Anomaly Detection with Switching Kalman Filter and Imitation Learning

Presenter: Zhanwen Xin
Co-author: James-A. Goulet

ABSTRACT: A reliable detection of anomalies in structural health monitoring (SHM) time-series is notoriously difficult, because one needs to separate reversible patterns caused by the environmental factors and loading, from the baseline irreversible degradation, whereas anomalies typically consist in long-term drifts that have an effect that is orders of magnitude smaller than the reversible patterns. The switching Kalman filter (SKF) has been shown to be an effective tool to address these challenges by quantifying the probability of a switch between different regimes corresponding to normal stationary conditions and abnormal non-stationary ones. Current applications of the SKF have relied on the probability of regime switch as a proxy for detecting anomalies. Despite its capacity to outperform threshold-based detection approaches, it remains prone to (1) missed alarms when the reversible and irreversible pattern separation fails, and (2) false alarms when the predictive capacity of the stationary model is too limited. Previous works have shown that these issues can be mitigated by replacing an alarm triggering policy based solely on the probability of regime switch, with imitation learning (IL) agents that build policies based on the probability of regime switch as well as the trend hidden state. This study further enhances the IL agents by including high-dimensional hidden states vectors as well as by leveraging multiple time steps from the data history. The IL agents that have learnt the underlying dependencies between the high-dimensional hidden state estimates and the labeled actions, which in our context, are whether to trigger an alarm or not, are tested on time series including synthetic and real ones collected on a bridge located in Canada. The results show that the new method provides anomaly detection agents that improve the detectability of anomalies, and are self-adaptive to new stationary conditions.

ABSTRACT: A reliable detection of anomalies in structural health monitoring (SHM) time-series is notoriously difficult, because one needs to separate reversible patterns caused by the environmental factors and loading, from the baseline irreversible degradation, whereas anomalies typically consist in long-term drifts that have an effect that is orders of magnitude smaller than the reversible patterns. The switching Kalman filter (SKF) has been shown to be an effective tool to address these challenges by quantifying the probability of a switch between different regimes corresponding to normal stationary conditions and abnormal non-stationary ones. Current applications of the SKF have relied on the probability of regime switch as a proxy for detecting anomalies. Despite its capacity to outperform threshold-based detection approaches, it remains prone to (1) missed alarms when the reversible and irreversible pattern separation fails, and (2) false alarms when the predictive capacity of the stationary model is too limited. Previous works have shown that these issues can be mitigated by replacing an alarm triggering policy based solely on the probability of regime switch, with imitation learning (IL) agents that build policies based on the probability of regime switch as well as the trend hidden state. This study further enhances the IL agents by including high-dimensional hidden states vectors as well as by leveraging multiple time steps from the data history. The IL agents that have learnt the underlying dependencies between the high-dimensional hidden state estimates and the labeled actions, which in our context, are whether to trigger an alarm or not, are tested on time series including synthetic and real ones collected on a bridge located in Canada. The results show that the new method provides anomaly detection agents that improve the detectability of anomalies, and are self-adaptive to new stationary conditions.