Hybrid CAE-DSVDD for unsupervised vibration-based damage detection in in situ steel truss bridge

Abstract

Recent advances in deep learning have enhanced vibration-based damage detection in civil infrastructure. However, the limited availability of labeled damage data remains a major obstacle for real-world deployment. To address this, this study proposes a novel unsupervised anomaly detection framework that relies solely on vibration data from intact structures. The framework integrates a convolutional autoencoder (CAE) with deep support vector data description (DSVDD), combining reconstruction-based and one-class classification approaches to improve sensitivity to subtle damage-induced changes. Time–frequency features are first extracted using the Fourier Synchrosqueezed Transform, which is transformed into images that amplify damage-sensitive patterns. The CAE-DSVDD model, tailored for multisensor structural health monitoring data under real-world operational conditions, is trained exclusively on undamaged-state data. A damage index is then computed by combining reconstruction error and latent-space deviation. Model performance is optimized through hyperparameter tuning, including a balancing parameter between the dual objectives. Unsupervised thresholds are evaluated to enable label-free classification. The framework is validated using field monitoring data from a real-world damaged truss bridge, demonstrating its high detection accuracy to real-world datasets. Additional experiments assess the framework’s robustness under varying loading conditions, input representations, sensor configurations, and its potential for damage localization. The proposed method achieved an F1-score of 98.46% under forced vibration, with consistent performance across low-severity damage cases. The overall results indicate the proposed framework provides an effective solution for unsupervised structural damage detection, offering a scalable and data-efficient approach for real-world infrastructure monitoring.