通信学报 ›› 2022, Vol. 43 ›› Issue (7): 134-142.doi: 10.11959/j.issn.1000-436x.2022145

• 学术论文 • 上一篇    下一篇

基于轻量化分布式学习的自动调制分类方法

杨洁, 董标, 付雪, 王禹, 桂冠   

  1. 南京邮电大学通信与信息工程学院,江苏 南京 210003
  • 修回日期:2022-06-20 出版日期:2022-07-25 发布日期:2022-06-01
  • 作者简介:杨洁(1980- ),女,江苏南京人,博士,南京邮电大学讲师,主要研究方向为分布式学习、边缘计算和智能无线通信等
    董标(1998- ),男,江苏淮安人,南京邮电大学硕士生,主要研究方向为基于分布式学习的自动调制信号分类技术
    付雪(1997- ),女,贵州遵义人,南京邮电大学博士生,主要研究方向为基于分布式学习的自动调制信号分类技术
    王禹(1996- ),男,江苏盐城人,南京邮电大学博士生,主要研究方向为基于分布式学习的自动调制信号分类技术
    桂冠(1982- ),男,安徽枞阳人,博士,南京邮电大学教授,主要研究方向为人工智能、深度学习、智能通信和智能物联网等6G技术
  • 基金资助:
    科技创新2030“新一代人工智能”重大基金资助项目(2021ZD0113003)

Lightweight decentralized learning-based automatic modulation classification method

Jie YANG, Biao DONG, Xue FU, Yu WANG, Guan GUI   

  1. College of Telecommunications and Information Engineering, Nanjing University of Posts and Telecommunications, Nanjing 210003, China
  • Revised:2022-06-20 Online:2022-07-25 Published:2022-06-01
  • Supported by:
    The National Key Research and Development Program of China(2021ZD0113003)

摘要:

为了解决集中式学习存在的问题,提出了一种基于轻量化网络的分布式学习方法。分布式学习利用边缘设备进行本地训练和模型权重共享的方法训练同一个全局模型,既充分利用了各边缘设备的训练数据,又避免了边缘设备数据泄露。轻量化网络是一种由多个轻量化神经网络块堆叠而成的深度学习模型,相较于传统的深度学习模型,轻量化网络以较低的空间复杂度和时间复杂度实现较高的调制分类性能,有效地解决了分布式学习在实际部署中存在的边缘设备算力不足、存储空间有限及通信开销较高的问题。实验结果表明,基于分布式学习的自动调制信号分类技术在 RadioML.2016.10A 数据集的分类准确率为 62.41%,相比于集中式学习,分类准确率仅降低了 0.68%,训练效率提高了近 5 倍。实验结果也证明了在分布式学习下,部署轻量化网络可以有效降低通信开销。

关键词: 自动调制分类, 分布式学习, 轻量化网络, 深度学习

Abstract:

In order to solve the problems in centralized learning, a lightweight decentralized learning-based AMC method was proposed.By the proposed decentralized learning, a global model was trained through local training and model weight sharing, which made full use of the dataset of each communication nodes and avoided the user data leakage.The proposed lightweight network was stacked by a number of different lightweight neural network blocks with a relatively low space complexity and time complexity, and achieved a higher recognition accuracy compared with traditional DL models, which could effectively solve the problems of computing power and storage space limitation of edge devices and high communication overhead in decentralized learning based AMC method.The experimental results show that the classification accuracy of the proposed method is 62.41% based on RadioML.2016.10 A.Compared with centralized learning, the training efficiency is nearly 5 times higher with a slight classification accuracy loss (0.68%).In addition, the experimental results also prove that the deployment of lightweight models can effectively reduce communication overhead in decentralized learning.

Key words: automatic modulation classification, decentralized learning, lightweight network, deep learning

中图分类号: 

No Suggested Reading articles found!