Memory-guided mask reconstruction with central contrastive learning for robust multivariate time series anomaly...

Mask reconstruction-based unsupervised multivariate time series anomaly detection (MTSAD) methods employ mask operations to model fine-grained local details in time series. However, existing methods introduce semantic biases during the masking process, which compromise the accurate extraction of temporal dependencies. Moreover, these methods are prone to overfitting contaminated data exhibiting strong local correlations. Meanwhile, existing contrastive learning-based MTSAD methods struggle to establish well-separated feature distributions between positive and negative samples, restricting the model's ability to capture high-level semantic information from normal data. This paper proposes a Memory-guided Mask Reconstruction method with Central Contrastive Learning (MMR-CCL). MMR-CCL constructs a global correlation-aware memory module to capture global commonality information across time windows, thereby guiding the reconstruction process. This mechanism mitigates the semantic distortions caused by mask operations, enhancing the model's comprehension of temporal contexts. Additionally, a memory regulation factor is designed to quantify global-local correlation discrepancies within the data, suppressing overfitting to contaminated data. Moreover, MMR-CCL advances a memory-anchoring central contrastive learning strategy to extract high-level semantic information precisely. This strategy employs unmasked data as the central-anchoring samples and masked data as positive samples. Negative samples are generated by filling the masked parts with noise. By maximizing the angular distance between vectors derived from positive and negative samples relative to the central-anchoring samples, MMR-CCL ensures that the mask reconstruction results consistently align with the normal pattern space, establishing a more rational discrimination boundary. Extensive experiments on six public datasets demonstrate that MMR-CCL outperforms 22 typical MTSAD methods.