Modern Defense Technology ›› 2026, Vol. 54 ›› Issue (4): 141-150.DOI: 10.3969/j.issn.1009-086x.2026.04.015

• PAPERS • Previous Articles    

DOA Estimation Based on Multi-scale Attention-Enhanced Deep Residual Network

Jinxuan ZHANG1,3, Huaxing KUANG1,2,3, Ben WANG1,3, Pengfei LENG1,3   

  1. 1.The Eighth Research Academy of CSSC,Nanjing 211153,China
    2.School of Information Science and Engineering,Southeast University,Nanjing 210096,China
    3.National Key Laboratory of Electromagnetic Effect and Security on Marine Equipment,Nanjing 211153,China
  • Received:2025-07-04 Revised:2025-09-17 Online:2026-08-28 Published:2026-09-01

基于多尺度注意力增强深度残差网络的DOA估计

张锦轩1,3, 匡华星1,2,3, 王犇1,3, 冷鹏飞1,3   

  1. 1.中国船舶集团有限公司 第八研究院,江苏 南京 211153
    2.东南大学 信息科学与工程学院,江苏 南京 210096
    3.海洋装备电磁效应及安全全国重点实验室,江苏 南京 211153
  • 作者简介:张锦轩(2001-),男,黑龙江哈尔滨人。硕士生,主要研究方向为阵列信号处理。

Abstract:

DOA estimation is the core task in array signal processing. Traditional methods perform well under high signal-to-noise ratio (SNR) and ideal conditions but exhibit significant performance degradation and reliance on prior information in non-ideal scenarios such as low SNR and limited snapshot counts. To address these issues, this paper directly utilized the in-phase (I) and quadrature (Q) component matrices of raw echo signals as inputs and proposed an end-to-end method based on multi-scale attention enhancement. The network combined double attention residual blocks and a global attention mechanism, dynamically calibrated channel and spatial weights, and enhanced the focus capability on critical features. Simulation results indicate that under the extreme conditions of an SNR as low as -5 dB and a snapshot count of only 10, the root mean square error (RMSE) and angle estimation error range of the proposed method are significantly lower than those of traditional algorithms (MUSIC, ESPRIT, MVDR, and DNN), which verifies its robustness and high precision advantages in complex scenarios.

Key words: direction of arrival estimation, deep learning, attention mechanism, residual network, limited snapshot number

摘要:

DOA估计是阵列信号处理的核心任务,传统方法在高信噪比与理想条件下表现良好,而在低信噪比(SNR)、有限快拍数等非理想场景中存在性能显著下降、依赖先验信息等问题。为此,直接利用原始回波信号的同相(I)和正交(Q)分量矩阵作为输入,提出了一种基于多尺度注意力增强的端到端的方法。该网络结合双注意力残差块和全局注意力机制,动态校准通道与空间权重,增强对关键特征的聚焦能力。仿真结果表明,在信噪比低至-5 dB且快拍数仅为10的极端条件下,所提方法的均方根误差(RMSE)显著低于传统算法(MUSIC,ESPRIT,MVDR,DNN),验证了其在复杂场景下的鲁棒性与高精度优势。

关键词: 波达方向估计, 深度学习, 注意力机制, 残差网络, 有限快拍数

CLC Number: