Aiming at two key problems of the difficulty in identifying the degraded state and the large remaining useful life （RUL） prediction error of rolling bearings， a method combined frequency domain correlation analysis with improved particle filter is pro? posed to address these problems. Based on the characteristic of short-term similarity and long-term difference of bearing vibration signal in frequency domain during the degradation process， the average correlation coefficient （ACC） curve is constructed by the correlation analysis of the amplitude spectrum in frequency domain. When the ACC reaches the preset threshold， the initial failure time （Degradation Initial Timepoint， DIT） is utilized to divide the bearing state into normal and damage stages. Then， the root mean square （RMS） values of vibration signals in damage stage is adopted as the input of observation samples. The dual factor par? ticle filter （DFPF） model， which considers both global exponential degradation tendency and local fluctuation of the measure? ments， is constructed to correct the particle distribution and predict RUL of bearings. The experimental results show that the pro? posed method can effectively identify the accurate initial failure time of bearings than traditional RMS and kurtosis method. More? over， compared with the conventional particle filter （PF） algorithm， the proposed method reduces the interference of abnormal ob? servations on the prediction trend and has higher RUL prediction accuracy.