A CONVOLUTIONAL NEURAL NETWORK-BASED APPROACH FOR FAULT DETECTION AND CHARACTERIZATION IN CONTROL SYSTEMS

Tyryshkin Sergey Yuryevich

Traditional methods of fault diagnostics in complex control systems are often focused solely on mechanical equipment failures. With the development of intelligent technologies and the Industrial Internet of Things, interconnections between control devices have become increasingly complex, while operating conditions have grown more diverse. This makes traditional approaches less effective, complicating the precise identification of problem areas. The present article is devoted to developing a new method for detecting and analyzing faults in control systems using convolutional neural networks. The research aims to develop an algorithm capable of accounting for both spatial and temporal characteristics of signals generated during the operation of control systems. This will significantly improve diagnostic accuracy and accelerate the process of identifying faults. To achieve this goal, we employ a hybrid model based on 2D-CNN-LSTM, combining a two-dimensional convolutional neural network (CNN) with a long short-term memory recurrent neural network (LSTM). This approach allowed us to uncover hidden patterns in signals characteristic of various types of faults, including defects that arise in the dynamics of production processes. The algorithm was tested on real data from industrial facilities, demonstrating high efficiency in determining critical moments and locations of faults. The results indicate the potential applicability of the proposed approach in modern manufacturing environments, where a high degree of automation and diagnostic precision are required. The developed method could serve as the foundation for creating new monitoring systems and early warning systems for faults, which would substantially enhance the reliability and safety of complex control systems.

Keywords:control, fault, diagnosis, industrial system, neural network, data, discretisation.