摘要
针对非合作通信条件下信号调制方式识别问题,提出一种基于Inception-V4模型的通信信号调制方式自动识别新方法。该方法首先将数字信号预处理成星座图,然后将星座图转换成灰度图,最后将灰度图作为神经网络模型的输入,利用Inception-V4网络模型对通信信号的调制方式进行自动识别。Inception-V4模型对信号星座图的形变具有不敏感性,自动学习各种数字调制信号的星座图特征,克服了传统信号调制识别方法中信号特征提取困难,通用性不强,抗噪声性能差等缺点,处理流程简单。仿真实验制作BPSK、4ASK、QPSK、OQPSK、8PSK、16QAM、32QAM、64QAM等8种典型信号的星座图,并选择在信噪比(SNR)为3.5 dB、3 dB、2.5 dB进行调制方式识别试验。在信噪比为2.5 dB时,BPSK、QPSK、4ASK、OQPSK、8PSK的识别准确率依然保持在99.5%以上,实验结果表明,基于Inception-V4模型的通信信号调制方式识别新方法是有效的,且在低信噪比环境下比传统的识别方法有更高的识别准确率。为了评估In-ception-V4模型在信号调制方式识别方面的性能,与AlexNet、InceptionResnetV2-TA模型在调制方式识别的效果进行对比,针对2种典型的调制信号集合进行仿真实验,Inception-V4模型识别准确率较InceptionResnetV2-TA模型提高5%,比AlexNet模型提高12%,进一步表明:将深度学习在图像分类识别领域的模型迁移到通信信号的星座图分类同样适用,对提高非合作通信条件下信号调制方式识别准确率有很大帮助。
Aiming at the problems of signal automatic modulation recognition in non-cooperation communication systems,a novel method of communi-cation signal modulation recognition which is based on the Inception-V4 model is proposed.Firstly,the received signal be generated con-stellation image,then converts the constellation image into a gray image,which used as the input of the neural network model which istrained to classify modulated signals.The Inception-V4 model can automatically learn the constellation diagram features of various digitalmodulation signals,which can simplify the processing procedures and overcome the weaknesses of traditional techniques,such as the diffi-culty in extracting the features,the absence of universal property,and the poor anti-noise performance.In addition,the deformation of theconstellation diagram is insensitive to the final classification performance by using Inception-V4.The simulation experiment produced theconstellation diagrams of eight typical signals including BPSK,4 ASK,QPSK,OQPSK,8 PSK,16 QAM,32 QAM,64 QAM,and selected theSignal-to-Noise Ratio(SNR)of 3.5 dB,3 d B,2.5 dB for the modulation identification test.The recognition accuracy of BPSK,QPSK,4 ASK,OQPSK,and 8 PSK still remains above 99.5%at 2.5 dB.The experimental results show that the new method of communication signalmodulation recognition is effective and has higher accuracy rate for digital signal modulation recognition at low SNR.In order to evaluatethe performance of the Inception-V4 model in signal modulation recognition,simulation experiments were carried out for two typical modu-lation signal sets,and compared with the AlexNet and InceptionResnetV2-TA models in the modulation recognition effect.The recognitionaccuracy of the Inception-V4 model is 5%higher than the InceptionResnetV2-TA model and 12%higher than the AlexNet model underthe same data set.It further shows that migrating the model of deep learning in the field of image classification and recognition to the con-stellation classification of communication sign
作者
邢科
吕泽均
XING Ke;LV Zejun(College of Computer Science,Sichuan University,Chengdu 610065)
出处
《现代计算机》
2021年第12期48-54,共7页
Modern Computer
