Diagnostic knowledge-based systems for batch chemical processes: Hypothesis queuing and evaluation.

摘要

Knowledge-based system (KBS) approaches to diagnosis have seen widespread application in the process plant domain. Much of the research focus, however, has been toward continuous processes. In this research, recognizing that effectiveness of a diagnostic system can be maximized by taking into account the characteristics of the process, batch and continuous chemical process characteristics are analyzed from a diagnostic viewpoint.;Based upon when diagnosis is initiated with respect to the cycle time of batch processes, two categories of diagnosis, namely after-cycle diagnosis and during-cycle diagnosis, have been identified. The computational demands in each category are characterized.;It has been shown that the Generic Task (GT) approach has the ability to capture knowledge from a variety of sources, and in a form that is directly useful in batch process diagnosis. It has been shown that after-cycle diagnosis involves a number of tasks, each of which perform a unique function. A framework for after-cycle diagnosis is developed consisting of three primary tasks: Qualitative Interpretation (QI), Hierarchical Classification (HC) and Hypothesis Queuing (HQ). HC is the core task incorporating all the process knowledge and the primary generic organizational and reasoning strategies.;A conceptual basis for knowledge organization in the form of hierarchy is evolved, in which the top-level hypotheses represent the batch procedures (steps), intermediate-level hypotheses represent specific fault categories of the top-level hypotheses and the tip-level hypotheses represent equipment faults or operator errors. Establish-refine strategy is shown to be the primary reasoning strategy. A new search strategy called hypothesis-invocation-by-elimination, where a hypothesis is evaluated due to the rejection of certain other hypotheses, is developed.;A new generic task, Hypothesis Queuing (HQ), is developed which uses the knowledge available in HC in a transformed manner. It queues the various hypotheses in the hierarchy based upon their "suggested" explanatory powers, which are evaluated by the use of initially available symptom data. This task is shown to be extremely useful in situations where diagnosis involves a number of field tests.;The conceptual developments have been applied to the after-cycle diagnosis of an industrial batch polymer process, in which diagnosis relies heavily on product quality data.