Wi-Fi fingerprint localization model via synergistic manifold and feature optimization

Abstract High-dimensional received signal strength indicator (RSSI) data often contain redundant and correlated features, leading to overlapping distributions and reduced localization accuracy. To address this, we propose a hybrid indoor localization framework that combines a convolutional autoencoder (CAE) with uniform manifold approximation and projection (UMAP) for hierarchical dimensionality reduction, followed by an improved weighted k -nearest neighbor (WKNN) algorithm for robust position estimation. The CAE extracts compact embeddings, while UMAP refines them by preserving intrinsic manifold structures, enhancing feature separability and noise suppression. The enhanced WKNN introduces feature-based weighting to stabilize accuracy under varying k -values. Experiments on two complex indoor sites demonstrate the effectiveness of the proposed method, achieving root mean square errors of 1.83 m and 2.66 m—consistently outperforming baseline approaches. The framework maintains real-time inference speed and exhibits strong robustness to signal degradation and noise. However, its performance may decline under sparse access point deployments or suboptimal hyperparameter settings, indicating a need for future improvements in adaptability and automated tuning strategies.