Feature Recovery And Anomalies Separation Of Electromagnetic Spectrum

Spectrum sensing is an important foundation of intelligent spectrum resource management.On the one hand,the accuracy of spectrum sensing is affected by environmental interference and noise,so it needs to realize accurate estimation of frequency usage by feature recovery.On the other hand,it is necessary to separate important information for abnormal frequency behaviors such as illegal use of spectrum from observation.Thesis carries out research on spectrum feature recovery and anomaly separation from the aspects of tensor optimization modeling,efficient algorithm design and real-time data processing.Main works of thesis are as follows:(1)In view of the problem that abnormal behaviors will damage the regularity of spectral characteristics in long-term spectrum observation,the high-dimensional tensor operation method is used to expand the two-dimensional matrix,and the kernel norm in the convex relaxation model is replaced by the non-convex Schatten-p quasi-norm to approach the tensor rank minimization problem more accurately,proposing an algorithm of spectral feature recovery and anomalies separation based on low-rank and sparse tensors.Computer simulation results show that the proposed algorithm can reduce the recovery error by 2?3 orders of magnitude compared with the traditional low-rank matrix recovery algorithm.(2)In view of the low operating efficiency of the Schatten-p quasi-norm model,a factorization model is obtained by rank decomposition of the tensor,and the iterative process to solve the factorization model is transformed into the forward propagation process of the Deep Unfolding neural network,proposing an algorithm of spectrum feature recovery and anomalies separation driven by model and data.Computer simulation results show that the operation time of the proposed algorithm is 1/10 of the factorization iterative algorithm with the same recovery effect.(3)In view of the large computation and poor real-time performance of batch algorithms when processing spectral data,convert the tensor into a product of a basis matrix and its coefficients.Continuously update basis matrix and solve the coefficients with stochastic online optimization method according to the latest input data samples,proposing an online algorithm of spectrum feature recovery and anomalies separation.Experimental results on real spectrum dataset show that the proposed algorithm can process the online spectrum data stream with millisecond-level delay.