【摘要】 针对在卫星信号拒止、视觉系统退化场景中无人机MEMS-IMU零偏无法准确估计并补偿导致导航误差迅速发散的问题，提出一种基于改进长短时记忆（LSTM）网络的零偏在线标定方法。首先，为解决MEMS-IMU零偏数据非线性强、传统循环时间网络训练效果差的问题，设计序列到序列的LSTM神经网络结构，引入教师强迫机制，提高了网络特征学习能力。然后，在导航过程中使用训练后的网络对MEMS-IMU零偏在线标定，补偿后的IMU量测与视觉信息联合优化，保证了导航定位精度。实验结果表明，在纯惯性导航实验中，所提方法的绝对位置误差比传统LSTM方法减小了6.5%；在EUROC数据集下进行的视觉惯性组合导航实验中，所提方法的平均绝对位置误差比传统LSTM方法减小了15%。更多还原

【Abstract】 Aiming at the problem that the MEMS-IMU bias of unmanned aerial vehicle(UAV) cannot be accurately estimated and compensated in the scenario of satellite signal rejection and vision system degradation, which results in the rapid divergence of navigation error, a bias online calibration method based on improved long short-term memory(LSTM) network is proposed. Firstly, to solve the problem of IMU bias nonlinearity and poor training effect of traditional recurrent neural network, a sequence to sequence LSTM and teacher forcing mechanism is added to improve network feature learning ability. Then, the trained network is used to calibrate MEMS-IMU bias online, and the compensated IMU measurement is optimized jointly with the visual information to ensure the navigation and positioning accuracy in the navigation process. The experimental results show that compared with the traditional LSTM method, the absolute position error of the proposed method is reduced by 6.5% in the pure inertial navigation experiment. The visual inertial integrated navigation experiments is conducted under the EUROC data set subsequence, and the absolute position error of the proposed method is reduced by 15% on average compared with the traditional LSTM method.更多还原