The aging of existing bridges and viaducts necessitates the development of more effective structural health monitoring techniques to support visual inspections, while addressing their inherent limitations. Direct structural health monitoring methods, which involve the installation of sensors on the structure itself, are widely recognized and established. However, these techniques are often characterized by high costs, low flexibility, and scalability, as well as significant safety risks. As an alternative, drive-by monitoring techniques exploits sensors mounted on moving vehicles, offering greater cost-effectiveness. Recently, an increasing number of studies have been focusing on the use of machine learning techniques for drive-by damage detection. This study presents a drive-by monitoring methodology that, starting from the vertical acceleration of the first bogie of the leading coach, leverages sparse autoencoder combined with continuous wavelet transform to detect damage affecting a Warren truss railway bridge. The approach takes advantage of the modular composition of the bridge to assess its condition. A series of numerical simulations is conducted to evaluate the robustness of the method under varying operational conditions. Promising results were obtained by exploiting only one sensing point, for large intensity damages.