Research On Communication Signal Modulation Recognition Algorithm Based On Interpretable Deep Learning

The existing communication signal modulation recognition algorithms based on deep learning lack interpretability and high redundancy.From the perspective of interpretability,this thesis studies the modulation recognition algorithm based on interpretable deep learning.Firstly,this thesis introduces the modulation principle of signal in detail,and describes the generation of ASK(Amplitude Shift Keying),PSK(Phase Shift Keying),FSK(Frequency Shift Keying)and QAM(Quadrature Amplitude Modulation)signals in theory.Then it introduces the modulation recognition model based on deep learning at this stage,theoretically expounds two types of deep neural networks CNN(Convolutional neural network)and RNN(Recurrent Neural Network),and uses five classical deep network models LSTM(Long Short-Term Memory),AlexNet,VGG16,ResNet50 and DenseNet to carry out modulation recognition simulation experiments on 11 types of modulation signals in the public data set of Radio ML 2016.10 a modulation signals,Next,the widely used hidden layer visual interpretation methods and live interpretable algorithms are introduced,and these algorithms are verified on the common data sets of images and texts.Then,an interpretable framework of modulation recognition depth neural network is proposed.The framework includes time-frequency feature extraction of modulation signal,training depth convolution neural network for modulation recognition of modulation signal and two visual interpretation methods: Based on Grad-CAM(Gradient-weighted Class Activation Mapping)and based on LIME-CAM(Local Interpretable Model-agnostic Explanations Class Activation Mapping).Grad-CAM,as a widely used visual interpretation method,is used to explain the time-frequency characteristics of modulated signals for the first time in this thesis.Aiming at the problems existing in Grad-CAM,this thesis proposes a visual interpretation method based on LIME-CAM.Then,simulation experiments are carried out to visualize the key depth features inside the depth model to promote signal recognition.Explain the decision-making basis that the depth network can correctly identify the modulation signal from different angles,reveal the internal mechanism of network performance degradation under the environment of low signal-to-noise ratio based on the extracted depth characteristics,analyze the reasons for the difference of network performance identified by different timefrequency modulation,establish the accuracy evaluation index of visual interpretation method,and verify the effectiveness of the method proposed in this thesisThen,based on explicability,this thesis proposes two pruning criteria for importance evaluation: Based on grad cam and based on lime-cam.Then the two pruning criteria are applied to the pruning of VGG16 and ResNet34 depth models respectively.The experimental results show that the two interpretable modulation recognition network pruning methods can maintain the recognition accuracy of the modulation signal and speed up the running speed of the model on the premise of massive compression of the model.In addition,the performance of pruning method based on LIME-CAM importance evaluation is slightly better than that based on Grad-CAM importance evaluation.Finally,this thesis designs an interpretable lightweight modulation recognition network(LCPI-Net).The network can visually interpret through the importance feature map m and life-cam,and the importance feature map m is added to the modulation recognition network structure and network training process,which improves the recognition accuracy of the modulation recognition network on the test set.In addition,the network prunes and optimizes the results of example interpretation based on the model,which greatly reduces the redundant modulation identification network structure and speeds up the operation speed of modulation identification network.The experimental results show that the accuracy of LCPI-Net on the modulated signal data set is higher than the current baseline network,and the network parameters and the memory occupied by the model of LCPI-Net are lower than the other seven classical deep convolution networks.