Detection of anomalous events from unlabeled sensor data in smart building environments

This paper presents a robust unsupervised learning approach for detection of anomalies in patterns of human behavior using multi-modal smart environment sensor data. We model the data using a Gaussian Mixture Model, where the features are weighted based on their discriminative ability and are simultaneously clustered. The number of clusters in this approach is automatically chosen using the Minimum Message Length (MML) criterion. The weight of non-discriminative features is driven towards zero which results in a form of dimensionality reduction. Our results indicate that, in practical applications involving unlabeled, high-dimensional multi-modal sensor data from smart building environments, feature weighting achieves higher accuracy in detecting anomalous events with lower false alarm rates compared to using traditional Gaussian Mixtures.