Robust detection and isolation of process faults using neural networks

The problem of robust model-based diagnosis of process faults is addressed by means of artificial neural networks. Different structures and learning methods are investigated for both approaches to function approximation and pattern recognition. Main emphasis is placed upon static and dynamic neural nets that are used as predictors of nonlinear models for symptom generation. Dynamic neural networks are properly integrated into a generalized observer scheme. The goal is to achieve an adequate approximation of process outputs for each known class of system behavior. Symptoms are then evaluated by means of pattern classification. Application to a laboratory process is presented. A diagnosing subsystem is designed to detect incipient faults in the components of a three-tank system. It is implemented in real-time by using the SIMULINL/MATLAB programming environment. Experimental results regarding the diagnosis of single and multiple faults are included in a comparative study. It demonstrates the effectiveness of the suggested approach.