This paper introduces an advanced AI-driven drive-by methodology for unsupervised damage detection on a Warren truss bridge, leveraging deep autoencoders for enhanced feature extraction and reconstruction. The methodology utilizes acceleration data collected from eight sensors mounted on a LAAGRSS-type freight wagon. Wavelet scattering coefficients derived from these acceleration signals serve as input features for the model. Deep autoencoders, trained on baseline condition data, are employed to reconstruct these coefficients, with the absolute reconstruction error acting as a damage-sensitive feature. Environmental and operational variations are mitigated through normalization techniques.
The proposed methodology involves a multi-step process implemented in MATLAB®2022a. Initially, wavelet scattering coefficients (WSC) are computed from the acceleration signals for each sensor, capturing time series properties sensitive to bridge damage conditions. Deep autoencoders are then trained to reconstruct the WSCs from a baseline scenario, consisting of data acquired when the bridge is undamaged. The absolute reconstruction error (ARE) of the autoencoders is computed individually for every scattering coefficient, serving as a new feature.
The methodology further integrates a robust data fusion approach, combining frequency, sensor, and time dimensions into a single damage indicator. This indicator is highly sensitive to damage and can detect various stages of damage without misclassification. The efficacy of the proposed approach is demonstrated through numerical simulations, highlighting its potential for early-stage damage detection. Four types of damage were simulated: one on the main diagonal, one on the main girder, and two on the bracings. The results show that the deep autoencoder-based methodology accurately detects all simulated damage scenarios, including those in their early stages, and distinguishes between different damage types.
Future work will focus on experimental validation and enhancement of the methodology for assessing damage severity. The proposed AI-driven drive-by monitoring system offers a promising solution for real-time structural health monitoring of railway bridges, providing an efficient and cost-effective alternative to traditional inspection methods. By leveraging deep learning techniques and advanced data processing, the methodology ensures accuracy and reliability in detecting structural damage, contributing to the safety and functionality of railway infrastructure.