PSI - Issue 5

Piotr Nazarko et al. / Procedia Structural Integrity 5 (2017) 460–467 Nazarko and Ziemianski / Structural Integrity Procedia 00 (2017) 000 – 000

464

5

a)

b)

Fig. 4. (a) Signals selected from the load history and (b) projection of two first principal components related to all signals received from the single bolt under cyclic load test.

In this case, three different ANNs training scenarios were adopted. First of all, patterns for testing and validation were selected with constant distribution. It means that every third signal was chosen for testing, one third of them was used for validation, while the others patterns have formed learning set. One of the best obtained ANNs training results were shown in Fig. 5. A mean testing error was equal to 1.88 kN (less than 2%) with standard deviation of 1.64 kN. a)

a)

b)

c)

Fig. 5. Results of ANNs training based on cyclic loads of a single bolt (constant patterns distribution): (a) learning error, (b) testing error, (c) validation error.

The next two scenarios assumed separation of patterns to investigate ANN's generalization ability. In the first of them it was assumed that the training set is separated to loading and unloading stages. A first set was then used for ANNs learning and the second for testing purposes. The last case concerns the situation where testing set consisted of patterns only from the third cycle. The other three cycles were used for ANNs learning. The obtained testing results from both simulations are shown in Fig. 6. Each unloading cycle on the first plots was marked with different colour, while on the second plot two colours are related to the loading and the unloading stages of the third cycle. In both cases the prediction errors are bigger than previously. In addition, these results indicate that signals recorded at the same load level differ from each other depending on the operating phase. Although the generalization ability of the ANN trained have decreased, the force identification errors remained at a relatively low level of 7.2 and 3.5%, respectively.