In an operating chemical plant, product quality is maintained by monitoring process variables and controlling their fluctuations within a desired range. When operating conditions vary outside these design limits, not only is the product quality in jeopardy, but if left uncorrected, these variations could result in a catastrophic event such as an explosion, fire, or the release of toxic chemicals. The objective of this project is research leading to the development of computer-based systems for on-line diagnosis of process malfunctions to be used by plant operators. Computer-automated diagnosis has been studied in many domains. However, several features distinguish the chemical plant diagnosis problem: (1) the state description involves time-dependent continues and discrete variables, (2) there are complex stream topologies with feedback and feedforward flows of material, energy and information, (3) complex dynamics are associated with malfunction response, (4) malfunctions include chemical effects such as side reactions, mechanical faults such as valves sticking and blockages in pipes, and information processing faults such as sensor vary, possibly producing a family of response types for the same malfunction, (6) the diagnosis is based on on-line measurements with fallible sensors, parameters are measured. This research will create new mechanisms for joining plant-specific and general behavioral knowledge with diagnostic methodology. The PI plans to do research in four subareas: (1) Create and refine formal representation of the objects involved in diagnosis and operator advising, including: a. Physical entities such as plant units, b. Abstract objects such as constraints on normal behavior and system intended functions, and c. Conceptual objects such as malfunction hypotheses, inferred malfunctions, abnormal events, and explanations. (2) Based on these representations, develop a library of unit operation models and a mechanism to link these components to form models of the causal topology and intended functions of specific plants. (3) Develop a mechanism for interpreting the basic flowsheet model and identifying potentially significant real-time events, their causes and relationships. (4) Conduct research on dynamic reasoning leading to the development of a general inference mechanism for interpreting real-time event sequences. Dynamic mathematical models of process plants incorporating realistic features and a board range of malfunctions will be developed. This will permit testing of the representations and methodologies, and the investigation of process/model mismatch, the effects of noise and of out-of-order alarms. Architecture will be created into which plant- specific knowledge can be incorporated by industrial design engineers.
