Research On Machine Learning Based Spectrum Sensing And Dynamics Spectrum Access Algorithms

As two key technologies of cognitive radio,spectrum sensing and dynamic spectrum access technology have important research significance to alleviate the contradiction between static spectrum allocation and dynamic spectrum demand.In order to improve the performance of existing spectrum sensing and dynamic spectrum access algorithms,this thesis combines the research methods of machine learning to study the spectrum sensing and dynamic spectrum access technology.Most of the traditional spectrum sensing schemes focus on the spectrum hole in the time dimension,but often ignore the spectrum hole in the space dimension.Therefore,in order to further improve the spectrum utilization,this thesis proposes a spectrum sensing scheme based on wireless fingerprint database.Firstly,in the target geographical area covered by cellular cognitive radio network,the secondary user equipments(SUEs)collect a large number of spectrum observation data,and process the spectrum observation data based on various machine learning algorithms,so as to obtain the joint transmission mode of primary user transmitters(PUTs)on the authorized spectrum.Secondly,in the different joint transmission modes of PUTs,the geographic location area is divided,and the grid label is obtained by using the method of spatial distance calculation,and the wireless fingerprint database is established.Finally,the SUEs with sensing needs obtains its wireless fingerprint according to the estimated time of arrival of the received base station reference signal,and determines its geographical location by matching with the wireless fingerprint in the wireless fingerprint database,and determines the access label of the authorized frequency band accordingly.Simulation results show that compared with the traditional energy detection scheme,the proposed scheme not only improves the spectrum sensing performance,but also increases the dynamic access opportunities of cellular cognitive network for licensed frequency band.In the traditional dynamic spectrum access problem,spectrum access completely depends on the result of sensing.In practice,due to the complexity and randomness of the wireless environment,the limited cooperation between SUEs and other practical factors,it may lead to false positives or missed detection of primary users(PUs)activities,and then lead to the wrong decisions of SUEs access to the channel.Therefore,this thesis proposes a dynamic spectrum access scheme based on the combination of upper confidence bound with Hoeffding algorithm and DQN algorithm.Firstly,the channel model is modeled as a discrete Markov chain,in which the state transition probability of the channel is unknown for SUEs.Secondly,the historical experience composed of access action and action feedback is established for each SUEs.SUEs obtains its own access action by learning the historical experience,thus ignoring the dependence on the result of spectrum sensing.Finally,in order to balance the exploration and development in the process of DQN learning and training,the algorithm based on the upper confidence bound with Hoeffding is proposed to improve the exploration efficiency of DQN,that is,to improve the efficiency of SUEs learning access strategy.Simulation results show that,compared with the traditional greedy strategy reinforcement learning method and Myopic and other traditional methods,the proposed algorithm has faster convergence speed,and SUEs can access the maximum number of channels while avoiding the interference to PUs to a large extent.