Enhancing Indoor Carrier Phase Positioning Through Integration of Neural Network and Robust Unscented Kalman Filter

Abstract

Multipath effects pose a significant challenge in precise indoor positioning using pseudolites, especially for low-cost receivers. This article addresses the issue using both functional and stochastic models. Specifically, by leveraging homologous array pseudolites, we construct an intersatellite differential precise point positioning (PPP) model, where the carrier phase observation equations predominantly contain multipath errors rather than other systematic errors. Subsequently, we establish a single-differenced (SD) phase multipath error prediction model utilizing a feedforward neural network, considering the mathematical relationship between residuals and true errors, as well as the intrinsic connection between the carrier-to-noise-density ratio (C/N0) and the phase multipath error. This model, independent of position information, employs SD phase residuals and original C/N0 measurements as feature parameters, bypassing iterative computations and adapting to environmental dynamics. After training with continuously sampled data from sparsely distributed indoor stations, we conduct static PPP experiments at three test points. The average prediction accuracy of multipath error reaches 0.08 cycles, corresponding to a normalized root-mean-square error of 8% . By error correction, the horizontal positioning accuracy improves by more than 70% . The remaining multipath errors demonstrate a more concentrated distribution closer to zero, which creates conducive conditions for further enhancing positioning accuracy through robust estimation techniques. Experimental results using a robust unscented Kalman filter corroborate this finding, which underlines the effectiveness of the proposed approach.