Graph Temporal Convolutional Network-Based WiFi Indoor Localization Using Fine-Grained CSI Fingerprint

WiFi-based localization has become the prevailing localization technology for indoors, due to minimal hardware requirements, high-precision, and ubiquitous access. However, traditional WiFi-based methods face challenges in the achievement of accurate and reliable localization due to their susceptibility to interference, high fluctuations of WiFi signals, and neglect of spatial and temporal relationships of fingerprint data. To address these challenges, we propose a WiFi indoor localization method based on multiscale principal component analysis (MSPCA) and a graph and temporal convolutional network (MSPCA-GTCN). The proposed method models the location relationship between access points (APs) as an undirected graph and uses fine-grained channel state information (CSI) as a fingerprint. To improve the localization accuracy, an MSPCA algorithm is proposed to preprocess the raw fingerprint, which uses wavelet decomposition to eliminate the noise and uses PCA to avoid the impact of redundancy on wavelet decomposition denoising. Moreover, to resolve the incapability of the graph convolutional network (GCN) to process temporal data in parallel, a GTCN is proposed for localization, which leverages the temporal correlation of fingerprint data to compensate for errors caused by environmental fluctuations and limited sampling durations, thus further achieving localization precision. Simulation results show that the proposed algorithm can reduce distance error by up to 52.9% and improve the cumulative distribution function (CDF) by up to 67.9% compared with some existing algorithms.