From Data to Alarms: Data-driven Anomaly Detection Techniques in Industrial Settings

This paper introduces a data-driven methodology for anomaly detection in industrial processes. Our focus is on minimizing misclassifications of normal operations and enhancing anomaly and outlier detection. This optimization is based on presumed ground truth (GT) labels associated with a dependent variable (isobutane concentration). Utilizing a moving-horizon approach on an extensive industrial dataset, we perform a comprehensive evaluation of filtering algorithms, and present a representative outlier classification. Secondly, effective anomaly detection, distinct from outlier detection, is achieved by integrating a regression model trained on measurements from independent process variables to fit the dependent variable. Trained regression models consistently achieve effective prediction, staying within an approved process tolerance.