融合张量分解的HRRP姿态角估计并行深度学习模型

Parallel Deep Learning Model for HRRP Attitude Angle Estimation Fusing Tensor Decomposition

摘要:

针对固定翼目标姿态角估计在宽角度、多目标场景下泛化不足、开销大的问题，提出基于并行Resnet1D-ConvLSTM-Tucker（RCLTu）网络的HRRP姿态角估计方法，该方法利用HRRP角度敏感性与类间相似性，通过双分支分别提取空间结构特征与角度序列相关性，融合层引入Tucker分解实现特征保留与模型轻量化，进而实现HRRP姿态估计；初步验证利用HRRP进行姿态估计的可行性；基于16类FEKO仿真HRRP数据集（6400样本）验证表明，模型方位角（Angle1）、俯仰角（Angle2）估计MAE均维持低水平，显著优于对比方案；消融实验证实Tucker分解有效降参，模型兼具高精度与轻量化；该方法为复杂空战场景提供轻量化解决方案，对雷达目标感知工程化具有参考价值。

Abstract:

To address the problems of insufficient generalization and high computational overhead in fixed-wing target attitude angle estimation under wide-angle and multi-target sce-narios, this paper proposes a High-Resolution Range Profile (HRRP) attitude angle estimati-on method based on the Parallel Resnet1D-ConvLSTM-Tucker (RCLTu) network. Leveragi-ng the angular sensitivity and inter-class similarity of HRRPs, the proposed method adopt-s a dual-branch structure to separately extract spatial structural features and angular seque-nce correlation. The fusion layer incorporates Tucker decomposition to achieve feature pre-servation and model lightweighting, thereby realizing HRRP attitude estimation. The feasib-ility of attitude estimation using HRRPs is initially verified. Experimental validation on a 16-class FEKO-simulated HRRP dataset (6,400 samples) shows that the Mean Absolute E-rror (MAE) of both azimuth angle (Angle1) and elevation angle (Angle2) estimation by t-he model remains at a low level, which significantly outperforms the comparison schemes. Ablation experime-nts confirm that Tucker decomposition effectively reduces the number of parameters, enabli-ng the model to achieve both high accuracy and lightweight performance. This method pr-ovides a lightweight solution for complex air combat scenarios and has reference value fo-r the engineering implementation of radar target perception.