Research On Spectrum Sensing Algorithm Based On Adversarial Autoencoder

Spectrum sensing means that the cognitive user can judge the current spectrum usage by detecting the signal of the target frequency band,which is the most basic technology in cognitive radio.Its purpose is to quickly find the spectrum holes in the complex wireless communication system,so that unauthorized users can access the idle spectrum and realize spectrum sharing.At present,most of the proposed spectrum sensing methods are based on hypothesis testing theory and use the characteristics of the model as test statistics,such as energy and eigenvalue,etc.However,these features cannot replace all the signal information.Based on this,many scholars introduce deep learning into spectrum perception.The core idea is to extract features through a series of nonlinear transformations at multiple levels in a data-driven way,so that features can be more expressive.At present,the proposed spectrum sensing algorithms based on deep learning are mostly supervised learning,which needs a lot of labeled samples for training,which is very difficult in practical situations.To solve this problem,two semi-supervised spectrum sensing algorithms are proposed.Firstly,a spectrum sensing algorithm based on Semi-Supervised Adversarial Autoencoder(SSAAE)is proposed.The model uses the idea of confrontation to restrict the display of the latent vectors,so that it conforms to the pre-set prior distribution,carries more information,and better performs classification tasks.This algorithm uses the original sampled signals as the input of the SSAAE model and does not need any prior knowledge.It is a blind detection algorithm.In the training process,it uses the generation description of unlabeled data to improve the classification performance obtained by using only labeled data.It reduces the cost of collecting training sets and improves the generalization ability of the model.When the signal-to-noise ratio(SNR)is low,the numerical experiments of OFDM signal under Gaussian white noise demonstrate that the detection performance of the SSAAE algorithm is significantly higher than other three traditional spectrum sensing methods.Aiming at the problem of slow convergence of SSAAE algorithm and poor feature learning effect,a semi-supervised spectrum sensing algorithm based on Convolutional Adversarial Autoencoder(CAAE)is proposed.Convolutional Neural Network(CNN)uses the concepts of local correlation and weight sharing,so changing the multi-layer perceptron in the anti-autoencoder to CNN.In addition to reducing the number of network parameters,the learning ability of the model is enhanced and the over-fitting problem is effectively reduced.This algorithm combines CAAE and logistic regression,using the real part,imaginary part,and mode information of the sampled signal to construct a three-channel training set for unsupervised training of CAAE and establishing an optimized feature extraction model.Then,a better classifier is obtained by using the extracted features to conduct logistic regression supervised training.For test the detection performance of the CAAE-based algorithm,the same training set is used to conduct experiments.The results demonstrate that the CAAE-based algorithm has an appreciable performance increase compared with the SSAAE.