A WiFi Indoor Localization Algorithm Based On Transfer Learning

As mobile terminal device technology matures and indoor positioning services play an important role in national security,economic development and civil applications,users are demanding more accurate and less costly indoor positioning systems.In outdoor environments,the use of satellite positioning systems and wireless base stations can help end-users achieve positioning results with accuracy down to the metre level.However,in indoor environments,satellite signals can be distorted to such an extent that satellite positioning is not possible,and on the other hand,the accuracy of satellite positioning is not high enough in indoor environments.Therefore,high-precision indoor positioning has long been a hot topic of research in positioning technology and is one of the main difficulties at present.WiFi fingerprinting based on Channel State Information(CSI)data can be used for indoor positioning.When using Received Signal Strength Indication(RSSI)data for indoor positioning,it is basically impossible to distinguish adjacent locations using RSSI due to random errors in RSSI caused by multipath shadowing,which fundamentally limits the ranging accuracy and positioning accuracy of RSSI.Compared to RSSI,CSI can characterise channel properties more finely and can be obtained over multiple subcarriers.When using CSI data for indoor positioning,we can obtain better results compared to RSSI.However,whether RSSI or CSI signals are used,the indoor environment usually changes after a certain period of time during indoor positioning deployments and fingerprint databases based on test data often deteriorate or even fail.This thesis presents a CSI-based signal pre-processing algorithm that allows the signal to be processed for migration learning,including an Automatic Gain Control(AGC)compensation algorithm,a clustering-based data selection algorithm,and a PCA-based data dimensionality reduction algorithm.The migration learning algorithms are then used to build a fingerprint database for indoor localisation,and a migration learning based WiFi indoor localisation system is built.The advantage of migration learning is that smaller data can be used to obtain better migration training results.In this thesis,we use migration learning to migrate the predictions of the fingerprint database and use Max Mean Discrepancy(MMD)to characterise the inter-domain distance between the source and target domains,thereby extending the lifetime of the fingerprint database and improving the robustness of indoor localisation.After experiments,the indoor localisation accuracy remained at 88% after one week and 80% after two weeks.At the same cost,the lifecycle and localisation accuracy of the model are higher than those of LSTM,CNN,SVM,DNN and other localisation systems.The main research results of this thesis include:· A signal processing algorithm suitable for targeting CSI data is proposed.For the problem that AGC destroys the distance information in CSI amplitude data,an al-gorithm is proposed to compensate for it using RSSI data.A clustering-based data selection algorithm is proposed for outlier data in CSI data.A data reduction al-gorithm is proposed for high-dimensional CSI data to centralise the localisation information and reduce the computational effort in migration learning.· A theoretical algorithm for migration learning of CSI data is proposed.It includes a feature extractor,a domain discriminator and a location predictor.The feature extractor is a multilayer fully connected neural network,the domain discriminator uses MMD to characterise the inter-domain distance and the location predictor is an LSTM network.· An optimisation algorithm is proposed for the fingerprint library.This includes an algorithm for the selection of the target domain data and an algorithm for the updating of the fingerprint library parameters.The life cycle of the localisation system is significantly extended and its robustness is enhanced.