The Analysis Of The Method For Modulation Recognition Based On Semi-supervised Learning

In traditional supervised deep learning,the models learn from a large number of labeled training samples to build themselves for inferring the labels of new samples.However,the number of labeled samples of modulation signals in the current electromagnetic space are seriously insufficient.Although it is quite easy to obtain a large number of unlabeled samples with the rapid development of data collection and storage technology,but it is still more difficult to obtain a large number of labeled samples,because it will take a lot of resources and time to fully label the unlabeled samples.Obviously,if the models only use a small number of labeled samples for training,it is difficult for themselves to obtain excellent analytical abilities;On the other hand,if only a few “expensive” labeled samples are used without a large number of “cheap” unlabeled samples,there will be a huge waste of data resources.At the same time,deep learning has a lot of room for improvement in modulation recognition of the electromagnetic space.Firstly,although deep neural networks have been successful in many fields,the corresponding proprietary model has not yet been born in the field of modulation recognition.Secondly,the electromagnetic space of the current world is very complex and changeable.As a result,deep networks cannot be “one-time-for-all”,and need to be retrained according to the changes of the modulation signal types in the monitored environment to maintain the credibility of themselves.However,this approach obviously has problems such as consuming resources and insufficient timeliness.According to the above background analysis,the core tasks that this paper needs to carry out are: To analyze the unique properties of modulation signal data that distinguish them from other data types in the field of deep learning in depth,and on this basis,find a breakthrough to improve the performance of modulation recognition under the condition of insufficient labeled samples,then research and design the innovative methods of modulation recognition suitable for semi-supervised learning scenarios.The main work and innovations of the paper are as follows:First of all,based on the frontier theory of the current deep learning field—Representation learning,the analysis draws the conclusion that there is a potential generalization ability between the modulation signals of different orders in the same type;Afterwards,with part of the Convolutional Neural Network of modulation recognition for high order signals as a feature extractor,the basically separated feature clusters of various types of high order and low order modulation signals are obtained in simulation experiment using the dimensionality reduction and visualization technology,thus verifying the correctness of the above conclusion.On this basis,a kind of semi-supervised modulation recognition method under high signal-to-noise ratio is designed combined with the Density Based Spatial Clustering of Applications with Noise under the experimental condition of the semi-supervised learning environment with sufficient labeled samples of higher order types and insufficient labeled samples of lower order types in the same type of modulation signal,and shows good performance in simulation experiment.Subsequently,in order to solve the problems of the density-based clustering method that the feature cluster interval is small under low signal-to-noise ratio,the confidence of the decision results decreases when the purity of modulation types within clusters is insufficient,and one single category may have multiple clusters,it is improved to the Graph Based SemiSupervised Learning method,and an innovative semi-supervised modulation recognition combined model is designed to get rid of the requirements for high signal-to-noise ratio and purity of types within clusters,and further realize the automation of modulation recognition.At the same time,the actual meaning and setting method of the key parameters in the Graph Based Semi-Supervised Learning are analyzed,and related simulations are given.Finally,the modulation recognition performance of the semi-supervised recognition combination model designed in this paper and the supervised learning method under the condition of completely labeling,incompletely labeling and incompletely labeling in conjunction with category balance method are compared through simulation experiments.It is confirmed that the correct classification probability of this model is higher than the latter two when dealing with the scene where the labels of part of samples are missing,and also the method with the closest performance to the supervised learning under the condition of complete sample labeling,which reflects this model is a feasible and advanced method for modulation recognition based on semi-supervised learning.