Abstract: To solve the problem of bad real-time performance of traditional bearing fault early warning and the accuracy affection of fault feature extraction on the early warning effect, transferring the idea of speech endpoint recognition, a double threshold method was used based on the MFPH feature to track the fault starting point. Firstly, in order to overcome the influence of parameter selection and endpoint effect of variational mode decomposition (VMD)on the feature extraction, based on the grid search method of energy difference, the parameters were optimized, and the breakpoint effect was suppressed by SVR. Then, combined with the advantages of MWPE in detecting the randomness of vibration signals, the ability of VMD to reconstruct the signals was fully utilized and the fault signal after the starting point was extracted. Finally, the effectiveness of this method in bearing fault warning was demonstrated by the experiment of bearing fault signal.