ABSTRACT
This paper introduces an innovative AI-driven methodology for unsupervised damage detection on Warren truss bridges using drive-by monitoring. By utilizing deep autoencoders, the method enhances feature extraction and reconstruction from acceleration data collected by sensors mounted on a freight wagon. Wavelet scattering coefficients from these signals serve as input features, with the absolute reconstruction error acting as a damage-sensitive feature. A multi-step process, implemented in MATLAB®, involves computing these coefficients, training autoencoders on baseline data, and calculating individual reconstruction errors. A robust data fusion approach combines frequency, sensor, and time dimensions into a single, highly sensitive damage indicator. The efficacy of the methodology, demonstrated through numerical simulations, shows accurate detection of various damage types in early stages. Future work will focus on experimental validation and severity assessment, offering a cost-effective, real-time monitoring solution for railway bridges that ensure structural safety and functionality.
