Modulation Recognition Algorithm Based On Lightweight Neural Networks

Modulation and demodulation is a key technology in the identification and demodulation of signals,it plays an initial role in the field of non-cooperative communication.Modulation identification is widely used in both military and civilian areas.How to realize accurate identification of modulation during the process of transition is an urgent and important problem to be addressed.In order to solve the problem of weak generalization ability and poor robustness,the thesis applies deep learning into the field of modulation identification,takes the light-weight neural network as the identification model,and thus improves the accuracy while decreases the quantity of calculation.The main work of this thesis is introduced as follows:1.A differential constellation diagram based on clustering algorithm is proposed as the data set of modulation images to realize modulation identification.The thesis applies clustering into the creating of constellation diagram,comparing with the recognition of common clustering density algorithm in constellation,reconstruct the center cluster and get the prepositive information of the center of the constellation.This constellation,which is constructed according to various colour degrees,performs better at frequency offset compared to the common ones,it judges and classifies QAM and PSK signals.And AlexNet is built up to complete the modulation identification under the circumstances of frequency deviation.2.Light-weight neural network is constructed for modulation of signals.CNN network and both MobileNet V3 and ReaNet 18 light-weight neural network models are also built while constellation,density constellation and differential constellation are used to complete the preprocessing of public datasets RML 2016.10 a in the modulation identification.Two neural networks are analysed in compared with each other based on the results and the most applicable one is chosen for the improvement and optimization.3.A modulation identification algorithm for light-weight neural networks based on SK-ResNext is proposed to complete the identification of public modulation datasets,and it is meanwhile compared with the newest modulation algorithm and analysed.Feed forward spontaneous attention mechanism is added based on the traditional ResNet18 network,and the traditional network convolution module is changed int parallel processing packet convolution in which image set and attention mechanism in paths are added while paths are endowed different power in modelling.According to the characteristics of modulation signals,the thesis adjusts the size of convolution kernels in the network models.The results show that improved light-weight neural networks have higher training speed and less parameters,overall retention rate in RML2016.10 a datasets has been improved 20% compared to traditional CNN network,and has been improved 10% compared to original ResNet 18 network.When SNR is 0b B,the identification rate is up to 90%,which is much better than common deep learning algorithm and this thesis proves the feasibility of light-weight neural networks used in modulation identification.