Neural observer-based approach to fault detection and isolation of a three-tank system

The simulated laboratory set-up Three-Tank System is investigated from the standpoint of fault-tolerant control. The problem of robust model-based diagnosis is therefore addressed. Dynamic neural networks with mixed structure are used to design different observer-based schemes. Symptom evaluation is based on static neural nets. They are used to classify the obtained residuals. Different classifiers and decision criteria are analysed. Experimental results of simulation are included into a comparative study. This refers to actuator, component and instrument fault detection and isolation.