Separating Sensor Anomalies From Process Anomalies in Data-Driven Anomaly Detection

Data-driven anomaly detection over time series data is studied from the perspective of separating data anomalies - corresponding to sensor failures - from process anomalies - that arise from equipment or operational failures. A semi-supervised approach is proposed that utilizes two predictive models trained on non-anomalous data using two different sensor groups as inputs, and a nested hypothesis test to reliably classify data or process anomalies. Conditions are derived on choice of sensor groups to guarantee reliable detection, and a case study is presented to demonstrate the proposed classification approach.