Latent and small fault detection and diagnosis for dynamic processes

Abstract

Compared to large process faults, the latent and small ones are difficult to be detected. However, the accumulation of these faults may even more harmful to the process. A novel fault detection and diagnosis method is proposed which is based on similarity factor and a variable moving window. The new method is based on the idea that a change of process can be reflected in the distribution of the data, which can be detected more easily by the proposed similarity factor. Meanwhile, it has no Gaussian distribution limitation of the process data, since the mixed similarity factor is introduced. The independent component analysis (ICA) factor and the principal component analysis (PCA) factor are used for similarity comparison for Gaussian and non-Gaussian information, respectively. Besides, in order to determine the dynamic step accurately and cut the computation cost, the conventional dynamic method is modified by using autocorrelation analysis. A case study of Tennessee Eastman (TE) benchmark process shows the efficiency of the new proposed method.