Gearbox Fault Classification Method Based on Multi-Scale Feature Fusion and Explainable Analysis

Gearbox fault classification poses significant challenges in industrial environments due to the complex interactions between multi-source signals and noise interference. This study proposes a novel multiscale feature fusion approach with interpretability analysis to overcome the limitations of conventional single-modality diagnostic methods. By systematically constructing a hybrid feature space incorporating time-domain statistics (e.g., variance, kurtosis), frequency-domain indicators (e.g., spectral centroid), and wavelet energy entropy, our method significantly enhances fault pattern characterization. Experimental results demonstrate that under 20Hz-0V and 30Hz-2V operating conditions, the Bayesian-optimized gradient boosting model achieves exceptional classification accuracies of 99.72% and 99.76%, respectively, outperforming conventional single-feature approaches by 18.6 and 17.4 percentage points in F1-score compared to time-domain-only LDA and AdaBoost. SHAP interpretability analysis reveals complementary mechanisms: time-domain parameters dominate initial fault screening (55% cumulative contribution), while wavelet energy entropy shows critical sensitivity to early wear faults (34.7% SHAP value proportion). The method achieves 100% accuracy for severe faults like tooth breakage and 98.4% for early root crack detection. Noise robustness testing establishes practical SNR thresholds for deployment, advancing intelligent fault diagnosis through interpretable multimodal fusion and providing theoretical and engineering solutions for critical rotating machinery in wind turbines and aerospace systems.