A novel hybrid of auto-associative kernel regression and dynamic independent component analysis for fault detection in nonlinear multimode processes. (August 2018)

• Record Type: Journal Article
• Authors: Yu, Jungwon; Yoo, Jaeyeong; Jang, Jaeyel; Park, June Ho; Kim, Sungshin
• Is Part Of: Journal of process control. Volume 68(2018)
• Page Start: 129
• Page End: 144
• Publication Date: 2018-08
• Subjects: Auto-associative kernel regression -- Dynamic independent component analysis -- Fault detection -- Nonlinear multimode processes; Process control -- Periodicals; Fabrication -- Contrôle -- Périodiques; Process control; Periodicals; Electronic journals; 660.281
• DOI: 10.1016/j.jprocont.2018.05.004
• Languages: English
• ISSNs: 0959-1524

Abstract

Highlights: A novel hybrid of two data-driven techniques is proposed for fault detection. Auto-associative kernel regression generates residuals of target processes. Dynamic independent component analysis is used for the residual evaluation. Detection indices are calculated from the extracted independent components. The synergies of hybrid method are demonstrated through three benchmark problems. Abstract: With modern industrial processes becoming larger and more complex, we should consider their nonlinear and multimode characteristics carefully for accurate process monitoring and fault detection. In this paper, a novel hybrid of two data-driven techniques—auto-associative kernel regression (AAKR) and dynamic independent component analysis (DICA)—is proposed for fault detection of nonlinear multimode processes. AAKR is a nonparametric multivariate technique; it can effectively deal with nonlinearity and multimodality of target systems by real-time local modeling in accordance with query vectors. Residuals obtained from AAKR usually deviate from Gaussian distribution (i.e., they are non-Gaussian), and there exist auto- and cross-correlations between them. The proposed method detects process faults by applying DICA to the residuals; DICA can capture useful statistical information hidden in the residuals. The validity and effectiveness of the proposed method are illustrated through three popular benchmark problems such as a three-variable multimodal process, a three-variable nonlinear process, and Tennessee Eastman process; the proposed method is also compared with several comparison methods The experimental results demonstrate the superiority of the proposed method, which achieves the best detection rates with reasonable false alarm rates.