摘要
针对水下环境信噪比低的特点,以及传统子空间算法计算复杂度高等问题,提出一种基于胶囊网络(capsules network,CapsNet)的波达方向(direction of arrival,DOA)估计模型。将水下矢量水听器阵列采集的协方差数据实虚部分离作为二维数据输入,利用胶囊结构构建向量神经元,通过动态路由的特征传递方法,得到相应矢量胶囊的分类输出,实现低信噪比条件下的DOA估计。为了验证胶囊网络模型的性能,在不同信噪比下条件下,与多重信号分类(multiple signal classification,MUSIC)和卷积神经网络(convolutional neural networks, CNN)的DOA估计结果进行对比分析。仿真结果表明,训练后的胶囊网络,具有更高的水下DOA估计准确率,抗噪性方面优势更加明显,并且在提升性能的同时,加快了方位估计速度。
Aiming at the low signal-to-noise ratio of underwater environment and the high computational complexity of traditional subspace algorithm, a direction of arrival estimation model based on capsules network is proposed. The real and imaginary part of covariance data collected by underwater vector hydrophone array is used as two-dimensional data input.Using the capsules structure to construct vector neurons, the classification output of the corresponding capsules is obtained through the feature transfer method of dynamic routing, so as to realize DOA estimation under the condition of low signal-tonoise ratio. In order to verify the performance of the capsules network model, the DOA estimation results of the capsules network model are compared with those of multiple signal classification and convolutional neural network under different signal-to-noise ratios. The simulation results show that the trained capsules network has higher underwater DOA estimation accuracy, more obvious advantages in noise resistance, and speeds up the azimuth estimation speed while improving the performance.
作者
余春祥
王彪
朱雨男
吴承希
徐晨
YU Chun-xiang;WANG Biao;ZHU Yu-nan;WU Cheng-xi;XU Chen(School of Electronic Information,Jiangsu University of Science and Technology,Zhenjiang 212100,China)
出处
《舰船科学技术》
北大核心
2023年第4期128-132,共5页
Ship Science and Technology
