Anomaly Detection for Multivariate Time Series with Cross-Range Modeling

Multivariate time series anomaly detection has received widespread attention due to its significant importance in practical applications. However, most methods are modeled based on fixed sliding windows, which limits their ability to capture comprehensive contextual information. As a result, existing approaches still face significant challenges in detecting anomalies with long durations or those distributed across windows. To address these issues, we propose a three branch attention fusion transformer framework. The dilated span attention, local global attention, and random mask attention cooperate with progressive segment-level decoding to model cross-range information. The dilated span attention module and local global attention module of this framework can capture multivariate time series associations and anomalies across range and granularities. In addition, we also designed a random mask attention module to improve the robustness of the model. Experiments on four real-world datasets demonstrate that our proposed method can obtain superior performance compared to state-of-the-art approaches.