State-space-guided neural networks for fault detection

Abstract

This article investigates the use of state-space models to enhance neural networks for fault detection in engineering systems. In modern control theory, it is well-established that a nonlinear system can be maintained at a setpoint using a linearized state-space model to approximate system dynamics. This concept is adapted to state-space-guided neural networks (SSGNNs), where a simplified state-space model provides an imperfect approximation of the system state, which is then utilized within a physics-guided neural network (PGNN) framework. By incorporating state-space model estimates into the feature space, the SSGNN can capture intricate patterns and relationships that purely data-driven models might miss. This augmented feature space allows the neural network to learn characteristic relationships between measurements and state-space model estimates, enhancing fault detection capabilities. The methodology emphasizes on guiding a machine learning model with simplified and easily discoverable governing equations while still achieving high fault detection accuracy. This study demonstrates that SSGNNs offer improved fault detection performance compared to benchmark neural networks, using both simulated and laboratory data. These findings encourage further research into hybrid physics-guided machine learning to enhance reliable fault detection in industrial systems.