Research On Indoor Localization Method Based On Federated Learning

Indoor localization is a process in which users determine which position they are in the indoor environment through some signals in the indoor environment.This demand is increasing with the popularity of mobile devices.In recent years,with the development of deep learning,it is a promising way to locate indoor based on fingerprint information generated by Wi Fi signal transmitter in indoor environment.The model training is carried out according to the indoor fingerprint information collected in advance.Then,the fingerprint information collected by the user is directly input into the model to predict the location of the user.However,due to the variability of indoor environment,in order to maintain location accuracy,service providers need to collect a lot of data frequently for model retraining,which is of high cost.The method of mobile crowdsensing can reduce the cost of data acquisition,but this method seriously leaks the location privacy of participants.Recently,researchers have proposed a federated learning method to solve indoor localization problems,which effectively protects the privacy of users.However,most of the work focuses on the localization of two-dimensional space plane,which cannot be applied to the large-scale 3D localization scene of large-scale buildings.At the same time,the federated learning method also faces the problem that the localization accuracy is significantly lower than the centralized method in the case of non independent and identically distributed data.In view of these problems,this paper proposes a federated learning based indoor localization system,introduces the building floor classification sub task and data augmentation means.The specific work of this paper is as follows:(1)A building floor classification method FedDNN-BFC based on federated learning is designed.By combining this method with the plane localization method,it can meet the needs of 3D localization scene.This method is based on multi label classification model design,through data preprocessing,centralized pre training,federated training three stages to complete the whole training process,effectively protect the privacy of participants.At the same time,on the basis of FedDNN-BFC method,combined with the distribution characteristics of RSS values in different positions,the FedCNN-BFC method is proposed,and the multi label classification model based on convolution is adopted to further improve the accuracy.Finally,the experiment on UJIIndoor Loc data set verifies the effectiveness of FedDNN-BFC method,and the classification performance of FedCNN-BFC method is better than that of FedDNN-BFC method.Under the condition of data independent and identically distributed,the overall classification accuracy of FedCNN-BFC is 1.2% higher than that of FedDNN-BFC,and the communication cost is 24.5% lower.Under the condition that the data is not independent and identically distributed,the overall classification accuracy is increased by 1.6% on average,and the communication cost is reduced by15.7% on average.(2)The FedADA-HCR is designed to embed data augmentation in federated learning localization method.By embedding data augmentation process and changing the original federated training process,this method can achieve the purpose of expanding the user's local data set without increasing the workload of users,so that each user can hold more data when participating in federated training,and then improve the final location accuracy.The experimental results on UJIIndoor Loc data set show that the localization error of FedADA-HCR method can be reduced by 1.48 m under the condition of independent and identically distributed,and 2.36 m under the condition of non independent and identically distributed compared with the most advanced methods available.At the same time,experiments show that the proposed method FedADA-HCR can achieve the best localization effect when the amount of enhanced data and the amount of data held by the client is about 1:1.