Research On Indoor Localization Algorithm Based On Federated Learning

With the popularization of 5G mobile communication technology,the Internet of Things(IoT)industry has been fully developed,but also put forward new requirements for precise indoor location services.In order to maintain the performance of localization model,location service providers require to frequently request location data from mobile device users for localization model update.However,location data is closely related to the behavior of mobile device users.If the location data of mobile device users is maliciously stolen during transmission,it will not only cause the privacy disclosure of mobile device users,but also reduce the willingness of mobile device users to participate in localization model update.With the development of edge computing,federated learning promises to be a solution to the privacy disclosure problem.As a distributed learning method,federated learning can combine multiple clients for distributed training of localization model,protecting their data privacy.However,the local model of the clients has to be uploaded periodically for model aggregation,which brings a serious communication traffic burden to the mobile devices of the clients.At the same time,the long-term update of the local model also leads to the problem of computation power consumption.In addition,due to the influence of non-independent identically distributed(Non-IID)data,the localization performance of the global model in federated indoor localization will ber affected to some extent.Therefore,this thesis studies WiFi indoor localization algorithm based on federated learning.The main research content of this thesis is as follows:(1)The technical characteristics of FedAvg algorithm in federated learning framework are studied.For the communication and computing process of the client,the influence factors of the communication energy and computation energy generated by the client mobile devices are discussed,and the energy consumption problem of the federated indoor localization is analyzed.For the workflow between the client and the server,the characteristics of the computation time and communication time of the client are discussed and the problem of the inconsistent synchronization time in each communication round are analyzed.For RSS fingerprint data,the heterogeneity of client's local fingerprint dataset is analyzed,and the problem of Non-IID data in the client local dataset is analyzed.(2)An improved federal indoor localization method is proposed.An improved strategy based on layerwise asynchronous aggregation is utilize to reduce the communication energy consumption caused by the traffic of upstream and downstream links in the process of federated indoor localization.This strategy improves the energy efficiency by performing asynchronous aggregation between the model layers.In addition,An improved strategy based on layerwise swapping training is proposed to mitigate the the impact of the Non-IID problem on the localization performance.This approach enables the deep layers to be adapted to the data distribution characteristics of other clients.Finally,experiments on standard database and measurement data verify that the proposed method can effectively improve the communication energy efficiency while maintaining the localization performance.(3)A reinforcement learning based computation resource adjustment method in federated indoor localization is proposed.Firstly,the relationship between client computation energy and synchronization time is analyzed under unstable communication environment.To reduce the computation energy consumption of the client,the CPU frequency of the client is dynamically adjusted by using the proximal policy optimization.Finally,experiments on standard database show that the proposed method can effectively reduce the synchronization time and computation energy consumption of each communication turn in federated indoor localization.