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

22 6

shows the comparison of prediction errors by machine learning methods, in particular, random forests and neural networks.

0 1 2 3 4 5 6 7 8

Wd-N Δε -N

MAPE, %

RF

NN

Machine learning methods

Fig. 7. The comparison of prediction errors by different methods of machine learning.

The parameters of the neural network method are given in Table 1. This table contains the topology, the respective learning algorithm, the error function, and the activation functions of the hidden and output layer. Table 2 presents the parameters of the random forest method.

Table 1. The parameters of neural network

Function of hidden activation Tangential Tangential

Function of output activation Exponential Tangential

Error function

Name of network

Algorithm of learning

Dependencies

W d - N Δ ε - N

MLP 1-4-1 MLP 1-8-1

BFGS BFGS

SOS SOS

Table 2. The parameters of random forest method Dependencies

Number of trees

W d - N Δ ε - N

520

1000

In particular, there was applied the neural network study algorithm of Broyden  Fletcher Goldfarb – Shanno (BFGS) in order to build the dependencies of dissipated energy and strain range on the number of loading cycles of NiTi alloy. The sum of error squares was chosen as the error function. 4. Conclusions There were built the dependencies of dissipated energy and strain range on the number of loading cycles by means of random forest and neural network methods. The obtained prediction results are in good agreement with the experimental data. The lowest errors, namely, 3.9 and 7%, were obtained for the dependencies of dissipated energy and strain range on the number of loading cycles by means of the method of random forest in all test sets. The methods of machine learning, for instance, random forests and neural networks, are powerful tools that can be used to assess the functional properties of NiTi alloy.