Fault detection using multiscale PCA-based moving window GLRT. (June 2017)

• Record Type: Journal Article
• Authors: Sheriff, M. Ziyan; Mansouri, Majdi; Karim, M. Nazmul; Nounou, Hazem; Nounou, Mohamed
• Is Part Of: Journal of process control. Volume 54(2017)
• Page Start: 47
• Page End: 64
• Publication Date: 2017-06
• Subjects: Multiscale principal component analysis -- Generalized likelihood ratio test -- Moving window -- Tennessee Eastman Process -- Fault detection; Process control -- Periodicals; Fabrication -- Contrôle -- Périodiques; Process control; Periodicals; Electronic journals; 660.281
• DOI: 10.1016/j.jprocont.2017.03.004
• Languages: English
• ISSNs: 0959-1524

Abstract

Highlights: A hybrid data-driven fault detection method is developed. It improves the performance of the GLRT chart using a moving window. It integrates the advantages of multiscale representation, GLRT, and PCA. Its advantages are illustrated via its application to the Tennessee Eastman Process. Abstract: The presence of measurement errors (noise) in the data and mode l uncertainties degrade the performance quality of fault detection (FD) techniques. Therefore, an objective of this paper is to enhance the quality of FD by suppressing the effect of these errors using wavelet-based multiscale representation of data, which is a powerful feature extraction tool. Multiscale representation of data has been used to improve the FD abilities of principal component analysis. Thus, combining the advantages of multiscale representation with those of hypothesis testing should provide further improvements in FD. To do that, a moving window generalized likelihood ratio test (MW-GLRT) method based on multiscale principal component analysis (MSPCA) is proposed for FD. The dynamical multiscale representation is proposed to extract the deterministic features and decorrelate autocorrelated measurements. An extension of the popular hypothesis testing GLRT method is applied on the residuals from the MSPCA model, in order to further enhance the fault detection performance. In the proposed MW-GLRT method, the detection statistic equals the norm of the residuals in that window, which is equivalent to applying a mean filter on the squares of the residuals. This means that a proper moving window length needs to be selected, which is similar to estimating the mean filter length in data filtering. The fault detection performance of the MSPCA-based MW-GLRT chart is illustrated through two examples, one using synthetic data, and the other using simulated Tennessee Eastman Process (TEP) data. The results demonstrate the effectiveness of the MSPCA-based MW-GLRT method over the conventional PCA-based and MSPCA-based GLRT methods, and both of them provide better performance results when compared with the conventional PCA and MSPCA methods, through their respective charts T 2 and Q charts.