Anomaly Detection for Multivariate Time Series with Multi-scale Feature Interactions

Fulu Kou, Lixing Yu, Kun Yue, Liang Duan, Zhongbin Li

Abstract: Efficient anomaly detection in multivariate time series (MTS) data holds paramount importance in the realm of monitoring intricate systems. However, numerous recent methods tend to disregard the vital intra-variable and inter-variable dependencies across distinct temporal and variable scales, leading to inadequate anomaly detection outcomes. To tackle these challenges, we propose a Multi-scale Feature Interactions (MSFI), a novel unsupervised model for anomaly detection in MTS. The MSFI model leverages multi-scale feature interaction blocks, which exploit the variance of each timestamp vector within various scale windows to capture the interdependency among neighboring timestamps. This model also facilitates a comprehensive exploration of variable correlations. Moreover, we propose a normalization-based method for computing attention coefficients to enhance inter-variable correlations and evidence its superiority over the softmax-based method under certain conditions. The resulting dependency features are merged, and a transformer is employed for time series data reconstruction. Furthermore, we introduce three types of noise and incorporate a dynamic noise regulation mechanism during model training to enhance robustness. Extensive experimental evaluations validate the superior performance of our model compared to state-of-the-art approaches.

External IDs:dblp:conf/dasfaa/KouYYDL24