Most industrial accidents occur during transient operations (start-up, shut down, etc.). During these periods, it is sometimes difficult for operators to determine abnormal operating conditions, identify causes of process trends (external load disturbances, degradation due to parametric changes, sensor and actuator failures, structural faults), and plan and schedule a sequence of operating steps. This research project will consist of studying a way to unify systems theory with human decision-making knowledge in the development of a generic and robust diagnostic methodology for fault diagnosis in chemical plants. The technique involves integrating experimental knowledge with expert systems based on dynamic models of the process. The system structure consists of embedding estimation-based techniques of fault detection and isolation within the framework of a knowledge-based system. Raw data are fed to a fault detector (a preprocessor) which performs statistical tests to identify the process conditions (normal or abnormal). The preprocessor triggers the initiation of the knowledge- based system. The task of the knowledge-based system is to either identify the false alarm or to determine the source and extend of the true fault. Both a state/parameter estimator and a statistical analyzer are included in the loop of diagnostic reasoning. The design methodology is a layered knowledge base that houses heuristic knowledge in high levels and process-general estimation knowledge in the low levels. The diagnostic reasoning alternates between the two domains of knowledge. The algorithm developed will be tested on a subsystem of a power plant that includes a deaerator. This is research planning grant under the Research Opportunities for Women Program.
