Vysotska, V.Lytvyn, V.Vladov, S. I.Владов, С. І.Yakovliev, R. P.Яковлєв, Р. П.Volkanin, Ye. Ye.Волканін, Є. Є.2024-08-282024-08-282024Neural network method for identifying potential defects in complex dynamic objects / Victoria Vysotska, Vasyl Lytvyn, Serhii Vladov, Ruslan Yakovliev, Yevhen Volkanin // CITI’2024: 2nd International Workshop on Computer Information Technologies in Industry 4.0 ( Ternopil, Ukraine, 12-14 June 2024). – Ternopil, 2024. – Vol. 3742. – Paper 4. – URL: https://ceur-ws.org/Vol-3742/paper4.pdfhttps://ceur-ws.org/Vol-3742/paper4.pdfhttps://dspace.univd.edu.ua/handle/123456789/22007Робота присвячена розробці нейромережевого методу ідентифікації потенційних дефектів складних динамічних об'єктів, таких як, наприклад, турбовальні двигуни вертольотів.The work is devoted to the development of a neural network method for identifying potential defects in complex dynamic objects, such as, for example, helicopter turboshaft engines. The proposed method is based on the use of the Transformer model, consisting of the encoder, decoder, positional encoding, and attention mechanism, instead of a generalized regression neural network. A modification of the ReLU activation function in the form of Smooth ReLU is proposed to make it smoother and more continuous, which leads to improved convergence and training stability. The analysis of the derivatives of the ReLU and Smooth ReLU functions showed that Smooth ReLU solves the problem of “dead neurons” by providing a non-zero gradient for all input data values, including negative ones, which ensures more stable training of neural networks and prevents neurons from stopping updating due to a zero gradient. As a neural network implementation of the Transformer model, the use of a graph neural network is proposed, the key advantage of which is its ability to model more complex dependencies and relationships between input data elements, which increases the efficiency of training and improves the quality of prediction in sequence processing tasks. As a neuron activation function, the use of a cross-entropy loss function between actual and predicted probability distributions has been proposed, the key advantage of which is its ability to provide efficient training of a classification model by minimizing the discrepancy between predicted and real class probabilities. The results of the computational experiment showed that the proposed method demonstrated almost 100 % accuracy in determining potential defects, such as the possible formation of cracks (burnouts) in the combustion chamber of helicopter turboshaft engines due to the predicted decrease in the gas temperature in front of the compressor turbine.Работа посвящена разработке нейросетевого метода выявления потенциальных дефектов в сложных динамических объектах, таких как, например, турбовальные двигатели вертолетов.enУкраїнаpublikatsii u Scopuspublikatsii u WoSneural networktransformer architecturegraph neural networktrainingidentifying potential defectshelicopter turboshaft engineactivation functionadaptive training rateнейронна мережатрансформаторна архітектураграфова нейронна мережанавчаннявиявлення потенційних дефектівтурбовальний двигун гелікоптерафункція активаціїадаптивна швидкість навчаннягелікоптервертольотhelicopterArticlehttp://orcid.org/0000-0001-8009-5254http://orcid.org/0000-0001-6417-3689http://orcid.org/0000-0002-9676-0180http://orcid.org/ 0000-0002-3788-2583http://orcid.org/0000-0003-3507-1987