@article{XIN2026104485,
title = {Time series anomaly detection with residuals stationarity intervention on state-space models},
journal = {Advanced Engineering Informatics},
volume = {72},
pages = {104485},
year = {2026},
issn = {1474-0346},
doi = {https://doi.org/10.1016/j.aei.2026.104485},
url = {https://www.sciencedirect.com/science/article/pii/S1474034626001771},
author = {Zhanwen Xin and James-A. Goulet},
keywords = {Structural health monitoring, Time series anomaly detection, State-space models, Intervention models, Dam, Infrastructure, Bayesian methods},
abstract = {Across the world, nations face challenges with maintaining aging infrastructure with limited resources. In order to address a part of this challenge with structural health monitoring (SHM), we need to enable the online detection of anomalies while limiting the number of false alarms; furthermore, we need to do it at scale on a large number of structures. State-space models (SSM) are prime candidates to address this challenge as they enable the separation of structural responses from environmental effects in an online manner. Various detection mechanisms, such as the switching Kalman filter and machine learning approaches, have been integrated with SSMs for identifying abnormal behaviours in time series. However, these methods face two main limitations: (1) the residual component characterizing the prediction errors may become non-stationary and mistakenly capture baseline shifts, thereby reducing anomaly detectability; and (2) following a detection, they may require months if not years of observations, to adapt to a new stationary regime. To address these challenges, we propose the Residuals Stationarity Intervention (RSI), which enhances the anomaly detection capacity by monitoring the stationarity of residuals and enforcing it through interventions predicted by a Bayesian neural network. Experiments on synthetic and real time series recorded on dams demonstrate that RSI outperforms existing approaches, achieving higher detection probabilities, shorter detection times, and faster adaptation, while maintaining a low false alarm rate.}
}