Vibration-based Bayesian anomaly detection of PC bridges

This study aims to investigate an effective damage detection for PC-box girders. A Bayesian anomaly detection is thus proposed. The Bayesian anomaly detection consists of two steps: one is to extract damage sensitive features from autoregressive model, and the other is to detect changes in these features by means of Bayesian hypothesis testing. Three half-scaled PC-box girders were tested to verify validity of the proposed anomaly detection method: one was used for the reference girder and remaining two were used for damaged girders which have unfilled grout zone around some of the post-tensioned tendons. In order to obtain the adoptable vibration data, ten accelerometers have been installed and impact tests have been conducted at the intervals of static loading steps. Observations showed that the proposed method detect changes in the damage sensitive feature when initial cracks propagate in the PC box girder. It also showed the possibility of detecting tendon-cut events.