Speaker
Description
Long-term dynamic monitoring of bridge structures is essential for ensuring their safety and serviceability under varying environmental and operational conditions, particularly under seasonal temperature fluctuations. However, distinguishing true structural anomalies from such variability remains a ma-jor challenge. This paper proposes a novel unsupervised learning frame-work, termed Reconstruction Error-based Density Clustering for Anomaly Detection (REDCAN). REDCAN combines a deep autoencoder with Densi-ty-Based Spatial Clustering of Applications with Noise (DBSCAN) to im-prove the robustness and accuracy of anomaly detection. In the first phase of the REDCAN framework, a deep autoencoder model is developed to re-construct dynamic responses (modal frequencies), and the resulting recon-struction errors are extracted as normalized features. These reconstruction errors are then clustered using DBSCAN, which identifies anomalies based on local density variations in the error space. To enable binary classifica-tion, the main hyperparameters of DBSCAN are tuned such that it forms two clusters: one corresponding to the normal condition and the other rep-resenting anomalies or damage. The algorithm outputs two labels, -1 and 1, where -1 denotes an anomaly or damaged state, and 1 indicates normal be-havior. The major innovation of this study is the introduction of a new ap-plication of unsupervised clustering to anomaly detection. In this regard, the anomaly detection task is transformed into a density separation prob-lem within the reconstruction error space, enabling REDCAN to effectively leverage both the global reconstruction behavior and local density irregular-ities without relying on handcrafted thresholds or multiple decision layers. The proposed method is validated using the long-term modal frequency da-taset of the Z24 bridge, which includes both normal and damaged structural states. Results demonstrate that REDCAN significantly enhances anomaly detection accuracy while maintaining low false alarm rates, even under complex environmental influences.
