针对低信噪比下基于实数卷积神经网络(RV-CNN)的阵列波达方向(DOA)估计方法对接收信号幅相特征提取不充分的问题,引入复数卷积神经网络(CV-CNN)进行DOA 估计。为进一步提高分类准确率,构建了一种基于复数卷积神经网络的非对称双通道DOA 估计模型(CV-DCNN)。该模型以阵列接收信号的复数协方差矩阵作为输入,分别输入由空洞卷积层组成的第一通道和由标准卷积层组成的第二通道中,其中空洞卷积在不损失角度信息的情况下,增大特征图的感受野。通过复数卷积神经网络(CV-CNN)独有的复数卷积方式提取和融合信号的幅值和相位特征,将双通道提取的特征融合后通过全连接层和sigmoid 函数实现角度分类结果输出。实验结果表明,CV-CNN 比RV-CNN 有更快的收敛速度,在低信噪比和少快拍条件下,CV-CNN 比RV-CNN 有更高的估计精度,而CV-DCNN 比CV-CNN 在收敛速度和估计精度上又有了进一步的提升。

Because array direction of arrival (DOA) estimation method based on real-valued convolutional neural network (RVCNN) under low signal-to-noise ratio has the problem of insufficient extraction of received signal amplitude and phase features, complex-valued convolutional neural network (CV-CNN) is introduced to estimate DOA. In order to further improve the classification accuracy, an asymmetric dual-channel DOA estimation model based on complex-valued convolutional neural network (CV-DCCN) is constructed. The model takes the complex covariance matrix of the array received signal as input, which is input into the first channel composed of the dilated convolutional layer and the second channel composed of the standard convolutional layer, in which dilated convolution increases the receptive field of the characteristic graph without losing angle information. The amplitude and phase features of the signal are extracted and fused by the unique complex convolution method of the complex convolution neural network, and the feature extracted by the two channels is fused and the angle classification result is output through the full connection layer and sigmoid function. The experimental results show that CV-CNN has faster convergence rate than RV-CNN, and CV-CNN has higher estimation accuracy than RV-CNN under the conditions of low signal-to-noise ratio and less snapshot, and CVDCNN has a further improvement in convergence speed and estimation accuracy than CV-CNN.

TN929

国家自然科学基金资助项目(61771182);中央高校业务费资助项目(B210202076);江苏省自然科学基金资助项目(BK20221499)