TWO SPECTRAL ANALYSES-BASED APPROACHES FOR DETECTION AND IDENTIFICATION OF FAILURE IN CONTROL SYSTEMS

This paper describes a comparative study between two current approaches to fault detection and identification in dynamic systems. The first approach uses a parametric model of the system. The main components of such techniques are residual and signature generation for processing and analyzing. The second approach is nonparametric, in the sense that the signature analysis is dependent only on the frequency or time domain information extracted directly from the input-output signals. Based on these approaches, two new fault monitoring schemes are developed where the feature extraction and fault decision algorithms employed are adopted from the template matching in pattern recognition. In the parametric approach, the on-line fault classification is accomplished by estimating the parametric input-output model of an incipient fault signature, extracting an appropriate feature set and matching it to one of the stored templates. In the second scheme, the template consists of Fourier descriptor and the on-line classification is performed by comparing Euclidean metric norm between a test signature and a set of stored templates. In both techniques the feature set is proved to be invariant under input scaling. Extensive simulation studies are performed to demonstrate the satisfactory performance of the proposed techniques. The advantages and disadvantages of each approach arc discussed and analyzed.