Abstract: In the fault diagnosis of various signals of power station, due to the complex working environment, the signals obtained by acquisition equipments are often interfered by strong background noises, which brings great difficulties to the subsequent identification and processing of the key features of the signals. In order to reduce the influences of noises, this paper combines empirical mode decomposition (EMD) and the modified wavelet soft-threshold de-noising method to process the signal of power station and verifies the validity of the method. The main process of this method is given as follows. Firstly, the signal is decomposed by EMD in order to separate the several intrinsic mode functions (IMFs) and the residual part. Secondly, the cross-correlation analysis is carried out between each IMF and the original signal to identify IMFs dominated by useful components or by noise. And IMFs dominated by noise are reconstructed with IMFs dominated by useful components and the residual part after processed by the modified wavelet soft-threshold de-noising method. Finally, the de-noised signal based on EMD and the modified wavelet soft-threshold de-noising method is compared with the de-noised signal only based on the traditional wavelet threshold de-noising method. The comparison results show that the de-noised signal based on EMD and the modified wavelet soft-threshold de-noising method has a higher signal to noise ratio, a smaller root-mean-square error, a larger correlation coefficient and a better smoothness, which can effectively eliminate the influences of background noises.