Deep learning based data anomaly detection considering data imbalance and extreme loadings

Manman Hu, Gao Fan, Jun Li, Yong Xia, Hong Hao,
Deep learning based data anomaly detection considering data imbalance and extreme loadings,
Journal of Building Engineering,
Volume 111,
2025,
113632,
ISSN 2352-7102,
https://doi.org/10.1016/j.jobe.2025.113632.
(https://www.sciencedirect.com/science/article/pii/S2352710225018698)
Abstract: Data anomalies caused by sensor faults or transmission failures can undermine the reliability of structural health monitoring (SHM). This study presents a robust deep learning-based anomaly detection approach that addresses key challenges of class imbalance and extreme loading conditions. The methodology involves generating a balanced multi-class dataset through synthetic anomaly modelling and incorporating a small proportion of real extreme loading responses into training to enhance robustness. An adapted one-dimensional DenseNet (1D-DenseNet) is developed to directly process time-series response data without signal transformations. Experimental validation using SHM data from the Guangzhou New TV Tower demonstrates that the proposed approach achieves an exceptional classification performance, with accuracies exceeding 99.5 % and 99.8 % under ambient and typhoon loading scenarios, respectively. The proposed approach exhibits strong adaptability to extreme events, low false alarm rates, and eliminates the need for manual anomaly labelling, making it highly practical for real-time SHM applications.
Keywords: Anomaly detection; Data imbalance; Extreme loadings; Deep learning; Structural health monitoring