A multi-level representation learning method for the classification with emerging new classes on power event monitoring data. Issue 1 (1st May 2022)

• Record Type: Journal Article
• Authors: Pan, Xiaohui; Liu, Yi; Meng, Fan; Xiang, Shuai; Zhou, Hang; Chen, Guang
• Is Part Of: Journal of physics. Volume 2233:Issue 1(2022)
• Publication Date: 2022-05-01
• Subjects: Physics -- Congresses; 530.5
• DOI: 10.1088/1742-6596/2232/1/012008
• Languages: English
• ISSNs: 1742-6588

Abstract

Abstract: For the classification problem of power event monitoring data, manual rules are mainly used to recognize the known classes of events, and new classes of events cannot be extracted by using existing manual rules. But new classes may emerge with the upgrading of power equipment and traditional representation of this time series data is not proper for the task. To solve this issue, we introduce an effective approach that includes three parts: i) embed the raw object of power event in semantic space ii) learn the representation of power event with multi-level. iii) detect new classes, classify old classes, and update models to classify both new classes and old classes. Experiments on real power event monitoring dataset demonstrate that the proposed method outperforms the state-of-the-art methods.