In order to extract effective features of complex signals, a fault diagnosis method based on empirical mode decomposition (EMD), singular value decomposition (SVD) and convolutional neural network (CNN) is proposed. First, the fault signal is decomposed into several intrinsic mode function (IMF) components by EMD. A time-domain and a frequency-domain spatial state matrix are constructed. Then, the matrix is decomposed to obtain an array of singular values by SVD ,constructing a time-domain and frequency-domain singular value feature matrix. Finally, the extracted singular value feature matrix is input into CNN for pattern recognition.The method is applied to the fault diagnosis of rolling bearing and gearbox, and has achieved good results in the data of the Case Western Reserve University and the PHM2009 dataset.The correct rate is better than the direct comparison of the original signal into CNN, which verifies the superiority of the method.