Research On Modulation Recognition Based On Deep Learning

Traditional manual modulation recognition methods have significant drawbacks,making it difficult to update the iterative changes of the signal in real-time.The modulation recognition method based on likelihood function needs prior information,requires high precision of likelihood function model and high computational complexity.Feature based methods have lower computational complexity and higher practicality,but if the extracted features do not match the classifier,the classification effect will be poor.Therefore,the thesis aims to find an automatic modulation recognition method with low computational complexity and strong robustness,and studies modulation recognition algorithms based on deep learning.Firstly,the thesis designs one-dimensional instantaneous features(phase features,amplitude features,and frequency features),twodimensional image features(constellation),and corresponding normalization algorithms used in modulation recognition preprocessing.Then,based on existing preprocessing methods,a data augmentation algorithm is designed to enhance the effectiveness of feature extraction.Secondly,the thesis designs a modulation recognition method based on a one-dimensional feature three input network.The algorithm inputs amplitude features,frequency features,and phase features into a three input neural network,and extracts the time characteristics of parameters by combining the one-dimensional CNN network with the LSTM network.The experimental results show that the modulation recognition method based on one-dimensional feature three input network proposed in the thesis has relatively low complexity,and achieves a classification accuracy of up to 94.6%on the RadioML2016.10a dataset,and up to 94.2%on the RadioML2016.10b dataset.Compared to most current deep learning methods,it has better classification performance.Thirdly,this thesis proposes a modulation recognition method based on multi feature fusion and multi input networks.The algorithm inputs onedimensional instantaneous features and two-dimensional image features after data augmentation into a neural network,and uses a multi input neural network structure for modulation classification.The neural network structure can be divided into two modules.The one-dimensional module takes amplitude features,frequency features,and phase features as inputs,while the two-dimensional module takes constellation maps as inputs.The network can extract the temporal and spatial features of signals from onedimensional and two-dimensional perspectives.The experimental results show that the complexity of the modulation recognition method based on multi feature fusion and multi input network proposed in the thesis is at a moderate level.However,due to the use of preprocessing methods,the training time is significantly reduced.At the same time,it achieves a classification accuracy of up to 99.1%on the RadioML2016.10a dataset,and a classification accuracy of up to 99.5%on the RadioML2016.10b dataset,which is superior to existing deep learning methods.