Fault diagnosis based on multivariate statistical techniques

Resumen

In this paper, multivariate statistical techniques such as Fisher Discriminant Analysis and Generalized Discriminant Analysis are utilized for fault diagnosis in an industrial process. The pair-wise FDA analysis is used to identify the fault, which determines the most related variable with the present fault. Therefore, the FDA is proposed to classify linearly separable faults and the GDA to classify faults where a nonlinear classifier is needed. A new procedure to study faults is proposed which include wavelet analysis in the extraction phase, to reduce and decorrelate the data. A continuous stirred tank reactor was simulated in presence of typical faults in order to study the proposed method.