Study On Online Bayesian Model Based On The Target Reognition Of High Range Resolution Profiles

Radar Automatic Target Recognition(RATR)is a combination of radar technology and pattern recognition technology,because of application value,it has been widespread concern in recent years.In the radar field,Radar High Resolution Distance Profile(HRRP)can contain a rich set of structural information,it is important for the identification of the target,In addition to its easy access,processing and storage features,making its research become a hot spot in the radar field.In the field of pattern recognition,the traditional machine learning classification algorithm,especially our commonly used support vector machine(SVM)algorithm,most of which are designed and worked in the batch learning mode.In practice,we often encounter the data set is too large,so that traditional batch learning will consume a lot of time and space resources,resulting in low learning efficiency.This thesis mainly focuses on the related projects of the National Natural Science Foundation,from the perspective of maximum interval supervision and learning theory,Bayesian statistical model,Bayesian model of online technology and other aspects of the relevant theoretical and technical research.The main contents of the thesis are summarized as follows:1.In this thesis,a Bayesian statistical(data dimensionality reduction)model,the largest interval factor analysis(MMFA)model,is proposed for the problem that the radar data is too high and there is characteristic redundancy.The model is to solve the model parameters by combining the generation model with the discriminant model in the same Bayesian framework.Here,the generated model is a factor analysis(FA)model,which can describe the information in the low-dimensional hidden space,to explore the potential common features which are often used for data feature extraction and dimensionality reduction.The discriminant model is a latent variable support vector machine(LVSVM)classifier.However,because the FA model is an unsupervised model,the extracted feature is not necessarily suitable for the back-end classification.In order to improve the separability of the feature,we use the hidden variable extracted by the FA model as the import of back-end discriminant model,the FA model and the LVSVM classifier are combined in the model solving process.Due to the constraint of the maximum interval of the back-end classifier,it is ensured that the FA model extracts the separability of the feature in the low-dimensional hidden space.In the solution of the model parameters,we use the variational Bayesian(VB)algorithm,w hich can estimate the model parameters with less computational complexity and good convergence performance.2.In view of the problem that the data set is too large and the learning efficiency is low,an online variational Bayesian algorithm is proposed.The traditional variational Bayesian(VB)algorithm requires the use of all the samples when learning the parameters of the model.When the amount of data is too large,the e fficiency of the algorithm will be very low,and when a new sample is added,The method needs to abandon the previous learning outcomes,collect all the samples again,re-train the model,which spends a lot of time and space resources.Therefore,we propose an online method for the Maximum Margin Factor Analysis(MMFA)model—the Online VB algorithm.In this method,the sample set is randomly divided into many sub-sample sets with the same size,and the objective function is optimized by stochastic gradient descent method.In the case of large sample,the method is faster than the traditional VB method.In the process of solving,the online VB algorithm divides the variables into global variables and local variables respectively.Only the global variable solution process is slightly different from the traditional VB.