This paper proposes a fault feature extraction method based on manifold and singular values entropy. First, on the basis of HHT time-frequency analysis, a two-dimensional manifold method was used to extract a signal manifold ingredient to reduce dimensions and extract the sensitive parameters of the bearing fault feature. Second, singular values entropy was defined to quantitatively measure the differences of the manifold ingredient under different fault statuses. This novel method differs from the general PCA method in terms of the global scope optimum value of European space, and from the one dimensional manifold method in terms of a vector as the research object. The method directly uses two-dimensional information as the research object and thus avoids information loss for a one-dimensional manifold algorithm in the necessary process that transforms two-dimensional information into a vector. Moreover, it can easily find more local data characteristics hidden in a high-dimensional data manifold structure compared with the PCA method. Finally, a manifold singular value vector combined with a probabilistic neural network was used to achieve bearing fault samples classification effectively. Engineering signal analysis verified the effectiveness of the proposed method. This paper provides a reliable method to accurately extract the rolling bearing fault feature.