【摘要】 针对传统子空间类波达方向(DOA)估计算法在处理不同相干性分布源定位时需依赖先验相干性信息的局限,文中提出一种基于深度神经网络(DNN)的水下分布源DOA估计方法。该方法利用部分相干分布源信号模型中时间和角度相干分量的可分性,通过分段均值归一化方法简化模型,并构建DNN模型，通过多组不同相干系数的样本数据训练,实现了对不同相干性分布源DOA角度的鲁棒性估计。仿真实验结果表明,该方法无需相干性先验知识即可有效估计不同相干系数下的分布源参数。将文中方法与4种传统子空间类方法和1种深度卷积神经网络算法进行对比,结果表明：在相干分布源情况下，文中方法在不同信噪比和快拍数条件下的均方根误差(RMSE)结果比其他方法平均降低0.42°；在非相干分布源情况下，当信噪比大于0 dB且快拍数大于600时，文中方法的RMSE结果比其他方法平均降低0.04°；在全相干系数范围内，文中方法均表现出更优的估计性能，验证了其在复杂水下环境中的适用性。更多还原

【Abstract】 In view of the limitation of traditional subspace-based direction of arrival(DOA) estimation algorithms that rely on prior coherence information for localizing distributed sources with varying coherences, a DOA estimation algorithm for underwater distributed sources based on deep neural network(DNN) was proposed in this paper. By leveraging the separability of temporal and angular coherence components in the partially coherent distributed source signal model and simplifying the model by segmented mean normalization, a DNN model was constructed and trained with multiple samples of different coherence coefficients, thus achieving robust DOA estimation for distributed sources with different coherence levels.Simulation results indicate that the proposed method can effectively estimate the distributed source parameters with different coherent coefficients without relying on coherence prior knowledge. The proposed method is compared with four traditional subspace-based methods and one deep convolutional neural network algorithm, and the results show that the root mean square error of the proposed method under the coherently distributed source case is 0.42° lower than that of other methods under different signal-to-noise ratios(SNRs) and snapshots; under the incoherently distributed source case, the RMSE of the proposed method is 0.04° lower than that of other methods with SNR greater than 0 dB and snapshots greater than 600. In the range of full coherence coefficients, the proposed methods all show better estimation performance, which proves their applicability in complex underwater environments.更多还原