Research On Spectrum Sensing Based On Deep Unsupervised Clustering

With the expansion of the application field of radio technology,the division of the ra-dio spectrum has become saturated,and it is difficult to find free frequency bands to deploy new services or improve existing services.In order to solve the problem of the shortage of spectrum resources,a cognitive radio method was introduced to allow perceptive users to access the allocated spectrum opportunistically.This method allows the spectrum to be fully utilized in time and space.By this way,we can improve the problem of the shortage of spectrum resources effectively.In cognitive radio,perceiving users need to be able to reliably obtain information about the unoccupied spectrum in time and space so that authorized users can be protected from interference by perceiving users.Therefore,how to improve the accuracy of spectrum sensing has become one of the research hotspots of cognitive radio.In this paper,we study the spectrum sensing algorithm based on deep unsupervised clustering.Traditional spectrum sensing algorithms need to rely on the hypothetical of signal noise model.Therefore,the accuracy of the noise probability function estimation and the correctness of the signal model will affect the result of the algorithm.Spectrum sensing algorithms based on machine learning are driven by data so they are not affected by this problem,however,most of the spectrum sensing algorithms based on machine learning are supervised learning and it needs to be trained through a large amount of la-beled data.However,in spectrum sensing scenarios,it is difficult to obtain this kind of data,which limits the application of supervised learning in practical problems.In this paper,a deep unsupervised clustering spectrum sensing algorithm is proposed based on the graph-embedding deep clustering model.First,the spectrum sensing problem is trans-formed into a clustering problem,and then the deep clustering network is trained to cluster the sensing signals,and finally Real-time spectrum sensing can be implemented through the trained network model.The research contribution and innovation of this paper are mainly to propose a deep unsupervised clustering spectrum sensing algorithm,it is based on unsupervised clustering without any hypothetical model of signal and noise.The algorithm does not need labeled data when it is training,and only needs a small amount of labeled data when determining the classification cluster category and designing the detector.Compared with other spec-trum sensing algorithms based on supervised learning,it reduces the need for labeled data.And this algorithm combines the advantages of graph embedding and Gaussian mixture auto-encoder,which can introduce the similarity information between perceptual signals into the variational Gaussian mixture auto-encoder,and clustering while learning features.Compared with the existing The unsupervised deep spectrum sensing algorithm has better spectrum sensing capabilities.At last,we completes theoretical simulation based on the proposed algorithm,ana-lyzes the impact of various hyperparameters in deep learning on the performance of deep spectrum sensing algorithms,and compares it with several existing spectrum sensing al-gorithms.The experimental results show that the accuracy of spectrum sensing is signifi-cantly improved by our algorithm.