Anomaly Pattern Recognition with Privileged Information for Sensor Fault Detection

摘要

Detection of malfunction sensors is an important problem in the field of Internet of Things. One of the classical approaches to recognize anomalous patterns in sensor data is to use anomaly detection techniques based on One Class Classification like Support Vector Data Description or One Class Support Vector Machine. These techniques allow to build a "geometrical" model of a sensor regular operating state using historical data and detect broken sensors based on a distance to the regular data patterns. Usually important signals/warnings, which can help to identify broken sensors, arrive only after their failures. In this paper, we propose the approach to utilize such data by using the privileged information paradigm: we incorporate signals/warnings, available only when training the anomaly detection model, to refine the location of the boundary, separating the anomalous region. We demonstrate the approach by solving the problem of broken sensor detection in a Road Weather Information System.