Abstract: The subspace decomposing algorithms based DOA estimation for coherent sources required the special array structure, and the estimation performance was poor or even broken. In addition, the compressed sensing (CS) based DOA estimation algorithms were able to achieve coherent sources DOA estimation under high SNR with excessive computational complexity. Aiming at above-mentioned drawbacks, a new DOA estimation method for coherent sources was proposed by using the deep learning approach, which was based on sparse representation of the array received signals. The deep learning network for the coherent sources DOA estimation was composed of deep convolutional network and fully connected network. Moreover, the strategies for training the deep learning network were also presented. Therefore, the DOA of coherent sources could be effectively estimated by using the trained deep learning network. The simulation experiments show that compared with the existing coherent sources DOA estimation algorithms, the proposed method is suit for any array structure with much lower time complexity, and the estimation error is superior to smooth de-coherent and L1-SVD algorithm, slightly worse than the OGSBI algorithm.