An improved method for empirical mode decomposition (EMD) is proposed to solve the end effect in empirical mode decomposition. The sequence of extreme value is symbolized and taken as main characteristics by considering the importance of extreme values in EMD. The characteristics of extreme value are taken to match the sequence and restored to the original signal. In addition, EMD and Hilbert Huang are transformed with broadened signal to eliminate the endpoint effect. This method is assessed by both simulated signal and engineering signal and evaluated, significantly restraining the effect of extreme points at both periodic and non-periodic signals, based on analysis of the variation of extreme values. The proposed method is compared with the ARMA model, BP neural network and mirroring broaden, respectively. The average RMS of each component is 19.64%, which is lower than that of other methods. The findings demonstrate that the proposed method provides a better way to restrain the low-frequency component.