用于磁传感器复杂误差补偿的深度模型训练与在线修正方法研究

Geomagnetic navigation and orientation are of great significance to military and civilian, high precision magnetic sensor is its core technology. Considering its sensitivity to external interferences, the measurement accuracy of the magnetic sensor is greatly affected by the error sources such as time varying current, soft magnetism and hard magnetism materials. So the positioning precision mostly depends on the calibration of the magnetic sensor. The traditional method of magnetic sensor calibration is to establish an explicit error model, but the problem is that it is difficult to describe all of the interferences in the explicit model. To overcome this problem, an implicit error model with deep architecture is proposed to improve the capability of nonlinear expression. And novel training methods and on line correction algorithms are developed for magnetic sensor calibration. Combining the two kinds of academic idea of extreme learning and deep learning, an implicit error model with deep architecture based on extreme learning machine (ELM) is established according to the characteristic of the magnetic sensor. And a novel training method that step by step training without iteration combining with reversed tuning, is developed to be used in this model. The thought of Kalman filter is applied in the on line correction of the calibration model, and the similar Kalman factors is designed according to the changing characteristics of the magnetic sensor error model. Nonmagnetic turntable and rotating coil are utilized in experiments which evaluate the compensation result and then gradually optimize the error model. The purpose of this project is to solve the problem of low compensation accuracy about complex nonlinear error by machine learning method. Moreover, a new idea of compensation for sensors modeling difficulty is available.

地磁导航定向具有显著的军事意义和民用价值，高精度磁传感器是其核心支撑技术。由于磁传感器易受周围变化的电流、软磁和硬磁材料等干扰源影响，误差来源众多，因此误差补偿和修正是决定地磁导航定位成败的关键问题之一。目前解决该问题主要采用显式误差模型，然而显式误差模型存在难以囊括所有干扰因素等问题。为解决这个问题，本项目拟建立深层隐式误差模型，提高非线性表达能力，并探索新的训练方法和在线修正算法。具体包括：融合超限学习和深度学习学术思想，针对磁传感器样本特点，建立基于超限学习机的深层隐式误差模型，研究采用“逐层无迭代训练和反向调优相结合”的训练方法；根据磁传感器误差模型变化特点，设计类卡尔曼因子，借鉴卡尔曼滤波思想对模型进行在线修正；通过“无磁转台”和“旋转线圈”实验，评估误差模型补偿效果，并逐步优化。本项目旨在利用机器学习方法解决复杂非线性误差补偿精度低的难题，为建模困难的传感器误差补偿探索新思路。

地磁导航定向具有显著的军事意义和民用价值，高精度磁传感器是其核心支撑技术。目前复杂磁传感器误差补偿主要采用显式误差模型，然而这类模型存在难以囊括所有干扰因素等问题。为解决这个问题，本项目构建了深层隐式误差模型，并研究其在线修正算法。主要研究内容和成果如下：..1、研制三轴磁传感器，采用各向异性磁阻传感器作为敏感元件，以基于ARM Cortex-M3内核的STM32F103芯片为处理器，作为地磁场数据采集的硬件基础。..2、构建了基于超限学习机的浅层和深层误差模型，并研究自主反向调优的训练方法，提高了误差模型非线性表达能力和训练精度。..3、磁传感器误差模型在长期使用过程中可能发生变化，但不同于常规在线修正又不会经常变化。借鉴卡尔曼滤波思想，我们研究了基于类卡尔曼因子的深层隐式误差模型在线修正算法，保证磁传感器误差情况发生变化时能及时发现并修正。..4、研究了车辆航向估计的数据融合算法，陀螺仪不易受环境干扰但长时间使用会积累误差，然而磁罗盘容易受外界电磁干扰但不存在积累误差，通过数据融合提高了车辆航向估计的抗干扰能力。..5、基于地磁场与重力场夹角不变以及修正角最小化原则给出了修正计算式,对磁传感器输出向量进行修正并，恢复地磁场与重力场所测姿态的一致性。..6、充分利用地磁场幅值不变性和方向不变性，研究了基于两阶段的粒子群标定算法，每个阶段分别设计粒子群算法的适应度函数。该方法具有更高的补偿精度，而且无需额外航向参考。..本项目的研究不仅提高了三轴磁传感器误差补偿精度，而且对其使用过程中可能发生的误差变化情况进行及时修正，为提高复杂非线性误差补偿精度提供了新的途径，所做研究工作具有重要的科学意义和实际使用价值。