随着全球导航卫星系统（Global Navigation Satellite System, GNSS）技术的发展，多星座融合显优化了空间几何构型并增加了可用卫星数量，提升了GNSS位置服务性能。然而在城市复杂环境中，GNSS信号极易受到非视距（Non-line of sight, NLOS）信号和多径效应的干扰，导致传统随机模型难以精确刻画实际观测噪声的分布。本文提出了一种基于可微状态估计架构的多系统GNSS随机模型精化方法，通过联合可微神经网络NetW与赫尔默特方差分量估计（Helmert vari-ance component estimation, HVCE）方法（NetW-HVCE），有效融合了神经网络提取的卫星观测量层级特征信息与通过HVCE优化的多GNSS系统级权重信息，实现了多层级GNSS随机模型精化。城市复杂环境下的车载动态实验表明，本文方法的定位性能得到提升，三维定位均方根误差（Root Mean Square Error, RMSE）较传统标准单点定位（Standard Point Positioning, SPP）方法和深度学习辅助定位方法SPP-TDL分别降低64.2%和32.1%。

With the development of Global Navigation Satellite System (GNSS) technology, multi-constellation fusion has greatly optimized the spatial geometric configuration and increased the number of available satellites, improving the perfor-mance of GNSS positioning services. However, in complex urban environments, GNSS signals are highly susceptible to interference from non-line-of-sight (NLOS) signals and multipath effects, making it difficult for traditional stochastic models to accurately characterize the distribution of actual observation noise. This paper proposes a multi-system GNSS stochastic model refinement method based on a differentiable state estimation architecture. By combining the differentiable neural network NetW with the Helmert Variance Component Estimation (HVCE) method (termed NetW-HVCE), the proposed method effectively fuses the observation-level feature information of satellites extracted by the neural network and the multi-GNSS system-level weight information optimized by HVCE, and realizes the multi-level refinement of the GNSS stochastic model. Vehicle-mounted dynamic experiments in complex urban environments veri-fy that the proposed method achieves improved positioning performance. The root mean square error (RMSE) of three-dimensional positioning is reduced by 64.2% compared with the traditional Standard Point Positioning (SPP) method and by 32.1% compared with the existing deep learning-assisted positioning method SPP-TDL.