Application Of Kernel Learning Methods In Wireless Sensor Networks

Due to the Global Positioning System(GPS)is not a good signal transmission in indoor environment,this thesis introduces the original method of sensor positioning in wireless sensor networks(WSN).Based on radio-location fingerprinting and machine learning,the method consists of defining a model whose inputs and outputs are the received signal strength(RSS)indicators and the sensors locations.Aiming at the influence of the change of the dynamic indoor environment on the positioning accuracy,a class of indoor positioning algorithms for wireless sensor networks using kernel-based learning method is proposed.The method of kernel-based learning methods includes Quantized Kernel Least Mean Square(QKLMS)method and Kernel Recursive Least-Squares(KRLS)method.The QKLMS algorithm used a simple vector quantization approach as an alternative of sparsification,to curb the growth of the radial basis function structure in kernel adaptive filtering.A class of kernel recursion least squares methods applying the sequential sparsification algorithm based on approximate linear dependence(ALD)and the algorithm of combined a sliding-window(SW)with L2-norm regularization as well as the fixed-budget kernel recursive least-squares(FB-KRLS)algorithm,construct the RSS mapping between the fingerprint information and the physical location of the non-linear mapping relationship.The ALD-KRLS algorithm using a sparsification procedure limited the size of kernel matrix and reduced the order of the problem,the SW-KRLS algorithm using sliding-window approach and conventional regularization can fixed the size of the kernel matrix.In contrast to a previous sliding-window based technique,the presented FB-KRLS algorithm does not prune the oldest data point in every time instant but it instead aims to prune the least significant data point in order to inhibiting the growth of the kernel matrix.These kernel learning methods are both kernel adaptive filtering methods.The ability of on-line learning enables the proposed indoor localization algorithm to adapt to the changes of dynamic environment,thus improve the positioning accuracy and computing speed.These kernel-based learning methods are both kernel adaptive filtering methods.The ability of online learning ability enables the proposed indoor localization algorithm to adapt to the changes of dynamic environment,thus improve the positioning accuracy and computing speed.The main contents of this thesis are as follows:(1)Study the positioning principle of position fingerprint positioning,through the analysis of the status quo at home and abroad to investigate the feasibility and practicality of position fingerprint positioning.This thesis studies the localization performance of traditional kernel-based learning methods,and studies the localization of Extreme Learning Machine(ELM)positioning method.(2)Study the theories of ALD-KRLS algorithm and SW-KRLS algorithm.Using the above two kinds of kernel adaptive filtering methods to define the position fingerprint localization model,a kind of wireless sensor network based on kernel adaptive filtering learning methods is proposed.The proposed positioning algorithm is applied in the simulation and physical examples.Under the same conditions,they will be compared with KPCA-SVM,KPCA-LSSVM,ELM,KELM,KPLS and OS-ELM algorithms.Experimental results show that the above two positioning algorithms proposed in this thesis have higher positioning accuracy than other positioning algorithms.(3)Study the theories of QKLMS algorithm and FB-KRLS algorithm.different positioning examples.Using the above two kinds of kernel adaptive filtering methods to define the position fingerprint localization model,a kind of wireless sensor network based on kernel adaptive filtering learning methods is proposed.The proposed positioning algorithm is applied in the simulation and physical examples.Under the same conditions,they will be compared with KPCA-SVM,KPCA-LSSVM,ELM,KELM,KPLS,OS-ELM,ALD-KRLS and SW-KRLS algorithms.The experimental results show that both QKLMS and FB-KRLS localization algorithms can shorten the training time of the model,and they all have higher positioning accuracy.More importantly,its ability to learn online sequences enables the proposed positioning algorithm to automatically adapt to changes in the dynamic environment.