Classification of Occupancy Sensor Anomalies in Connected Indoor Lighting Systems

Abstract

We consider the problem of classifying anomalous occupancy sensor behavior in connected indoor lighting systems. Anomalous occupancy sensor behavior may occur in the form of either a high number of false alarms (type-1 anomalies) or missed detections (type-2 anomalies). Two anomaly discovery scenarios are considered: one, in which no anomalies exist post-deployment, and two, in which both anomaly types are found together with normally functional sensors. We address the problem of classifying anomalies that may occur subsequently using a machine learning approach. Under scenario 1, we consider a one class random forest classifier to determine whether an occupancy signal is normal or not. In scenario 2, we consider a supervised random forest classifier to determine whether the detection signal of an occupancy sensor is normal, or exhibits type-1 or type-2 anomalies. We devise occupancy signal features in time and frequency domains to perform 2-class classification in scenario 1, and 3-class classification in scenario 2. The proposed method is evaluated using motion sensor data from an office building, and is shown to have higher true positive rate and a lower false positive rate in comparison to an unsupervised $k$ -means method and a random forest classifier with a single signal energy feature.