Abstract
Real-time structural health monitoring (SHM) of cable-stayed bridges under dynamic operational loads presents significant challenges due to ambient excitation noise, sensor degradation, and inherent structural uncertainties. This study proposes a novel Bayesian Deep Learning (BDL) framework utilizing a 1D Bayesian Convolutional Neural Network (B-CNN) integrated with Variational Inference for real-time damage detection and spatial localization. By leveraging continuous vibration acceleration time-histories, the proposed approach quantifies both epistemic (model) and aleatoric (data) uncertainties, enabling robust classification even under low signal-to-noise ratios. The framework was evaluated using synthetic vibration data from a fine-grained finite element model of a long-span cable-stayed bridge as well as experimental data from a laboratory-scale cable-stayed bridge benchmark subject to stay-cable tension loss and girder stiffness reduction scenarios. Results demonstrate that the proposed B-CNN achieves a damage localization accuracy of 98.4% under noisy ambient conditions (SNR = 10 dB), significantly outperforming standard deterministic deep learning models. Crucially, the epistemic uncertainty estimates effectively flag out-of-distribution states caused by unexpected operational loads or sensor faults, thereby preventing false-positive damage alarms. The proposed probabilistic framework provides civil engineers with a trustworthy computational tool for continuous automated health assessment and predictive maintenance of long-span bridge infrastructure.
