PSI - Issue 59

Victor Aulin et al. / Procedia Structural Integrity 59 (2024) 444–451

445

2

Victor Aulin et al. / Structural Integrity Procedia 00 (2019) 000 – 000

1. Introduction The technical condition of each joint, system, and unit of vehicles that undergo technical maintenance and repair (TM&R) is characterized by a set of controlled parameters (Aulin et al., 2020; Aulin et al., 2019). The values of these controlled parameters are determined during the initial diagnosis stage (Nguyen, 2019, Maruschak et al., 2016). They significantly depend on the presence or absence of defects in the components of units, systems, and assemblies (Hevko et al., 2018). If the number of TM&R complexes is a known value, then some deviations in the controlled parameters of the units, systems, and assemblies of vehicles that undergo TM&R are observed (Nguyen, 2019; Aulin et al., 2018). Such deviations from nominal values indicate the presence of defects. The design of the TM&R complex of operations for units, systems, and assemblies of vehicles is associated with the formation of the most efficient production-technical base, which ensures a significant reduction in the company's in-house production costs. To solve the problem of defect recognition in units, systems, and assemblies of vehicles during the full TM&R complex of operations, some artificial neural networks (ANNs) are used (Bodyanskyi et al., 2004; Setlak 2004). ANNs are mathematical models of an ordered set of artificial neurons that are interconnected in a specific way. The choice of the ANN method is caused by the following factors: – the ability to learn and memorize, as well as tochange the adaptive parameters of artificial neurons that make up the network; – avoidance of the process of accumulating the statistical information for calculating the probabilities of defect occurrence aimedat the optimal distribution of units, systems, and assemblies within TM&R complexes of operations; – verification of the adequacy of the constructed TM&R complexes of operations based on test samples formed during experimental research. The aim of this work is to provide the theoretical substantiation of production losses associated with the presentation of the TM&R complex of options for units, systems, and assemblies in accordance with the errors in the distribution of defects in their components. To enhance the reliability of defect recognition and prediction, an algorithm has been developed. 2. Methods and results of the experimental study When applying Artificial Neural Networks (ANNs) (Mikhaylenko et al., 2017; Tereykovsky 2007; Nuzzolo et al., 2016), the first question that arises is the choice of network architecture, which includes the number of hidden layers and the quantity of artificial neurons in each of them, related to the specific task set. From the mathematical perspective, an artificial neuron is a function of a single variable, i.e. a linear combination of all input signals with an activation function that generates the neuron's output signal (Vishnukumar 2017; Kim 2017). In its general form, the mathematical model of an artificial neuron is a well consideredadder, and can be represented as follows (Shuklin 2003):       n i i i S x w x w 1 0 0 ) ( (1) where S – is the weighted sumof the neuron's input signals; i x – the value at i -inputof the neuron; i w – is the specific weight of i- th synapse; n – is the number of inputs; 0 x and 0 w – are the values of the additional input signal and its specific weight respectively. The neuron’s output value is the function of its state: ( ) Y f S  , (2)