Research On Modulation Recognition Technology Based On Deep Neural Networks

Modulation recognition technology is a technique used to identify the modulation type of a target signal in non-cooperative communication scenarios.It is widely used in civilian radio monitoring and military electronic warfare,and has important research significance in the field of communication.Traditional modulation recognition methods rely on decision theory,signal features,and classifier selection,and the recognition accuracy is also limited by complex electromagnetic environments.Modulation recognition technology based on deep learning can automatically extract signal features to determine the modulation type of a signal,has become an important research direction at present.but it still faces challenges such as poor real-time performance,difficulty in distinguishing similar signals,and low accuracy under low signal-tonoise ratio conditions.To address these issues,this thesis conducts in-depth research on deep learning-based communication signal modulation recognition technology to effectively improve modulation recognition performance,mainly including the following aspects of work:(1)To address the problem of real-time requirements and small dataset batches in practical applications of modulation recognition,an improved algorithm based on the CNN-GRU network model is proposed.Firstly,data augmentation operations are performed on the dataset in the data preprocessing stage to increase its diversity.Then,the hyperparameter combinations of convolutional neural networks(CNN)and gated recurrent units(GRU)are optimized using grid search.Finally,the signal time-domain and frequency-domain information enhanced by data augmentation are input into the CNN-GRU network to enhance the network’s feature generalization ability.Experimental results show that compared with existing classical networks,the proposed network model has lower parameters and achieves an accuracy of 61.5% on the dataset.(2)To address the problem of difficulty in identifying partially similar signals,a modulation signal recognition method based on CA-VGG is proposed.Firstly,the original I/Q data is converted into a constellation diagram as a two-dimensional image feature input of the signal.Secondly,the traditional VGG-16 network model is improved to speed up the network training process.Then,the coordinate attention mechanism(CA)is added to enhance the important features of the signal and suppress irrelevant features,thereby improving the data generalization ability.Finally,Experimental results show that the CA-VGG network improves the accuracy of similar QAM16 and QAM64 signals by 6.4% compared to the CNN-GRU network proposed in this thesis.(3)To address the problem of low accuracy of deep learning modulation recognition models under low signal-to-noise ratio,a multi-feature fusion(MFF)network model is proposed.Firstly,the time-frequency features and constellation diagram features of the modulation signal are used as inputs to enrich the data representation of each modulation method and realize the complementarity between different types of data features.Then,the CNN-GRU network and CA-VGG network proposed in this thesis are fused,where the CNN-GRU module based on convolutional neural networks and gated recurrent unit networks is used to extract the timefrequency features of different modulation signals,and the CA-VGG module based on convolutional neural networks and coordinate attention mechanism is used to extract the constellation diagram features of signals.Finally,the output results of the two modules are effectively combined through a decision layer to improve the accuracy of signal recognition,and the weight allocation is optimized through distance information in the feature layer.Experimental results show that the MFF network improves the recognition accuracy to 63.9% by fusing the two models.