Learning Triple-View Representation Discrepancy for Multivariate Time Series Anomaly Detection with Multi-Scale Patching

Wei Liu, Wanying Zhang, Yating Jiang, Shan Chang, Sun Zhang

Abstract: Multivariate time series anomaly detection is a longstanding but crucial technology, holding significant potential for system security and stability. Prior studies focus on designing sophisticated architectures, integrating advanced modules (e.g., CNN, LSTM, and Transformer), and identifying anomalies based on point-wise reconstruction errors, as it assumes anomalies cannot be correctly reconstructed. However, reconstruction-based methods are risk of over-generalization, making the assumption untenable. Moreover, the performance of complicated architectures is catastrophically overestimated by the flawed point-adjust protocol. In this paper, we propose a simple but efficient architecture comprising only feed-forward layers. The input time series is hierarchically transformed into multi-scale patches to discover complex temporal information. Triple-views are constructed to capture representation discrepancies among different views as the anomaly criterion, circumventing the overgeneralization issue of reconstruction-based methods. To further amplify the discrepancy, a loss function is designed to encourage the consistency among different views during the training phase. The proposed method is evaluated through quantitative and qualitative experiments, demonstrating its competitive performance on four benchmarks, and providing new baselines to the community without point-adjust protocol.