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

19

3

2. Material and methods One of the methods of machine learning that can be used to model and predict the properties of a material is neural networks. In particular, the structure of neural networks is tightly connected with the application of learning algorithms. Therefore, supervised learning algorithms are used to perform well with target values corresponding to certain input signals. The most common model is a multilayer neural network, which contains neurons on the layers from the input to the output (Fig. 1) by Haykin (1999). In general, there exist weighted connections in the neural network from the neurons of the given layer to the neurons of the next layer. One of the advantages of neural networks is the ability to get the result based on incomplete information or one that contains noise. As a result, this network can generalize the rules from the given cases on which it was trained and apply the obtained rules to the new data. It is known that a neural network needs the training set for building an efficient model together with its topology, learning algorithms and some other parameters. Generally, neural networks are a powerful tool for solving tasks when the data are presented appropriately.

Fig. 1. Full connected feedforward network with one hidden layer and one output layer by Haykin (1999).

The method of the random forest allows the processing of data effectively. The random forest consists of trees (Fig. 2). Particularly, the algorithm builds the set of decision trees and then averages their prediction results by Alpayndin (2010).

Fig. 2. Algorithm of random forests by Aldrich (2020).