Research And Implementation Of Communication Signal Modulation Recognition Based On Hybrid Neural Network

As a key technology in wireless communication,modulation recognition is widely used in various military and civilian fields.By identifying the modulation mode of the received unknown signal,the relevant signal processing work is coordinated,so as to improve the performance of the communication system.With the comprehensive digitalization and intelligence of the world,the rapid development of wireless communication puts forward higher requirements for modulation recognition:1)Accurate and efficient identification under diversified modulation modes;2)Lightweight identification on the intelligent hardware side;3)Robust identification in complex electromagnetic environments.Therefore,by in-depth study of hybrid model modulation recognition,lightweight modulation recognition and robust modulation recognition under noise mismatch,this paper aims to combine signal modulation recognition problems with deep learning,simplify feature extraction,improve recognition efficiency and accuracy,futher lightweight the network model,facilitate the deployment and implementation of edge devices,finally consider the influence of non-Gaussian noise on signal recognition,improve the model with related methods,reduce the network model and improve the robustness of the algorithm under the premise of ensuring the accuracy.The main work is:(1)Aiming at the problem that the operation of traditional recognition algorithms is complex and the recognition rate is insufficient,Chapter 3 designs a modulation recognition method based on convolutional and cyclic hybrid models.Firstly,according to the rich spatial and temporal characteristics of complex signals,the signal data is amplified.Then,by combining spatial modules on the timing module,maximized learning of the original signal is facilitated.Experiments show that this method can effectively improve the recognition accuracy.(2)Aiming at the problem that the recognition model is large and difficult to embed deployment,Chapter 4 further combines pruning and deep separable convolution operations to give a lightweight modulation recognition method.First,the model is wholly and locally pruned to remove redundant weights,then the depth separable convolution is replaced to reduce the amount of computation.Experiments show that this method can significantly reduce the model without degrading the recognition performance.(3)Aiming at the unstable identification in the actual complex electromagnetic environment,Chapter 5 considers the noise mismatch scenario and proposes a robust modulation recognition method based on K dimensional tree and network optimization.Firstly,the modulated signals in different noise environments are generated,then the K dimensional trees are used for preprocessing to enhance the separability of different signals,GoogLeNet and squeeze-and-excitation blocks are further mixed to assign more weights to important features.Experiments show that the method not only has high recognition accuracy under Gaussian and non-Gaussian noise,but also has good robustness under noise mismatch.