open access

https://hdl.handle.net/2115/100474

Anomaly Detection in Each Modal Response Using Bayesian Hypothesis Testing and Stochastic Subspace Identification

英語

Bayesian hypothesis testing

2026-07-01

Proceedings of the 10th International Symposium on Reliability Engineering and Risk Management (ISRERM 2026)

This study proposes a two-step anomaly-detection method for the upper structures of bridges using vehicle-induced vibration acceleration. The first step trains a structural model from the reference dataset. Bayesian inference is applied to a vector autoregressive (VAR) model. The autoregressive order is automatically adjusted with the Bayesian information criterion. Covariance-driven stochastic subspace identification (SSI-COV) is applied to find the damage-sensitive parametric subspace in which the parameters represent modal properties in the trained model. The dimension of the parametric subspace is adjusted by the signal-to-noise ratio, comparing the power of each modal response to the noise. The second step detects anomalies by comparing the test dataset to the trained model. Bayesian hypothesis testing is applied to detect changes in the parametric subspace. SSI-COV enables anomaly detection in each modal property directly from measured acceleration. Therefore, we can expect the proposed method to shorten the screening process before further analysis. A field experiment on a truss bridge with severed members was examined as a case study.

会議発表論文

Version of Record