Research On Signal Modulation Recogination Based On Convolutional Neural Networks

Signal modulation recognition plays an important role in non-cooperative communication.It is critical to develop more proven methods to accurately distinguish the modulation type used by the received signal.Therefore,this thesis combines convolutional neural network(CNN)to study signal modulation recognition,and the main contents are as follows:1.Modulation recognition methods based on CNN take effect without manual experience.However,training deep learning models requires massive volume of data.An insufficient training data will cause serious overfitting problem and degrade the classification accuracy.To cope with small dataset,data augmentation has been widely used in image processing.However,in wireless communication areas,the effect of different data augmentation methods on radio modulation classification has not been studied yet.Therefore,this thesis studies a signal modulation recognition method based on data augumentation and Alex Net.Firstly,the method preprocesses the modulated signals in the public dataset Radio ML2016.10 a into constellation representations,and divides the constellations into a training set and a test set according to the ratio of 8:2.The divided training set is then data augmented using random erasing,Cut Mix,and rotation,and an augmented training set is obtained.Finally,the image features of the augmented training set are automatically extracted through the Alex Net model,and the model is tested according to the test set.The simulation results show that,compared with the modulation recognition method without data augumentation,the proposed method improves the performance within a certain range of SNR.And the performance of rotation is better than random erasing and Cut Mix.2.The similarity between Many Phase Shift Keying(MPSK)signals is higher,so the intra-class recognition is more difficult.And there is currently no intra-class recognition method specifically for MPSK(M=2,4,8,16)signals.Therefore,this thesis studies an automatic modulation recognition method for signals based on ISPP-Res Net18 and constellation in Gaussian channels.Firstly,the method converts the signals into constellation diagram representations.Then,aiming at the problem that the traditional spatial pyramid pooling(SPP)is difficult to integrate the multi-scale local characteristic information of the same convolution layer,the SPP is improved so as to obtain an improved spatial pyramid pooling(ISPP).Then ISPP is introduced into the classic convolutional neural network Res Net18,forming the ISPP-Res Net18 network.Finally,combining the constellation diagram and ISPP-Res Net18 network,automatic modulation recognition is completed.And the impact of the number of symbols,the size of selected area,the image resolution and the kind of activation function on performance is researched.The simulation results show that compared to the modulation recognition method with traditional spp or without ISPP,the proposed method improves the performance when SNR is less than 6d B,which proves that ISPP is more effective than SPP.