Joint Subsequence Anomaly Detection and Fault Classification for Multivariate Time Series

To ensure the stability and reliability of a large-scale system, each device is typically monitored with a large number of Key Performance Indicator (KPI) streams. The detection of anomalies in these streams is vitally important. However, diverse anomalous patterns, high dimensionality, and the complex relationships of multivariate time series pose significant challenges. This work proposes a method to achieve effective subsequence anomaly detection and fault classification for multivariate time series. The idea is to capture the abnormal patterns by generating graphs that consider the correlation among different KPIs, calculate the similarity matrix based on the generated graphs, and use the similarities to identify anomalous subsequences. In this method, we explicitly utilize the correlation among different KPIs to make the detection results interpretable and can conduct both anomaly detection and fault classification. Extensive experiments using the real-world Exathlon benchmark dataset show that our algorithm can accurately recognize subsequence anomalies and classify faults with a high F1-score on two-level standards (F1-score = 0.92, 0.91). Our algorithm significantly outperforms the best-performing baseline method.