Key Technologies Of Modulation Classification For Non-Cooperative Signals Based On Deep Learning

With the continuous development of radio technology and computer technology,communication technology has shown great development potential in civil and military use,and a large number of communication equipment has been put into use,but this has also led to the complexity of the electromagnetic environment,and the existence of noise and malicious interference has aggravated the deterioration of communication conditions.In order to make better use of spectrum resources,improve communication quality and communication efficiency,cognitive radio technology has received extensive attention from researchers at home and abroad.Cognitive radio technology obtains the current communication environment and communication state based on perception recognition technology,and then makes reasonable adaptive adjustments to the occupation of spectrum resources to ensure high-quality communication status.Therefore,the perception recognition of signals is the basis of cognitive radio,especially in the non-cooperative electromagnetic environment and the presence of malicious interference,accurate and efficient perception of signal types is particularly important.Automatic modulation recognition is an important part of cognitive radio technology in analyzing signals and sensing signals,so it is of great significance to study modulation recognition under non-cooperative conditions.In this paper,an intelligent recognition method of signal modulation style type is realized by deep learning under non-cooperative conditions where label samples are difficult to obtain or contain all classes of training sets that are difficult to construct are investigated.The main contents of this paper are as follows:(1)An semi-supervised modulation recognition method for small-sample conditions was studied.For most of the unlabeled data,the feature extraction network is pre-trained by contrastive learning,and the classification network is designed and constructed downstream to complete the classification task by fine-tuning the small-sample label signal,which is verified in the public dataset.(2)An intelligent unsupervised clustering method for modulating types is proposed.Similar to the semi-supervised method,the unsupervised method also uses contrastive learning as a pre-training method,and then constructs the clustering network in the downstream clustering task,and uses the clustering loss to complete the clustering task,and also designs a self-labeling loss to optimize the parameters of the entire clustering network by using reliable negative sample labels,and finally confirms the reliability of the method in the public dataset(3)A method to solve the problem of zero-sample signal recognition in complex electromagnetic environment is given.This paper argues that even if sufficient label data is obtained,it is difficult to construct a training set containing all categories of cases under actual conditions.Therefore,a two-stage feature transfer algorithm is designed to solve the zero-sample problem.Therefore,a two-stage feature transfer algorithm is designed to solve the zero-sample problem.Firstly,the feature extraction network is trained to obtain the initial feature vector of the class,then the adjacency graph is constructed and then fed into the graph convolutional network to obtain the final features,and finally an adjacency graph loss is designed to optimize the network model.The effectiveness of this method is demonstrated in both the real source dataset and the public modulation style dataset...