Abstract: | The performance of traditional high-resolution direction-of-arrival (DOA) estimation methods is sensitive to the inaccurate knowledge on prior information, including the position of ar-ray elements, array gain and phase, and the mutual coupling between the array elements. Learning-based methods are data-driven and are expected to perform better than their model-based counter-parts, since they are insensitive to the array imperfections. This paper presents a learning-based method for DOA estimation of multiple wideband far-field sources. The processing procedure mainly includes two steps. First, a beamspace preprocessing structure which has the property of fre-quency invariant is applied to the array outputs to perform focusing over a wide bandwidth. In the second step, a hierarchical deep neural network is employed to achieve classification. Different from neural networks which are trained through a huge data set containing different angle combinations, our deep neural network can achieve DOA estimation of multiple sources with a small data set, since the classifiers can be trained in different small subregions. Simulation results demonstrate that the proposed method performs well both in generalization and imperfections adaptation. |