Device-Free Localization Approaches Based On Extreme Learning Machines

Device-free localization(DFL)is a wireless localization technique that the target does not require any electronic device,which has great potential application value in elderly care,emergence rescue and intrusion detection,etc.DFL technique can estimate the target's location by analyzing the differential signal strength of the links in the monitoring area.In practice,the radio frequency(RF)signals propagate in a complex environment and are subject to multiple factors,including time-varying environment,non-line-of-sight,multi-path signal propagation,etc.,which leads to the collected data contained much noise,and directly degrades the localization accuracy.Mechanism based DFL approaches can reveal the signal propagation,and also can achieve satisfactory localization accuracy.However,it is sensitive to the changing of environment with relatively poor robustness.Thus,how to build robust DFL models in the noisy environment is necessary.This thesis proposes fast,efficient,accurate and robust DFL approaches based on extreme learning machine(ELM)theory.Firstly,it explores the machine learning based DFL from the aspects of ignoring the effects of noise distribution,assuming the noise follows Gaussian distribution,and the noise follows non-Gaussian distribution,and then extends the traditional small area DFL to the large-scale complex area DFL;finally,it realizes the automatic and efficient feature extraction of the collected data by proposing a multilayer ELM based on deep learning theory,and also converts DFL into probability problem,which improve the localization accuracy further.The main research contents are as follows:(1)Single feature usually cannot guarantee to establish an accurate DFL model Thus,an ELM based parameterized geometrical f-eature extraction(ELM-PGFE)is proposed to extend the number of features for strengthening the robustness of the created model and improving the localization accuracy.(2)During the propagation of radio frequency signals,it will cause a large number of noise due to the influence of uncertainties.Therefore,a residual compensation ELM(RC-ELM)is proposed based on hierarchical residual compensation mechanism,for building a DFL model with better robustness and generalization performance by remodeling the modeling errors iteratively.(3)The noise contained in RF signals is usually very complex and may be subject to Gaussian.Laplacian or mixed distributions,so the conventional machine learning approaches may be difficult to achieve satisfactory results.In order to tackle this issue,a robust ELM(R-ELM)is proposed based on mixture of Gaussians(MoG),which can improve the localization accuracy by approximating of arbitrary continuous noise distributions,and make fast and accurate DFL under random disturbance.(4)Most of the existed DFL approaches are only suitable for small-scale area,which are lim ited.In order to expand the application range of DFL,and further improve its practicability,a hierarchical framework for large-scale DFL is proposed based on ELM theory and K-means clustering.Firstly,the large-scale area is divided into some small sub-regions using K-means,and then creating the corresponding DFL models;furthermore,considering the high noise contained in the data collected from large-scale and complex area,an extended RC-ELM(ERC-ELM)for non-Gaussian noise is proposed based on RC-ELM.Compared with RC-ELM,ERC-ELM abandons the extension in depth,and instead of expansion in width.Thus,it is more efficient and robust in dealing with non-Gaussian noise.(5)There are many redundant data in links of a wireless sensor networks,and shallow machine learning approaches sometimes fail to extract and learn the features provided by those links.Simultaneously,artificial feature extraction is not only time-consuming and labor-intensive,but also only suitable for specific scenarios.Thus,a multilay'er probability ELM(MP-ELM)is proposed for DFL based on deep learning theory.The proposed MP-ELM can automatically extract features of links fast,efficiently,and accurately,and converts DFL into a probability problem.Therefore,it can reduce the modeling error accumulation,and improve the localization accuracy further.