Combining pattern classification and assumption-based techniques for process fault diagnosis

Assumption-based approaches have been proposed in recent times for the diagnosis of process malfunctions. These methods are systematic in the derivation of the knowledge base and are general in approach. However, the quantitative approaches that have been proposed in the past are compiled, difficult to develop and lack generality. Furthermore, the standard assumption-based approaches that use Boolean logic have problems with the completeness and resolution requirements. To circumvent these problems, and to improve upon pattern classification techniques, we propose the tuples method which combines assumption-based approaches with the pattern recognition techniques, using neural networks to perform real-time diagnosis. The tuples method is based on deep-level quantitative models of the process and its knowledge base is developed with relative case. The method is robust in the sense that it allows for modeling inaccuracies. It is also general in that the process model and the diagnostic method are separated. This integrated approach was found to successfully diagnose single and multiple faults including sensor faults and parameter drifts. It has good real-time speed, and gives early diagnosis of malfunctions. Our study also suggests that the generalization characteristics of a neural network can be improved by using a fully-connected network.