Research On Sensed Signal Classification And Prediction Using Machine Learning Techniques

With the development of wirless communication systems,the spectrum resource is increasingly saturated.Cognitive Radio(CR)technique emerges as a promising solution to this saturation by accessing the channels that are not occupied by the licensed users.The cognitive users select the best available channels dynamically and tune their tranceviers to the channel they choose.The best-known CR architecture was referred to as the Radiobot which accomplish self-learning,self-configuration and spectrum awareness.In order to be aware of its electromagnetic environment,a CR node needs to be able to classify and,if possible,predict sensed signals.However,it is often impossible for a CR node to know the ideal signal feature vectors and collect sufficient training samples as a prior while entering an alien RF environment,as well as having a prior knowledge of the number and types of wirless systems.Lack of these information,we resort to nonparametric unsupervised learning techniques to achieve classification and prediction.In particular,the Dirichlet Process Mixture Model(DPMM)requires neither the prior knowledge of the clusters,nor the training samples.The model allows the structure to adapt to the complexity of observed data and infer the number of signal classes from the data itself.The classification problem in this model is equivalent to estimate the posterior of classification label,which,however,the posterior is not analytically tractable because of its highly complicated form.In this paper,we resort to the approach of Variational Inference(VI)approximation approach to estimate the posterior with analytically tractable form.The classification problem then is solved by estimating a classification label for each observation through calculating the maximum posterior probability(MAP)of this approximated posterior distribution.Simulation results validate the proposed classification model and show that it can achieve way faster convergence than the existing non-parametric unsupervised classification algorithms at the cost of incurring an acceptable reduction of the classification accuracy.Based on the proposed signal classification,we further develop a new method for predicting the posterior of various wireless systems parameters and that of the sensed signal.Based on the probability theory,we derive the posterior distribution of future signal based on a set of sensed signals.We further point out that,in the context of CR networks the precise prediction of future signals is not necessary.Instead,the classification label and its parameter of corresponding wireless system are preferred.Following this clue,we further develop another prediction algorithm for classification label such that the wireless system,to which the forthcoming signal belongs,can be effectively predicted.Based on the posterior of certain wirless system parameters,such as carrier frequency and signal bandwidth,the CR node can effectively avoid the channel to be occupied by certain wireless system in advance and save its overhead of channel switching.Finally,numerical simulation verifies that,compared with the existing ones,the proposed two prediction algorithms have a better prediction accuracy.Moreover,the more samples of sensed signal,the more accurate the prediction of the wireless system to which the signal samples belong.