Research On Modulation Recognition Of Communication Signals Based On Machine Learning

As the key technology of wireless communication system receiver,communication signal modulation recognition has important application value in cognitive radio,spectrum monitoring and other fields.With the complexity of communication scenarios and the diversification of modulation methods,the traditional modulation recognition methods based on likelihood function and signal characteristics become more and more complex,which is difficult to meet the needs of high reliability and low delay in wireless communication.How to apply machine learning method to modulation recognition task,combined with the characteristics of data distribution,to achieve efficient and accurate recognition,is the focus of this paper.It is difficult to obtain a large number of training samples.However,in the practical communication scenario when machine learning is applied to modulation recognition.For example,in the MIMO spatial-correlated channel,the correlation between antennas brings many difficulties to the model establishment and sample acquisition.Therefore,how to make full use of the existing samples,to achieve accurate classification of modulated signals is the current main tasks.To solve this problem,this paper combines transfer learning with active learning,and proposes a modulation recognition method based on active transfer learning.Firstly,in the source domain,the convolutional neural network model is pre-trained by the sufficient data samples in the Additive White Gaussian Noise channel.Then,the pre-trained parameters are transferred to the MIMO system of the target domain to realize parameter sharing.Combined with the active learning method,the unlabeled sample data with the most information is selected by the query function,and then the labeled sample data is added to the training set,and remove the data samples that do not conform to the characteristics of the target domain.The proposed method not only makes the model training more valuable by the selection of important samples,but also promotes the model convergence by knowledge transfer.Experimental results show that the proposed method effectively overcomes the problem of insufficient data samples,the recognition accuracy is significantly better than other traditional methods.Currently,the modulation recognition methods based on machine learning often have the problem of uneven separability,that is,when the average recognition accuracy is high,some modulation methods which are easy to be confused are difficult to distinguish,resulting in model defects.To solve this problem,this paper proposes a modulation recognition method based on semi-supervised structure.This method combines the unsupervised clustering with supervised classification and recognition.Firstly,sparse selfencoder is used to reduce the dimension of signal samples.Then,fuzzy C-means clustering is used to divide the samples with large similarity into a cluster,and the samples with large difference are distinguished from each other.Finally,the shallow features are combined with the deep features,and CNN is used to classify the mixed clusters.Simulation results show that compared with the supervised learning method,the proposed algorithm can break through the suboptimal performance of neural network,effectively overcome the problem of classification nonuniformity,and has certain application value.