Reasearch And Implementation Of MPSK Signal Activity Detector Based On Deep Learning

The current shortage of spectrum resources in communication systems is becoming more and more serious,and static spectrum allocation schemes have been unable to meet the requirements of higher data rate equipment.Under the special spectrum allocation mechanism,many licensed spectrums are idle,which is a waste of spectrum resources.Cognitive radio technology can solve the shortage of available spectrum resources through dynamic spectrum access.The essence of cognitive radio is spectrum sensing,the core idea of which is to accurately detect"spectrum holes" through signal activity detection,which plays a vital role in spectrum access and spectrum sharing.In the actual communication system,the signals of authorized users are various,the system has multiple channel parameters,and the channel environment interferes with the signals inconsistently.Therefore,the spectrum sensing technology is confronted with great difficulties.The majority of current spectrum sensing algorithms utilize heuristic algorithms,which manually customize a series of rules and assumptions in traditional detection algorithms.They cannot learn the characteristics of signals autonomously,and there is no generalization between multiple methods.To solve this problem,this paper proposes two types of deep learning-based Multiple-Phase-Shift-Keying(MPSK)signal activity detectors:A Signal Activity Detector Based on Neutrul Network and Dropout(NND-SAD)And A Signal Activity Detector Based on Attention Covariance Matrix Convolutional Neural Network(ACM-SAD).There are three points about the innovations in this paper,which can be described as followings:(1)The NND-SAD algorithm fuses the neural network and Dropout,and uses the received original signal as feature,and uses the memory and learning capabilities of the deep neural network to find the relationship between the input signals.Simulation results show that compared with other commonly,the NND-SAD algorithm,can achieve better performance.(2)The convolutional neural network model is introduced to the spectrum sensing algorithm,in order to make better use of deep learning.In this paper,we use the strong ability of the convolutional neural network to extract data features and extract the covariance matrix of the received signal as a feature to design test statistics to improve the performance of the signal activity detector.(3)This paper also improves the convolutional neural network algorithm and introduces innovative points of attention in the mixed domain.The introduction of attention can make the convolutional neural network more intelligent to improve the problem of insufficient representation.Therefore,this algorithm will help identify the activity of the detection spectrum in a complex background.Compared with three other commonly used signal detection algorithms,the simulation results show that the ACM-SAD algorithm can obtain better detection results in the detection probability,false alarm probability,and ROC curve.