Research On Spectrum Intelligent Perception Method Based On Deep Learnin

In the civilian sector,5G technology is driving the development of diversified wireless services,which require more spectrum resources to support in the context of high speed,low latency and large connection services.In the military field,especially after the "decoupling of weapon systems and dynamic reorganization of elements",the proliferation of frequency-using devices in high mobility and strong confrontation battlefield environment makes the wireless communication in limited area face serious internal and external interference,and the spectrum resources are especially tight.Cognitive radio can achieve multi-dimensional reuse of spectrum resources through spectrum sensing and dynamic access to improve the utilization of spectrum.However,the existing spectrum sensing methods still face the problems of insufficient sensing performance at low SNR,large amount of multi-user collaborative sensing data,cumbersome processing,and insufficient reliability of the fusion law.we investigate the single-user sensing and multi-user collaborative sensing fusion methods at low signal-to-noise ratios,and the main work and innovation points are as follows:(1)Aiming at the problem that existing spectrum sensing methods are difficult to distinguish signal from noise in low SNR environment and the sensing performance is insufficient,a single-user spectrum intelligent sensing method based on SNR classification is proposed.The method takes advantage of the obvious difference between the time domain waveforms of sampled signals in high and low SNR environments,and classifies the signals to be measured as high SNR or low SNR through an efficient SNR grading network,which can directly judge the presence of the primary user for signals classified as high SNR,and further sense them through a low SNR sensing network for signals classified as low SNR.The experimental results show that the method can effectively improve the detection probability of spectrum sensing at low SNR.(2)To address the problems of lack of reliability due to the lack of communication mechanism among sub-users in centralized collaborative spectrum sensing and large data transmission due to the adoption of soft merge fusion criterion,a self-attention sensing fusion network model for a few sub-users and a hypergraph learning sensing fusion network model for most sub-users are proposed.The self-attention fusion network model establishes a perceptual user exchange mechanism in the fusion center of centralized collaborative sensing,and the hypergraph learning sensing fusion network model uses the idea of subclustering to achieve information exchange between near-neighbor sub-users.The experimental results show that the self-attention fusion network model improves the detection probability of collaborative sensing with a few secondary users involved in sensing,and the hypergraph learning fusion network model reduces the false alarm probability of collaborative sensing with most secondary users involved in sensing.