Decision making in fault detection

Abstract

Every model based fault diagnosis scheme that utilizes an analytical, a neural or a fuzzy model consists of the decision part, in which the evaluation of the residual signal takes place and, subsequently, the decision about faults is made in the form of an alarm. The residual evaluation is nothing else but a logical decision making process that transforms quantitative knowledge into qualitative statements [4, 9, 7]. It can also be seen as a classification problem. The task is to match each pattern of the symptom vector with one of the pre-assigned classes of faults and the fault-free case. This process may highly benefit from the use of intelligent decision making. A variety of well-established approaches and techniques (thresholds, adaptive thresholds, statistical and classification methods) can be used for residual evaluation. A desirable property of decision making is insensitivity to uncontrolled effects such as changes in inputs and a state , disturbances, model errors, etc. The reasons why decision making can be sensitive to the mentioned uncontrolled efects are as follows [189]: Sometimes it is impossible to completely decouple disturbances and effects of faults; Unmodelled disturbances or an incorrect model structure implies that the performance of the decision making block is decreased; Even though noise terms are included in the model, it is impossible to prevent the noise from affecting the decision making process. © 2008 Springer-Verlag Berlin Heidelberg.