Knowledge-aware Multi-scale Time Series Abnormal Segment Detection

Abnormal segment detection of multivariate time series is of great importance in both data mining research and real-world applications. While recent approaches have achieved significant progress in this area, they often overlook the extensive knowledge information associated with multivariate time series and the diverse temporal patterns among them, which is crucial for time series abnormal segment detection. In this paper, we propose a novel Knowledge-aware Multi-scale Time Series Abnormal Segment Detection (KMTAD) to address this issue. The main idea is to decompose the multivariate time series into the trend and seasonal components and subsequently model each component separately. The resulting knowledge-aware trend component module and multi-scale seasonal component module constitute the key elements of the proposed KMTAD method. The knowledge-aware trend component module builds a time series-oriented knowledge graph based on the extensive knowledge information associated with time series to capture relationships among trend components across different time series. Meanwhile, the multi-scale seasonal component module splits the input seasonal component into segment-level patches, serving as input tokens for the Multi-scale Transformer. Then, the Multi-scale Transformer is used to capture the diverse patterns in the seasonal component of the time series. Extensive experiments are conducted on two real-world datasets, and the results have confirmed the superiority of the proposed KMTAD method.