无监督领域自适应调制识别方法

Unsupervised Domain Adaptation Based Automatic Modulation Recognition

针对现有自动调制识别(AMR)方法中,由于训练样本和实际应用的数据存在较大的不同,而导致识别率恶化的问题,提出无监督领域自适应(DA)的调制识别方法。该方法在传统识别网络中加入域分类子网络,在训练的代价项中加入域分类代价,使得网络能够同时适应目标域和源域。通过开源软件仿真的数据集证明,相比于无迁移学习和基于参数精调的方法,在CNN为基础网络的条件下,识别率分别提高了41%和7%;在ResNet为基础网络的条件下,识别率分别提高了43%和9%。

Existing automatic modulation recognition(AMR)methods could have deteriorated recognition rate because that the training samples and the actual data differed enormously.In this paper,an unsupervised domain adaptation AMR method was proposed,which added a sub domain discriminator network to existing network,and added the cost of domain discriminator,which made the new network feasible for both source and target domain.According to experiments based on simulated data,compared with the method without transfer learning and fine tuning based method,the proposed method had a recognition rate improvement of about 41%and 7%with CNN structure,the recognition rate were 43%ad 9%with ResNet structure.