Abstract: Elevators are a critical component of urban rail transit systems, and rolling bearings play a vital role in their stable operation. Addressing the challenges of nonlinear and non-stationary vibration signals in rolling bearings—which hinder the extraction of effective feature parameters—and the impact of penalty factors and kernel function parameters on classification accuracy when using Support Vector Machines (SVM) for fault classification, this paper proposes a fault diagnosis method based on Variational Mode Decomposition (VMD), Rat Swarm Optimizer (RSO), and SVM. The method first adaptively decomposes rolling bearing vibration signals using VMD to obtain Intrinsic Mode Functions (IMFs). The average Multiscale Permutation Entropy (MPE) of the IMFs is calculated to construct feature data samples, which are then used to train the SVM model. To enhance the classification model performance, the RSO algorithm is employed to optimize the input parameters of the SVM, using classification accuracy as the RSO's fitness function to construct the corresponding classification model, thereby achieving classification of rolling bearing vibration signals. Finally, the rolling bearing dataset from Case Western Reserve University (CWRU) is used to establish feature data samples, validating the effectiveness of the method. The results demonstrate that the VMD-RSO-SVM model achieves a classification accuracy of up to 96.67%, outperforming SVM, VMD-SVM, and RSO-SVM models by improvements of 9.17%, 4.17%, and 3.34%, respectively. Additionally, the proposed model exhibits the fastest convergence speed, further verifying its effectiveness.