Research On Electromagnetic Signal Analysis Method Of Incomplete Samples Based On Deep Learing

Electromagnetic spectrum is the most valuable natural resource in the information age,it is also an important scarce strategic resource.In the military field,it is both a carrier of information transmission and an important means of reconnaissance.Take advantage of electromagnetic spectrum space will directly determine the initiative of war.With the increase of access equipment,electromagnetic spectrum resources become much scarce.Spectrum sensing can provide the cognitive wireless network with the information of the surrounding environment and the basis for decision-making,which is the basis and important method to improve the efficiency of electromagnetic spectrum utilization and manage electromagnetic spectrum resources efficiently.Due to the large frequency monitoring range,the spectrum sensing technology is limited by the ADC Effective Number of Bits in the hardware implementation process,so some scholars put forward to compressed sensing to solve the problem in communication signal process.The signal recover is the key to compressed sensing technology,it determines the effect of compressed sensing in practical application.Therefore,how to be precisely,efficiently,instantaneously recover signal is the key to spectrum analysis.Traditional signal reconstruction algorithms exist high complexity,long recover time and high recover error problems.In response to the above problems,the research content of this thesis is as follows.The aim of this thesis is to precisely,efficiently,instantaneously achieve spectrum analysis.The modulated wideband converter as a hardware implementation system for spectrum analysis to achieve under-sampling and precisely,efficiently,instantaneously recover frequency hopping signal in the monitoring frequency band,thereby achieving spectrum sensing task.The simulation experiments analyze the relationship between the selection of different parameters of the modulated wideband converter and the effect of multi-band signal recover,which has certain reference significance in the actual application.The undersampled signal reconstruction is completed through the greedy algorithm.According to the simulation result of original signal and reconstructed signal,the greedy algorithm has high reconstructed error problem when recover the signal.According to the shortcomings of the greedy algorithm,this thesis research a deep learning signal recover algorithm based on modulated wideband converter,including a signal reconstruction algorithm based on convolutional neural network and a signal reconstruction algorithm based on a variational autoencoder with detailed steps and flowcharts.The frequency hopping signals of common modulation types is generated by MATLAB software.According to the corresponding relationship between the under-sampled signal of the modulated wideband converter and the original signal in the frequency domain,a new data preprocessing method is proposed,which is beneficial to the signal reconstruction network to extract signal characteristics and improve the learning ability of the network.Finally,this thesis designs and optimizes the signal reconstruction network from three aspects: hyper-parameter tuning,network structure optimization,and optimization algorithm selection to design the optimal signal network structure.The over-fitting problem reasons and solutions in the process of network design and optimization is summarized.In order to verify the superiority of the signal reconstruction network in this thesis,a simulation experiment is set up to compare the method proposed in this thesis with the SAMP and OMP algorithms from reconstruction error,reconstruction time and the recover signal bit error ratio.The influence of modulation type and the number of sub-bands in multi-band model on the bit error rate of demodulation signal is analyzed.According to the simulation,the signal recover algorithm studied in this thesis shows better performance,it is precise,efficient,and instantaneous,which can efficiently complete spectrum analysis tasks.Under the same simulation conditions,compared with the SAMP and OMP algorithm,the threshold of CNN and VAE on the reconstruction decrease to 6% and 8%.When the number of reconstructed samples is 12000,the reconstruction time of CNN is about half of SAMP and OMP algorithm,the reconstruction time of VAE is about one-tenth of SAMP and OMP algorithm.When the number of samples is larger,the performance of CNN and VAE is better.