Research On Modulation Signal Recognition Based On Data Preprocessing And Convolutional Neural Network

Automatic modulation recognition technology is crucial and is widely used in military and civilian applications.In recent years,modulated signal recognition based on deep learning methods has attracted attention because of the powerful self-learning capability of deep learning,but there is still a problem of low recognition rate.In order to improve the recognition rate,preprocessing the signal from various aspects with convolutional neural network as the base network is proposed in the thesis,and the main research includes the following three aspects.1.The modulation recognition algorithm based on phase correction and convolutional neural network is proposed for the problem that the channel affects the I/Q phase shift to make the recognition accuracy not high.The received signal is first learned by using a deep neural network with linear activation function for phase offset parameter,then the phase transformer is performed to transform the I/Q signals to reach 90° phase difference again,and the transformed signals are fed into the CNN for training and classification.The results show that the recognition rate of the algorithm with phase correction is improved by about 5% over that of the uncorrected algorithm,with the highest recognition rate of 89%.2.To address the problem of limited signal timing features extracted due to the limited perceptual field of CNN,a truncated migration data preprocessing algorithm is proposed,which enables CNN to extract more sampling points and compare more symbolic information by truncating the distance units at one end of the sampling matrix and then migrating them to the other end and merging them into a new matrix in turn.In addition,an improved parallel residual neural network is proposed to focus on features in both horizontal and vertical directions simultaneously through two parallel branches.The results show that the truncated migration preprocessing of the signal improves the accuracy by about 10% over the input original signal,and the improved parallel residual network further improves the recognition rate with the highest accuracy of 93.78% when the signal-to-noise ratio is 14 d B.3.A multi-task learning algorithm is investigated for the three groups of signals that are always confused,QPSK and 8PSK,QAM16 and QAM64,AM-DSB and WBFM.Three auxiliary tasks are used to train the confusion categories separately and then pass the learned features to the main task for training.A gated recurrent unit network is used for the auxiliary task model,while the main task model uses a one-dimensional multiscale convolutional network cascaded with a gated recurrent unit network in order to extract both signal space and timing features.To better fit the input of the network,a one-dimensional timing format is used for the input signal.The results show that the multi-task learning approach improves the recognition rate of confused signals,with 99% for QPSK and 8PSK,96% for QAM16 and QAM64,and 80% for AM-DSB and WBFM when the signal-to-noise ratio is 18 d B,improving recognition accuracy up to 95%.In addition,the algorithm complexity has low complexity,which is conducive to practical application deployment as well as migration learning.