Research On Modulation Recognition Method Of Communication Signal Based On Lightweight Network

With the development of intelligent communication,modulation recognition of communication signals has far-reaching significance in many application scenarios such as electronic reconnaissance,electronic countermeasures,software radio and cognitive radio.Current modulation recognition methods based on deep learning can effectively improve the recognition accuracy under low SNR by relying on the powerful and complex fitting ability of neural networks,but the network structure is too complex.Aiming at the problems of complex network,high computation and high demand for hardware platform in the process of recognizing signals with different modulation modes using deep learning,this paper presents a modulation recognition method using signal constellation diagram in an improved MobileNetV3 lightweight neural network.The main work completed in this paper includes:1.The common principles of single carrier modulation and OFDM are introduced in detail,and three methods of modulation recognition are mainly introduced,which are based on decision theory,statistical pattern and deep learning.This paper presents the application of lightweight network MobileNetV3 in modulation recognition,introduces the composition and structure of MobileNetV3 network,compares deep separable convolution with standard convolution,and introduces the cross-layer structure of ResNet into MobileNetV3 network.2.Based on the open source data set,five modulation signals of 2PSK,QPSK,8PSK,16QAM and 64QAM in the RML2016.10b modulation recognition data set disclosed by Timothy J O’Shea were extracted,and the baseband complex signals were converted into signal constellation maps and further image processing was carried out.Including gray level image extraction,gray level image enhancement and gray level image convolution,the generated image data set is input into the lightweight network MobileNetV3,the parameters are optimized,and the optimal network parameters are selected for training.Comparing the performance of the three image processing methods,it is concluded that gray-level image convolution can improve the recognition rate of high-order modulation signals more effectively than other methods.Graylevel image convolution is used to input the data set into the improved MobileNetV3 network,and the effectiveness of the proposed method is verified by comparing with several classical modulation recognition networks and the overall recognition rate of MobileNetV3 network.3.Six signals,2PSK,QPSK,8PSK,16QAM,64QAM and OFDM,were generated through MATLAB 2018a simulation,and image data sets were generated using gray level image convolution processing.The generated data sets were input into the trained lightweight network MobileNetV3 for classification and recognition.Simulation results show that when SNR is 10dB,the overall recognition rate of various types of modulated signals is close to 99.7%,among which OFDM signal has the best classification effect and the strongest anti-interference ability to noise.When SNR is 0dB and-3dB,the recognition accuracy of OFDM signal is 99.7%and 98.3%,respectively,and the number of network parameters and calculation amount are significantly reduced compared with the network that uses deep learning to realize modulation recognition.