抄録
A Bayesian framework for anomaly detection in monitoring data using the relevance vector machine (RVM) is developed. Actual monitoring data of building response including anomalous data obtained due to sensor malfunction and/or impacts of human activities are investigated. The most significant feature of the relevance vector machine is that a model class which has very sparse models in it is automatically selected by taking the automatic relevance determination (ARD) prior and maximizing the marginal likelihood, which is called evidence in the context of Bayesian model selection. The optimized RVM successfully classifies the data into the categories of normal and anomaly with extremely high reliability.
