Integrated Fault Detection Using Fuzzy Soft-Sensor and Autoencoder Techniques for Nonlinear Dynamic Systems

This paper focuses on the fault detection (FD) problems for a class of nonlinear dynamic systems, particularly in scenarios where conventional methods are prone to failure. Specifically, when a Takagi-Sugeno (T-S) fuzzy model is utilized for system approximation, the residual signal generated exhibits complex non-zero dynamics even under healthy operating conditions, which leads to poor FD performance of traditional residual-based methods. To address this problem, this paper proposes an integrated method that combines a T-S fuzzy soft-sensor with an autoencoder. The method first utilizes the fuzzy soft-sensor to generate residuals, and then the autoencoder is employed to learn the residual patterns under normal operation states. Ultimately, the FD is achieved by monitoring the reconstruction error of the autoencoder, which is quantified as the squared prediction error (SPE) statistic. The final case study on a ship propulsion system validates the feasibility and superiority of the proposed FD method in detecting both actuator and sensor faults.