Radio Signal Recognition Based On Hybrid Neural Network

Modulation recognition is a key link between signal detection and demodulation.The feature engineering method adopted in the traditional modulation recognition algorithm is an artificial feature extraction method based on expert prior knowledge,which requires high requirements for researchers and signal analysis equipment and is generally robust.In recent years,the popular modulation recognition based on neural network can effectively simplify the process and improve the robustness of the algorithm.However,there are still some problems,such as insufficient recognition accuracy under low SNR.In addition,the traditional method model and some deep learning models have high complexity,which is difficult to meet the needs of lightweight deployment in practical applications.In view of the above problems,this thesis designs two different deep learning models to apply to radio signal recognition,and deploys them under the deep learning reasoning framework to solve the problems of over reliance on manual design features and insufficient applicability and robustness in radio signal recognition,and realizes end-to-end radio signal recognition.The main research contents and achievements of this thesis are as follows:1.Based on the feature that attention mechanism can extract more valuable information from data,this paper uses the original IQ signal as the model input,residual network as the spatial feature extraction module,and LSTM layer as the temporal feature extraction module.Combining the two in series,a hybrid neural network model based on attention mechanism is designed to integrate different types of attention mechanisms with residual network,Explore the attention mechanism applicable to this problem,and finally propose a hybrid neural network model based on Se attention mechanism.When the SNR is greater than or equal to zero,the average accuracy is close to 90%,which is verified to be better than the existing methods.2.Aiming at the problem that the recognition accuracy is still low when the SNR is low,in order to improve the accuracy under the condition of low SNR and further extract the signal features,a hybrid neural network method based on filtering noise reduction is proposed on the above basis.The spatial feature extraction module uses Densenet network with stronger feature extraction ability,combined with the noise reduction ability of Savitzky Golay filtering algorithm,The original IQ signal is denoised as the input of the model.The network still concatenates the output of the Densenet network with the LSTM layer and sends it to the classifier.Finally,Bayesian algorithm is used to optimize the super parameters of the model.Compared with the above methods,it is obviously improved at low SNR and high SNR.3.Aiming at the problem that the model parameters in the previous chapter are large and the training and reasoning time is long,the model in the previous chapter is improved by using the deep separable convolution block network to reduce the calculation amount of the model effectively without losing performance.Finally,the model is deployed in the onnxruntime framework of the C + + environment,which verifies the effectiveness and real-time performance of the method...