Public concern regarding bridge safety has intensified amid natural hazards and material deterioration. Conventional damage assessment approaches — including field instrumentation and modal analysis — are often costly, time-intensive, and susceptible to substantial inaccuracies. This study presents five novel Convolutional Neural Network (CNN) architectures of increasing depth to jointly predict bridge damage location and severity. This approach overcomes key limitations of traditional methods (high computational cost, sensitivity to environmental noise, and reliance on dense sensor arrays) by enabling accurate, single-sensor assessment directly from vehicle vibration data. A systematic exploration of architectural design choices and hyperparameters is conducted to optimize performance under varying road conditions. In addition, a Baseline-Referenced Residual Transformation (BRT) of training signals is introduced, which markedly improves predictive accuracy and computational efficiency, particularly when accounting for road-surface roughness. Validation on test datasets demonstrates accurate localization and quantification of damage. The framework is readily extensible to more complex bridge systems with multiple damage sites, providing a scalable, efficient, and accurate solution for structural health monitoring and damage prognosis.
Huang et al. (Sat,) studied this question.