In engineering applications, roller bearing signals are often submerged in strong background noise, which makes it difficult to extract weak fault features. To solve this problem, a novel weak fault feature extraction method based on contribution rates-singular value decomposition method (CR-SVD) with periodic modulation intensity (PMI) is proposed in this paper. Firstly, singular value contribution rates are introduced to determine the dimension of Hankel matrix, which can consider the information of all singular values comprehensively. Secondly, the matrix is decomposed by SVD method. PMI of the SVD components are obtained with three types of theoretical fault frequency calculated by the geometric parameters of the roller bearing and the rotating speed. Thirdly, part of PMI contained significant information are retained and the SVD components are weighted by filtered PMI. Then unimportant components are eliminate by weighting. Finally, the correct denoising signal is selected from the correlation coefficients of the three groups of signals. The simulation signal analysis and the vibration signal analysis show that the proposed method can extract weak fault feature and eliminate noisy components effectively.