MTS anomaly detection based on temporal and feature correlation

Anomaly detection is a crucial task in monitoring the diverse states within Industrial Control Systems (ICS), which are typically characterized by Multivariate Time Series (MTS). Accurately detecting anomalies, along with identifying a set of highly anomalous metrics to explain these anomalies, is essential for facilitating effective troubleshooting. This paper introduces a sophisticated deep learning method that comprehensively addresses both temporal and feature correlations within multivariate time series data. During training, this method integrates models based on both forecasting and reconstruction. The forecasting model adeptly captures the temporal and feature dependencies within the data. Meanwhile, the reconstruction model enhances the system’s robustness, accommodating the stochastic nature of the time series and better resisting perturbations and noise. By synergizing these approaches, the method capitalizes on their combined strengths. The efficacy of this model is demonstrated through its evaluation on the SWaT water treatment dataset, where it achieves an impressive average anomaly detection F1-Score of 0.93, outperforming other methodologies in the field.