PSI - Issue 36

Iryna Didych et al. / Procedia Structural Integrity 36 (2022) 166–170 Iryna Didych et al. / Structural Integrity Procedia 00 (2021) 000 – 000

169

4

training method was Broyden – Fletcher – Goldfarb – Shanno (BFGS) by Gurney (1997), Richard (1998), Goodfellow et al. (2016). In particular, the hidden activation function is tangential and the function of output activation is logarithmic. The stop parameter of learning network was the number of epochs, which in this study was equal to 1000. The prediction error was Mean Absolute Percent Error (MAPE):

true y y y 

n

100% 1   

(1)

prediction

MAPE

i n 

1

t

rue

where y prediction is the predicted element of sample, y true is true value of the sample element; n is volume.

3. Experimental approach Smooth cylindrical specimens of AMg6 alloy were subjected to tensile loading on STM-100 electrohydraulic machine at a temperature of 293 K. The specimens with a diameter of diameter 10 and a working part 25 mm long were manufactured by turning bars in the delivery state. The tensile loading was performed with a rate that was equal to 1.6 MPa/s. The magnitude of jump-like creep of AMg6 aluminum alloy was predicted by method of neural networks according to the experimental data obtained in Fedak (2003). While training, the dataset was divided into two unequal parts, that is, training and test sets. The dataset consisted of 89 elements. The input parameters were the stress parameter  p (  i ), while the strain jump  (  i ) was chosen as the output parameter for the training set. Whereas, the input parameters for test set were the stress parameter  p (  i ), while the magnitude of jump-like creep  p was chosen as the output parameter in order to estimate the predictions quality. It was found that the obtained models can make predictions according to the data that were not used in the training set. Hence, these results contain useful information for the investigation of their quality. The input and output parameters were normalized using the decimal log function in order to decrease the prediction error. 4. Results and discussion There were plotted the dependences of the experimental jump-like creep log10(  p true ) on the predicted values of  p log10(  p prediction ) by method of neural networks (Fig. 3).

-1,4

-2,5 -2,4 -2,3 -2,2 -2,1 -2,0 -1,9 -1,8 -1,7 -1,6 -1,5

Exper/pred

-1,6

-1,8

-2,0

log10(  p prediction ), mm/mm

-2,4 log10(  p), mm/mm -2,2

Exper_test Pred_test Exep_train Pred_train

-2,6

-2,5

-2,3

-2,1

-1,9

-1,7

-1,5

2,3

2,4

2,5

2,6

log10(  p true ), mm/mm

log10(  ), MPa

Fig. 3. Predicted log10 (  p prediction ) and experimental log10(  p true )

Fig. 4. Predicted and experimental dependences of the jump-like creep