MambaTAD: A Mamba-Based Contrastive Learning Framework for Time Series Anomaly Detection

Time series anomaly detection (TAD) is essential for ensuring the security, reliability, and efficiency of systems across various domains. Reconstruction-based methods in TAD, while dominant, often restrict model representational capacity, leading to suboptimal performance. This paper introduces MambaTAD, a novel contrastive representation learning framework designed to address these limitations. MambaTAD leverages multi-scale patching, dual Mamba encoders, a shared matrix mechanism, and a rewriting module to capture both global and local dependencies within time series data. By focusing on the discrepancies between Inter-Patch and Intra-Patch views, MambaTAD significantly improves the differentiation between normal and anomalous patterns. Extensive experiments across nine diverse datasets have demonstrated that MambaTAD surpasses state-of-the-art performance, highlighting its practical utility and its potential to significantly advance TAD research and applications.