Study on Applications of Advanced Fault Detection Methods to Complex Automatic Control Systems

Due to the fast development of information technology and computer science in the past decades, the automatic control systems have been widely applied in the industrial systems, which has concurrently raised higher demands for production safety. In general, the focus of safety in production mainly falls into two categories: i) malicious human-induced attacks, ii) naturally occurring faults within systems. This thesis is dedicated to investigating advanced fault detection methods for addressing the aforementioned challenges. First, two novel kernel-attack detection methodologies in Linear Time Invariant (LTI) Cyber-Physical Systems (CPSs) are studied. These methodologies leverage the eigen dynamic characteristics of both the system and controllers to encrypt control and residual signals within the feedback loop. The results demonstrate that residual encryption based on the Youla parameterization framework achieves robust kernel-attack detection while preserving operational integrity under adversarial conditions. Second part of this work studies a projection-based fault detection methodology for LTI systems. This approach projects the input-output pairs to the image space of system, where the resulting projection residuals can be interpreted as errors caused by faults and uncertainties. By computing the upper bound of system uncertainties with the corresponding method, an adaptive threshold for fault detection can be established. Both theoretical derivations and experimental validations demonstrate that this approach exhibits superior fault detection capability compared to the observer-based fault detection methods. Finally, this work investigates a performance-based fault diagnosis method for a nonlinear (affine) system. By solving the Hamilton-Jacobi Equation (HJE) and designing the feedback and feedforward controllers, the system is transformed into an inner system with lossless properties. Based on Control-Theory-Informed-Neural-Network (CTIML), the primary contribution of this part is the proposal of an online method for solving the HJE, while the lossless property will be continuously monitored online for faults and potential performance degradation detection. Compared to traditional Physics-Informed Neural Networks (PINNs), the proposed approach enables real-time acquisition of system states and online application.