Unsupervised Deep Learning for Enhanced Damage Detectability with Small Vibration Data
Wenmiao Gao, Zheng-Han Chen, Alireza Entezami, Hassan Sarmadi
TL;DR: An unsupervised deep learning methodology that integrates generative and discriminative models for enhanced damage detectability under small vibration data conditions is proposed and demonstrates the ability to enhance data diversity, improve class separability, and increase the sensitivity of damage indicators to structural damage.
Bridges, as critical components of transportation networks, demand reliable structural health monitoring (SHM) programs that enable quantitative assessment of their structural states and long-term performance under varying environmental and loading conditions. However, in many practical cases, vibration-based SHM projects, while effective and efficient, are short-term due to budget, accessibility, or operational constraints, leading to small vibration datasets that limit the reliability and detectability of structural damage. Such data scarcity hinders the use of machine learning models, which typically rely on abundant labelled data for robust pattern recognition and anomaly detection. To overcome this limitation, this study proposes an unsupervised deep learning methodology that integrates generative and discriminative models for enhanced damage detectability under small vibration data conditions. First, a generative neural network is employed to perform data augmentation by learning the underlying distribution of limited modal frequency samples of normal structural states and producing synthetic yet physically consistent features. Then, an unsupervised anomaly detection network, trained only on augmented healthy-state data, is utilized to identify damage-induced deviations. This combined generative–discriminative approach enables learning rich feature representations while preserving the unsupervised nature of the problem. The framework is validated using a modal frequency dataset of a cable-stayed bridge monitored within a short-term program. Despite limited modal frequency instances, the proposed method demonstrates the ability to enhance data diversity, improve class separability, and increase the sensitivity of damage indicators to structural damage.