DUMA: Dual Mask for Multivariate Time Series Anomaly Detection

As a major category of unsupervised anomaly detection methods for multivariate time series, autoregression-based methods train a predictor to model the normal pattern from only normal time series, and then detect anomalies by prediction error. However, we find that the discrepancy of input data between the training and inference stages is a crucial challenge, which may lead to volatile results when detectors take abnormal time series as input. Furthermore, the correlations among multiple sensors are intricate, where irrelevant sensors may bring noise dependencies. This article proposes an autoregression-based time series anomaly detection method named DUal Masked self-Attention (DUMA). First, we propose a block-mask mechanism to enhance the robustness of the predictor for abnormal input data. Then a max-mask self-attention is proposed to reduce the noise dependencies between irrelevant sensors. Experiments on three cyber-physical systems (CPSs) datasets show that DUMA outperforms the state-of-the-art baseline methods.