The individual heterogeneity and working condition difference of bearings lead to the different performance degradation tracks of bearings, which results in the mismatch between the deep learning model established by full-life bearing and the test bearing. Aiming to this problem, this paper proposes a remaining useful life(RUL) grey prediction method based on convolutional autoencoders(CAEs) and self-organizing maps(SOMs). In this method, a batch normalized CAE driven is introduced to extract the deep features of bearing performance degradation, and the self-organizing maps algorithm is used to construct the performance degradation indicator(DI). The algorithm of dynamic time warping(DTW) is applied to match the bearing degradation trajectories, and the DI curve of the bearing with similar degradation trajectory is used as the reference to set the failure threshold of the test bearing. Driven by the history DIs of the test bearing, the gray forecasting model with full order time power terms(FOTP-GM) is used to predict the RUL of the test bearings. The experimental results show that the proposed method not only retains the individual differences of bearing degradation trend, but also realizes the independent and reasonable setting of bearing failure threshold and improves the prediction accuracy of RUL.