PSI - Issue 59

Oleh Yasniy et al. / Procedia Structural Integrity 59 (2024) 17–23 Oleh Yasniy et al. / Structural Integrity Procedia 00 (2019) 000 – 000

20 4

The random forest method is the most widely used algorithm to model the data since it is quite simple and versatile. Each tree is learned on the subset of data. Therefore, the algorithm of random forest relies on the power of the crowd, and in this case, the general bias of the algorithm decreases. The advantage of this algorithm is its stability. For instance, if a new data point is being introduced into the dataset, then such changes do not greatly affect the general algorithm since the new data can significantly change one tree , though not the entire random forest. The functional dependencies were modelled for the experimental data of NiTi alloy obtained in the paper Iasnii and Yasniy (2019). In the learning process, the dataset was split into two unequal parts, namely, the training and testing set. The sample contains 719 elements, 70% of which were sampled randomly for the training set, and 30% were preserved to estimate the prediction quality. The input parameter was W d , and the number of loading cycles N was treated as the output parameter. Remarkably, the sample to predict the strain range of NiTi alloy on the number of loading cycles comprised 760 elements. Here, the input parameter was the stress range Δ ε , and the number of loading cycles N was also considered as the output parameter. The prediction error was calculated using the Mean Absolute Percent Error (MAPE) formula: ∑ | | | | (1) 3. Results and discussion The modelling of dependency of dissipated energy and strain range of NiTi alloy on the number of loading cycles using the various methods of machine learning, in particular, by neural networks and random forests. The predicted and experimental dependences of dissipated energy per cycle on the number of loading cycles are shown in Fig. 3. Particularly, the intensive decrease of dissipated energy is observed during the first 10-20 loading cycles. a b

0 0,5 1 1,5 2 2,5 3 3,5 4 4,5

0 1 2 3 4 5 6

Exp Pred

Exp Pred

Wd, MJ/m3

Wd, MJ/m3

0

200

400

600

800

0

200

400

600

800

N, cycles

N, cycles

Fig. 3. The predicted and experimental dependencies of the dissipated energy of NiTi alloy on the number of loading cycles obtained through the random forest (a) and neural network (b)

The dependencies of the experimental number of loading cycles on the predicted ones were built using the methods of machine learning (Fig. 4). The points are close to the bisector of the first quadrant, which confirms the good agreement of the predicted and experimental data, as can be seen from Figure 4. The error of the random forest method was 3.9% for the test set, and the error of the neural network method comprised 5.5%. The dependencies of the stress range of NiTi on the number of loading cycles are shown in Fig. 5. The dependencies of an experimental number of loading cycles of the strain range of NiTi alloy were built by means of random forests and neural networks (Fig. 6).