Shortwave localization in complex scenarios plays a crucial role in battlefield communications. Existing deep-learning-based intelligent shortwave direction-finding methods mainly rely on manual annotation, and there is a scarcity of high-quality shortwave signal samples. To address the aforementioned challenges, a shortwave autonomous direction-finding and localization method is proposed, based on a space-time high-resolution time-frequency representation of shortwave signals and an attention-enhanced deep deterministic policy gradient (DDPG) algorithm. A reinforcement learning environment is first designed based on space‑time high‑resolution time‑frequency representation, along with an agent possessing a hybrid action space. Then, a reward function is designed based on the estimated localization results, thereby establishing an observable Markov decision process. Subsequently, by enhancing deep reinforcement learning with an attention mechanism for multi‑level non‑fine evaluation, automatic direction finding and positioning are achieved. This design reduces the need for manual labeling and enables autonomous evolution through simultaneous work and learning, gradually improving the intelligent direction‑finding and localization capabilities for shortwave signals. Experimental results show that, comparable localization accuracy and parameter estimation performance are maintained, while the direction‑finding and localization time is improved by approximately 77.7%.
短波通信是远距离通信的主要方式,具有安装简单、通信距离远等优点,是军事上重要的通信方式[1-2]。对短波信号进行测向定位,在态势分析方面具有重大意义。常用的短波信号测向方法通过多重信号分类(Multiple Signal Classification, MUSIC)算法获得仰角和方位角二维角度信息[3]。目前,短波定位中最常用的方法是多站测角交会定位。该方法首先需要在每个观测站部署阵列天线,利用波达方向(Direction of Arrival, DOA)估算入射信号的方位角与仰角;进而,基于DOA参数确定目标位置[4-5]。但传统的测向定位工作往往分为两步进行,即先对短波信号测向后定位,而且基于参数估计的定位方法计算量较大,因此难以实时测向与定位[6-7]。
由上述研究可以看出,当前短波信号的协同测向定位研究较少,深度学习方法存在样本依赖强、在线自主学习能力不足的短板,经典强化学习算法的网络结构则缺乏动态筛选和聚焦的能力。注意力机制在处理任务时,可以根据重要性为其他元素分配不同的权重。因此,本文考虑将注意力机制融入强化学习,进一步提升协同定位的智能化水平。基于此,本文提出了基于注意力增强DDPG的短波协同测向定位(Direction-Finding and Localization Based on DDPG, DL-DDPG)方法。首先,引入对阵列数据进行空时高分辨处理的方法[11],得到空时高分辨时频图,以此为基础设计了具有混合动作空间的强化学习测向环境;其次,设计可变矩形滤波窗为智能体,对所选信号利用MUSIC算法测得方位角和仰角进行DOA估计,根据定位结果的准确度设计奖励函数评价定位精度;最后,利用DL-DDPG训练智能测向定位环境,实现自动测向定位,可以减少人工标注,实现边工作边学习的自主进化,逐步提升短波信号的智能测向定位能力。
所提DL-DDPG算法是基于注意力增强DDPG的深度强化学习改进算法。在深度强化学习中,智能体通过与环境的持续交互,记录其观察到的状态 s 、所执行的动作 a 以及获得的即时奖励,并利用这些积累的经验数据来学习策略函数。在DL-DDPG算法中,既要考虑价值函数,也要考虑策略近似,学习近似策略的网络被称为策略网络(Actor),学习近似价值函数的网络被称为价值网络(Critic)。
FENGQ Y, TANGT, ZHANGY P, et al. Multistation target localization via direction-of-arrival: an efficient multiagent deep reinforcement learning approach[J]. IEEE Internet of Things Journal, 2025,12(19):40130-40145.
XUX X, HUANGQ H. MD-DOA: a model-based deep learning DOA estimation architecture[J]. IEEE Sensors Journal, 2024,24(12):20240-20253.
[4]
ZHENGS L, YANGZ, SHENW G, et al. Deep learning-based DOA estimation[J]. IEEE Transactions on Cognitive Communications and Networking, 2024,10(3):819-835.
[5]
ZHOUC W, GUY J, SHIZ G, et al. Structured nyquist correlation reconstruction for DOA estimation with sparse arrays[J]. IEEE Transactions on Signal Processing, 2023,71:1849-1862.
ZHENGH, ZHOUC W, VOROBOYVS A, et al. Decomposed CNN for sub-nyquist tensor-based 2-D DOA estimation[J]. IEEE Signal Processing Letters, 2023,30:708-712.
[8]
CONGJ Y, WANGX P, YANC G, et al. CRB weighted source localization method based on deep neural networks in multi-UAV network[J]. IEEE Internet of Things Journal, 2023,10(7):5747-5759.
[9]
TOMICS. Secure AoA-based localization in the presence of malicious nodes[J]. IEEE Internet of Things Journal, 2026,13(7):12800-12810.
[10]
ZHOUM, WANGR J, XIEC J, et al. ReinforceNet: a reinforcement learning embedded object detection framework with region selection network[J]. Neurocomputing, 2021,443:369-379.
JIANGB Q, DUJ, JIANGC X, et al. Underwater searching and multiround data collection via AUV swarms: an energy-efficient AoI-aware MAPPO approach[J]. IEEE Internet of Things Journal, 2024,11(7):12768-12782.
[13]
YUY, LIUM. Deep reinforcement learning-based multisensor control for labeled multi-bernoulli filtering[J]. IEEE Transactions on Aerospace and Electronic Systems, 2025,61(5):13548-13564.
[14]
XUW R, YUJ, MIAOZ J, et al. Spatio-temporal deep Q-networks for human activity localization[J]. IEEE Transactions on Circuits and Systems for Video Technology, 2020,30(9):2984-2999.
[15]
WEIW, FUS, TANGY J, et al. Joint optimization of UAV aided covert edge computing via a deep reinforcement learning framework[J]. Chinese Journal of Aeronautics, 2025,38(10):No.1103218.
SHAKYAA K, PILLAIG, CHAKRABARTYS. Reinforcement learning algorithms: a brief survey[J]. Expert Systems with Applications, 2023,231:No.120495.
[18]
MNIHV, KAVUKCUOGLUK, SILVERD, et al. Human-level control through deep reinforcement learning[J]. Nature, 2015,518(7540):529-533.
[19]
VAN HASSELTH, GUEZA, SILVERD. Deep reinforcement learning with double Q-learning[J]. AAAI Press,2016,30(1). DOI:10.1609/aaai.v30i1.10295 .
[20]
LILLICRAPT P, HUNTJ J, PRITZELA, et al. Continuous control with deep reinforcement learning. [J]. Computerence, 2015. D01:10.1016/S1098-3015(10)67722-4.
[21]
BELLVERM, GIRO-I-NIETOX, MARQUESF, et al. Hierarchical object detection with deep reinforcement learning[DB/OL]. (2016-11-11)[2026-04-09].
[22]
WANGY, WUJ D, HEH W, et al. Data-driven energy management for electric vehicles using offline reinforcement learning[J]. Nature Communications, 2025,16:No.2835.