A MOBO-VMD and Riemannian SPD ProSwinNet for Bearing Fault Diagnosis

Reliable fault diagnosis of rolling bearings is essential for ensuring the safe and stable operation of rotating machinery. However, existing diagnostic models are often hindered by the strong non-stationarity and weak fault characteristics of vibration signals under complex operating conditions, as well as uneven sample distributions among fault categories, which compromise discriminative representation learning and lead to biased fault recognition. To address these challenges, this paper proposes a novel rolling bearing fault diagnosis framework, termed MOBO-VMD-SPD-ProSwinNet, which integrates adaptive signal decomposition, geometry-aware fault representation, and semantic-guided deep feature learning. Specifically, multi-objective Bayesian optimization (MOBO) is introduced to adaptively determine the key parameters of variational mode decomposition (VMD), while the selected high-energy intrinsic mode functions are transformed into Riemannian symmetric positive definite (SPD) recurrence images to preserve local statistical dependencies and temporal recurrence characteristics. On this basis, a convolutional neural network (CNN)-Swin Transformer architecture is combined with prompt-guided semantic prototype learning to capture complementary local–global fault information while enhancing class-oriented representation and reducing class-biased feature learning. Experiments on the Case Western Reserve University (CWRU), Jiangnan University (JNU), and Southeast University (SEU) bearing datasets demonstrate the effectiveness of the proposed framework. On the CWRU benchmark, the proposed method achieves mean Accuracy, Macro-Precision, Macro-Recall, and Macro-F1 of 99.59%, 99.61%, 99.60%, and 99.59%, respectively, improving the mean accuracy by 1.75 percentage points over the strongest evaluated comparison method. Moreover, mean accuracies of 99.64% and 98.86% are obtained on the JNU and SEU datasets, respectively, further demonstrating the robustness of the proposed framework across different operating conditions and data distributions.