Spatiotemporal attention-based trend and residual decomposition model for multivariate time series anomaly detection

Detecting anomalies in multivariate time series (MTS) data is essential for ensuring stability in critical domains such as industrial monitoring, healthcare, and financial transactions. Traditional methods often fail to capture complex spatiotemporal patterns and are sensitive to noise interference. This study proposes an innovative Spatio-Temporal Attention with Decomposition (STAD) model that addresses these limitations. STAD effectively captures global spatial structures and local temporal fluctuations through an independent-channel, patch-based attention mechanism and a trend-residual decomposition approach. A dynamic sample selection loss function is introduced to enhance model robustness by dynamically adjusting training samples and minimizing the impact of anomalous data during training. An innovative information entropy-based scoring method effectively filters out noise and redundant information by identifying and selecting the most relevant features for anomaly detection. Experiments conducted on benchmark datasets, including Server Machine Dataset (SMD), Secure Water Treatment dataset (SWaT), and Mars Science Laboratory Rover dataset (MSL) show that STAD achieves state-of-the-art performance.