To effectively address the challenges in fault diagnosis of axle box bearings in urban rail transit vehicle bogies and to prevent severe traffic safety incidents caused by bearing failure or seizure, this study innovatively proposes a method that integrates convolutional neural networks (CNNs) with recombined particle swarm optimization (RPSO) to construct an advanced diagnostic model. This approach not only optimizes the structural parameters of CNNs to enhance diagnostic accuracy but also introduces a recombination mechanism from RPSO to effectively avoid the premature convergence issues associated with traditional particle swarm optimization (PSO), thus preventing particles from becoming trapped in local optima. Utilizing this innovative model, vibration data from axle box bearings of Guangzhou Metro Line 3 vehicles were collected and analyzed, enabling accurate identification of abnormal bearing conditions and timely implementation of corresponding measures. This method not only averts significant safety risks that could arise from bearing failures during operation but also markedly improves the reliability of vehicle operation services, demonstrating the immense potential of combining advanced algorithms with practical applications.