Early Multi-damage Classification in Railway Bridges Using Drive-by Numerical Measurements with Piecewise Aggregate Approximation and Convolutional Neural Networks

Early detection of structural damage is critical for ensuring the safety and longevity of infrastructure. In bridge systems, multiple damage types can occur simultaneously, posing significant challenges for effective monitoring and diagnosis. This study introduces an innovative methodology for early multi-damage classification in railway bridges using drive-by monitoring data, specifically considering a limited number of vehicle passages. The analysis incorporates responses from sensors positioned on the car body and the front bogie of the leading vehicle. Three damage types are evaluated: cracks, bearing device damage, and scour. The proposed data processing methodology follows a two-step approach. First, vertical acceleration data are pre-processed using Min–Max normalization. Dimensionality is then reduced using the Piecewise Aggregate Approximation (PAA) technique. Second, a Convolutional Neural Network (CNN) model extracts features from the reduced data and classifies various damage scenarios. The optimal CNN architecture and hyperparameters are determined using Bayesian Optimization. To account for data variability and neural network randomness, classification accuracy is evaluated through confusion matrices and boxplots. Additionally, a sensitivity analysis is conducted to assess the influence of each operational and environmental variability on classification performance. The results show that the proposed methodology achieves high performance in classifying early-stage multi-damage scenarios, making it a robust and scalable approach for railway bridge monitoring. The PAA technique significantly reduces processing time while maintaining high classification accuracy. Both the car body and front bogie sensors demonstrate strong classification performance despite the effects of all Environmental and Operational Variabilities (namely, track profile, vehicle velocity, signal noise, as well as the temperature and mechanical model property variations). Finally, the sensitivity analysis shows that mechanical property variability significantly contributes to reducing accuracy for the sensor positioned on the car body.