Deep weighted joint distribution adaption network for fault diagnosis of blast furnace ironmaking process. (June 2022)

• Record Type: Journal Article
• Authors: Gao, Dali; Zhu, Xiong zhuo; Yang, Chunjie; Huang, Xiaoke; Wang, Wenhai
• Is Part Of: Computers & chemical engineering. Volume 162(2022)
• Publication Date: 2022-06
• Subjects: Data-driven -- Fault diagnosis -- Blast furnace -- Joint distribution adaption; Chemical engineering -- Data processing -- Periodicals; 660.0285
• DOI: 10.1016/j.compchemeng.2022.107797
• Languages: English
• ISSNs: 0098-1354

Abstract

Highlights: DWJDAN reduces the impact of class prior distribution on migration performance by adding the class prior distribution adaptation module. DWJDAN considers both the joint distribution and the prior distribution, which further reduces the distribution difference between the source domain and the target domain. DWJDAN improves the classification accuracy on the target domain. Abstract: The blast furnace (BF) ironmaking process has the following characteristics: (1) few fault data (2) large fluctuation in data distribution. For robust fault diagnosis, some transfer learning-based methods are proposed. They mostly aim at aligning only the marginal and conditional distribution discrepancy of datasets. However, in actual application, ignoring the prior distribution will lead to degraded domain adaptation performance. Hence, we propose a new fault diagnosis method called deep weighted joint distribution adaption network (DWJDAN) to align the probability distributions of the dataset more comprehensively. The DWJDAN consists of three modules: condition recognition, class prior distribution adaptation, and joint distribution adaptation. The condition recognition is constructed by a convolutional network (CNN) to extract features and generate health labels of BF. The class prior distribution adaptation generates class-specific auxiliary weights to exploit the prior probability on the source and target domains. The joint distribution adaption facilitates the CNN to learn domain-invariant features by minimizing maximum mean discrepancy (MMD) with weighted data. Experiments using actual industrial BF data show that the proposed method can obtain good results in diagnosing the abnormal conditions of the BF ironmaking process.