Probabilistic anomaly detection in structural monitoring data using a relevance vector machine

A method for classifying monitoring data into two categories, normal and anomaly, is developed to automatically remove anomalous data included in large sets of monitoring data. A Relevance Vector Machine (RVM) is applied to a probabilistic discriminative model with basis functions and their weight parameters whose posterior distribution conditional on the learning data set is given by Bayes' theorem. One of the significant features of the RVM is that through the optimization process where the evidence, the marginal likelihood, is maximized, the terms which do not have much relationship with the optimal model are removed, resulting in a very sparse model. The proposed framework is applied to actual monitoring data sets containing anomalous data collected at two buildings in Tokyo, Japan. The trained models with large values of evidence distinguish anomalous data from normal data very clearly.