Multivariate Time Series Anomaly Detection for Over-reconstruction of Anomalous Data

Anomaly detection in multivariate time series (MTS) data is vital for maintaining system stability and safety. Current reconstruction-based methods aim to learn the normal patterns of data. However, these methods often suffer from the model's over-generalization and the polluted training set, leading to the over-reconstruction of anomalous data. To address these issues, we propose MTAD-ORe, an unsupervised MTS Anomaly Detection framework for Over-Reconstruction of anomalous data. Firstly, to alleviate the model's over-generalization, we design the Gated Diversity Memory Module (GDMM) and the Cross Fusion Module (CFM). GDMM designs a memory block to store typical and diverse features of normal data, which is used for bidirectional updating with the input data. This makes the reconstructed results of anomalous data resemble normal data. Meanwhile, to prevent information loss during the bidirectional updates, CFM is designed to integrate feature information. Secondly, to calibrate the learning bias caused by the polluted training set, we design the Pollution Calibration Strategy (PCS). PCS creates a more reasonable objective for MTAD-ORe, which leads MTAD-ORe to concentrate more on plausible normal data. Experimental studies on four public datasets demonstrate that our MTAD-ORe outperforms several existing competitive baselines.