HRRP稀疏自编码器深层特征与散射中心特征的关联性研究

利用稀疏自编码器网络对典型目标一维高分辨距离像(high resolution range profile, HRRP)进行了学习训练, 基于各层权重系数矩阵定义了一种综合权重系数, 通过综合权重系数和降维特征与散射中心特征的对比分析, 发现稀疏自编码器深层特征与散射中心特征之间具有一定的关联性, 并对综合权重系数和深层降维特征的物理意义进行了解释。首先针对HRRP构建稀疏自编码器网络, 经过深层学习后获取训练后的权重系数和降维后的特征, 并与散射中心的位置特征和强度分布特征进行关联性分析。结果表明, 综合权重系数矩阵为与散射中心密切相关的类字典系数矩阵, 反映了距离域强散射中心位置随角度变化的可能的分子集; 降维特征能够实现对强散射中心的学习和提取, 反映了强散射中心位置和强度随角度的变化。最后分析了网络训练层数和降维维数对学习训练结果的影响, 可指导后续网络参数的选择。文章首次针对雷达HRRP数据开展深度学习特征的可解释性研究, 为后续深度学习在雷达数据处理中的广泛应用提供了有益的导引。

Correlation research between deep features of HRRP sparse auto-encoder and scattering center features

A sparse auto-encoder network is used to learn and train one-dimensional high resolution range profile (HRRP) of typical targets. A comprehensive weight coefficient is defined based on the weight coefficient matrix of each layer. By comparing the weight coefficient and dimension reduction feature with the scattering center feature, it is found that there is a certain correlation between the deep feature of sparse auto-encoder and the scattering center feature. And the physical meaning of the comprehensive weight coefficient and deep dimension reduction feature is explained in this paper. Firstly, a sparse auto-encoder network is constructed for HRRP. After deep learning, the weight coefficient after training and the feature after dimension reduction are obtained, and the correlation with the position feature and intensity distribution feature of the scattering center is studied. The results show that the comprehensive weight coefficient matrix is a dictionary-like coefficient matrix closely related to the scattering center, which reflects the possible molecule set of the strong scattering center position changing with the angle in the range domain; and the dimension reduction feature can realize the learning and extraction of the strong scattering center, which reflects the change of the strong scattering center position and intensity with the angle. Finally, the influence of the number of training layers and the dimension reduction dimension on the learning and training results is analyzed, which can guide the selection of the subsequent network parameters. For the first time, this paper studies the interpretability of deep learning features for radar HRRP data, which provides a useful guidance for the subsequent extensive application of deep learning in radar data processing.