Research On Modulation Pattern Recognition Algorithm Based On Neural Network

Neural networks modulation recognition is a novel technology that uses the powerful nonlinear approximation ability of neural networks to identify signal modulation patterns.The performance bottleneck of traditional signal recognition technology is expected to be broken.However,the interpretability of neural network classifiers is still lacking in the field of wireless communication.Therefore,the investigation into the characteristics,parameter influence,and interpretability of convolutional neural networks(CNN)has been performed in this thesis.The major contributions are as follows:(1)For the modulation recognition algorithm based on signal characteristics,the performance of K-Nearest Neighbor(KNN),Support Vector Machine(SVM),and ensemble learning are compared on the same dataset,where the dataset is composed by different characteristic parameters signals and different modulation types signals.The experimental results show that the SVM method can achieve the best recognition accuracy.(2)A deep investigation into the theoretical and realization of CNN recognition technology has been performed.And the transformation characteristics of the wireless signal through the convolutional neural network have been analyzed and deduced.The analysis shows that different signals will cause different additional frequency components due to the nonlinearity nature of the activation function.The characteristic difference of the modulated signal is increased to realize the accurate modulation recognition.In addition,the CNN network structure can be optimized by evaluating the impact of additional feature differences generated by different CNN network parameters on the performance of the model.The constructed CNN network achieves95.3% recognition accuracy when SNR=18d B.Compared with SVM,the average recognition accuracy of the CNN algorithm is improved by 7%.More importantly,a well-trained CNN model exhibits strong robustness to the effects of Gaussian noise.(3)A CNN modulation recognition visualization technique is proposed.The function of the convolution kernel in the network and the output features of the hidden layer are analysed in the perspective of the frequency domain,where the convolution kernel will retain different frequencies of information.The average pooling layer is visualized by the weight of the fully connected layer.Meanwhile,this dissertation explores different hyperparameter setting via extensive numerical evaluations of Gradient-weighted Class Activation Mapping(Grad-CAM++).Features relating to modulation reference points is extracted by Grad-CAM++.And Grad-CAM++ captures the transitions between modulation reference points to identify different modulation pattern.In addition,the t-Distributed Stochastic Neighbor Embedding(t-SNE)algorithm is used to visualize the output of the fully connected layer,which intuitively shows the degree of confusion and modulation recognition performance of the neural network.