Research On Intelligent Spectrum Sensing Technology Under Non-cooperative Conditions

With the expansion of wireless communication,the demand for wireless resources is increasing and the frequency spectrum resources are becoming increasingly scarce.In order to completely explore the available frequency range and avoid the spectrum usage conflict amongst users,cognitive radio develops at the historic moment.Spectrum sensing is one of the basic technologies of cognitive radio.Cognitive user Through spectrum sensing,the user access status in the channel is assessed to give a basis for subsequent dynamic spectrum access.Traditional recognition techniques generally rely on past information and cannot make use of multi-source data features.Intelligent recognition algorithms based on machine learning offer great multi-source data fitting capabilities and became a research hotspot in this subject in recent years.However,the existing intelligent detection research rarely analyzes the reliability of recognition under non-cooperative conditions.Artificial intelligence methods under noncooperative conditions will encounter the following issues.Firstly,it is difficult to gather adequate user data in non-cooperative conditions,thus the target data is scarce,and the model is prone to over-fitting problems in the training process.Secondly,the noise in space will introduce channel interference,impact the signal-to-noise ratio,and subsequently affect the performance of the intelligent perception algorithm.Thirdly,it is difficult for the receiver to correctly follow the frequency and phase of the carrier under non-cooperative conditions,especially under the condition of high speed frequency hopping.In view of the foregoing challenges,this study recommends the following methods:1.Aiming at the influence of frequency offset in non-cooperative conditions,a frequency offset correction approach based on constellation density clustering is given,and the signal recognition insensitive to offset is realized by developing a fine-tuning VGG-16 network.2.On this premise,a convolutional neural network approach based on domain knowledge embedding is studied.The correction operation is built into the neural network structure,which offers greater real-time performance compared with frequency deviation correction based on density clustering.3.Aiming at the problem of the scarcity of target samples,a meta-learning recognition method based on siamese networks is presented,which can conduct category matching based on a small number of samples and efficiently differentiate unfamiliar tags.In addition,by incorporating rotation-insensitive features in the network,the recognition efficacy is improved when the phase cannot be synchronized adequately.Based on open source data and software radio platform,the experiment proves the effectiveness of target recognition under small sample settings.4.Simulating the actual application situations of time,phase and frequency asynchronization under non-cooperative conditions,a hardware-in-loop simulation experiment platform is constructed based on software radio,which can support numerous transmitters,receivers and jammers.The software development and hardware debugging are accomplished.The radio frequency database is produced based on different types of software and wireless hardware in diverse electromagnetic spectrum situations,and the suggested algorithm is confirmed based on the observed data.In this study,three signal identification methods are proposed,emphasizing on the difficulties of phase,clock and frequency synchronization and the problem of small samples under non-cooperative settings.In addition,a non-cooperative electromagnetic environment was developed based on ETTUS,DEEPWAVE and other multi-brand software radio gear combined with GNURadio tools to evaluate the effectiveness of the algorithm.This study is of tremendous significance for spectrum recognition under noncooperative conditions.