(19c) Distributed Output-Feedback Fault Detection and Isolation of Cascade Process Networks

Authors: 
Yin, X. - Presenter, University of Alberta
Liu, J., University of Alberta
Due to the demands for improved efficiency and profitability, modern process networks that are comprised of tightly interconnected units have attracted substantial research attention within the chemical industry. These process networks are normally featured by complex and highly integrated structures, which pose significant challenges to the design of satisfactory automated control systems. For these process networks, centralized approaches are not favorable due to the issues of computational complexity, fault tolerance, etc. As one common architecture in process control, decentralized control systems may not provide desired control performance. Specifically, the subsystem interaction is completely neglected or treated in a relatively conservative way, which may lead to much degraded control performance compared to the centralized counterpart. Alternatively, distributed control frameworks can give overall improved performance by enabling information exchange among local regulators, and serve as a very promising candidate for automatic control of modern processes.

However, the control performance of a distributed control system is subject to potential faults which are commonly encountered in complex and integrated process networks. Particularly, faults (e.g., actuator faults and sensor faults) in some key components may result in control performance degradation, a higher risk of system failures, reduction in economic profits, or even catastrophic impacts on operation safety as well as the environment. Since these undesired consequences may propagate from one single subsystem to the entire plant very quickly in modern process networks, fault detection and isolation (FDI) is especially important for these plants. During the past two decades, we have witnessed rapid progress in FDI. Model-based Fault detection, diagnosis, isolation and fault tolerant control issues have been studied within distributed frameworks to account for spatially distributed nonlinear processes with medium to large sizes. However, most of these results are based on the availability of measurements of the entire system states, which may not be satisfied in many applications. Therefore, state estimation based distributed FDI approaches are highly desirable from an application perspective.

In this work, we focus on cascade nonlinear processes that have been widely encountered in various manufacturing processes. We propose a distributed output-feedback FDI approach accounting for both actuator and sensor faults. First, a distributed state estimation network is designed for the cascade process based on an approach developed in [1]. It is assumed that for each subsystem, there already exists a decentralized estimator of exponential convergence when its subsystem interaction is at steady-state values. A compensator is designed for each subsystem (except the first subsystem) to better handle subsystem interaction. Each developed compensator is then connected to the associated decentralized estimator to form an augmented estimator. The estimators for the subsystems are required to communicate continuously to exchange subsystem state estimates and the output measurements. In a fault-free case, we show that the estimation error of the developed distributed state estimation scheme converges to zero using the integral input-to-state stability (iISS). Subsequently, we design a state predictor to generate two sets of reference state predictions for each subsystem. Each predictor is re-initialized with the subsystem estimate given by the distributed state estimation system after every prediction horizon in two paralleled manners. For each subsystem, we also design a subsystem residual generator, which gives two residual signal sequences for FDI. The thresholds on the residuals are also characterized. A distributed output feedback FDI mechanism is established. A fault can be detected and further isolated by examining all the residuals of the subsystems and the associated thresholds. It is demonstrated that the proposed approach can address FDI problems for medium-sized systems. The proposed method also has the potential to handle processes with larger sizes.

[1] X. Yin, J. Zeng, J. Liu. From decentralized to distributed state estimation. In proceedings of American Control Conference, Seattle, WA, May 2017.