Research On RF Fingerprinting Localization Based On Metrics And Transfer Learning

RF(Radio Freqency)fingerprinting localization is one of the most promising indoor localization technologies,and its main advantage is to avoid analyzing complex electro-magnetic propagation environment by artificial modeling.How to analyze and process the information contained in the complex and rich spatial RF fingerprint is both an op-portunity and a challenge for the positioning task.At present,the main problems of RF fingerprinting localization in practical applications include: high dimensional and nonlin-ear,complex distribution,high cost of acquisition and maintenance,large variance,etc.Aiming at these problems,this dissertation mainly studies the key problems of RF finger-printing localization from the point of view of RF fingerprint metric learning and transfer learning.The main contributions and innovations of this dissertation are as follows:1.The theoretical performance limitations of RF fingerprinting localization are ana-lyzed.The existing RF fingerprint channel models and classical RF fingerprinting local-ization algorithms are summarized.Based on the multi-wall model and ray-tracing model,the simulation environment of fingerprinting localization is designed.The performance of the classical fingerprinting localization algorithm is compared on four large public datasets and the advantages and disadvantages of the classic algorithms is analyzed.2.Aiming at the problem of solving the problem of optimal metric in classical RF fingerprint location algorithm,this dissertation studies the metric learning method in RF fingerprinting localization,and proposes two Mahalanobis distance metric learning meth-ods driven by KNN fingerprinting localization,called LMNN-RF and NCA-RF method,respectively.LMNN-RF optimizes the neighborhood relation of fingerprint by using the hinge loss,and makes the neighborhood of fingerprint points as its real physical space neighborhood.NCA-RF learns the metrics by minimizing the expected regression error of the KNN fingerprinting localization.In order to enhance the constraint of the fingerprint adjacent relation and the learning ability of the model,the second-order nearest neighbor loss term and the kernel technique are introduced into the loss function of the methods.The proposed methods can be used as supervised fingerprint feature extractors or dimen-sion reducers.The proposed measurement learning methods are simulated and tested on public data sets.The results show that their performance in fingerprinting localization,feature extraction and dimension reduction is better than that of the traditional methods.3.Aiming at the fingerprint invalidation caused by the change of distribution in the RF fingerprint location,the transfer learning method is studied.Based on the Optimal Transport(OT)theory,a RF fingerprint transfer learning method is proposed.By min-imizing the Earth Mover's Distance(EMD)between the source domain and the target domain,the fingerprint is mapped from the source domain to the target domain to realize the transfer of the fingerprint distribution.In addition,Laplace regularization and fin-gerprint feature transformation are introduced into the loss function to obtain a smoother fingerprint mapping.Through the simulation of logarithmic attenuation model and multi-wall model,it is found that the traditional method has negative transfer,and the OT-based transfer learning method has positive transfer.Numerical simulation and open dataset experiments show that the proposed method has better transfer performance than the tra-ditional method.4.Aiming at the problem of dynamic RF fingerprinting localization,the method of learning space-time pattern of RF fingerprint using sequence model is studied.Two mod-els based on Recurrent Neural Network(RNN),named RFed RNN and RFmp RNN,are proposed for sequential fingerprinting localization.In the models,encoding,decoding,attention and path prediction are introduced to realize the end-to-end filtering and pre-diction from fingerprint to path.The simulation results show that the proposed method is superior to the existing methods in performance.The simulation results show that the proposed method is feasible and effective.Test results on public datasets have reached the same conclusion.