Deep Learning Based Indoor Fingerprint Localization With RSSI

In recent years,location-based applications are developing rapidly which has become an important feature of the Internet of Things era.Wi-Fi fingerprint localization has shown strong competitiveness for indoor localization because of simple deployment,low cost,and easily available signal.Unfortunately,the complicated and ever-changing indoor environment incurs reflection,scattering,and refraction of signal,which leads to highly time-varying and non-linear characteristics of signal.It is difficult to capture the signal features using traditional indoor localization methods,thus causing poor localization accuracy.In this thesis,deep learning methods are adopted in indoor localization,even in the case where some APs malfunction.First of all,based on the measured dataset,we conduct experiments to verify the feasibility of machine learning for localization.In a classroom of Beijing Jiaotong University,we set up signal collection fingerprint point 0.6 m apart each.In total,6,720 pieces of data are collected at 224 reference points.Then we performed localization experiments using machine learning algorithms such as k NN and random forest,and localization accuracy up to 1.62 m can be achieved.Secondly,according to UJIIndoor Loc public dataset,data are constructed to construct RSSI image.A CNN-IMG localization scheme is proposed.The results showed that the scheme had an average performance in the localization of buildings and floors.The localization accuracy of the latitude and longitude of the room reached 8.6 m.Considering the instability of RSSI single reading,we proposed a 1D-Res Net localization scheme based on the one-dimensional time sequence,which can achieve 100% classification accuracy on building and floor localization.And the localization accuracy of the latitude and longitude of the room reached 3.45 m,which outperforms the state-ofthe-art counterparts.Furthermore,the localization scheme is also used on the aforementioned measured dataset,and the localization accuracy of 1.53 m can be acquired,which is slightly better than machine learning algorithms.Finally,data incompleteness caused by AP failure is a challenge in localization.The detection of faulty APs as well as data completion is proposed.The lightweight network Shuffle Net V1 can detect faulty APs at accuracy rates of 77.8% and 85.07% for the nontime sequence and time sequence,respectively.The average RSSI is filled in lieu of the faulty AP data to complete the dataset.The Res Net is then used in subsequent localization process.The results show that the localization error after data completion can be reduced up to 0.9 m in comparison with the scheme of incomplete dataset.