MEMS-IMU error model and calibration method based on LSTM deep neural network

The performance of MEMS inertial sensors is significantly affected by the environment such as heat and mechanics and it is difficult to effectively reduce measurement errors by the traditional polynomial models. An error model based on Long Short Term Memory (LSTM) deep neural network is established to improve the representation ability of MEMS-IMU errors. The input of the model includes angle velocity, acceleration, and temperatures, and the output includes the errors of angle velocity and acceleration. A MEMS-IMU calibration scheme is presented for collecting data and forming the training set that includes the effects of thermal, linear motion and angular motion. The validation experiment is carried out and the proposed method reduces the residual mean value of acceleration and angular rate compensation by about 70% compared with the traditional method and mean square deviations of acceleration and angular rate are decreased by 39% and 64% respectively. The results show the model can be a suitable method for error compensation and the calibration of the MEMS-IMU. © 2020, Editorial Department of Journal of Chinese Inertial Technology. All right reserved.