ConDa-CorA: A Context-and-Data-driven Correlation Analysis Mechanism for Complex System's Multi-Senor Time-series Anomaly Detection
Jinhan Zhou, Jinsong Yu · 2023
Correlation analysis plays an important role in the multi-sensor time-series treatment, such as anomaly detection and time-series prediction. However, the common-accepted methods are almost based on pure statistical mechanism, which risks of suspicious correlation while analyzing the variously-interrelated data of industrial systems. Meanwhile, the contextual knowledge, such as the industrial documentation, is another valuable source which can deduce the physical or structural correlations among the parameters and thus, provides a complementary valuable perspective on correlations among industrial parameters. In this paper, we propose a context-and-data-driven correlation analysis (ConDa-CorA) mechanism for multi-sensor parameters' correlation analysis, which is constructed in a general industrial application condition. Firstly, the proposed method builds a keyword-oriented semantic analysis module, which takes advantage of contextual knowledge represented by the parameters' brief descriptions to achieve a semantic correlation correspondent to some physical-or-structural relations. Secondly, a repeatability-based statistical analysis is integrated to ensure a more robust correlation mining on the operating correlation via the statistical knowledge, particularly accumulated data. Finally, an evidence combination method based on evidence theory is proposed to combine the semantic and the statistical correlations respectively based on the contextual and the statistical knowledges and improve therefore the representability of the mined correlation coefficients. Experiments on a public ALFA UAV control dataset are conducted, by which the proposed hybrid method is demonstrated to achieve state-of-the-art results on its effectivity on anomaly detection task.