A modular integration of knowledge-based systems and artificial neural networks for process fault diagnosis

The development of intelligent computer-aided systems for process monitoring, control, scheduling and optimization is evolving towards integrating the different techniques for improved performance. The Neural-Expert hybrid approach, consisting of Knowledge-Based Systems (KBSs) and Artificial Neural Networks (ANNs), is one of the promising developments. Areas of research and development include the fundamentals of hybrid systems, as well as theoretical and practical integration techniques. From an engineering point of view, the research focus is to develop a practical integration system that is most appropriate to apply to a given problem. If the problem is in a domain where expertise can be easily obtained, we can use ANN to improve the KBS performance. The integration should emphasize the verification and refining of expertise by data or obtaining a heuristic evaluation function through the ANN to reduce the search efforts in the KBS. On the other hand, if data are abundant for the target problem, the ANN should be core in the integration. We can use a KBS to improve the quality of data and to guide and verify the training of the ANN. In situations where both knowledge and data are available but incomplete, such as most of fault detection and isolation (FDI) problems, an integration must effectively manipulate the information sources and use the KBS and ANN to complement one another.; This research proposes a modular design to integrate a KBS and an ANN for process fault diagnosis. Both techniques serve as inference components and help control the functioning of one another. Each module can be either a KBS, which utilizes a known set of causal relations, or an ANN, which implicitly extracts knowledge from archived data. For those parts of a diagnostic problem where data and knowledge are available, a decision integrator is used to improve the system's performance by considering and combining the diagnostic conclusions from the ANN and the KBS. The decision integrator provides ranks for the fault candidates selected by KBS and ANN to reflect the confidence of occurring. The ranking process is based on the records of past performance of the tools.; The principal advantage of the modular integration system is its structured methodology that takes into account both qualitative reasoning and empirical modeling. In addition, the modular approaches require less effort in acquiring knowledge for the KBS since the ANN can better handle events of a quantitative nature. Conversely, the KBS will complement the ANN when the events require logical representations. A case study using a simulated recycle reactor process to demonstrate the successful application of this methodology is presented.