The rolling bearing vibration signal can provide timely and accurate information on the status characteristics of electromechanical equipment and can realize online or offline monitoring， which is widely used in fault diagnosis of rolling bearings. Due to the complex and changing working environment of rolling bearings， they are often mixed with a lot of noise， which can drown out the useful feature information of electromechanical equipment status. As the traditional wavelet thresholding function cannot greatly reduce the noise in bearing signals， this paper proposes a wavelet soft-thresholding algorithm optimized by differential evolution （DE） to suppress the noise in bearing vibration signals. The algorithm performs wavelet decomposition on the noisy signal， uses the generalized cross-validation （GCV） function as the new thresholding function to process the decomposed wavelet coefficients， and employs the DE algorithm to seek the optimal threshold. The experiments use the bearing data of Case Western Reserve University for simulation analysis. Compared with the commonly used denoising methods， the method in this paper can effectively improve the denoising effect by removing the noise to a large extent while well retaining the characteristic signal.