Algorithm For Concept Drift Detection In Autonomic Smart Buildings

In order to support the self-awareness, self-configuration, self-description, and self-optimization autonomic properties, autonomic systems need to be able to detect concept drift in their external environment and in their internal operating characteristics. For some autonomic systems, such as smart buildings, this is a very difficult problem because their state is described by multivariate time-series data containing thousands of features generated by thousands of building sensors. Data generated by each sensor typically contains trend, seasonality, and cycles, as well as a significant amount of noise. In this paper we present a new statistical ensemble algorithm for detecting changes in noisy multivariate time-series data. Our algorithm can detect concept drift with up to 100% accuracy, and important change points with up to 92% precision, and 8% false-positive rate. Our algorithm was observed to reduce required features up to 5.4x, reducing the required on-line computational effort.