A multivariate time series anomaly detection method with Multi-Grain Dynamic Receptive Field

The multivariate time series (MTS) anomaly detection methods based on masked reconstruction pose challenges in model training by setting unknown areas to the data, compelling the model to explore deeper patterns to enhance its performance. Due to the low information density of MTS, point-masked methods relying on timestamps can only capture a limited amount of data information, while patch-masked methods based on segments can more effectively uncover advanced semantic features of underlying trends in MTS. However, patch-masked methods process MTS with either fixed or random masks, which may not only sacrifice the known information but also impose restrictions on the size of mask blocks during the reconstruction process. In this paper, a multivariate time series anomaly detection method with Multi-Grain Dynamic Receptive Field (MGDRF) is proposed. MGDRF designs multi-grain mask strategies to excavate semantic features of MTS ranging from lower to higher levels. The dynamic receptive fields are specifically crafted to mitigate information loss encountered in existing methods, thereby facilitating learning of temporal and dimensional relationships of the data. Furthermore, MGDRF incorporates the receptive-field-based and model-based layered losses. It establishes primary losses for each single-grain receptive field, enabling the extraction of different semantic features. Based on ensemble learning, MGDRF constructs a model-based loss through the fusion of outcomes from multiple grains of dynamic receptive fields, thereby further learning the interaction characteristics among different grains of MTS. Extensive experiments on five representative public datasets demonstrate that the proposed algorithm exhibits more advanced performance compared to 18 typical MTS anomaly detection methods. •The multi-grain mask strategies are designed to excavate semantic features of data.•Dynamic receptive fields are established to alleviate the information loss problem.•The hierarchical losses are proposed to enhance anomaly detection performance.