Rolling bearings are fundamental components in mechanical systems, and the precise prediction of their Remaining Useful Life (RUL) is pivotal for ensuring the dependable operation of the system while concurrently reducing maintenance expenditures. Nevertheless, the intricate degradation behavior of rolling bearings gives rise to the issue of feature redundancy, which presents considerable challenges for the development of predictive models. In response to this issue, the present study introduces an innovative approach grounded in the Transformer framework. Specifically, kernel mutual information (KMI) is initially incorporated as a soft constraint within the Random Forest (RF) model to identify the most optimal features for selection. The selected features are subsequently fed into a dual-channel Transformer model, which integrates a cross-attention mechanism to facilitate the accurate prediction of RUL. The proposed methodology is validated using the PHM2012 dataset and compared against existing state-of-the-art approaches. Experimental results substantiate that the proposed method effectively captures the intricate degradation characteristics of rolling bearings, providing enhanced interpretability and superior predictive accuracy in RUL forecasting tasks.