Detecting faults in information poor systems using neurofuzzy models

Abstract

Considers the problem of detecting faults in information poor systems where an accurate mathematical model is difficult to produce, the data available for training a black-box model are incomplete, and measurements are sparse and of poor quality. The problem of detecting faults in the cooling coil of an air-conditioning system is used as an illustrative example. Results are presented which demonstrate the advantages of using a neurofuzzy model-based detection scheme with a variable threshold. The performance is compared to that of an ideal model-based fault detector and that of detectors with fixed thresholds. The sensitivity of the diagnosis to the type and magnitude of the fault is also examined. Experimental data collected from a full-scale air-conditioning system are used to design and test a fault detector.