Multi-Source Big Data Indoor Positioning Technology under the Federated Learning Framework

In view of the problems that traditional indoor positioning technology relies on centralized data processing, resulting in high risk of user location privacy leakage and difficulty in effectively fusing multi-source heterogeneous data, this study proposes a multi-source big data indoor positioning method based on the federated learning framework. First, this paper designs an asynchronous federated learning architecture based on the FedAvg protocol, which allows multiple edge devices to use LSTM networks to process their heterogeneous sensor time series data locally, and fuse PDR constraints to build a local positioning model. Secondly, this paper introduces an adaptive weighted aggregation algorithm on the server side, dynamically evaluates the reliability of each local model update through the entropy weight method, Finally,the method achieved an average positioning error of 1.24 meters using 1.2TB of multi-source data collected by 20 types of equipment.At the same time, the model convergence reduced the number of training rounds by 40%, and the differential privacy mechanism only introduced an additional error increment of 0.15 meters. The results show that the framework successfully resolves the contradiction between privacy and accuracy, and its adaptive fusion mechanism and lightweight privacy protection technology provide an innovative solution for large-scale indoor positioning.