首页|期刊导航|电讯技术|融合混合注意力与动态调度的残差网络DOA估计

融合混合注意力与动态调度的残差网络DOA估计OA

DOA Estimation with a Residual Network Integrating Hybrid Attention and Dynamic Scheduling

中文摘要英文摘要

针对常规深度学习模型在阵列信号波达方向(Direction of Arrival,DOA)估计中易受跨界误差和过拟合影响的挑战,提出一种融合混合注意力和动态学习率调度的残差网络模型.该模型以阵列信号的空间协方差矩阵为输入,输出对应信号源的 DOA 估计值.模型整体构建于残差神经网络框架之上,集成了角度限制策略以有效解决周期边界误差问题,引入混合注意力机制以增强关键特征提取能力,结合标签平滑技术缓解模型过拟合风险,并采用余弦退火调度策略动态调整学习率以进一步提升训练性能.实验结果表明,在单源输入下的均匀线阵与非均匀线阵条件下,所提模型性能均优于传统方法,在多个典型仿真场景中均保持最低误差.

To address the challenges of boundary-crossing errors and overfitting commonly encountered in conventional deep learning models for direction-of-arrival(DOA)estimation in array signal processing,the authors propose a residual network model that integrates hybrid attention and dynamic learning rate scheduling.The model takes the spatial covariance matrix of array signals as input and outputs the estimated direction of the signal sources.Built upon a residual neural network framework,the model incorporates an angular constraint strategy to effectively mitigate periodic boundary errors,employs a hybrid attention mechanism to enhance the extraction of key features,applies label smoothing to alleviate overfitting,and adopts a cosine annealing schedule to dynamically adjust the learning rate for improved training performance.Experimental results demonstrate that the proposed model consistently outperforms traditional methods under both uniform and non-uniform linear array conditions with single-source input,achieving the lowest estimation errors across various typical simulation scenarios.

杨翰琨;杨文铁;王作帅;徐以东

哈尔滨工程大学 烟台研究院,山东 烟台 265500武汉第二船舶设计研究所海洋电磁探测与控制湖北省重点实验室,武汉 430064武汉第二船舶设计研究所海洋电磁探测与控制湖北省重点实验室,武汉 430064哈尔滨工程大学 烟台研究院,山东 烟台 265500

信息技术与安全科学

DOA估计残差网络混合注意力标签平滑动态学习率调度深度学习

DOA estimationresidual networkhybrid attentionlabel smoothingdynamic learning rate schedulingdeep learning

《电讯技术》 2026 (9)

1464-1471,8

国家自然科学基金资助项目(52101383)

10.20079/j.issn.1001-893x.250319001

评论