Combined approach to fault detection in complex technical systems based on bond-graph model

Dmitriev V.A., Marusina M.Ya. Combined approach to fault detection in complex technical systems based on bond-graph model. Scientific and Technical Journal of Information Technologies, Mechanics and Optics, 2025, vol. 25, no. 3, pp. 536–544 (in Russian). doi: 10.17586/2226-1494-2025-25-3-536-544

A new fault detection approach for complex technical systems has been developed and investigated, enabling the identification and classification of single and multiple simultaneous faults. The challenge of reliable and timely identification of both single and multiple simultaneous faults under conditions of limited access to labeled data has been addressed. A threat to the safe operation of autonomous equipment is a common challenge in the field operating conditions where traditional model-based or data-driven approaches, used individually, prove to be ineffective. This work presents a hybrid approach to fault detection. The proposed solution combines an analytical bond-graph model and a Convolutional Neural Network (CNN). The bond-graph generates residuals — the difference between values calculated based on the system physical laws and sensor measurements. The residuals are then analyzed by the CNN which is trained to detect and classify faults based on their characteristic features. Linear Fractional Transformation is employed to account for parameter uncertainties (e.g., resistance or capacitance). This approach allows combining a priori knowledge of the system physics with the capabilities of deep learning. The effectiveness of the approach was evaluated on a simulator of a hydraulic steering control system for autonomous equipment. Gaussian noise was added to the simulation to simulate real-world conditions. The experiments included incipient, abrupt, single, and multiple faults. Tests with varying amounts of training data, using sample sizes less than 128, demonstrated the higher effectiveness of the proposed hybrid approach compared to classical machine learning methods (such as Random Forest or K-Nearest Neighbors). A solution is proposed for fault detection in hydraulic control systems of autonomous equipment. The developed approach is particularly effective with limited data, making it suitable for field conditions. It allows for timely detection and classification of faults (e.g., valve leaks or solenoid valve failures), which reduces the risk of failures and ensures the safety of autonomous equipment. The results can be adapted and implemented for electrical, mechanical, and other complex technical systems.

This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License