Research On Key Techniques Of Spectrum Situation Generation Based On Deep Learning

Recently,various applications relying on wireless transmission are emerging,result in that the demand on spectrum resources is increasing.Meanwhile,most of the allocated spectrum may be not fully utilized due to the traditional static spectrum allocation strategy.Therefore,there is a contradiction between the shortage of spectrum resources and the waste of spectrum resources.In order to solve this contradiction,with the development of deep learning,the related technology of neural network is more and more mature,which has the advantages of nonlinear,adaptive,high error tolerance,high-speed parallel,and self-learning,etc.Therefore,spectrum inference is designed accordingly.It allows the secondary user to access the spectrum authorized by the primary user,whilst the wireless transmissions of the primary users are not hampered,leading to effectively improve the spectrum efficency,and alleviate the contradiction between resource shortage and spectrum resource waste.In this thesis,a spectrum prediction strategy relying on deep learning is designed,which can predict the future spectrum situation in advance by analyzing the historical spectrum data.The designed spectrum prediction strategy includes two modules,which are adaptive threshold spectrum quantization and offline online joint spectrum prediction.Specifically,considering the effect of noise randomness in spectrum in actual situations,a spectrum quantization scheme based on adaptive threshold is proposed to convert the power spectral density to spectrum occupancy.In this scheme,the power spectral density of different channels is analyzed independently,and the average noise power of each channel is estimated by kernel density estimation method.The decision threshold is set adaptively to the noise fluctuation to improve the accuracy of quantization.Considering that the spectrum is essentially a time series,a spectrum occupancy prediction model based on long short-term memory networks is constructed to realize real-time prediction of future spectrum occupancy.The long short-term memory network,which is a recurrent neural network,is chosen as the model.The design of activation function,loss function,optimization algorithm and other details are analyzed,and the hyperparameters of the model are evaluated and optimized through the combination of grid search and cross-validation.Additonally,considering that the spectrum situation may change with time,the neural network model itself cannot automatically capture this change,thus an offline-online joint spectrum prediction scheme is also proposed.In this scheme,the average accuracy of the model is monitored in real time.When the accuracy is lower than the predefined minimum accuracy threshold,the latest spectrum data is actively collected and the model is retrained and evaluated to achieve the recognition of the changes in spectrum situation and the adaptation to them.Experimental results show that the proposed adaptive threshold scheme can effectively eliminate the influence of noise randomness and correctly convert the power spectral density value into the spectrum occupancy,while compared with the traditional fixed threshold scheme.The correct quantization results can provide reliable data for the design of the neural network model and avoid the loss of model performance due to wrong quantization results.With the preprocessed spectrum data,the constructed spectrum occupancy prediction model is trained,and the hyperparameters are designed.The experimental results show that the model can effectively predict the future spectrum occupancy,and the performance is always better than the baseline model which is based on the traditional fixed threshold scheme.Finally,in order to verify the proposed offline-online joint spectrum prediction scheme,a new dataset is artificially constructed based on the original dataset.The experimental results under this dataset show that the scheme can identify the changes in spectrum situation in time,and automatically adapt to the changes to ensure that the average accuracy will not gradually deteriorate due to the changes in spectrum situation.