Subsequence heterogeneity contrastive learning for time series anomaly detection

Time series anomaly detection is widely applied across various real-world scenarios. Recently, contrastive learning has shown remarkable ability in learning discriminative representations for detecting anomalies. However, most existing contrastive-based methods rely on complex contrastive mechanisms and specially designed model architectures, which make it difficult to maintain efficiency and flexibility across various application scenarios. To address this limitation, we introduce Subsequence-Heterogeneity that defined as the discrepancies in variation patterns and statistical characteristics between subsequences obtained through fixed-interval sampling, which are more pronounced in anomalous sequences than in normal ones. It can serve as a natural discrimination criterion and eliminate the need for complex contrastive mechanisms and specialized model architectures. Specifically, we adopt an efficient temporal hierarchical masking strategy with linear complexity to construct two branches for learning representations at different granularities. The Subsequence-Heterogeneity Contrastive Learning (SHCL) is implemented with different neural networks and enables flexible application to anomaly detection across diverse scenarios. Experiments on eight benchmark datasets demonstrate that SHCL not only achieves state-of-the-art performance with reduced time and resource costs but also significantly improves the ability of different neural networks to distinguish normal from anomalous patterns. The source code is publicly available at https://github.com/Zhangzzbzzb/SHCL/.