The noise from electrical power system itself and introduced by the sampling of fault recorder（FR）have adverse effect on data compression of electrical power system signals and the extraction of fault information.The traditional approach such as wavelet soft/hard-threshold denoising method is not concerned with the nature of noise itself，because it is difficult to distinguish the outset time and the duration time for fault detection after denoising，and it will retain excessive redundant coefficients at the same time which is not conducive to data compression.This paper presents a denoising method for transient data based on intra and interscale dependencies of wavelet coefficients.In an unknown interval with definite variance，the wavelet coefficients are modeled as independent zero mean Gaussian random variables and the variance of them which is determined by related interval changes slowly.By using the intra and interscale dependencies of wavelet coefficients and estimating the noise variance，the noise-free coefficient used to extract the fault information time can be obtained.This method can reduce the noise effect of fault analysis，obtain high frequency coefficients with extremely few quantities and be advantageous to fault data compression.