大规模测量平差分布式计算技术及应用研究

发布时间：2018-04-14 16:35

本文选题：分布式计算 + 分块最小二乘平差　；参考：《解放军信息工程大学》2013年硕士论文

【摘要】：信息技术的飞速发展推动现代测量数据处理方式的变革，大规模测量平差的分布式处理为当前规模化的测量数据处理提供一种新型的高性能处理方式，本文对此进行了比较详细的研究与实践，主要研究内容和创新点如下： 1．分析了以大型GNSS网络数据为主的大规模测量数据处理的研究背景和现状，，指出了当前存在的主要挑战，提出了大规模测量平差分布式处理的理论和方法，明确了本文的研究范围和基本内容。 2．研究了适合多核多机分布式计算环境下的分布式计算技术和测量平差计算模型。本文采用多核计算和Web服务技术作为分布式计算环境下的测量平差分布并行计算主要技术途径，从经典法方程层面的最小二乘参数平差计算模型入手，分析了组合解与序贯解的等价性，并讨论了参数固定与参数可变情况下的分块最小二乘平差的组合解法和序贯解法，得出组合解更适合分布式环境下的并行平差计算的结论，并详细阐述了利用求解系数矩阵广义逆进行观测方程层面平差模型直接解算的原理，为平差模型的并行分解奠定基础。 3．提出了测量平差模型分块并行计算的理论方法。通过分析测量平差计算涉及的密集计算任务，利用分块理论建立了测量平差计算涉及的各种矩阵运算的分块并行算法。在此基础上，进一步提出测量平差计算涉及的法方程层面和观测方程层面平差模型的分块并行处理算法，其中法方程层面重点实现了高斯消去法、Cholesky分解法、Jacobi迭代法和共轭梯度迭代法的分块并行算法，观测方程层面利用索引交换序列实现了系数矩阵奇异值分块并行分解，在多核多机的分布式环境下验证了分块并行算法的高效性。 4．将测量平差分块并行计算理论应用到大型GNSS网数据处理中。首先从时间复杂度角度分析了采用非差与双差网解处理GNSS数据的差异，利用分布式计算技术实现了非差PPP的分布式处理和基于子网划分法的双差网解的分布式处理。详细推导了基于“不动点理论”的Ambizap算法的网解新方法，通过PPP的分布式处理和独立基线双差固定解的分布式处理，再利用双差固定解约束PPP解进行整网分布并行平差，既保证了大型GNSS网数据解算精度，又突破了计算规模的限制，具有良好的可移植性和扩展性，为大型GNSS网数据处理提供一种高效、经济的处理方法。 5．以测量平差分块并行计算的理论与方法为基础，编制了以GNSS数据处理为主的大规模测量平差分布式计算软件原型，对基线解算或网平差的SINXE文件进行分布并行融合，采用坐标模式进行网平差的并行计算，验证了测量平差分布并行计算理论与方法的正确性与高效性。
[Abstract]:The rapid development of information technology promotes the transformation of modern measurement data processing. The distributed processing of large-scale measurement adjustment provides a new high performance processing method for the current large-scale measurement data processing.This article has carried on the more detailed research and the practice to this, the main research content and the innovation point are as follows:1.This paper analyzes the research background and present situation of large-scale measurement data processing based on large GNSS network data, points out the main challenges, and puts forward the theory and method of large-scale measurement adjustment distributed processing.The research scope and basic contents of this paper are clarified.2.This paper studies the distributed computing technology and the measurement adjustment calculation model suitable for multi-core and multi-computer distributed computing environment.In this paper, multicore computing and Web service technology are used as the main technical approaches for parallel computing of measurement adjustment distribution in distributed computing environment, starting with the calculation model of least squares parameter adjustment in the level of classical normal equation.This paper analyzes the equivalence between combinatorial solution and sequential solution, and discusses the combinatorial solution and sequential solution of block least square adjustment in the case of fixed parameters and variable parameters. It is concluded that the combined solution is more suitable for parallel adjustment calculation in distributed environment.The principle of using generalized inverse of coefficient matrix to solve the observation equation layer adjustment model is described in detail, which lays a foundation for parallel decomposition of adjustment model.3.A theoretical method for block parallel computation of measurement adjustment model is presented.By analyzing the dense computing tasks involved in the calculation of measurement adjustment, a block parallel algorithm of various matrix operations involved in the calculation of measurement adjustment is established by using the block theory.On this basis, the block parallel processing algorithm of the normal equation level and the observation equation level adjustment model involved in the calculation of measurement adjustment is further proposed.In the normal equation level, the cholesky decomposition method Jacobi iterative algorithm and conjugate gradient iterative algorithm are implemented in detail, and the coefficient matrix singular value block parallel decomposition is realized at the observation equation level using the index exchange sequence.The efficiency of the block parallel algorithm is verified in the distributed environment of multi-core and multi-computers.4.The measurement adjustment block parallel computing theory is applied to the data processing of large GNSS nets.In this paper, the difference between GNSS data processing using non-differential and double-difference net solutions is analyzed from the point of view of time complexity, and the distributed processing of non-differential PPP and double-difference network solutions based on subnet partitioning is realized by distributed computing technology.A new network solution method of Ambizap algorithm based on "fixed point theory" is derived in detail. Through the distributed processing of PPP and the distributed processing of independent baseline double-difference fixed solution, the parallel adjustment of the whole network distribution is carried out by using the PPP solution constrained by the double-difference fixed solution.It not only guarantees the precision of large GNSS net data resolution, but also breaks through the limitation of calculation scale, and has good portability and expansibility. It provides an efficient and economical method for data processing in large GNSS nets.5.Based on the theory and method of parallel computing of measurement adjustment block, the distributed computing software prototype of large-scale measurement adjustment based on GNSS data processing is developed. The SINXE files of baseline solution or network adjustment are distributed and parallel fused.The coordinate model is used for parallel calculation of network adjustment, which verifies the correctness and efficiency of the theory and method of parallel calculation of measurement adjustment distribution.
【学位授予单位】：解放军信息工程大学
【学位级别】：硕士
【学位授予年份】：2013
【分类号】：P207.2

【参考文献】