Remaining Useful Life (RUL) prediction is of significant importance for the safe operation of machines. However, the Health Indicators (HIs) constructed by a single feature of machines cannot precisely describe the non-stationary and accelerating degradation operation. Moreover, poor identification of Time to Start Prediction (TSP), and the limited adaptability of models restrict the accuracy of RUL prediction. A Dual-Model Adaptive Kalman Filter (DMAKF) RUL prediction method based on feature fusion and online TSP recognition is proposed. Firstly, a feature-fusion strategy is proposed, integrating various information to construct distinguishing HIs. Secondly, a statistic-based change-point detection approach is investigated to online recognize TSP, balancing the sensitivity and interference immunity. Finally, the DMAKF is conducted for adaptive adjustment of RUL prediction in real-time. The proposed method is validated on two sets of bearing datasets from different sources. Comparison with four other methods demonstrates the effectiveness of the proposed method.