基于因子图的“热插拔”全源导航机理研究

Most current integrated navigation systems rely on a combination of GNSS and an IMU, and often one or more extra aiding sensors, to mitigate the inevitable drift in single inertial navigation solution. Nonetheless, these systems are optimized with custom filter solutions for their specific sensors and measurement sources operating at different frequencies, leading to point solutions that are inflexible to new capabilities or mission challenges. The quality and robustness of the navigation solution can be significantly enhanced through the use of information extracted from different combinations of aiding sensors, such as, LiDARs, laser rangers, cameras, barometers and magnetometers. Signals of opportunity (SoOps) including cellular, Wi-Fi, ASTC, DVB-T, DAB/ DMB and AM/FM transmissions from the RF background infrastructure and other natural phenomena have also proven useful, and new sensors and measurement sources are being explored constantly. Therefore, it is essential to develop new navigation filtering algorithms, abstraction methods, and an overall navigation system architecture, to accommodate any combination of a large and rapidly expanding array of sensors and measurements, in a plug-and-play fashion. Success in these objectives enables robust positioning and navigation in the face of new conditions and missions, while reducing the cost of bringing new sensors and new capabilities to the users. It is in this regard that a new information fusion approach is explored for the aforementioned ASPN (All Source Positioning and Navigation) system. In this paper a factor graph approach is applied to process all available measurements into a navigation solution. Using factor graphs allows handling different sensors at varying frequencies in a simple and intuitive manner. The factor graph scheme also provides plug-and-play capability, since measurement updates are just additional sources of factors added to the graph and vice versa; no special procedure or coordination is required. The outcome of the proposed research will be (i) a modelling technique for all possible sensors, based on Factor Graph, (ii) an incremental non-linear filter/smoother for the aforementioned model.

全源导航，是一种面向未来新型导航传感器信息融合的创新理念，可实现传感器的热插拔和导航算法的实时调整。尽管在多传感器信息融合方面，已经建立了一套完备的理论且具有多年行之有效的技术方案，但现有的信息融合只针对固定的几种传感器组合，缺乏一种糅合所有传感器的普适信息融合方法；一旦使用新型传感器，必须重新设计算法，且缺乏现场热插拔的能力。本项目研究重点是通过使用因子图理论建立所有已知和在研的导航传感器信号模型，规避现有测量模型固有的不利因素，对导航新信号的融合具有重要意义，为全源导航奠定基础。在此模型基础上，发展增量非线性优化自适应导航算法，进行导航算法的动态调整、结果的计算和优化。最后，通过载体传感器实验验证模型的可信性和算法的可行性。本项目预期研究成果是：（1）基于因子图的普适导航传感器测量模型建模方法，（2）基于增量非线性优化的全源导航算法。

近年来由于导航领域的传感器制造成本的下降及性能的阶段化提升，多传感器信息融合再一次成为了军民应用技术研究的热点。本项目的研究内容以全源导航（All Source Positioning and Navigation, ASPN）为大背景，涉及多传感器信息融合的两个方面：建模和解算。其中，用以因子图为代表的概率图模型（Probabilistic Graphical Models, PGM）被用于解决建模问题；基于非线性最优化的推断/推理（inference/reasoning）方法被用于解决解算中的问题。前者着重于研究建模方法，且主要是状态空间向量和测量向量之间的概率关系，也可以称为描述组合问题的方法；后者着重于研究解决前者留下的模型的输入输出关系，即如何将这一概率关系实例化为各种已知的滤波、预测、平滑或非线性最优化方法。当前的多传感器融合技术的转机在于：（1）已有传感器的性能相对于上一个10年具有至少一个数量级的提升。这其中的代表是MEMS陀螺和芯片级原子钟，及基于卷积神经网络的图像处理技术。（2）之前处于实验室状态或小批量生产的传感器，现在正在或已经开始大批量成体系供应。这其中的代表是LiDAR及民用多频GNSS芯片（大众市场已有L1/L5双频芯片等方案，2017年年中投入量产）。基于以上两点，研究新型传感器融合的理论和方法就势在必行。研究表明，本项目采用的概率图模型方法具有以下2大优势：（i）概率图模型通过其极强的表示属性，能够方便地以条件概率语言描述绝大多数估计问题，并因此将问题的建模和问题的解决分离开来，从而使得（ii）经概率图模型表示的问题，可以几乎不受限制地（当然，得考虑每个问题的实际情况酌情选择）选择最优的估计方法。这两个特性，将使这种方法在全源导航（ASPN）领域具有广泛的应用前景。