Application of evolutionary and genetic algorithms in forming the architecture of neural network models for forecasting the state of a technical object

In order to reduce the time for eliminating critical situations in the operation of a technical facility, it is necessary to respond in a timely manner to disturbances in its functioning. This raises the problem of predicting the state of an object and possible disturbances in its operation based on the results of studying a set of controlled parameters of the object. In this study, the problem is solved based on neural network models. However, when constructing a neural network model, it is necessary to carefully select the model architecture to ensure the best accuracy of predicting the state of objects. In this paper, a new technique is proposed for the automatic design of neural network models, which consists of the sequential use of three evolutionary algorithms: the Cartesian genetic programming (CGP) algorithm for primary initialization, the multi-criteria evolutionary algorithm NSGA-II for tuning neural network architectures, and the evolutionary algorithm CMA-ES for refining the architecture parameters. To assess the quality of forecasting using neural network models, the mean absolute error (MAE) and the determination coefficient (R2) are used. Cross-validation is used to exclude the possibility of fitting the model to the optimal forecast characteristics. It allows obtaining unbiased estimates of quality metrics. To implement these methods and models in the Python programming language using the tensorflow and keras libraries, a special program was written. The objects of the study were a turbojet engine, a lithium-ion battery, and bearings. The AutoKeras library was used to compare the effectiveness of the proposed method. The study showed that the use of the proposed approach significantly improves the quality metrics of neural network models for all technical objects compared to models found using the AutoKeras library: the value of the MAE error function for all data sets during forecasting decreases by an average of 4 times, and the value of the determination coefficient increases by 1.9 times. This approach can be used by specialists to predict the technical condition of objects in various fields of technology, especially in aviation.