Fault tolerance in non‐linear systems: A model‐based approach with a robust soft sensor design

Abstract

A novel multiple Kalman filtre (KF)-based scheme is proposed, as a generalisation of conventional gain-scheduling techniques, for fault diagnosis and tolerance in a large class of multiple-input and multiple-output non-linear systems. The outputs are corrupted by unknown stochastic disturbance and measurement noise. A reliable and computationally efficient, two-stage identification of a piecewise linear parameter-varying Box–Jenkins dynamic model that better approximates the non-linear system, at each operating point, and the design of the associated KFs are proposed. Novel emulators, whose induced parameter changes mimic likely and predictive operating scenarios, are used to provide an accurate model identification and robustness to noise, disturbance, non-linearity errors and model perturbations. These crucial emulators generate missing representative data, aid predictive analytics, and improve the reliability and accuracy of the identified KF model. A novel formulation of the KF is used for fault isolation, and the Bayes strategy is used to isolate difficult-to-detect incipient faults in noisy environments. The proposed scheme leads to the design of a novel robust soft sensor aimed at replacing the maintenance-prone hardware sensor in practical applications including product quality assessment, performance monitoring, condition-based maintenance, fault diagnosis and fault-tolerant control. The proposed soft sensor was successfully evaluated on simulated and laboratory-scale physical control systems.