近年来，铁路已成为国家经济和社会发展的重要支柱。中国铁路的迅猛发展，不仅展示了国家的工程实力，更将中国的铁路技术推向了世界的前沿。已有的铁路建设多采用全球定位系统（Global Positioning System，GPS）观测，缺乏对北斗卫星导航系统（BeiDou Navigation Satellite System，BDS）数据在铁路中的应用研究。BDS是我国自主建设、独立运行的全球卫星导航系统，能够提供高精度、高可靠性的导航、定位、授时和通讯服务。铁路BDS连续运行基准站(Continuously Operating Reference Station，CORS)为带状线路工程提供服务，结合大高差高海拔山区的特殊环境条件，对铁路工程的基线解算精度、对流层延迟估计以及高程拟合等关键技术产生显著影响。因此，利用BDS在大高差高海拔地区的铁路工程中的技术优势，对解决西部山区峡谷和高原铁路高精度快速定位问题、铁路行业摆脱对GPS的依赖具有重要意义。论文以西部高原山区某铁路为研究对象，依托铁路沿线建设的带状稀疏北斗CORS站观测数据，开展大高差高海拔铁路BDS观测数据质量、BDS带状控制网解算的影响因素、解算精度、大高差对流层延迟建模以及高程拟合等方面的研究。论文的主要研究成果如下：

（1）利用高原山区铁路沿线的带状稀疏2020年在年积日248-261的7个CORS站观测数据，从星座运行轨迹、可见卫星数、数据可用历元、观测值类型及卫星数、伪距多路径效应以及信噪比等方面对BDS的数据质量进行检核；针对带状控制网解算的多个影响因素，通过实验对比不同的基线解算策略，利用质量指标筛选出适用于铁路带状控制网的优化解算策略，分析BDS在大高差高海拔的高原复杂环境下的解算精度；研究基于CORS的BDS在铁路工程静态、快速静态和网络RTK测量中的精度。结果表明：铁路CORS站BDS观测数据质量较高。在引入IGS站点进行坐标约束时，建议对坐标分量设置 0.05m的约束，观测值类型选择LC_HELP和LC_AUTCLN，并结合实际的地理环境和工程的具体要求加入海潮和大气负荷改正，可提高带状铁路控制网基线解算精度。在铁路带状控制网基线解算精度方面，BDS与GPS精度相当。基于CORS的BDS在铁路工程静态、快速静态和网络RTK测量试验中，精度均符合铁路规范要求，与GPS相当。

（2）针对在高原山区、高差较大的带状区域对流层延迟受高程因素影响较大，常规插值模型难以适应这种区域的问题。利用铁路沿线7个CORS站的天顶对流层延迟(Zenith Tropospheric Delay，ZTD)，将线性内插法(Linear Interpolation Method，LIM)、距离相关线性内插法(Distance Based Linear Interpolation Method，DIM)、最小二乘配置法(Least Squares Collocation，LSC)、克里金插值法(Kriging Interpolation Method，Kriging)应用于该区域进行对流层延迟建模，评估其插值精度；针对大高差因素的影响，构建引入衰减函数的附加高程改正的LSC对流层延迟模型；评估附加高程改正的LSC在网络配置方面的有效性，即基准站的数量、基准站间距和基准站高度分布。结果表明：附加高程改正的LSC模型优于传统的LSC模型和其他模型。处理高程变化明显的区域表现更为出色，能够充分利用对流层延迟的空间相关性和高程相关性，适用于高程变化较大的区域。然而，对于基线较长的网络，由于空间的相关性减弱，插值精度会显著下降。但是，由于空间相关性依然存在，仍可以对基准站高度范围之外的流动站进行外推。

（3）为有效地拟合大高差高海拔地区的地形高程，构建一种基于深度学习的高程拟合方法，该方法使用多层感知器（Multi Layer Perception，MLP）作为核心模型，并根据不同的优化器和激活函数的特点，选择合适的组合，来捕捉大高差高海拔地区的地形特征和高程变化规律，从而实现高精度的高程拟合。同时分析不同的优化器和激活函数组合对模型性能的影响。结果表明：深度学习模型在大高差高海拔地区高程拟合中表现出了较好性能，其MSE最低，MAE最小，R²最接近1，显著优于BP神经网络和遗传算法改进的神经网络方法，且RAdam优化器和GELU激活函数的组合表现较好。

In recent years, railroads have become an important pillar of national economic and social development. The rapid development of China's railroads not only demonstrates the country's engineering strength, but also pushes China's railroad technology to the forefront of the world. Most of the existing railroad construction adopts Global Positioning System (GPS) observation, and lacks the research on the application of BeiDou Navigation Satellite System (BDS) data in railroads. BDS is a global satellite navigation system independently constructed and operated by China, which can provide high-precision, high-quality, and high-quality data. BDS is a global satellite navigation system built independently and operated independently by China, which can provide high-precision and high-reliability navigation, positioning, timing and communication services. Railroad BDS Continuously Operating Reference Station (CORS) provides services for the belt line project, which, combined with the special environmental conditions of high-altitude mountainous areas with large height difference, has a significant impact on the baseline solving accuracy, tropospheric delay estimation, and elevation fitting and other key technologies of the railroad project. Therefore, it is of great significance to utilize the technical advantages of BDS in railroad projects in large height difference and high altitude areas to solve the problem of high-precision and fast positioning of railroads in the western mountainous canyons and plateaus, and to get rid of the dependence on GPS in the railroad industry. The thesis takes a railroad in the mountainous area of western plateau as the research object, and relies on the observation data of the band sparse BDS CORS station constructed along the railroad to carry out the research on the quality of the BDS observation data of the railroad in large height difference and high altitude, the influencing factors of the solution of the BDS band control network, the solution accuracy, the modeling of the tropospheric delay in large height difference, and the fitting of the elevation, and other aspects of the research. The main research results of the thesis are as follows:

(1) Using the observation data from seven CORS stations along the railroads in the plateau mountainous area with band sparse 2020 in the annual cumulative date 248-261, the data quality of BDS is checked in terms of the constellation trajectory, the number of visible satellites, the available calendar elements of the data, the type of observation values and the number of satellites, the pseudorange multipath effect, and the signal-to-noise ratio, etc.; in view of the multiple influencing factors of the solution of the banded control network, the different baseline solution strategies are compared through experiments, and the quality indexes are used to screen the optimized solution strategy applicable to the railway banded control network. Different baseline solution strategies are compared, and the optimized solution strategy for railroad strip control network is screened out by using quality indexes, and the solution accuracy of BDS is analyzed in the complex environment of plateau with large height difference and high altitude; the accuracy of CORS-based BDS in static, fast static and network RTK measurements of railroad engineering is investigated. The results show that the quality of BDS observation data from railroad CORS stations is high. When introducing IGS station coordinate constraints, it is recommended to set 0.05m constraints for coordinate components, select LC_HELP and LC_AUTCLN for observation value types, and add tide and atmospheric load corrections in combination with actual geographic environments and specific requirements of the project, which can improve the baseline solution accuracy of the strip railroad control network. In terms of the baseline solving accuracy of the railroad strip control network, the accuracy of BDS is comparable with that of GPS. In the static, fast static and network RTK measurement test of railroad engineering, the accuracy of CORS-based BDS meets the requirements of railroad specification and is comparable with GPS.

(2) Aiming at the problem that the tropospheric delay is greatly influenced by elevation factors in highland mountainous areas with large altitude differences, it is difficult to adapt the conventional interpolation model to such areas. Using the Zenith Tropospheric Delay (ZTD) from seven CORS stations along the railroad, the linear interpolation method (LIM), distance based linear interpolation method (DIM), least squares collocation (LSC), and LSC are applied to the ZTD. LIM, DIM, LSC, and Kriging are applied to the region for tropospheric delay modeling to evaluate the interpolation accuracy. The LSC tropospheric delay model with additional elevation correction is constructed by introducing the attenuation function for the effect of large height difference; the effectiveness of the LSC with additional elevation correction is evaluated in terms of the network configuration, i.e., the number of reference stations, the spacing of reference stations, and the distribution of the height of reference stations. The results show that the LSC model with additional elevation correction is superior to the traditional LSC model and other models. It performs better in dealing with regions with significant elevation changes, and is able to fully utilize the spatial correlation and elevation correlation of tropospheric delays, which is suitable for regions with large elevation changes. However, for networks with long baselines, the interpolation accuracy decreases significantly due to the weakened spatial correlation. However, since the spatial correlation still exists, it is still possible to extrapolate to mobile stations outside the altitude range of the base station.

(3) In order to effectively fit the terrain elevation in large difference high elevation areas, an elevation fitting method based on deep learning is constructed, which uses the Multi Layer Perception (MLP) as the core model and selects a suitable combination of different optimizers and activation functions according to their characteristics to capture the terrain features and elevation in large-difference high-elevation areas. The MLP is used as the core model, and according to the characteristics of different optimizers and activation functions, a suitable combination is chosen to capture the terrain characteristics and elevation change rules in high-altitude areas with large differences, so as to realize high-precision elevation fitting. At the same time, we analyze the effects of different combinations of optimizers and activation functions on the model performance. The results show that the deep learning model shows better performance in fitting elevation in high altitude areas with large difference, with the lowest MSE, the smallest MAE, and the closest R² to 1, which is significantly better than the BP neural network and the neural network improved by the genetic algorithm, and the combination of the RAdam optimizer and the GELU activation function has the best performance.