基于改进AdaBoost.M2算法的自动调制识别方法  

Automatic modulation recognition method based on improved weight AdaBoost.M2 algorithm

作　　者：王沛[1,2] 刘春辉 张多纳[3] WANG Pei;LIU Chunhui;ZHANG Duona(School of Electronic Information Engineering,Beihang University,Beijing 100191,China;Institute of Unmaned System,Beihang University,Beijing 100191,China;School of Information Engineering,North University of Technology,Beijing 100144,China)

机构地区：[1]北京航空航天大学电子信息工程工程学院,北京100191 [2]北京航空航天大学无人系统研究院,北京100191 [3]北方工业大学信息工程学院,北京100144

出　　处：《北京航空航天大学学报》2023年第8期2089-2098,共10页Journal of Beijing University of Aeronautics and Astronautics

基　　金：北京市自然科学基金(4204102)。

摘　　要：针对同族调制类型通信信号识别难度大、深度学习模型普遍存在泛化能力弱的问题,基于经典AdaBoost.M2算法,提出改进样本权重的AdaBoost.M2算法,用于解决大样本情况下学习率与加权后样本数据难以相适应的问题。改进后的新样本权重确保训练样本数据的数量级在加权后不变,并使算法更迅速地关注到难分类样本,提高了弱分类器综合性能,降低了加权投票模型中弱分类器重要性之间的差异。针对部分样本的统计特性易淹没于噪声中造成难分类问题,提出随机特征裁剪方法,使算法避免过度关注异常特征,降低了极难分类样本对AdaBoost.M2算法性能的负面影响,提升了算法的泛化能力,并以低信噪比数据进行实验验证。针对调制类型同族信号难分类的问题,选取同族调制类型的通信信号开展模型训练和测试。实验结果表明:相比于单一卷积长短时记忆全连接深度网络(CLDNN)算法,改进AdaBoost.M2算法对低信噪比PSK族类和QAM族类通信信号的测试集准确率分别提高了8.5%和11.25%,相比于直接集成CLDNN的经典AdaBoost.M2算法,测试集准确率分别提高了8.25%和6.5%。A signal modulation recognition method is proposed based on AdaBoost.M2 algorithm to address difficult identification of signals from the same family of modulation types and the poor generalization of the deep learning model.An improved method of sample weight is proposed to solve the problem that the learning rate of the weak learning algorithm is difficult to adapt to the weighted sample data in the case of large samples.The improved sample weight ensures that the order of magnitude of the training sample data remains unchanged after weighting,so that the algorithm pays more attention to the difficult classification samples,improving the comprehensive performance of the weak classifier.In addition,in view of the difficult classification problem caused by the statistical characteristics of some samples easily submerged in noise,a random feature clipping method is proposed to avoid much attention given to abnormal features.This method reduces the negative impact of extremely difficult classification samples on the performance of AdaBoost.M2 algorithm,improving the generalization ability of the algorithm.Experimental verification with low signal-to-noise ratio data is conducted.Finally,given the fact that the signals of the same family of modulation types are difficult to classify,the signals of the same family are selected for model training and testing.Results show that the improved AdaBoost.M2 algorithm increases the test set accuracy of PSK family and QAM family by 8.5 and 11.25%respectively compared with the single CLDNN algorithm,and by 8.25 and 6.5%respectively compared with the classical AdaBoost.M2 algorithm.

关 键 词：AdaBoost.M2算法 深度学习 调制分类 样本权重 过拟合 

分 类 号：TN911.3[电子电信—通信与信息系统]