Consensus-based anomaly detection for efficient heating management

TL;DR: A consensus-based anomaly detection approach that exploits the power of the Symbolic Aggregate approXimation (SAX) and the specificity of such time series and the normalization of the signal becomes a proper element of the modeling.

Abstract: The analysis of data to monitor human-related activities plays a crucial role in the development of smart policies to improve well being and sustainability of our cities. For several applications in this context anomalies in time series can be associated to smaller timeframes such as days or weeks. In this work we propose a consensus-based anomaly detection approach that exploits the power of the Symbolic Aggregate approXimation (SAX) and the specificity of such time series. In our approach, the normalization of the signal becomes a proper element of the modeling. In fact, we conjecture that different normalization horizons allow to include in the shape of the timeseries patterns an additional, variable, component from a longer period trend. To support the analysis phase, a calendar can be used as an additional source of information to discriminate between really unwanted anomalies and expected anomalies (e.g. weekends), or even to signal a possible anomaly whenever a “normal” behavior is not expected. Preliminary experiments on temperature analysis in an indoor environment, with the scope of thermal energy saving, showed that our approch effectively identifies all known anomalies, and also pointed out some unexpected, but clear, anomalies.