An algorithm combining particle swarm algorithm and twin support vector machines is proposed. Secondly, in an attempt to suppress the vibration signal noise and enhance the signal features, a rolling bearing signal feature extraction model based on improved variational mode decomposition is proposed. Then, in an attempt to reduce the data dimensionality to improve the computational speed, the study introduces the kernel principal component analysis to reduce the dimensionality of the data, and at the same time, the excavator distance algorithm is introduced so as to construct a condition monitoring model. Finally, the proposed algorithm is applied to the monitoring model to construct a rolling bearing fault diagnosis model based on the algorithm and data dimensionality reduction. The suggested approach outperformed the comparison algorithm in terms of average accuracy rate and loss value, with 97.2% and 1.39, respectively, according to a comparative performance analysis. The bearing defect diagnostic model underwent performance comparison study as well, and outcomes confirmed that the model's average diagnosis accuracy rate was 94.7%. The identification accuracy of inner ring pitting fault, outer ring pitting fault, outer ring fault and rolling element pitting fault are 96.2%, 94.7%, 94.5% and 84.9%, respectively, much higher than the comparison model.