| With the rapid development of spatial information networks,the kinds and the number of terminals increase correspondingly,result in a shortage of available spectrum resources.In spatial information networks,spectrum sharing can alleviate the shortage of the available spectrum.However,the spectrum-sharing scenarios may be complex and dynamic,since the high-speed moving of satellites in orbit.Therefore,spectrum sharing based on real-time spectrum-sensing resultsapproaches as an effective way.However,there are some challenges: 1)How to transmit the mass sensing data to the gateway for further processing without consuming much transmission capacity of the link from the satellite to the gateway;2)How to obtain the future occupancy states in advance based on the historical spectrum data,in the presence of the long distance and big propagation delay of the considered links;3)How to make efficient use of spectrum resources based on the spectrum sensing situation.At present,the artificial intelligence technology is developing rapidly,and has been widely used.Different from the traditional methods,the artificial intelligence can realize rapid deployment based on a unified platform.Therefore,this article respectively studies spectrum reconstruction,spectrum prediction and spectrum resource utilization methods based on deep learning,to improve the efficiency and capability for processing spectrum data.The specific work and innovations of this article are introduced as follows:First,to address the problem of mass spectrum-sensing data and limited channel capacity of the links from spectrum-sensing satellites to the gateway,this article proposes a method called joint anomalous data repairing and deep convolutional neural network based spectrum reconstruction,Specifically,the satellites are allowed to downsample the spectrum-sensing data,and then transmit these data to the gateway.Then,the gateway preprocesses the incomplete data using the low-rank characteristics of the spectrum sensing data,and constructs a deep convolutional neural network to reconstruct the preprocessed spectrum data.Compared with the traditional compressed-sensing reconstruction method,the proposed method can achieve a lower reconstruction error.Second,to address the problem of obtaining the future spectrum situation based on the historical spectrum-sensing data,this article proposes a deep learning,based prediction model to generate the spectrum situation,termed as the convolutional neural network and bidirectional long short-term memory.Specifically,the historical spectrum data are first preprocessed.Then,a prediction model,combined convolutional neural network and bidirectional long and short-term memory neural network,is constructed to predict spectrum situation relying on the preprocessed historical data.Performance evaluations show that the proposed model outperforms the traditional neural network models in terms of both the accuracy and the mean absolute error of the spectrum prediction.Finally,to solve the problem of using spectrum resources based on the spectrum situation efficiently.this article studies a method to dynamicallyallocate spectrum and power resources using deep reinforcement learning.To maximize the wireless transmission throughput,the co-frequency interference imposed on satellite users and ground users,relying on efficient utilization spectrum and power resources.Simulation results show that comparing with the Q-learning method,the designed method can achieve a higher throughput,and a quiker convergence speed. |
