Real-time anomaly detection with Bayesian dynamic linear models

A key goal in structural health monitoring is to detect abnormal events in a structure's behavior by interpreting its observed responses over time. The goal is to develop an anomaly detection method that (a) is robust towards false alarm and (b) capable of performing real-time analysis. The majority of anomaly detection approaches are currently operating over batches of data for which the model parameters are assumed to be constant over time and to be equal to the values estimated during a fixed-size training period. This assumption is not suited for the real-time anomaly detection where model parameters need to be treated as time-varying quantities. This paper presents how this issue is tackled by combining Rao-Blackwellized particle filter (RBPF) with the Bayesian dynamic linear models (BDLMs). The BDLMs, which is a special case of state-space models, allow decomposing time series into a vector of hidden state variables. The RBPF employs the sequential Monte Carlo method to learn model parameters continuously as the new observations are collected. The potential of the new approach is illustrated on the displacement data collected from a dam in Canada. The approach succeeds in detecting the anomaly caused by the refection work on the dam as well as the artificial anomalies that are introduced on the original dataset. The new method opens the way for monitoring the structure's health and conditions in real time.