Mobile sensing of bridge damage with raw acceleration data from random passing vehicles using unsupervised deep learning

Mobile sensing has emerged as a cost-effective alternative for bridge condition assessment by exploiting vehicle–bridge interaction responses. However, existing physics-based methods require specially designed test vehicles and prior vehicle information, while supervised learning approaches depend on labeled damage data that are difficult to obtain in practice. To address these limitations, this study presents an unsupervised deep-learning framework for indirect bridge damage identification using raw acceleration data collected from random passing vehicles. The proposed framework integrates time–frequency signal processing with a convolutional block attention module-enhanced convolutional autoencoder (CBAM-CAE) to automatically learn damage-sensitive representations from vehicle responses. Damage detection, localization, and quantification are achieved through reconstruction-error-based feature analysis without requiring labeled damage samples or explicit modal identification. Numerical investigations demonstrate that the framework can reliably distinguish healthy and damaged bridge states and quantify damage severity under varying road roughness conditions. While localization performance decreases with increasing roughness, damage corresponding to 40% element stiffness loss remains identifiable under rough surface conditions. Laboratory experiments using a vehicle–bridge interaction system further validate the framework, where structural damage equivalent to 3.69% mass addition is successfully localized using only vehicle acceleration measurements. The results demonstrate the effectiveness, robustness, and practical applicability of the proposed computational framework for data-driven infrastructure diagnostics.