Authors: Mr. Ashish Kumar, Dr. Satender Kumar
Abstract: The study of anomaly detection in time series has become one of the key topics in intelligent monitoring systems such as industrial automation, cybersecurity, healthcare, finance, IoT. The traditional approaches to anomaly detection primarily focused on detecting any signs of anomalous behaviour following their occurrence. However, in many cases, reactive anomaly detection does not allow for timely response to detected anomalies. Recently, some researchers have suggested the novel idea of Precursor-of-Anomaly (PoA) detection to detect and analyse warning signs prior to anomalies' occurrence. The present paper provides a review and experimental framework of PoA detection in time series. The paper outlines approaches to traditional anomaly detection, deep learning based forecasting models, uncertainty-aware models, and early warning approaches. Also, the paper outlines a practical framework of PoA analysis using industrial SWaT dataset and Isolation Forest approach. Experimental results prove that uncertainty-aware PoA detection is capable of delivering early warning signals before critical anomalies occur. The paper considers modern limitations and challenges in designing proactive anomaly prediction systems.
