Incorporate active learning to semi-supervised industrial fault classification. (June 2019)

• Record Type: Journal Article
• Authors: Yin, Lili; Wang, Huangang; Fan, Wenhui; Kou, Li; Lin, Tingyu; Xiao, Yingying
• Is Part Of: Journal of process control. Volume 78(2019)
• Page Start: 88
• Page End: 97
• Publication Date: 2019-06
• Subjects: Active learning -- Semi-supervised -- Fisher discriminant analysis -- Fault classification; Process control -- Periodicals; Fabrication -- Contrôle -- Périodiques; Process control; Periodicals; Electronic journals; 660.281
• DOI: 10.1016/j.jprocont.2019.04.008
• Languages: English
• ISSNs: 0959-1524

Abstract

Highlights: A new method incorporating active learning to semi-supervised industrial fault classification is proposed to solve the lack of labeled training data. The limited capacity of semi-supervised FDA can be effectively improved by introducing the active learning, and the semi-supervised FDA can utilize the information of unlabeled data to help active learning reduce the amount of labeled data. At each iteration, the data having the greatest entropy will be labeled manually to update current model. The iteration of active learning will stop when current model is confident of all unlabeled data. Compared with semi-supervised FDA, ALSemiFDA can get better discriminant projection vectors and can handle more situations. Compared with active learning based FDA, ALSemiFDA needs less manually labeled data. Abstract: The performance of Fisher discriminant analysis (FDA) method is highly depended on the labeled data. While obtaining the true labels of the industrial data is often time-consuming and expensive in practice. Although there are some researches on semi-supervised methods based on the FDA to solve this problem, they will fail when the labeled data are not satisfied with some special conditions. To improve the methods' applicability, this paper proposes an active learning based semi-supervised FDA model for industrial fault classification. This method bridges labeling important unlabeled data of active learning and learning from unlabeled data of semi-supervised learning. The performance has been greatly improved with the information from the new labeled data, and the amount of new labeled data has been reduced by additional unlabeled data. The experimental analysis on the two dimensional data sets indicates that active learning and semi-supervised learning are complementary for each other. Finally, the experiments carried out on the UCI benchmarks and Tennessee Eastman process (TEP) prove the effectiveness of the proposed method.