Abstract: In the context of predicting the remaining life of rolling bearings, traditional models struggle to adaptively integrate local and global features, and their hyperparameter optimization frequently relies on empirical methods prone to settling in local optima. To address these limitations, this paper proposes an improved Grey Wolf Optimization (GWO) algorithm combined with a CNN-Transformer adaptive fusion framework. First, a nonlinear convergence factor and an elite reverse learning strategy are introduced to enhance the global optimization capability and convergence speed of the GWO. Second, a dual-stream feature extraction architecture is constructed, where convolutional neural networks capture local spatial features from vibration signals while Transformer encoders model long-range temporal dependencies. An adaptive feature fusion module based on a gating mechanism is then employed to dynamically weight and integrate these features, thereby improving the representation capability of the degradation process. Finally, the improved GWO is implemented for end-to-end joint optimization of the critical hyperparameters and fusion weights for the CNN-Transformer model. Experimental validation on the PHM2012 and XJTU-SY datasets demonstrates the superiority of the proposed method, achieving R² values of 0.991 8 and 0.985 2, respectively. These results significantly outperform those of traditional single models and fixed fusion strategies, confirming the method’s superior predictive accuracy and robust generalization performance across different operational conditions.