Automated reasoning about chemical plants from first principles: applications to troubleshooting and design
S.D. Grantham, Lyle Ungar · Scholarly Commons (University of Pennsylvania) · 1990
Computer systems which exhibit the flexibility and range of performance of a chemical engineer will need the ability to reason with an understanding of the fundamental physical and chemical phenomena which underlie the behavior of chemical plants. This thesis takes a step in this direction by developing prototype systems for troubleshooting and design which work by automatically creating and manipulating qualitative models of a simple chemical plant. A library of basic physical and chemical phenomena such as reaction, heat flow and liquid vapor equilibrium has been developed in the Qualitative Process Theory (QPT) representation of Forbus and used to automatically build models of the units of a simple chemical plant. The phenomenon definitions specify the conditions under which they will become active and the constraints they contribute to a model. Given a description of the process conditions and materials present, the system determines which phenomena are relevant, builds the associated qualitative model and solves it to determine the behavior of the unit. In order to analyze realistic models, a system was developed to focus the model building and solution mechanisms on specific aspects of unit behavior. To analyze the behavior of chemical units once models are built, one must determine how changes to the operating conditions affect the unit behavior. Existing comparative analysis techniques to do this were extended to predict the effects of changes in the qualitative equations as well as changes in parameter values. The troubleshooting and design systems exploit QPT's explicit representation of the conditions under which phenomena hold and the constraints they contribute to the model. The troubleshooter identifies which changes in a unit description could account for a set of qualitative discrepancies between expected and actual sensor readings by identifying assumption changes which can account for a single discrepancy, determining the resultant changes in the model, simulating the effect on the unit behavior, and comparing with all discrepancies. The designer uses a similar analysis to aid in the conceptual development of batch processing designs, suggesting which phenomena would help achieve a set of design goals, determining the processing conditions required, and pointing out the positive and negative consequences of the changes.