Enhancing Multivariate Time Series Anomaly Detection With an Inference Stacked Recurrent-Autoencoder in Strong Mechanistic Contexts

Existing self-supervised multivariate time series anomaly detection methods struggle with interference among variables during reconstruction. They also tend to miss capturing critical anomaly information, resulting in unsatisfactory performance, especially in scenarios with strong mechanistic contexts. To this end, we propose a targeted anomaly detection algorithm called inference stacked recurrent autoencoder (ISRAE). Its key contribution lies in the design of a specific inference kernel, derived from specialist knowledge, which captures the strong mechanistic relationships among variables. This kernel is then fused with the multidimensional anomalies predicted by the SRAE, which mitigates interference among variables through the stacking technique. Furthermore, a novel differential constraint is introduced into the loss function, which not only highlights anomaly reconstruction errors, but also smooths the reconstructions, enhancing overall detection performance. Comprehensive comparison experiments and ablation studies show that ISRAE achieves superior anomaly detection performance under strong mechanistic contexts and highlight the importance of each key module in ISRAE.