Application of neural network technologies to solving the problem of materials classification of two-layer structure by hardness parameter

. The focus of the article is on utilizing neural networks, a form of artificial intelligence, to address the task of categorizing mechanical characteristics of diverse materials. Brinell hardness was chosen as the considered characteristics of materials for the study, the choice of this property was justified. The study simulates a finite element model of the impact of an indenter on a two-layer structure in an Ansys environment. The difference in the properties of the construction materials is determined by the application of a strengthening coating or the accumulation of multiple defects in the surface layer. Using the model, a set of data for training a neural network was obtained. As part of the experimental part, the structure of the neural network was developed, its hyperparameters were adjusted. A comparative analysis is presented that examines two different methods for neural network calculations based on the nature of the input impact.


Introduction
To date, any industry activity has high requirements for the informatization of this sphere. Constantly developing information technologies make it possible to modernize production processes, methods and means of their implementation, and are also a source of increasing the efficiency of a particular area. Artificial intelligence tools have been very popular in recent years in terms of information technology. The most promising direction of artificial intelligence is neural networks. Neural networks are capable of addressing a wide range of tasks, including but not limited to classification, recognition, prediction, data approximation, and many more. A typical data mining task for a neural network is a classification task. The popularity of the classification method is explained by the simplicity of implementation and easily interpreted results. This area has been extensively explored in the scientific community, with a considerable focus on technical industries, as evidenced by numerous research papers [1][2][3][4][5]. The authors set and solve classification problems in the field of transport systems, road construction, and structural mechanics.
One approach to building classification systems is to rely on image processing and analysis to recognize and categorize images. So, in the study [6], the issue of improving intelligent transport systems is considered by the example of solving the problem of classifying ships of various types. Classification is based on processing by a neural network of images obtained in different weather conditions, from different distances and shooting angles, as well as from different international and sea harbors. A convolutional deep learning neural network with a data set from Kaggle was selected for calculations. The author has proposed a new method of image classification based on the improvement of the resnet 152 architecture. A comparison of the resulting values of the developed algorithm and existing methods of classification of water transport is presented. It is established that the proposed method has better showing.
In terms of traffic monitoring, researchers in the article [7] carried out a classification of vehicles whose images were collected from traffic cameras and dashboard cameras. The classification algorithm that was suggested is grounded on a self-produced data set utilizing a pre-existing VeRi set. The algorithm demonstrated the accuracy of calculations, which is considered a high indicator.
Classification problems solved by a neural network in the construction field deserve special attention [8][9][10]. First of all, they are aimed at developing systems for assessing the strength, rigidity, and stability of the structure. A special place is occupied by studies of the stress-strain state of structures under the influence of shock loads [11,12]. In the article [13], the classification analysis of the properties of such materials as mild steel, stainless steel and aluminum of various thicknesses was carried out by means of neural networks. Experimental data were obtained by the authors using modal tests, specifically the vibration method. The input data for training the neural network were the natural frequencies of the material, since the natural frequency varies depending on the thickness and type of material, which makes it quite easy to classify mechanical properties. The research results have a high level of computational accuracy, the highest classification index is almost 100%. In scientific papers [14,15], high accuracy rates of the neural network have also been achieved in solving classification problems.
Concluding from the above analysis, it can be inferred that utilizing neural network technologies as a means for categorizing material properties is a feasible, relatively straightforward, and impactful approach, thereby underscoring the relevance of this investigation.
The aim of the study is to implement a neural network capable of classifying the properties of materials of layers of a two-layer structure. It should be noted that the thickness of the upper layer of the structure under consideration does not exceed 1 mm, which makes it possible to talk about a spray-coated structure.
To solve the problem of classifying the properties of materials, the Brinell hardness parameter (HB) is selected. The reason for this is that hardness serves as a universal parameter that is connected with various other mechanical properties, including strength, tensile strength and endurance. Knowing the hardness, with the help of simple transformations, you can always go to the required parameter.
There are 5 classification groups presented in Table 1. Meanwhile, the hardness values of the upper layer are fixed, and the characteristics of the base vary in the range of 200-650 HB.

Materials and methods
As part of the research, the impact of a conical-shaped indenter on a two-layer structure in ANSYS was simulated (Figure 1). The classical models presented in the theory of elastic-plastic deformations are used in numerical calculations. To describe the behavior of the material in the field of plasticity, the option of Multilinear Isotropic Hardening (MISO) was chosen. This model uses the von Mises plasticity criterion.
Based on the data in Fig. 1, it is possible to observe fluctuations associated with the transition to the plasticity zone, elastic vibrations occur below this zone, which fade with time.
It describes the oscillatory process of a dynamic system with two masses: 1 -the mass of the indenter; 2 -the mass of the striker, coupled with linear springs with stiffness 1 и 2 respectively [16].
To prepare for the modeling process, deformation diagrams were constructed for each hardness group (Fig. 2). According to Hooke's law, the modulus of elasticity is calculated by the formula:   = 20 values). The simulation data was processed using a fully connected neural network. Preliminary data preparation, normalization, and correlation analysis were carried out. During the simulation, the velocity values obtained (as shown in Fig. 3) serve as input parameters for the neural network, which then determines if the signal belongs to any of the categories at the output. The created neural network has a relatively simple structure. It is a recurrent neural network containing 1 hidden layer. The Rely function is used as an activation function in the hidden layers of a multilayer network. The softmax function is used to represent the categorical distribution over the selected classes in the output layer of the network.  The training and test sets are randomly separated from the original data set in a percentage ratio of 80% to 20%.

Results and discussions
Within the framework of the study, 2 approaches to training a neural network were carried out: the first assumed the use of a full signal as the input of the network, and the secondthe supply of point kinematic characteristics to the input of the network. When implementing the second approach, 4 characteristic signal points were identified: the first found maximum and minimum points and two distances between them 1 and 2 (Fig. 4). Based on the results obtained, it can be concluded that it is extremely important to supplement and expand the original set of training data in order to achieve complete training.

Conclusions
In summary, the study's findings suggest that the neural network has the capability to detect variations in the stress-strain condition of materials caused by an impact. To enhance the precision of neural network calculations, a more extensive data set is required for training purposes. Additionally, it was discovered that the neural networks accuracy in determining coating hardness using point kinematic features is inferior to that obtained by using the complete time characteristic. Based on the results of practical research, it can be argued that the use of neural network technologies as a means of evaluating the mechanical properties of materials is effective.