Speaker
Description
Ensuring the safety and integrity of bridge structures requires effective damage detection methods, which should be capable of handling real-world uncertainties in monitoring data. In particular, modal frequency data obtained from vibration-based monitoring are often affected by irregularities and outliers that can compromise the reliability of damage detection. This study presents a novel unsupervised hybrid anomaly detection method tailored for vibration datasets with high levels of outlier contamination. The proposed approach combines an autoencoder-based reconstruction neural network with a local outlier factor (LOF) model to enhance the robustness and accuracy of structural condition assessment. In the first phase, the autoencoder is trained exclusively on undamaged-state frequency data to learn their latent representation. Reconstruction errors are then quantified and compressed into a univariate feature vector. In the second phase, the LOF algorithm is applied to this compressed feature space, where deviations from normal patterns are detected as potential indicators of damage. This two-stage framework effectively suppresses spurious variability while highlighting structural anomalies. Validation on modal frequency data of a real-world post-tensioned concrete bridge subjected to realistic settlement at one of its piers demonstrates that the proposed method can properly distinguish the bridge damage condition from its normal behavior, thereby significantly enhancing decision-making reliability in the presence of outlier-contaminated vibration data.
