【作者】 谭红力；

【导师】 黄新生；

【作者基本信息】 国防科学技术大学 ， 控制科学与工程， 2007， 博士

【摘要】 采用GPS/INS制导方式的航空炸弹成本低、精度高,在现代战争中大量使用。但GPS信号易受干扰,实战中仍需要具备纯惯性制导能力。在保证纯惯性制导精度的条件下,实现低成本是制导组件研制的难点。本文对制导组件的重要子系统:低成本捷联惯导系统展开研究,涉及的关键技术包括低成本硬件设计、惯性测量组件的标定与补偿、非线性滤波方法和传递对准技术。在硬件方面,设计了以DSP和FPGA为核心的一体化弹上计算机。根据协方差分析得到的惯性器件精度指标,选择了低成本挠性陀螺和石英挠性加速度计。并针对所选器件,设计了高速过采样A/D采集电路,输出数据的频率达到200Hz,有效分辨率达到19位,实现了低成本高精度的数据采集。研究了一种利用圆锥误差标定陀螺动态误差(标度因数误差和交叉耦合误差)的方法。利用转台双轴摇摆作圆锥运动,激励陀螺的动态误差产生圆锥误差,圆锥误差会引起姿态漂移;通过测量姿态漂移,来估计并补偿陀螺的动态误差。仿真结果表明,这种标定补偿方法,可将圆锥误差引起的姿态漂移减小一个数量级以上。研究了一种利用神经网络标定和补偿陀螺动态误差不对称性的方法。网络的输入为陀螺输出的角速率、输出为陀螺动态误差的补偿量。标定时只要求转台由静止开始作单轴摇摆,停止后回到初始位置,以导航解算的最终姿态漂移率最小作为目标,训练神经网络。由于最终的姿态漂移不是网络的期望输出,无法采用有导师的训练方法,为此采用了微粒群优化算法。仿真结果表明,这种神经网络补偿方法可将陀螺动态误差的不对称性减小一个数量级。实际实验中,利用我们研制的低成本惯导系统进行不同幅度的摇摆试验,经过神经网络补偿后,姿态漂移率平均降低到0.8°/h以下。低成本惯性测量组件(IMU)受温度影响大,针对这一问题研究了一种基于神经网络的温度补偿算法。网络的输入是温度测量数据。在静态条件下对IMU的输出作滤波抽取,并去除信号初值,得到随温度变化的IMU零偏误差作为期望输出,来训练网络。实验结果表明,神经网络补偿后惯性测量组件零偏变化的幅度降低了60%。低成本捷联惯导系统对准时初始姿态误差的不确定度大,往往不能满足线性化假设条件。为此提出了一种非线性大失准角模型,采用欧拉角表示姿态误差,不对姿态误差作任何小角度假设,可以准确描述惯导系统误差的传播规律。在此基础上设计实现了扩展卡尔曼滤波(EKF)和代表点卡尔曼滤波(SPKF)等非线性滤波算法。针对滤波器在方差更新过程中会出现方差阵负定的问题,采用奇异值分解方法对SPKF算法进行了改进。仿真结果表明,初始姿态误差较大时,基于大失准角模型的非线性滤波算法的对准精度优于传统的线性模型和大方位误差模型。将大失准角模型应用于快速传递对准算法中,并针对姿态观测方程复杂的非线性特性,采用无需求导的SPKF算法。建立了传递对准的仿真环境,通过仿真实验比较了基于大失准角模型的非线性滤波算法和传统线性卡尔曼滤波算法的对准精度。仿真结果表明,基于大失准角模型的非线性滤波算法对准精度优于线性卡尔曼滤波算法,尤其是在安装角较大的情况下,方位对准精度提高了一个数量级,受杆臂误差和陀螺动态误差的影响也比较小。最后利用跑车实验验证了快速传递对准算法,五次实验结果表明对准后位置误差由原来的40米下降到10米以内。更多还原

【Abstract】 The aerial munitions guided by global position system (GPS) and inertial navigation system (INS) have been widely used in modern battles for their low-cost and high accuracy. However, the signal of GPS is easily disturbed and the munitions must have the capacity to maintain the homing accuracy just based on the guidance of INS. The low-cost strapdown inertial navigation system (SINS), which is an important subsystem of guidance assembly, is studied in this dissertation. The key technology of that involved include low-cost hardware design, calibration and compensation of inertial measurement unit (IMU), nonlinear filter and transfer alignment.In hardware aspect, an integrative missile-borne computer is designed using DSP + FPGA structure. The low-cost flex gyroscopes and quartz flex accelerometers are chosen for the IMU based on the anticipative precision which was achieved by the analysis of error covariance. Low-cost and highspeed oversample A/D converter is designed for the IMU. This converter can provide 19-bit accuracy at data rate up to 200Hz.A novel method of calibrating and compensating gyros dynamic error (scale factor error and misalignment error) based on the coning error is proposed. When the coning motion of the shaking apparatus occurs, the coning error induced by gyros dynamic error may cause attitude drifts. Then the gyros dynamic error can be estimated and compensated by measuring the attitude drifts. The results of the simulation experiments show that this compensation method can reduce the attitude error caused by coning error by 90 percent.A method of calibrating and compensating the asymmetry of gyros dynamic errors based on neural network (NN) is presented. The angular rate of gyros output and compensation of gyros dynamic error are the input and output of the neural network respectively. In calibration test, the shaking apparatus was required to do single-axis shake from static, and then stop at the initial position. The terminal attitude drifts were used as the network performance function to train NN. Unlike the supervised training, the terminal attitude drifts were not the target outputs of NN. In this condition, the particle swarm optimization (PSO) algorithm was applied to train the network. The simulation experiment results demonstrate that the asymmetry of gyros dynamic errors reduce to about ten percent of those without the NN compensating. By using the low-cost SINS we designed, the shaking tests of different amplitude are performed. The mean attitude drifts after compensated is less than 0.8°/h.As low-cost IMU is sensitive to temperature, a neural network is designed to compensate the influence of temperature. Temperature measurement is used to be the input of NN. In static condition, the bias of the IMU, which is varied by different temperature and can be gotten by filtering the outputs of IMU, extracting and minus the initial value, is used as the anticipative output to train the NN. The results of the experiments show that the bias of IMU can reduce by 60% compared with those without compensation.The attitude errors of low-cost SINS may be too large to meet the hypothesis of linear model in initial alignment. A nonlinear model for large angle error was deduced to solve this problem. The Euler angles were introduced to present the attitude errors. To achieve accuracy propagation of SINS error model, none little attitude error hypothesis is made. Based on the large misalignment angle model, the extended Kalman filter (EKF) and the sigma-point Kalman filter (SPKF) are designed. Singular value decomposition (SVD) is used to improve SPKF, in the condition that the updating covariance matrix is negative. The simulation experiments results demonstrate that the filter based on the large misalignment angle model has better accuracy than any other traditional filters based on linear model or large heading uncertainty model, under large attitude error of initial alignment.The large misalignment angle model was used in rapid transfer alignment. The SPKF was applied to avoid the derivation calculus of the attitude measurement equations which were with complex nonlinear characteristic. The simulation platform is designed to compare the alignment accuracy of the nonlinear filter based on the large misalignment angle model and conventional Kalman filter. The results of simulation experiments show that the alignment accuracy achieved by the nonlinear Kalman filter base on the large misalignment angle model is higher than that of linear Kalman filter. In large misalignment error condition, the heading alignment accuracy achieved by the nonlinear filter is 10 times higher than that of Kalman filter. The nonlinear filter is not sensitive to the lever arm error and gyros dynamic error. The results of mobile tests validated the alignment algorithm. Five tests results revealed that the position errors were limited up to 10m compared to 40m before alignment.更多还原