摘要
针对大量有标签样本的数据驱动模型方法存在数据分布不完备问题,结合实际环境中通信信号样本差异大的特点,提出一种对抗域适应迁移算法.通过类判别器和域判别器对抗训练,使特征提取器能够提取到既具有类差异性又具有域不变性的特征.以无监督学习方式对目标域信号进行分类,以提升调制识别算法在实际环境中拟合存在分布差异数据集的自适应能力.对实际信号样本集中调制方式相近的9类调制信号在不同信噪比条件下进行测试,域适应迁移方法通过对抗训练有标签高信噪比的源域样本和无标签低信噪比的目标域样本,结果发现该算法的平均识别准确率较以往直接训练的平均识别准确率大幅提升.
Aiming at the problem of incomplete data distribution in the data-driven model approach with a large number of labeled samples, by combining the characteristics of large differences in communication signal samples in the actual environment, an adversarial domain adaptation migration algorithm was proposed.The classifier and domain discriminator were trained against each other so that the feature extractor could extract features with both discriminative for labeled and indiscriminative for domain.Unsupervised learning was used to classify the target domain signals to improve the adaptive ability of modulation recognition algorithm to fit datasets with distribution differences in actual environments. Nine types of modulated signals with similar modulation methods in the actual signal sample set were tested under different SNR(signal to noise ratio) conditions. Domainadaptive transfer methods trained source domain samples with labeled high SNR and target domain samples with unlabeled low SNR by adversarial training. Results show that the average recognition accuracy of the proposed algorithm is significantly improved compared to the average recognition accuracy of previous direct training.
作者
许华
苟泽中
蒋磊
冯磊
XU Hua;GOU Zezhong;JIANG Lei;FENG Lei(Institute of Information and Navigation,Air Force Engineering University,Xi'an 710077,China)
出处
《华中科技大学学报（自然科学版）》
EI
CAS
CSCD
北大核心
2021年第4期127-132,共6页
Journal of Huazhong University of Science and Technology(Natural Science Edition)
基金
国家自然科学基金资助项目(61601500)。
关键词
数据驱动模型
分布不完备
对抗域适应
类差异性
域不变性
data-driven model
incomplete distribution
adversarial domain adaptation
discriminative for labeled
indiscriminative for domain
