A knowledge-based framework for process and malfunction diagnosis in chemical plants

Diagnosis of malfunctions and abnormal operating conditions in the chemical process plant (CPP) domain is a problem that is complex, difficult to solve and important to the safe and economical operation of processes. This dissertation presents a knowledge-based approach to the diagnosis problem, that takes advantage of the human diagnostician's expertise and understanding of process behavior. The approach addresses the issues of ease of development of diagnostic systems, computational efficiency, and effectiveness of problem solving.;The contributions of this dissertation are as follows. Firstly, the overall CPP diagnosis problem is decomposed into two distinct types of problems, called fast response and slow response diagnosis. Further, the slow response problem is characterized in terms of its inputs and outputs; and a high-level, general diagnostic algorithm is presented for mapping the inputs into the outputs. Two distinct types of slow response scenarios, called process and malfunction diagnosis are identified, and the problem-solving demands of each are characterized.;Secondly, a knowledge-based computational framework in terms of generic tasks, specific to the slow response diagnosis problem, is presented. This framework is shown to consist of three generic tasks: hierarchical classification, intelligent data abstraction, and abductive assembly, each of which accomplishes a portion of the high-level diagnostic algorithm. The framework is shown to be capable of handling both the process and malfunction diagnosis scenarios, and addresses the identification of hardware malfunctions and operator error, as well as inappropriate operating parameter settings. A working, prototype diagnostic system for the fluid catalytic cracking process, built using the computational framework, is presented as proof of principle of the research.;Lastly, the dissertation presents an improved characterization of the three generic tasks in the framework as they apply to the slow response diagnosis problem. This includes extensions to the tasks as well as improved knowledge engineering guidelines. The extensions allow the framework to (1) diagnose causal interactions between malfunctions that are known a priori; and (2) augment a systematic hypothesis-driven reasoning strategy by using symptom-driven reasoning to advantage. The knowledge engineering guidelines facilitate the building of diagnostic expert systems, based on the framework, for chemical processes.