米波雷达在探测低空目标时存在严重的多径效应,直达波和反射波相当于距离较近的两个强相干点源,目标回波信号协方差矩阵存在噪声子空间与信号子空间相互渗透的问题,经典的超分辨物理算法仰角估计精度会急剧变差。为解决上述问题,文中基于米波雷达经典镜像多径反射模型,利用深度神经网络和全连接网络构造了一个深度学习网络用于低仰角目标波达方向(DOA)估计,将子空间相互渗透的原始协方差矩阵数据实部、虚部及相位特征作为深度学习网络输入,利用智能学习方法解决了多径反射条件下DOA 估计问题。相比于基于子空间分解或信号拟合类的超分辨估计方法,文中所提方法仰角估计精度更高且计算量更小。仿真实验验证了新方法的优越性和有效性。

Meter-wave radar faces serious multipath effect in detecting low-altitude targets, where the direct wave and the reflected wave are equivalent to two close-range strongly coherent point sources, the noise subspace and signal subspace in covariance matrix of target echo signal will interpenetrate, and the elevation estimation accuracy of classical super-resolution physical algorithms will decrease dramatically. To address the above problem, the idea of deep learning is employed. With the real part, imaginary part and phase of the raw target echo signal covariance matrix taken as input, a deep neural network for direction of arrival (DOA) estimation of low elevation targets is constructed, which can achieve high accuracy DOA estimation under multipath conditions. Compared with the super-resolution estimation method based on subspace decomposition or signal fitting class, the elevation estimation method proposed in this paper has higher accuracy and less computation. Simulation results show the superiority and effectiveness of the new method.

TN953