Multi-scale representation contrastive learning for multivariate time-series anomaly detection

In recent years, multivariate time series anomaly detection has received growing attention due to its broad relevance in real-world applications. Nevertheless, accurately identifying anomalies in large-scale time series remains challenging. Many existing methods partition time series into patches to model temporal information, but single-scale patching cannot simultaneously capture complex dynamics at multiple granularities. Moreover, multivariate time series often exhibit strong coupling and complex interactions across sensor channels, such that anomalies may propagate across channels and disrupt dependency structures. To address these challenges, we propose a multivariate time series anomaly detection method based on Multi-scale Representation Contrastive Learning (MsRCL). Specifically, MsRCL performs multi-scale patch decomposition and learns complementary representations from two views, multi-scale intra-patch and multi-scale inter-patch, to capture local dynamics at different temporal scales, cross-segment dependencies, and temporal patterns such as trends and periodicity. In addition, we design a channel-aware attention mechanism that strengthens salient channel correlations and suppresses weakly correlated noise, enabling adaptive modeling of cross-channel dependency structures and improving the representation of cross-channel coupled anomalies. Experiments on five real-world multivariate time series datasets show that MsRCL delivers effective performance gains over the mainstream alternatives.