In order to solve the problems of poor adaptability of segmentation threshold and poor separation effect in the process of separating bearing fault signals by the traditional maximum between class variance（MBCV），a fault diagnosis method for rolling bearings based on the improved MBCV threshold is proposed. In this method，the fixed threshold value of the traditional MBCV method is changed to a dynamic threshold curve，and the detailed expression of the threshold curve is strengthened，which is more suitable for bearing fault diagnosis. The signal spectrum is divided into several subsections，and the segmentation thresholds of each section are calculated by the traditional MBCV method. The thresholds are fitted with high order to get the threshold curve for the whole frequency domain. The signal spectrum is divided into high and low parts according to the magnitude of the threshold curve. The low amplitude part is inverted by Fourier transform to get the final signal to be diagnosed. The fault is diagnosed by analyzing the obvious frequency component in the square envelope spectrum of the signal. In addition，by adjusting the number of segments of the spectrum，and the RMS error between the separated signal and the original signal is minimized to determine the optimal threshold curve for the target. The new method can effectively eliminate the strong interference components and maximize the extraction of bearing fault information systems. The experimental results also show that compared with the traditional MBCV method，the fault feature extracted by this method is more obvious.