Research On Classification Of Signal Adaptive Modulation Based On Dataset Enhancement And Knowledge Distillation

The development of technology in the field of communication facilitates communication between people,which results in a large increase in the number of wireless devices and demand for spectrum resources.At the same time,the research direction of scholars is not only how to improve the utilization rate of spectrum resources,but also how to ensure secure communication.To avoid eavesdropping and other situations,the sender can randomly change the modulation method of the signal to ensure the safety of communication.Meanwhile,the receiver needs to perform signal modulation adaptive classification,which plays an important role in the field of civil and military wireless communication.At present,the research direction of signal adaptive modulation classification is to use neural network for feature-based classification,and the classification accuracy can be further improved by modifying the dataset and network structure.First of all,in this paper,an improved Triple-GAN network is used to expand the original dataset,which solves a problem that the convolutional neural network was limited in signal modulation classification due to insufficient training data and over-fitting.Furthermore,the newly generated samples provide additional auxiliary information to train the classification network.The accuracy of signal modulation classification can be improved by using the expanded dataset to train the signal modulation classification network.The IQ vectors in the expanded dataset are calculated to obtain the corresponding instantaneous amplitudes and phases,and the calculated results are taken as the input of the network to verify whether the generated data has a real meaning.Secondly,in this paper,by improving the classification network,the ResNeXt with excellent performance in classification is applied to the adaptive classification network to improve the classification performance of the network.However,some portable devices are limited by hardware,and cannot deploy large-scale networks.So the improved network is compressed through knowledge distillation,and through migration learning,the simple network achieves higher classification performance than before.Finally,different signal modulation classification networks are used for experimental comparison.The experimental results show that the expanded dataset using the improved Triple-GAN can promote the classification accuracy of the signal.In addition,the improved network achieves higher classification accuracy,and under the guidance of the improved network,the simple convolutional network also achieves higher classification performance.