Research On Communication Signal Modulation Recognition Based On Deep Learning

Modulation recognition technology refers to a technology that aims to recognize the modulation type of the target signal in non-cooperative communication scenarios.Modulation recognition technology is widely used in civil radio monitoring,military electronic countermeasures and other fields.It is an important research branch in the field of communication.The recognition performance of traditional classifiers relies on the choice of decision theory,signal features and classifiers,and the recognition accuracy is limited by the complex electromagnetic environment.However,deep learning methods can give full play to the feature extraction capabilities of the original data,avoiding manual feature extraction,and the recognition accuracy is higher than that of the classical algorithms.In this thesis,the modulation recognition technology of communication signals based on deep learning is deeply studied.The main works are as follows:(1)For the modulation recognition of analog and digital mixed modulation signals,a deep learning network based on multi-task network with attention mechanism is proposed.Firstly,an auxiliary learning model combining main task and auxiliary task is designed.The time domain information and frequency domain information of the modulated signals are used as the signal input of the main network and the auxiliary network respectively.By making full use of the potential information of the auxiliary task,the characteristic structure of the main network is adjusted to increase the generalization ability of the network.Secondly,the convolutional block attention mechanism module is introduced into the auxiliary network.Adaptive feature refinement is performed on the feature images extracted from the convolutional layer in spatial dimension and channel dimension.The local spatiotemporal feature extraction is realized and the ability of feature representation is enhanced.In addition,aiming at the above multi-task learning model,a two-step training scheme is designed to reduce the training complexity of the model effectively.The simulation results based on RML2016.10a data set show that the auxiliary network can effectively improve the identification accuracy of the main network in PSK type modulation signals.At the same time,compared with the existing classical networks,the proposed network can effectively improve the recognition accuracy at low SNR and has strong robustness.(2)For the modulation recognition of digital modulation signals,a deep learning network based on multi-channel separable complex structure is proposed.Firstly,the multi-channel network structure is used as the basic architecture to make full use of the modulation characteristic information of different channels of the digital modulation signals.Secondly,based on the multi-channel architecture,a complex convolutional layer structure is proposed to realize two-dimensional feature extraction,which effectively utilizes the two-dimensional feature information of digital signals.In addition,in the one-dimensional channel,a deep separable structure is used to reduce the complexity of multi-channel structure and realizes the effective extraction and fusion of multi-dimensional features.Finally,the dimensionality reduction of multi-dimensional features is realized by the fusion feature extraction module,and the time series feature processing is realized by the long short-time memory network to further realize the classification and recognition.The simulation results show that the complex network structure and the depth-separable model can effectively improve the recognition accuracy of the digital modulated signal,and the recognition accuracy can reach 97.5%at high SNR.Compared with the basic multi-channel model,the proposed method not only improves the accuracy of network identification,but also realizes the model lightweight processing.At the same time,the comparison with the existing classical networks shows that the proposed network has outstanding advantages in the recognition accuracy and robustness.