HybridAD: A Hybrid Model-Driven Anomaly Detection Approach for Multivariate Time Series

Anomaly detection, in recent years, has gained increasing attention in the research and practice of time series processing. However, the task is particularly challenging with multivariate time series which complicates the temporal dependency between observations and introduces complex inter-channel correlation. Meanwhile, in order to fit a broader range of applications, robustness during both training and detection is also a critical aspect. In this paper, we propose an unsupervised, hybrid model-driven anomaly detection scheme capable of (1) transforming sequences into a fused representation of temporal dependency embeddings and inter-channel correlation embeddings, and (2) achieving robust anomaly detection using a temporal prediction network for sample-wise posterior estimation combined with a data reconstruction network to assess the source of prediction. On this basis, we develop a probability density-based anomaly scoring mechanism for online detection in multivariate time series, where the anomaly score for each observation is rectified by the reliability of the prediction source. The results of extensive experiments on five publicly available datasets show that our proposed solution outperforms various state-of-the-art anomaly detection algorithms (including DL-based and non-DL-based), achieving a performance improvement (in F1-Score) by up to 10.42%.