Multivariate Time Series Learning for GNSS NLOS Intelligent Detection

The reception of non-line-of-sight (NLOS) signals significantly degrades the positioning precision of the global navigation satellite system (GNSS) in urban canyons. Recently, deep learning-based methods have shown significant potential in NLOS detection. However, existing methods suffer significant performance degradation across different scenarios, primarily due to an insufficient understanding of multivariate dependencies within GNSS measurement sequences. To address the above challenge, we propose a multivariate time series learning-based NLOS intelligent detection method, which improves detection accuracy by treating nonlinear temporal changes in GNSS measurement sequences as a multivariate dependency issue. Firstly, a transformer-based multivariate time series learning (TMTSL) module is designed to capture multivariate dependency features from GNSS measurement sequences effectively. It utilizes a decoupling-to-aggregation mechanism to extract univariate temporal information and aggregate multivariate temporal information, enhancing the representation of multivariate dependency features. Moreover, a graph transformer-based environment information extraction (GTEIE) module is developed to capture diverse environmental information. Specifically, a sky satellite graph is constructed to represent the spatial geometric distribution of satellites at a given epoch. The graph transformer is then used to extract environmental representations from this graph. Consequently, a multivariate time series learning-based NLOS detection method is proposed using a cross-attention-based fusion network to combine the environment information and multivariate dependency features for accurate NLOS detection. Experimental results demonstrate that the proposed method outperforms existing approaches, achieving a detection accuracy of over 96% while also exhibiting enhanced generalization performance.