Research On Automatic Modulation Classification Based On Deep Learning

With the rapid development of wireless communication technology and the continuous expansion of its application fields,the Radio spectrum resources are increasingly scarce,and the electromagnetic environment is increasingly complex.Therefore,according to the different external electromagnetic environments,the idle spectrum is automatically searched and used,achieving high spectrum utilization and optimal communication performance is extremely important.As a key technology of cognitive radio,modulation identification is an indispensable part of analyzing,processing,detecting,and demodulation signals.This thesis aims to use deep learning technology to study the problem of digital signal modulation classification,to improve the accuracy and robustness of digital signal modulation classification.The main content of this thesis is as follows:1.Automatic modulation classification using cascaded convolutional neural networkIn the scenario where a single antenna receives a single signal,we propose a signal modulation classification method based on a cascaded convolutional neural network(CasCNN).This method uses feature optimization engineering to extract feature of high quality from received signals,and the density of the constellation diagram(DCD)are input into the network.CasCNN is a cascaded convolutional neural network,which can decompose a complex modulation classification task into multiple simple sub-classification tasks and takes advantage of the simple task easier to train and optimize.In addition,to balance complexity and accuracy,extensive simulations are conducted to select the optimal network combination from the candidate cascade network combinations.2.Automatic modulation classification of overlapped sources using shuffling data augmentation based bidirectional gating residual networkIn the scenario where a single antenna receives multiple timefrequency overlapping signals,to solve the problems existing in traditional feature-based methods that require prior information,complex preprocessing,and deep learning-based methods are highly dependent on the training dataset,we propose a data augmentation method which is shuffling data augmentation to effectively expand the number of training samples.In addition,a novel bidirectional gated residual network(BGRN)is proposed by fusing the residual structure,the bidirectional gating unit,and the fully connected structure.The method automatically learns the discriminative feature representation of received signals which are processed by the bidirectional gating unit,and finally improves the classification and recognition ability of the BGRN classifier.Extensive simulations are conducted to verify the superiority of the proposed cascaded convolutional neural network and shuffling data augmentation-based bidirectional gating residual network.