Research On Indoor Device-free Passive Fingerprinting Localization Via Transfer Learning

With the emergence of Io T applications,the demand for location information has increased dramatically and it plays a crucial role in the application implementation process.Applications such as smart healthcare,smart home,and object tracking all require accurate indoor location information.And with the continuous development of wireless technology and the wide deployment of wireless devices,Wi Fi,5G/4G,FM,TV,and other wireless signals cover almost every corner of our lives.When a target is in a different position,it inevitably affects the surrounding wireless signal differently,so it is feasible to estimate the target location by analyzing the affected wireless signal.This localization technique to estimate the target location without carrying any localization device or actively participating in the positioning process,we call it indoor device-independent passive wireless location technology.Due to the coarse-grained nature of many wireless signals in the indoor environment,they are affected by indoor multipath effects and cannot perceive the target accurately,resulting in unsatisfactory indoor device-free passive localization performance.With the release of IEEE 802.11 a/g/n protocol,the OFDM technology adopted by Wi Fi provides fine-grained channel status information for wireless positioning,which can inscribe indoor multipath propagation characteristics,providing a new direction for the development of indoor fine-grained and high-precision device-free passive localization,which has become the current research hotspot for indoor wireless positioning.At present,most of the literature on indoor device-free passive wireless fingerprinting localization requires the arrangement of multiple pairs of transmitting and receiving devices,which increases the overhead of the system.Besides,they ignored the fact that the fingerprints will change with time,resulting in a drastic decrease in positioning accuracy with time.Therefore,this dissertation,in conjunction with transfer learning,focuses on the device-free passive wireless fingerprinting localization methods based on a fine-grained single communication link in the context of localization fingerprints change with time.The main work and contributions of this dissertation are as follows.1.Firstly,to address the high system overhead and the dramatic degradation of localization accuracy over time that exists in most current work,we present MSDFL(Matrix Similarity-based Device-Free Localization),a device-free passive fingerprinting localization system based on relational transfer.It completes the classification based on the amplitude information of the channel state information by comparing the test samples and the fingerprint database samples by the proposed matrix similarity algorithm and finally estimates the target location.Because it utilizes rich channel state information as well as Multiple Antenna Information,accurate localization results can be provided using only a single communication link.To reduce noise and extract the subcarriers that contribute the most,we propose a novel data preprocessing scheme to process the collected amplitude data and sample alignment for the next step of relational transfer.Finally,the fingerprint change relationship is obtained through an artificial neural network,and the fingerprint database is updated using this relationship,solving the problem that the location fingerprints change over time resulting in a drastic drop in location performance.2.Secondly,to address the problem that the transfer of the MSDFL system through fingerprint variation relationships can potentially destroy the separability of data in the fingerprint database,we propose TLLOC(Transfer Learning-based Device-Free Localization),a device-free passive fingerprinting localization system based on a weighted transferable discriminative dimensionality reduction method.In TLLOC,we draw on the idea of common feature subspaces in transfer learning to propose a novel weighted transferable discriminative dimensionality reduction method.It aims to construct a lowdimensional potential space from the fingerprint database and a small number of samples of fingerprints after they have changed,which can simultaneously improve the separability of training samples and reduce the difference in distribution between the fingerprint database and samples of fingerprints after they have changed.In this low-dimensional potential space,TLLOC can utilize the amplitude information of a single communication link to ultimately obtain satisfactory positioning accuracy based on the support vector machine algorithm,while saving calibration costs due to the need to recollect the fingerprint database.3.Thirdly,to address the problem that the amplitude response of channel state information is low for indoor target perception,we calibrate the phase information extracted from the channel state information and comprehensively analysis their characteristics and their impact on the indoor device-free passive fingerprinting localization when used as the fingerprints.Based on this,we propose DFPhase FL(Device-Free Fingerprinting Localization using Phase),the first system that uses only the phase information to complete the device-free passive fingerprinting localization.DFPhase FL first extracts the original phase information from the channel state information measurements,then removes the phase offset to obtain the filtered calibration phase information.Due to the unpredictability of the calibration phase over time,we treat the problem of inconsistent distribution between the training sample and the test sample as a transfer learning problem and propose a transfer depth supervised neural network(TDSNN)method that combines deep neural networks with transfer learning.Through this method,we can learn both transferable and discriminative feature representations as fingerprints from the calibrated phase.Then,using the support vector machine algorithm,the DFPPhase FL system achieves satisfactory localization accuracy using only phase information from a single communication link,while saving the cost of recollecting the fingerprint database.4.Finally,MSDFL,TLLOC,and DFPhase FL systems first needed to train a transfer model to obtain new feature representations,and the localization model is then trained in the new feature representation space to locate the target,leading to an overly tedious training process.To address this problem,we propose LTLoc(Long-Term Effective Device-Free Localization).It requires only a Wi-Fi access point and a receiver and uses the amplitude and calibrated phase extracted from the channel state information together as a fingerprint to train a deep neural network(DNN)based regression model to obtain the target location.The rich fingerprint information further enhances the system's performance.Besides,to address the problem of sharp performance degradation of the localization model over a long period due to fingerprints change,we propose an adaptive deep neural network method based on a meta-network.It can use meta-networks to learn which layers and features in the localization model need to be transferred to automatically adapt to the change in the fingerprints.In this way,LTLoc eliminates the cumbersome training process of the MSDFL,TLLOC and DFPhase FL systems and achieves satisfactory positioning accuracy using only channel status information from a single communication link,which can be maintained over time without the cost of recapturing the fingerprint database.