A Combined Model WAPI Indoor Localization Method Based on UMAP

Abstract

With the rapid advancement of the Internet, indoor localization technology has gained increasing importance across various fields. However, the complexity of indoor environments presents significant challenges for achieving precise positioning using GPS or BeiDou systems. As a result, there is a growing demand for innovative localization methods that deliver high accuracy, improved security, and cost-effectiveness. In this study, a dataset comprising 9291 fingerprints collected from a building was processed and split into training and test sets in a 7:3 ratio. To facilitate feature extraction, four algorithms—UMAP, LDA, PCA, and SVD—were employed. Subsequently, six machine learning models (KNN, Random Forest, ANN, SVM, GBDT, and XgBoost) were trained on the training set and evaluated on the test set to compare their performance with different feature extraction algorithms. The objective was to identify the most effective feature extraction method. Model performance was assessed using three metrics: average error, coefficient of determination, and accuracy. Finally, a stacking ensemble model was developed, incorporating the six models as primary learners and selecting the five models with superior predictive performance as secondary learners. This approach aimed to enhance the localization accuracy. UMAP feature extraction significantly improved the prediction accuracy of the indoor localization model, whereas the stacking ensemble model, combining KNN, GBDT, XgBoost, ANN, Random Forest, and SVM as primary learners and Random Forest as the secondary learner, achieved the highest localization accuracy, with an error of approximately 1.48 m.