Abstract: Measurement signal is often distorted by noise,which seriously influences signal analysis in magnetotelluric sounding.One of important current questions is that how to filter and select a filter time scale in the data processing.In this paper de-noising,feature preservation and shape restoring of signal are reasonable balanced by anisotropic diffusion(AD) and diffusion balance(DB).Simulation and field data show:(1) filtered signal is accuracy approximated to the original signal and SNR(signal noise ratio) is up to 69.25 dB from 14.19 dB by AD and DB;(2) diffusion balance time scale is closely related to noise intensity, the more noise,the more diffusion balance time scale;(3) "outliers" are effectively removed and basic shape of signal is restored in the field data.