Recognition Of Communication Signal Modulation Types Based On Deep Learning

Signal modulation type recognition is a technology that determines the modulation method of a wireless communication signal based on the received signal,which has wide applications in both civilian and military communication fields.Traditional modulation type recognition methods mainly include maximum likelihood-based methods and feature extraction-based methods.The former relies on accurate channel model parameters and has high algorithm complexity,while the latter relies on expert-designed features and is difficult to guarantee recognition performance in practical applications.In recent years,with the mature application of deep learning in the fields of computer vision and natural language processing,there has been increasing attention on deep learning-based modulation recognition methods,which have demonstrated superior performance advantages over traditional methods.This thesis focuses on the research of deep learning-based signal modulation type recognition and proposes the following:1.Communication signals may be in a silent period or have a short signal sequence length,which can lead to issues such as misidentification or confusion during modulation recognition.To solve this problem,a multi-channel feature fusion(MFF)network composed of a feature pyramid network and a long short-term memory network in parallel is proposed.First,communication signal data is preprocessed to obtain the amplitude,phase,and fractional Fourier transform of the signal.Then,a feature pyramid network is used to extract the temporal information of the raw signal and the results of the signal’s fractional Fourier transform.A long short-term memory network is used to extract the amplitude and phase information of the signal.Finally,the features extracted by the two networks are merged,and the overall characteristics of the acquired signal are utilized for the purpose of modulation type recognition.The experimental results indicate that the proposed network can to a certain extent solve the problem of misjudgment of wide band frequency modulation(WBFM)signals and the confusion problem of multiple quadrature amplitude modulation(MQAM)signals; compared with classical neural networks,the proposed network has a higher recognition rate.2.To address issues such as overfitting and poor recognition performance caused by a small number of actual communication signal samples,a modulation recognition method based on a siamese convolutional long short-term deep neural network(SCLDNN)is proposed.This method sends original sequence signals with the same length and representation into the twin network input end,performs feature extraction and weight sharing through a convolutional long short-term deep neural network,uses the Euclidean distance for measurement,and a nearest-neighbor classifier is used to achieve modulation type identification of multiple signals.The experimental results demonstrate that,compared with classical neural networks and generative adversarial networks,the proposed network has superior recognition performance.3.When the actual data differs from the training data,traditional neural network models usually require retraining the data,which can lead to high training time costs.To address this issue,a transfer learning method is proposed,which combines fine-tuning theory with center loss cost function.The base network used in this method is a multi-channel feature fusion network,and the main weights of the network are frozen after pre-training.The fully connected layer is then modified to incorporate the center loss cost function for secondary training.The experimental results demonstrate that the proposed method effectively reduces the number of convergence iterations in secondary training,thereby reducing the training time of the network.Compared to domain adaptation transfer methods,the proposed method exhibits superior recognition performance.