Research On Modulation Recognition Method Based On Deep Learning In Jamming Environment

Modulation recognition has always been an important research area in communication systems.Regardless of electronic reconnaissance,identification of enemy and friend in military field,or cognitive radio in civil field,modulation recognition has its important applications and it is one of the most core technologies.In recent years,with the application of 5G and the development of AI technology,the research on modulation recognition technology has become more popular.At present,modulation recognition is mostly studied under the condition of Gaussian white noise,and some research results have been found.However,there is still a long way from the application.Because the electromagnetic environment of communication equipment is becoming increasingly complex,and there are various forms of jamming.Most of the original modulation recognition schemes are sensitive to jamming.Therefore,the design of modulation recognition schemes in the jamming environment has become a hot and difficult point of research.Based on this,the thesis considers applying deep learning algorithms to modulation recognition,and proposes a modulation recognition scheme based on jamming recognition to improve the performance of the system in the jamming environment.First,the thesis summarized the jamming signal models commonly used in anti-jamming communication and several classic neural network structures.The back-propagation algorithms of fully connected networks and convolutional neural networks are also derived.Next,based on the RML2016.10 a data set,the thesis analyzed and compared the modulation recognition performance under different jamming environments of four networks,CNN,CLDNN,Res Net and Inception.The results show that the CNN network has the lowest complexity and the best performance.Then,a modulation recognition scheme based on jamming recognition is proposed.The scheme selects the modulation recognition deep learning networks according to the results of jamming recognition.Based on the characteristics that the singular values of different jamming signals have large differences,a jamming recognition method based on singular value decomposition is proposed.The simulation results show that the proposed method has an accuracy advantage of 15% ? 25% when compared with the traditional method at low jamming to signal ratio.Finally,a simulation experiment was carried out on the modulation recognition system based on jamming recognition,which is in the case of random jamming applied to the RML2016.10 a data set.The results show that the scheme has a greater performance improvement than the deep learning scheme,especially when the signal to jamming ratio is lower than-10 dB,the recognition accuracy is improved by about 10%.Finally,for the RML2018.01 a data set with a large amount of data,we choose four deep learning networks such as ResNet,VGGNet,SE-ResNet and DenseNet,and give the simulation and comparison of the modulation recognition performance of each network.The results show that ResNet has a better performance and lower complexity.Under the ResNet algorithm,the influence of sampling rate and different signal interception lengths on the recognition accuracy is analyzed.The results show that properly reducing the sampling rate will not affect the recognition performance of the network,but also reduce the input signal dimension which can reduce the amount of network parameters.Finally,a simple sampling parameter selection strategy is proposed,which can give the optimal parameter selection under the consideration of performance,delay and complexity,which can be a reference in the design of deep learning network.