| 摘要: |
| 为了能够准确高效地对离格信号的波达方向(Direction of Arrival,DOA)进行估计，利用卷积神经网络来提取信号协方差矩阵中的深度特征信息，并采用改进型标签策略来确保网络的估计精度和效率。具体来说，通过带小数的标签来注释协方差矩阵构成的张量，并配合上改进后的二进制交叉熵损失函数来使得所提出的小数标签能够用于网络训练。针对DOA估计对应的多标签—多分类的问题，使用了包含6层结构的卷积神经网络的输出单元类别以及幅度来分别对离格信号的DOA整数部分与小数部分进行重构。通过与6种现有典型方法的均方根误差（Root Mean Square Error,RMSE）仿真对比，所提方法能够在信噪比为-10 dB的情况下保持着RMSE<0.5°的优秀表现。虽然无法在较少快拍下正常工作，但该方法在快拍数大于8的条件下仍然保持着RMSE<1°的表现性能。同时，在信号数量为5时，所提方法依然具有较高的估计稳定性，且计算速度能够达到毫秒级，用时明显低于其他方法。 |
| 关键词: 离格DOA估计 人工智能 卷积神经网络 监督学习 |
| DOI：10.20079/j.issn.1001-893x.240206002 |
|
| 基金项目: |
|
| Off-grid DOA Estimation via Convolutional Neural Network with Improved Label Strategy |
| YUAN Ye,LYU Zhao,WANG Miao,XU Buyun,LI Pan |
| (Unit 32806 of PLA,Beijing 100091,China) |
| Abstract: |
| In order to estimate the direction of arrival(DOA)of off-grid signals accurately and effectively,a convolutional neural network(CNN) is utilized to extract the depth feature information in the covariance matrix of signal,and an improved labeling strategy is employed to ensure the accuracy and efficiency of the estimation network.Specifically,the tensor composed of the covariance matrix is annotated by labels with decimals,and an improved binary cross-entropy loss function is used to make the proposed decimals labels available for network training.For the multi-label and multi-classification problem corresponding to DOA estimation,the output unit categories and magnitudes of convolutional neural network containing 6-layer structure are used to reconstruct the integer and fractional parts of the DOA of the off-grid signal,respectively.By comparing the simulation results of root mean square error(RMSE) with six typical methods,the proposed method have an excellent performance with RMSE less than 0.5° when SNR=-10 dB.Although it is unable to work properly with fewer snapshots,the proposed method still maintains the performance of RMSE<1° under the condition that the number of snapshots is greater than 8.Meanwhile,the proposed method still has high estimation stability when the number of signals is 5,and the computation speed can reach milliseconds,which takes significantly less time than other methods. |
| Key words: off-grid DOA estimation artificial intelligent convolutional neural network supervised learning |
