Neural Network Architecture Search Algorithm for Technical Object State Prediction

The paper presents a new algorithm for automated search for neural network architecture for predicting the state of technical objects. Prediction is a key element of monitoring, allowing us to prevent accidents and improve the reliability of complex systems. The proposed approach combines the methods of Cartesian genetic programming, multi-criteria evolutionary algorithms and Bayesian optimization. This algorithm adapts neural network architectures to different types of data, speeds up the search for a suitable model configuration, and improves forecast accuracy. The developed algorithm was implemented as a standalone program in the Python programming language using the TensorFlow and Keras libraries. The effectiveness of the developed approach has been tested in practice in predicting the remaining service life of bearings, a turbojet engine, and a Li-ion battery. The AutoKeras library was used to compare the results. The study showed that the use of the developed algorithm significantly improves the performance quality of neural network models compared to the models obtained using the AutoKeras library. The algorithm demonstrated high accuracy of predictions, stability to various conditions, and potential for application in monitoring systems of technical objects.