PSI - Issue 42

Iryna Didych et al. / Procedia Structural Integrity 42 (2022) 1344–1349 Author name / Structural Integrity Procedia 00 (2019) 000 – 000

1347

4

Fig. 3. The predicted and experimental date of the strain for T = 343,413  C obtained by method of neural networks (a) and boosted trees (b).

Fig. 4. The predicted and experimental dependences of the strain on the stress at T = 343 and 413  C obtained by methods of neural networks (a) and boosted trees (b). The parameters of the constructed neural network and boosted trees are given in Table 1 and 2, respectively.

Table 1. Neural network parameters

Function of hidden activation

Function of output activation

Name of network

Error function

Algorithm of learning

Temperature

MLP 2-9-1

BFGS

SOS

Logarithmic

Tangential

T = 343,413  C

Table 2. Parameters of boosted trees Temperature

Number of trees

2000

T = 343,413  C

In this investigation, the neural network is constructed by means of multilayer perceptron (MLP 2-9-1), which contains 2 input, 9 hidden, and 1 output neurons. Broyden-Fletcher-Goldfarb-Shanno (BFGS) learning algorithm by Fletcher (1987) is applied and the sum of error squares (SOS) is chosen as the error function by Richard (1998). The activation signal is converted by activation functions, which are determined experimentally for the neurons of the hidden and output layers and, as a result, the output signal of the neuron is obtained. The parameter of stopping the learning network is the number of epochs, which in this investigation is equal to