Research On Modulation Recognition Based On Autoencoders And Convolutional Neural Networks

Automatic Modulation Recognition(AMR)is a complex and important technology before demodulation.It has been widely used in military fields such as electronic reconnaissance and electronic countermeasures,and civil fields such as radio spectrum management.In recent years,automatic modulation recognition based on machine learning and deep learning has attracted much attention.Compared with traditional algorithms,modulation recognition based on this kind of algorithm has higher recognition accuracy and stronger robustness.This paper studies the classification and recognition of modulated signals based on autoencoder and convolutional neural network,and does the following work:1.The modulation signal recognition algorithm based on feature extraction is studied,and the modulation signal model under complex channels is established.In addition,the statistical characteristics of the modulation signal are extracted to complete the pre-processing of the statistical characteristics of the signal.2.In the case of complex channels,the modulation signal is disturbed,which results in low recognition accuracy.A modulation recognition algorithm based on Anti-alias Linear Discriminant Analysis(A-ALDA)and Stacked Sparse Denoising Autoencoders(SSDAE)is proposed.In this algorithm,the A-ALDA algorithm reconstructs the signal cumulant features into new features,which have better separability;the original features and new features are input into the SSDAE for classification,and SSDAE has the ability to extract key information and anti-noise Stacked neural network.Experiments show that this algorithm has a higher recognition accuracy rate for modulated signals under complex channels,and it has a better recognition accuracy rate under different signal lengths and phase and frequency error interference.3.In order to solve the problem of single dimensional feature learning in traditional neural networks,a variety of convolutional neural networks with two-dimensional feature learning ability are designed for comparative study.Using the data set generated by the real speech signal and channel effect,the in-phase and quadrature components of the modulated signal are input into the network for training and testing,which avoids the tedious calculation steps of feature extraction.On this basis,the influence of parameters of convolutional neural network on recognition accuracy and training time is also studied.The experimental results show that the recognition accuracy of the algorithm based on different types of convolutional neural network is higher than that of the traditional algorithm based on feature extraction,KNN and linear SVM.4.In order to solve the problem of separating the local and global features of signals in traditional research,a modulation recognition algorithm based on features fusion and convolutional neural network is proposed.The statistical characteristics of modulated signals were input into DLDA(Deep Linear Discriminant Analysis)for optimization,and then the optimized statistical characteristics were combined with the same phase and orthogonal components of modulated signals through double channels.The combined channel is given a weight that can be changed automatically according to the training,so that the network can learn the importance of local features and global features,thus improving the recognition accuracy.