A Scalable Approach to Time Series Anomaly Detection & Failure Analysis for Industrial Systems

Modern industrial systems are complex and require continuous monitoring for smooth operation. Even a small anomaly in an important variable could lead to suboptimal performance, or worse, a system failure. In critical systems, anomalies that go unaccounted can lead to increased maintenance and operating costs. For this reason, industrial systems opt for algorithms that can predict these anomalies. Modern industrial systems have tens or hundreds of variables with potential correlation with an anomaly. For this reason, a method of detecting anomaly and key failure is developed. By finding the key factors for failure, we can get a better insight about that anomaly, avoiding it in the future. Creating a scalable anomaly detection and key factor analysis framework for different industrial systems is difficult as the systems are very dynamic and varying. In our work, we have proposed a scalable stochastic anomaly detection and key factor analysis framework that is scalable across industries reducing downtime costs, maintenance overheads and increasing system efficacy. We have used a combination of Bayes' theorem and Bitmap detection to detect anomalies in time series data. Then, we have aggregated the anomalies and built a mapping tree to find key factors of the anomalies. We have successfully scaled our work achieving high accuracy anomaly detection and precise key factor analysis for different industries.