To solve the problem that the separation and the fault diagnosis of rotating mechanical compound fault signal always difficult to obtain desired results under the condition of single-channel, first, the method of ensemble empirical mode decomposition (EEMD) was applied to build the virtual channels and the method of linear local tangent space alignment (LLTSA) was applied to reduce the dimension, which solved the problem of underdetermined blind source separation well. Then, training the over-complete dictionary and using the method of signal sparse decomposition to extract the sparse characteristics of rotating mechanical compound fault signal. Finally, the support vector machine was employed to evaluate the effect of signal separation and fault diagnosis method. Moreover, the proposed method was applied to the experiment of rolling bearing fault diagnosis, and it's found that the separation and classification of compound fault signal was completed efficiently.