A discrete event systems approach to failure diagnosis.

Abstract

Failure diagnosis in industrial systems is a crucial and challenging task. Accurate and timely diagnosis of system failures can enhance the safety, reliability, availability, quality, and economy of industrial processes. The complex nature of today's technological systems and the limited availability of sensors make failure diagnosis a difficult problem and motivate the need for automated diagnostic mechanisms. We propose in this thesis a new approach to failure diagnosis in the framework of discrete event systems (DES). This approach is most appropriate for "sharp" failures that involve significant changes in the status of the system components but do not necessarily bring the system to a halt, and is applicable not only to systems that fall naturally in the class of DES, such as manufacturing systems and communication networks, but also to systems traditionally treated as continuous variable dynamic systems, such as process control systems and heating, ventilation and air conditioning (HVAC) units. In this thesis, we first present a systematic methodology to build discrete event models of industrial systems for the purpose of failure diagnosis. Starting from finite state machine (FSM) models of the individual system components and from discrete valued sensor maps, we present a procedure to generate a global model for the system that captures the interaction among the components and also incorporates in it the sensor information. This model accounts for both the normal and the failed behavior of the system. Next, we introduce the notion of diagnosability of a system. Simply speaking, a DES (or the language it generates) is diagnosable if it is possible to detect, with a finite delay, occurrences of the (unobservable) failure events from observed event sequences. We present necessary and sufficient conditions for a language to be diagnosable and show how an FSM called the diagnoser can be used to check if a given system is diagnosable. The diagnoser is built from the FSM model of the system and can be thought of as an extended observer for the system; in addition to estimates of the system state that an observer provides, the diagnoser provides information about past occurrences of failures in the system via labels attached to the state estimates. Given a diagnosable system, we next show how the diagnoser can be used to perform on-line failure diagnosis. Finally, given a non-diagnosable system, we present a systematic procedure for the design of diagnostic controllers for the system. These controllers ensure that the closed-loop system, in addition to satisfying other control objectives, is also diagnosable. We illustrate the theory developed in this thesis by means of several examples, primarily of HVAC systems. We conclude the thesis with a comparison of the proposed approach with other approaches to this problem that have appeared in the literature, and also outline several directions for future work.