Attention Mechanism-Based Feature Fusion and Degradation State Classification for Rolling Bearing Performance Assessment

Rolling bearing failure poses significant risks to mechanical system integrity, potentially leading to catastrophic safety incidents. Current challenges in performance degradation assessment include complex structural characteristics, suboptimal feature selection, and inadequate health index characterization. This study proposes a novel attention mechanism-based feature fusion method for accurate bearing performance assessment. First, we construct a multidimensional feature set encompassing time domain, frequency domain, and time-frequency domain characteristics. A two-stage sensitive feature selection strategy is developed, combining intersection-based primary selection with clustering-based re-selection to eliminate redundancy while preserving correlation, monotonicity, and robustness. Subsequently, an attention mechanism-driven fusion model adaptively weights selected features to generate high-performance health indicators. Experimental validation demonstrates the proposed method's superiority in degradation characterization through two case studies. The intersection clustering strategy achieves 32% redundancy reduction compared to conventional methods, while the attention-based fusion improves health indicator consistency by 18.7% over principal component analysis. This approach provides an effective solution for equipment health monitoring and early fault warning in industrial applications.