Multi-Phase Adaptive Snow Goose Algorithm-Driven Frequency-Enhanced Residual Networks for Bearing Fault Diagnosis

To address the problems of low diagnostic accuracy and weak robustness in bearing fault diagnosis, this paper establishes a theoretical framework for frequency-domain-enhanced residual network diagnosis driven by a multi-stage adaptive snow goose algorithm (SGA), breaking through the fundamental bottleneck of structural rigidity in traditional methods. The dual-flexible residual network achieves adaptive matching between network topology and fault characteristics through variable convolutional layers and dynamic skip connections. The frequency-domain attention module, based on Fourier transform and magnitude-phase separation techniques, provides in-depth analysis of the frequency-domain discrimination mechanism in vibration signals. The improved snow goose algorithm integrates multi-stage adaptation, opposition-based learning, elite memory management strategies, and cooperative foraging mechanisms, enhancing the optimization performance of SGA. Validation on the Jiangnan University and Case Western Reserve University (CWRU) bearing fault datasets demonstrates that the proposed method achieves diagnostic accuracies of 99.81% and 99.78%, respectively, representing improvements of 4.62%, 6.41%, and 8.76% over mainstream methods such as ResNet, WDCNN, and Transformer. Furthermore, when faced with changes in data structure, the proposed method exhibits a performance degradation of only 0.43%, which is significantly better than other methods, demonstrating excellent robustness and cross-domain generalization capability, and it has certain practical significance.