Dissipativity Based Fault Detection and Diagnosis for Process Systems

Abstract

Modern industrial processes are often operated close to the design constraints, and this can make the plants susceptible to faults. Failures in process components can lead to a considerable reduction in the efficiency of the process and affect the quality of the product. In some cases, faults may even result in fatalities. Therefore, fault detection and diagnosis (FDD) is becoming an essential part of modern control systems. In this thesis, a novel fault detection and diagnosis scheme based on dissipativity theory is developed. The dissipativity properties represent the features of process dynamics, which will change when certain faults occur. Intuitively, faults can be detected and diagnosed by detecting changes in the dissipativity properties of the process.

Quadratic difference forms (QdF) are used to represent supply rates and storage functions in the dissipativity based analysis. This can provide a tight description of the system dynamics. These dissipativity conditions are represented as quadratic functions of the input/output trajectories of the process, which do not require observers for online implementations and lead to a simple approach.

The main development of this approach contains three parts. Firstly, dissipativity properties are used for robust fault detection. The dissipativity property is adopted to describe the dynamic features of system uncertainty. A fault is detected if the model uncertainty is sufficiently large such that the dissipation inequality is violated. The uncertainty bound can be estimated either from system identifications or robust control design methods, which leads to a simpler design procedure. The proposed approach does not require detailed faulty model structure.

The second aspect is the design of a fault diagnosis scheme using dissipativity properties. For a given process, dissipativity is not a unique property, with different supply rates reflecting different dynamics features. In this approach, the dissipativity of a process is ``shaped'' such that it is fault-sensitive (i.e., no longer valid when faults occur) and fault-selective (i.e., no longer valid when one particular fault occurs). These dissipativity properties are determined offline by solving an optimization problem with linear matrix inequality (LMI) constraints.

Furthermore, preliminary studies on dissipativity based fault diagnosis based on plant data was carried out. The concept of dissipative trajectories is introduced. The features of process behavior are captured by the dissipativity of the process input/output trajectories. A fault diagnosis method using process data is developed, and a training algorithm is developed to search for the dissipativity property that is sensitive to certain fault.

The fault detection and diagnosis methodologies proposed in this thesis are illustrated using the examples presented in the respective chapters. The simulation results show the effectiveness of the proposed approach.