| 摘要: |
| 基于信号特征进行模式识别的调制识别方法需要先计算信号的高阶特征、高阶累积量再进行模式识别，整体设计复杂，特征不易计算。机器学习技术由于其强大的特征提取能力和分类能力，被广泛应用到模式识别领域中。针对调制识别问题，提出了一种基于欠完备自编码器的调制识别技术，使用欠完备自编码器进行调制信号的特征自动提取，再使用神经网络分类器进行分类识别。整体模型更为简洁，运算复杂度较低，有利于部署在硬件上进行实时识别。对常见的BPSK、QPSK、2ASK、2FSK、16QAM数字调制方式进行的识别实验表明，算法在信噪比10 dB时平均识别率高于0.97，并且在信噪比为0 dB时仍然有0.92以上的平均识别率。 |
| 关键词: 认知无线电 调制识别 神经网络 欠完备自编码器 特征提取 |
| DOI： |
|
| 基金项目:北京市科技计划课题资助（Z191100001419001）；北京市优秀人才资助计划青年拔尖项目（2016000026833ZK08）；北京市属高校高水平教师队伍建设支持计划“青年拔尖人才培养计划”（CIT&TCD201704065）；国家自然科学基金资助项目（61620106001）；北京信息科技大学学科建设类项目（5121911006） |
|
| An Undercomplete Autoencoder Modulation Recognition Technology |
| ZHANG Peiyue,XU Zhan,ZHAO Yiyang,CHEN Jinhui,ZHI Ruxin |
| (1.School of Information and Communication Engineering，Beijing Information Science & Technology University,Beijing 100101,China;2.Simpleedu Co.,Ltd.,Beijing 100190,China) |
| Abstract: |
| In modulation recognition methods based on signal feature,high-order features or high-order cumulants of the modulated signal usually need to be calculated,which is complex in general design and difficult in calculating.Machine learning technology has been widely used in pattern recognition due to its powerful feature extraction and classification capabilities.Therefore,for the problem of modulation recognition,a technology based on undercomplete autoencoder is proposed.Undercomplete autoencoder is used to automatically extract features of modulated signal,and a neural network classifier is used for recognition.The overall model is more concise and has lower computational complexity,which facilitates real-time recognition on the hardware.Experiments show that five common digital modulation methods can be identified by this method,including BPSK,QPSK,2ASK,2FSK,and 16QAM.The average recognition accuracy is higher than 0.97 when the signal-to-noise ratio(SNR) is 10 dB,also the average recognition accuracy can be above 0.92 at 0 dB SNR. |
| Key words: cognitive radio modulation recognition neural network undercomplete autoencoder feature extraction |
《电讯技术》编辑部 