Wavelet denoising based on soft thresholding is a commonly used approach for feature extraction of nonstationary signals. A new wavelet denoising meth od based on soft thresholding and the dyadic wavelet transform modulus maxima is proposed in order to improve the performance of the wavelet denoising method based on soft thresholding. Firstly, the proposed method performs the dyadic wavelet transform on the signal, and the wavelet coefficients are processed by soft thresholding.Then，the modulus maxima of wavelet coefficients are selected Finally, the denoised signal is reconstructed by the alternating projection algorithm . Theoretical analysis shows that the proposed method can effectively reduce the lower bound of denoising error of the wavelet denoising method based on soft thresholding. Experiments prove that the new method improve signal-to-noise ratios of the denoised results and well reserve the singularities of the orig inal signal. Both the proposed method and the soft thresholding denoising method based on the dyadic wavelet transform are used to denoise the vibration signal of a roller bearing. Results show that the proposed method more effectively extracts the impulse feature of the signal.