Industrial process monitoring based on Fisher discriminant global-local preserving projection. (September 2019)

• Record Type: Journal Article
• Authors: Tang, Qiu; Chai, Yi; Qu, Jianfeng; Fang, Xiaoyu
• Is Part Of: Journal of process control. Volume 81(2019)
• Page Start: 76
• Page End: 86
• Publication Date: 2019-09
• Subjects: Fisher discriminant -- Global-local preserving projection -- Dimension reduction -- Fault diagnosis -- Process monitoring; Process control -- Periodicals; Fabrication -- Contrôle -- Périodiques; Process control; Periodicals; Electronic journals; 660.281
• DOI: 10.1016/j.jprocont.2019.05.010
• Languages: English
• ISSNs: 0959-1524

Abstract

Highlights: FDGLPP is more sensitive to fault with small magnitude attribute to the great feature extracting ability. FDGLPP method seeks the optimal directions for both representation and discrimination, improves fault detection rate and reduces false alarm rate. Kernel density estimation is introduced to improve process monitoring performance with less false alarm and detection delay. Abstract: A novel data-driven process monitoring method named Fisher discriminant global-local preserving projection (FDGLPP) is proposed and applied to diagnosis fault in industrial process. Integrating Fisher discriminant analysis with global-local preserving projection to solve the optimal projection direction, then transforming original data set with the projection matrix can not only preserve global manifold structure and local neighborhood structure of the data set but also promise the discrimination of the projection subspace, which can afford better fault identification performance. On the other hand, Kernel density estimation is introduced to calculate a more accurate control limit of monitoring indexes, which improves process monitoring performance with less false alarm and detection delay. A numerical simulation is introduced to validate the better dimension reduction performance of FDGLPP algorithm. And a case study on Tennessee Eastman process demonstrates the advantages of proposed method in fault detection and identification contrast with GLPP, LPP and PCA method. Then, t-SNE is applied to visualization the discrimination of FDGLPP.