以 模 式 方 法 為 基 礎 之 程 序 錯 誤 診 斷

Abstract

The focus of this dissertation is on developing novel model-based approaches for additive and multiplicative fault diagnosis (FD). The identified process diagnostic models can be extended to have varying fault diagnostic capabilities, from simple fault detection to detailed fault isolation and identification. Some frequently used multivariate statistical methodologies for FD are reviewed, and their major limitations in fault isolation are demonstrated. Novel modifications of conventional statistical techniques are proposed, and extended to monitor time dependent processes and to isolate some specified faulty types under open-loop conditions. Some basic static and dynamic process model identification approaches are reviewed. An efficient model identification method by merging PLSR sub-models is presented; and a numerical application is used to illustrate the practicality of this method. An alternative decentralized FD scheme is also proposed based on the sub-models and merged global model. Moreover, the parameter variance and covariance structures are investigated analytically for dynamic process representation. A novel and unified sensor FD approach is constructed to arbitrary multiple sensor failure scenarios. Based on the proposed methodology, the faulty sensors can be easily detected, isolated and identified. A variety of parameter similarities for dynamic processes are defined based on the derived parameter variances. With the use of these similarities, the multiplicative faults of processes can be detected and isolated. For some multiplicative faults, e.g. changes in gain and deadtime, the faulty parameter can be specified, and the fault magnitude can be identified. Illustrative case studies are included to demonstrate these theoretical ideas in this thesis.