Hierarchical Attention Transformer-Based Sensor Anomaly Detection in Structural Health Monitoring

Structural health monitoring (SHM) is vital for ensuring structural integrity by continuously evaluating conditions through sensor data. However, sensor anomalies caused by external disturbances can severely compromise the effectiveness of SHM systems. Traditional anomaly detection methods face significant challenges due to reliance on large labeled datasets, difficulties in handling long-term dependencies, and issues stemming from class imbalance. To address these limitations, this study introduces a hierarchical attention Transformer (HAT)-based method specifically designed for sensor anomaly detection in SHM applications. HAT leverages hierarchical temporal modeling with local and global Transformer encoders to effectively capture complex, multi-scale anomaly patterns. Evaluated on a real-world dataset from a large cable-stayed bridge, HAT achieves superior accuracy (96.3%) and robustness even with limited labeled data (20%), significantly outperforming traditional models like CNN, LSTM, and RNN. Additionally, this study visualizes the convergence process of the model, demonstrating its fast convergence and strong generalization capabilities. Thus, the proposed HAT method provides a practical and effective solution for anomaly detection in complex SHM scenarios.