Research On Key Technologies For Satellites Spectrum Cognition Via Deep Learning

With the development of spatial information network,more and more frequency-using devices and communication services are emerging,but the available spectrum resource is increasingly limited,which puts forward a higher requirement for efficient utilization of spectrum resource.At present,as the core of cognitive radio technology,spectrum sensing is an important basic of dynamic spectrum resource allocation.As a common method of satellite spectrum sensing,the energy detection is affected by the preset noise threshold and the noise uncertainty leads to its poor performance in low SNR environment.Deep learning is data-driven and independent of prior information.Therefore,to improve the performance of satellite spectrum sensing,this thesis introduces deep learning into satellite spectrum sensing.The specific research contents and innovations are as follows:Firstly,to overcome the problem that the performance of single-satellite spectrum sensing using energy detection is affected by SNR-wall,this thesis designs a satellite spectrum sensing method based on deep learning,which uses temporal convolutional network to reduce the impact of noise uncertainty.Get the corresponding data set by processing received signal and noise,and obtain the well-trained model of temporal convolutional network through extracting the spatial features of samples.The well-trained model will be applied in on-line detection to further improve the satellite spectrum sensing performance under circumstance of low SNR.Secondly,in the overlapping coverage area,the spectrum sensing performance of different satellites varies greatly and the computing capability of satellite is limited,a low complexity satellite cooperative spectrum sensing method based on deep learning is designed.First,a spectrum sensing model which integrates convolutional neural network and long short-term memory network is proposed.Through optimization training,it can obtain higher spectrum sensing performance with fewer model parameters.Then,the well-trained model will be deployed on a satellite in the satellite collection of overlapping coverage area which works as the fusion center and receives the forwarded signal from other satellites through inter-satellite chains to complete cooperative spectrum sensing.This method can improve the performance of satellite spectrum sensing and reduce the deployment complexity of the model without requiring that all satellites in the collection have enough computing capability.Finally,considering that it is difficult to support the security control and interference identification of electromagnetic spectrum space only by judging spectrum occupancy state,this thesis designs a satellite signal interference detection method based on object identification,which can detect interferences like noise fluctuation,single-tone interference and wideband interference.In addition,a spectrum cognition demonstration platform is constructed and verified by real-time acquisition of spectrum data from the downlink frequency band of Tiantong-1 satellite which ranges from 2170 MHz to 2200 MHz.The designed functional modules can work well on the platform.