Research On Automatics Modulation Recognition Technology Of Communication Signal Based On Deep Learning

With the rapid development of communication technology,the electromagnetic environment has become increasingly complex.In military and civilian fields such as radio resource management,spectrum monitoring,radio fault detection,UAV(Unmanned Aerial Vehicle)monitoring,communication detection,electronic military countermeasures,and electromagnetic situational awareness,higher requirements are placed on the technology of signal detection,identification and classification.At present,the traditional modulation identification algorithm requires a large amount of prior information,and the calculation amount is large and complex,and the identification accuracy is not high under various types of signals and complex channel environments.In order to further to improve the performance of modulation recognition,this thesis applies deep learning technology to the field of modulation recognition.In order to realize the identification and classification of various types of high-order modulation signals in complex electromagnetic environments,this thesis first designs an IRBL(Inception-Residual Bi-directional Long-Short-Term Memory Neural Network)modulation identification network,which uses the in-phase quadrature of the signal.IQ(In-phase Quadrature)data onto modulation identification.The IRBL network to consist of an InceptionResidual unit and a bidirectional LSTM(Bi LSTM)network.The network is broadened and deepened by multiple Inception-Residual units,and the Bi LSTM structure is used to improve the extraction of timing features of the modulated signal.Compared with other modulation recognition networks of comparable computational complexity,the IRBL network to improve the recognition accuracy by 1.2% on the Radio ML2018.01 dataset.In the IQ data onto the modulated signal,the individual I and Q data are continuously changing in time and have unique characteristics.In order to extract this feature,a multichannel module is designed in this thesis to extract the data information on I and Q channels separately.In addition,there is a lot of noise in the modulated signal data.In order to reduce noise interference and strengthen the weight of important features,this thesis introduces the CBAM(Convolutional Block Attention Module)attention module.Combining the multichannel module and the CBAM attention module in the IRBL network,a CBAM-MIRBL(CBAM-Multi-IRBL)network is designed,and the recognition accuracy on the Radio ML2018.01 dataset is further improved compared to the IRBL network.Most of the existing modulation recognition networks are improved on the basis of image classification networks,and their parameters and floating-point calculations are large,and the feature extraction efficiency of IQ signal data is low,which is difficult to apply in practice.Therefore,this thesis designs a lightweight CBAM-Res LW(CBAM-Residual Lightweight Neural Network)network.The network reduces the computational complexity of the network by using dimensionality-reducing convolutional units and a small number of filters.At the same time,by connecting the feature outputs of the convolutional layer across layers,the features of different depths and different scales of the signal data are extracted,and the diversified expression of higher semantic level information is realized,thereby ensuring the recognition accuracy of the network.The experimental results for the Radio ML2018.01 dataset show that the recognition accuracy of the CBAM-Res LW network is comparable to that of the IBRL network,but the amount of parameters and floating-point calculations of the network is only about 1% of the IRBL network,which has higher practical application values.In addition,the network designed in this thesis performs well on both Radio ML2016.10 a and Radio ML2016.10 b datasets.