PSI - Issue 28

Oleh Yasniy et al. / Procedia Structural Integrity 28 (2020) 1392–1398 Oleh Yasniy et al. / Structural Integrity Procedia 00 (2019) 000–000

1397

6

a

b

225 235 245 255 265 275 285 295 305

300

Exper Pred

Exper Pred

290

280

270

260

 , MPa

 , MPa

250

240

230

0,003

0,008

0,013

0,018

0,003

0,008

0,013

0,018

 (  i ) , mm/mm

 (  i ) , mm/mm

c

d

225 235 245 255 265 275 285 295 305

225 235 245 255 265 275 285 295 305

Exper Pred

Exper Pred

 , MPa

 , MPa

0,003

0,008

0,013

0,018

0,003

0,013

 (  i ) , mm/mm

 (  i ) , mm/mm

Fig. 3. Predicted and experimental dependences of the jump-like strain on the tensile stress obtained by the methods of (a) neural networks, (b) boosted trees, (c) support-vector machines, and (d) k -nearest neighbors.

It was discovered that the method of neural networks gives the least prediction error equal to 6.9%. The errors of the methods of boosted trees, support-vector machines and k -nearest neighbors are 9.1, 7.7, and 9.4% in the test set, respectively. In this study, NN were the most accurate among the applied machine learning algorithms.

5. Conclusions The obtained results show that method selection and dataset are important for the high accuracy of the method. In addition, methods of supervised learning usually show better results than the unsupervised learning methods. It was found that the predicted data are in good agreement with the experimental data, in particular, the accuracy of the NN prediction is 93.1%, which is the largest among the applied methods. The performance of the methods of boosted