Research On Machine Learning Based Wideband Spectrum Sensing Algorithms

With the rapid development of wireless communication networks and application services,spectrum resources have become increasingly scarce.Fixed wireless spectrum allocation mechanisms have led to low utilization of a large number of spectrum resources.As a spectrum sharing technology,cognitive radio can significantly improve spectrum utilization,so it is considered to be a key technology to alleviate the problem of spectrum resource shortage.Wideband spectrum sensing is an important part of cognitive radio,which can realize dynamic tracking and monitoring of the status of all sub-bands in the wide licensed frequency band.Aiming to overcome the drawbacks of traditional wideband spectrum sensing and further improve the spectrum utilization efficiency,this thesis explores the possibility of applying machine learning classification methods in spectrum sensing.The specific content includes the following two aspects:Wideband spectrum sensing not only needs to identify whether primary user's signal exists on the target frequency band,but also accurately locate the sub-bands occupied by the primary user.This thesis proposes a dual target detection scheme based on convolutional neural network for wideband cooperative spectrum sensing,which divides the wideband cooperative spectrum sensing into two parts: judging the width and the starting position of the sub-bands occupied by the primary user.First,we collect a large number of spectrum energy observation data under multiple spectrum occupancy patters in different environments,and design a convolutional neural network model.Then we train the designed model based on the spectrum observation data.Finally,the trained classifier is utilized to detect the unknown signal to identify the spectrum occupancy pattern,thereby achieving wideband cooperative spectrum sensing.Simulation results prove that the proposed scheme has greater advantages in detection performance than the traditional wideband cooperative spectrum sensing scheme based on energy detection technology and typical machine learning classification algorithms,and especially in low signal-to-noise ratio the effect is more prominent in the environment.The supervised learning classifier requires label set corresponding to the training data set,which is difficult to obtain in practice.Aiming at tackling the issue that the label set is inaccessible,a wideband cooperative spectrum sensing scheme based on clustering algorithm is proposed.In order to decrease the computing complexity in the sensing process and raise the sensing efficiency,this scheme first performs dimensionality reduction for the spectrum observation data using 2D principal component analysis algorithm,and trains the classifier based on the low-dimensional training data.Then the trained classifier is used to determine the current set cardinality of the sub-bands occupied by the primary user.Finally,part of the sub-bands with the largest energy observation average value among all the frequency bands are found by calculation.Simulation results demonstrate that this scheme can perform well without the label set corresponding to the training data set,and the training and detection time can be greatly shortened after dimensionality reduction.