A Research And FPGA Implementation Of Broadband Spectrum Prediction Based On Deep Learning

With the sustained development of wireless communication technology,the fixed spectrum allocation strategy can not meet the increasing frequency demand,the phenomenon of imbalance between supply and demand of spectrum resources appears.In order to solve the problem,some scholars come up with the idea of cognitive radio,which can flexibly allocate spectrum resources,and spectrum prediction technology can make it play a better role.With the progress of deep learning,the technology of neural network gets more and more mature,and deep learning technology has been deeply applied to the field of spectrum prediction.Based on deep learning,channel occupancy state prediction in spectrum prediction is studied in this thesis,including the following aspects:Firstly,how to convert the received time domain data into channel occupancy state is studied in this thesis.At first,the broadband signals are divided into sub-channels through the digital channelization structure based on polyphase filter banks to obtaine the timedomain signals of each channel.In the second place,the preliminary occupancy states of each channel are obtained through the simple and efficient adaptive threshold detection algorithm based on time-domain autocorrelation.Then,a false signal decision method based on short-time Fourier transform is designed to eliminate the influence of digital channelized prototype filter transition band.Finally,the modified channel occupancy states are acquired through morphological processing.Then,how to use the historical channel states to predict the channel states in a certain time in the future is studied in this thesis.Since the channel states are time-dependent,three neural networks including Recurrent Neural Network(RNN)which has a better processing ability for time series problems,Long Short-Term Memory(LSTM)which is improved for the gradient disappearance problem of RNN and the sequence to sequence(Seq2Seq)network structure based on LSTM are respectively applied to this problem.According to the characteristics of cognitive radio,the evaluation indexes of the performance of the spectrum prediction model are defined,and the experimental simulation and comparative analysis of the designed prediction algorithm are carried out.The final experimental results show that the prediction model based on LSTMSeq2Seq structure can remember longer input and has a better prediction performance.Finally,the prediction model based on RNN is implemented in hardware in this thesis.At first,the realization schemes of matrix multiplication based on multiplier adder and multiply accumulator are designed.Then,the architecture of hidden layer and output layer of RNN model are designed respectively.The simulation results are consistent with the fixed-point results of Matlab,which shows the correctness of the hardware implementation.Finally,the prediction model achieved by FPGA is compared with the performance of CPU and GPU in speed and power consumption.In terms of speed,FPGA reaches96.0875 GFLOPS,which is about 3 times of CPU.In terms of power consumption,the total power consumption of FPGA is 3.923 W,which is less than one percent of GPU.FPGA implementation is more suitable for embedded scenario with low power consumption and high real-time performance.