Bogie axle box bearings, as critical components of Electric Multiple Units (EMUs), operate in complex environments, resulting in complex signal components. Symplectic Geometry Mode Decomposition (SGMD) demonstrates excellent performance in handling nonstationary signals but lacks effective component selection criteria. To address the challenges in processing nonstationary signals, a unique approach is presented, integrating the Symplectic Geometry Mode Decomposition (SGMD) with Graph Similarity (GS). SGMD breaks down the vibration signal into several Symplectic Geometric Components (SGC). The purpose of introducing GS is to optimize the components of SGMD and select the effective components as the input for subsequent diagnosis. Finally, fault classification is performed using Convolutional Neural Networks (CNN). Experimental results demonstrate that SGMD-GS successfully selects the desired component signals. SGMD-GS-CNN achieves a classification accuracy of 94.3% on the CWRU bearing dataset, enabling accurate identification of fault types and severity, thereby demonstrating strong practicality and robustness.