Hybrid fault characteristics decomposition based probabilistic distributed fault diagnosis for large-scale industrial processes. (March 2019)

• Record Type: Journal Article
• Authors: Li, Wenqing; Zhao, Chunhui
• Is Part Of: Control engineering practice. Volume 84(2019)
• Page Start: 377
• Page End: 388
• Publication Date: 2019-03
• Subjects: Distributed fault diagnosis -- Probability -- Large-scale process -- Hybrid fault characteristics decomposition; Automatic control -- Periodicals; 629.89
• DOI: 10.1016/j.conengprac.2018.12.009
• Languages: English
• ISSNs: 0967-0661

Abstract

Abstract: The performance of fault diagnosis is highly dependent on the representation of fault characteristics. However, for large-scale industrial processes with high-dimension variables, treating the whole process as a single subject will degrade the representation accuracy. It may result from the following reasons: First, fault may disturb a part of variables rather than the whole process where the fault information may be buried by the unaffected variables. Second, fault characteristics may be hybrid, in which linear fault patterns and nonlinear fault patterns coexist. Therefore, an effective process decomposition mechanism is of great demand to well describe the complex fault characteristics of large-scale processes. This paper proposes a fault characteristics decomposition based probabilistic and distributed fault diagnosis method. First, process is decomposed into different subsets by evaluating fault effects from linear and nonlinear aspects. Based on the decomposition result, distributed diagnosis models are developed where different fault modeling strategies are implemented for different subsets to closely describe fault characteristics. For online application, probabilistic fault diagnosis is implemented at two levels. At the lower level, distributed diagnosis models are adopted to reveal the underlying characteristics of new sample in each subset; at the upper level, the final affiliation can be revealed by integrating the results from each subset in a probabilistic way. The effectiveness of the proposed algorithm is tested by both the numerical example and industrial processes. Highlights: The whole process is decomposed into different subsets with affected variables distinguished from unaffected variables. A distributed fault modeling strategy is proposed to extract the underlying characteristics from different subsets. Fault diagnosis results are calculated in a probabilistic way to tell more information. The efficacy of the proposed method is verified by numerical examples and industrial applications.