Research Of Signal Modulation Recognition Based On Deep Learning

As one of the most important techniques in the field of signal recognition,signal modulation recognition provides essential information for radiation source work pattern recognition and individual recognition,therefore plays an important role in applications such as wireless spectrum monitoring and electromagnetic countermeasures.Compared with traditional modulation recognition methods,modulation recognition based on deep learning facilitates automatic feature learning of electromagnetic signals.It also effectively improves recognition performance under the conditions of complex modulation modes and low signal-to-noise ratio.Therefore,deep learning based modulation recognition has attracted both academic and industrial attentions.On the one hand,researches of deep learning based modulation recognition mainly focused on the optimization of neural network models and improvement of recognition accuracy;On the other hand,in practical applications such as electromagnetic environment monitoring and information collection,both the lightweight design and network design are vital,which demand modulation recognition algorithms that can be deployed on low-cost electromagnetic sensors and data fusion methods that make full use of network gains.This research is motivated by two practical requirements in electromagnetic sensing networks,namely,lightweight modulation recognition algorithm design and data fusion method design,and the major contributions are as follows.Firstly,the signal modulation methods in the literature are analyzed in depth.The methods may be categorized into decision theory based,handcrafted feature design and automatic feature learning.Their research paradigm,recognition performance and practical application considerations are compared.Several recent research hotspots are also identified,namely,the mixing input of handcrafted and automatic learned features,electromagnetic signal feature guided neural network model,and signal modulation recognition design under constrained computing resources.Secondly,a novel lightweight modulation recognition algorithm is proposed.As the basis for algorithm training,a data set including 10 communication modulation methods is constructed based on software radio platforms,on which the signal transmission,reception and storage modules are implemented.Afterwards,the computational complexity of the convolutional neural network is analyzed,and a Le Net4 based lightweight recognition algorithm is designed and deployed on edge computing chip.Under the premise that the recognition accuracy is comparable to that of CNN2 and CLDNN,the proposed lightweight algorithm significantly reduces the computational cost.Results implemented on actual data sets show that the recognition accuracy rate exceeds95% when the signal-to-noise ratio is 0d B.Thirdly,data fusion methods for lightweight electromagnetic sensor networks are designed in order to deal with severe recognition accuracy degradation under low signalto-noise ratio conditions.Firstly,a signal data set is constructed that considers frequency offset,phase offset,and multipath effect.Then,the influence of channel distortion on the modulation recognition performance is analyzed.Afterwards,data fusion methods based on classifier confidence and Bayesian theory is designed,which utilizes the mapping from average recognition accuracy and confusion matrix to signal-to-noise ratio.Finally,simulation experiments under different network topology conditions are implemented,which verifies that the data fusion method may improve the accuracy of signal recognition.The results may provide a quantitative reference for the design and deployment of electromagnetic sensor networks in actual environments.