In order to suppress the noise interference and extract the useful fau lt feature from the measured signal, an adaptive wavelet denoising method was p roposed according to the inverse characteristics of useful signal and noise in d ifferent wavelet scales and the limitation of the traditional threshold methods. Firstly, the measured fault signal was decomposed by discrete wavelet transform . Then, the threshold formula was improved to make this method has energy distri bution adaptive denoising ability. At every decomposition level, an optimal ad a ptive threshold, which made the class variance be the maximal one and the varian ce within clusters be the minimal one, was chosen according to the class varianc e. Lastly, the noise in the signal was denoised by wavelet reconstruction. The method was applied to simulation experiment and bearing fault diagnosis experime nt. The results reveal that this method considerably improves the capability of feature extraction and incipient fault diagnosis for rolling bearing in strong n oise backgrounds.