Multi-strategy Improved BKA Algorithm Optimization for Transformer-KELM Rolling Bearings

The issue of fault diagnosis in rolling bearings has become increasingly complex and challenging in modern industry, and traditional fault diagnosis methods are no longer sufficient to improve diagnostic efficiency and accuracy. To address this issue, this paper proposes a time-domain and variational mode decomposition (VMD) multi-feature fusion method, combined with Transformer feature processing and kernel-based extreme learning machine (KELM) classification, to create a highperformance bearing fault diagnosis model. Furthermore, traditional optimization algorithms suffer from slow convergence speeds and are prone to getting stuck in local optima. This paper fuses and improves the Black-winged Kite Algorithm (BKA) with the Sparrow Search Algorithm (SSA), and proposes adaptive weight coefficients and mirror imaging reverse learning to enhance the algorithm’s ability to optimize parameters for the Transformer-KELM model. Specifically, ISBKA performs adaptive optimization on the number of multi-head attention mechanisms, kernel function parameters, and learning rate. When trained with the optimized Transformer-KELM model on a bearing fault dataset, the accuracy rate reaches $100 \%$, indicating that the proposed model achieves superior optimization performance.