine learning is random forests. This ensemble algorithm uses the concepts of bagging and random subspaces, and the basic algorithms are decision trees by Alpaydin (2010). In the regression problem, their answers are averaged; in the classification problem, the decision is made by majority vote (Fig.2).

Fig. 2. Algorithm of random forests

A tree is built, as a rule, until the sample is exhausted (until representatives of only one class remain in the leaves). Still, in modern implementations, some parameters limit the tree's height, the number of objects in the leaves, and the number of things in the subsample at which the splitting is. 3. Results and discussion The stress-strain diagram of the 6061-T651 aluminum alloy is predicted by machine learning methods according to the experimental data obtained in the paper by Aakash et al. (2019). In the learning process, the data set was divided into two unequal parts - training and test samples. In addition, the stress-strain diagram was divided into two regions, that is, linear and non-linear, to improve the prediction quality. As a result, two networks were built using different machine learning methods. The sample contained 2018 elements in the linear region and 643 elements in the non linear region, of which 80% were randomly selected for the training sample, and 20% were left to assess the quality of the prediction. It is important that stress and temperature were chosen as the input parameter, while deformation is the output parameter. Several numerical experiments were conducted using input-output pairs to obtain the best algorithm architecture. It is found that the obtained results are in good agreement with the experimental data. The prediction error was calculated using the Mean Absolute Percent Error (MAPE) formula: = 100% ⋅ 1 ∑ | − | | | =1 (1) By using the methods of machine learning, we plotted the dependences of the experimental data on the predicted values of the linear region (0.01-0.08) of the strain for different temperatures (Fig. 3).