Incoherently distributed sources localization using convolutional neural network

To solve the problem that the traditional subspace methods for incoherent distributed sources location are difficult to select the effective dimension of the subspace, and rely on the model assumption, an incoherently distributed source localization method based on convolutional neural networks is proposed. As a robust spatial power density distribution feature extractor, convolutional neural networks realize the mapping from the covariance matrix to the direction angle power density distribution. On this basis, the key parameters can also be extracted from the estimated spatial spectrum. In addition, transfer learning technique is employed to solve the mismatch problem between the real signal source distribution and the training model, and improve the generalization performance of the model. Simulation results demonstrate the proposed method is robust to different distributed source models and has better parameter estimation performance than the traditional subspace methods. The real data from microphone array shows that the estimation error of the central angle and the distributed angle with this method is less than 1°.