The fault signal of rolling bearing is a kind of non-linear and non-stationary signal, and it is often mixed with noise signals and interference components. Therefore, it is important to effectively separate the bearing fault source signal and diagnose it. A blind source separation method for rolling bearing faults based on manifold learning is proposed. Firstly, the measured single channel signal is subjected to empirical mode decomposition (EMD) to construct a multi-channel test signal, and then the source number is determined by the deceleration ratio of the singular value of the multi-channel test signal covariance matrix. Then the kurtosis and other indicators are used to select the optimal observation signal. After that, Kernel principal components analysis (KPCA) is used to extract the stream formation of the signal, and finally fast independent component analysis (FastICA) is used to restore the source signal. This method not only solves the problem of underdetermined blind source separation of fault signals, but also proposes the criteria for determining the optimal observed signals. An example is executed to verify the effectiveness of the method.