| With the increase of wireless communication types and the rapid development of wireless technology,spectrum resources can no longer meet the needs of the times.Spectrum sensing(SS),as a key technology in cognitive radio(CR)to solve the shortage of spectrum resources,is of great significance for the full utilization of spectrum resources.In recent years,with the rapid development of artificial intelligence,technologies such as machine learning have begun to be used to solve spectrum sensing problems.Due to the characteristics of multiple samples and high dimensions in large data environments,support vector machines(SVM)have long training times,insufficient feature parameter extraction,and poor detection results;At the same time,due to the large bandwidth and other characteristics of 5G and other related technologies,the sampled signal has sparse time domain,low recognition rate under low signal-to-noise ratio,complex compression sensing reconstruction algorithms,and large computational complexity.In this thesis,convolutional neural networks(CNN),long short term memory(LSTM),and non-reconstructed compression methods are used to carry out research on spectrum sensing technology.The content and contributions of the thesis are conclude as follows:First,aiming at the problems of insufficient feature parameter extraction and low recognition rate under low signal-to-noise ratio,a CNN collaborative SS scheme based on covariance matrix decomposition was proposed through covariance matrix decomposition and CNN perceptual classification.Firstly,the covariance matrix of the sub user acceptance signal sampling matrix is decomposed by Cholesky to obtain a lower triangular matrix.Based on its element characteristics under different conditions,statistics are constructed to fully extract the differences between PU and noise signals,thereby enhancing the impact of PU signals.Secondly,the statistics obtained by multiple secondary users are fused to form a statistical feature matrix as a single training sample to improve collaboration between secondary users and reduce the perception time of a single SU.Finally,by using the feature extraction of the CNN for high-dimensional matrices,the spectrum sensing results are obtained through CNN training and testing.Experimental results show that at the false alarm probability of 0.1 and the signal-to-noise ratio of-15 d B,the detection probability of the proposed scheme is 60% and 69% higher than that of the traditional CNN and SVM classification methods,respectively.Second,in the classification of high-dimensional samples,with long operation time,low training accuracy,and poor classification performance of SVM classifiers,a CNN sensing scheme under Adam is adopted through two-dimensional convolutional neural networks and an Adam gradient optimizer.Firstly,the Adam essentially combines the advantages of Momentum and RMSprop algorithm.After correcting its bias,the learning rate of each iteration has a fixed range,making the parameter fluctuation smaller.Secondly,when updating parameters,compared to the RMSprop and the SGD algorithm,it dynamically adjusts the adaptive learning rate of different parameters,and synchronously updates parameters with accelerated descent and resistance attributes,resulting in high computational efficiency.Finally,the algorithm is selected to optimize the gradient of the training set,fully improving the training efficiency of the model.The simulation shows that it has the fastest speed in gradient descent.When the number of iterations reaches 60,the accuracy value is stable at 100%,and the loss value is basically stable at about 0.Moreover,its training accuracy is 5% and 10% higher than that of the RMSprop and the SGD,respectively.Third,aiming at the sparse time domain of sampled signals and the complexity of traditional compressed sensing signal reconstruction algorithms,an LSTM collaborative SS scheme based on non-reconstructed compression was proposed through non-reconstructed compression and LSTM prediction.First of all,under the condition of ensuring that the amount of original signal information is not lost,the sampled signal is multiplied by the measurement matrix to obtain a low dimensional compressed signal to reduce the dimension of the PU signal and reduce the overall computational complexity.Secondly,the compressed signals of multiple secondary users are fused as a matrix,and the main diagonal elements of the compressed covariance matrix are taken as time series statistics based on the characteristics of each element of the covariance matrix before and after compression,in order to reduce the dimension of single sample and improve the correlation between the elements.Finally,using the prediction characteristics of LSTM for timing signals,the input vectors of different time steps are correlated with each other,and ultimately the spectrum sensing results are predicted.The simulation results show that the overall detection accuracy of LSTM is improved by 10% compared to CNN at-20 d B to-14 d B and under the same sample conditions;When the false alarm probability is 0.1 and the signal-to-noise ratio is-20 d B,the detection probability of the proposed scheme is 22% higher than that of traditional SVM methods.This thesis mainly studies the application of deep learning SS.The proposed CNN cooperative SS scheme with covariance decomposition has short operation time,high training accuracy,and high detection accuracy at low SNR.The non-reconfigurable compression LSTM cooperative SS scheme fully extracts the characteristics of PU signals and improves the overall perception accuracy.Therefore,the proposed scheme is suitable for 5G multi noise interference signal spectrum detection applications. |
