Accurate remaining useful life (RUL) prediction for rolling bearings remains difficult because vibration degradation features are highly nonlinear and show complex temporal dependence. To improve prediction accuracy and reliability, an interpretable CNN-BiGRU-KAN method was proposed. First, 12 time-frequency degradation features were extracted from horizontal vibration signals, and supervised samples were constructed with a sliding time window. Then, local degradation information was extracted using a one-dimensional CNN, forward and backward temporal dependence was captured with BiGRU, and a Kolmogorov-Arnold network was introduced to enhance nonlinear mapping between deep features and RUL. SHAP was used to quantify feature contributions. The results show that, compared with CNN, Transformer, TCN, and Transformer-BiLSTM, the proposed model reduces MAE and RMSE by 17.06% and 15.75%, respectively, and improves R2 by 1.94% on average. SHAP results show that time, vibration RMS, vibration peak value, and spectral flatness contribute greatly to RUL prediction.