| Blind recognition of modulation signal is the basis of many fields,such as signal demodulation and radio signal recognition.Current modulation recognition algorithms can be roughly divided into three categories.The first is the blind recognition of modulation based on feature extraction,which is the most mature recognition technology at present,but the complexity will increase with the increase of modulation signal types;the second is the blind recognition of modulation based on likelihood estimation and probability theory;the third is the blind recognition of modulation based on depth learning.Because deep learning has a better ability of automatic feature extraction,this thesis uses deep learning as the basic recognition algorithm,the main work is as follows.First of all,the research of the current deep learning modulation blind recognition algorithm focuses on the recognition of all modulation signals into the network,without judging whether the signals with poor recognition effect are suitable for the network.In this thesis,the feature extraction and convolution neural network are combined to separate the modulation mode,and the small network is trained for different modulation modes.On the one hand,the recognition effect of convolutional neural network is improved to some extent.On the other hand,it solves the problem of long training time of many signals recognized by convolutional neural network alone.Secondly,the current research of deep learning modulation recognition algorithm focuses on the amount of input data and the number of layers of convolution neural network,without analyzing whether different data formats and data types will affect the recognition rate of the network.In this thesis,the input data format of convolutional neural network is transformed to enhance the feature extraction of IQ signal,so as to improve the recognition rate of the algorithm.On the other hand,this thesis transforms the data type of input network,and uses different input types to identify different modulation signals.This thesis improves the recognition effect for noisy signals without frequency offset and phase offset.Finally,when selecting parameters for the network,this thesis applies the Bayesian optimization parameter algorithm to the modulation recognition algorithm to realize the selection and optimization of super parameters.Through the effective selection of parameters for the deep convolution neural network,the selection time of network parameters is reduced and the rationality of the network is enhanced.Then the SGDM optimizer,RMSProp optimizer and Adam optimizer are used to optimize the network respectively,and finally the optimizer suitable for blind recognition algorithm is found out;finally,the signal source and receiver are used to build a platform for the actual signal to realize the modulation recognition of the actual collected signal. |
