Addressing the common challenges of feature space relationship loss and reliance on tedious manual parameter tuning in deep learning for rolling bearing fault diagnosis, this paper proposes a fault diagnosis method grounded on Bayesian optimized capsule networks. First, the standard capsule network is optimized by designing multi-scale convolutional modules to capture fault information under different receptive fields. Then, a channel attention mechanism is introduced to dynamically weight the extracted multi-scale features. Furthermore, a Bayesian optimization framework is introduced to globally optimize the complex hyperparameter space of the improved model. Findings denote that the enhanced capsule network model achieves an accuracy of 99.8% under standard operating conditions, with its accuracy rapidly increasing to over 90% after 20 iterations. The proposed rolling bearing fault diagnosis method achieves an average F1 score of 99.7% in 10 repeated experiments, and only requires an average of 24 iterations to complete one hyperparameter optimization, verifying the comprehensive superiority of the method. The designed method, through the deep integration of innovative model structure and intelligent parameter optimization, effectively improves the accuracy and efficiency of diagnosis, providing a new paradigm for predictive maintenance of critical equipment under complex operating conditions.