基于空洞卷积神经网络的姿态估计算法

Attitude estimation algorithm based on hole CNN

针对姿态估计中存在微机电系统(MEMS)陀螺仪漂移和误差发散的问题,提出了一种基于空洞卷积神经网络(CNN)补偿微惯性测量单元(MIMU)误差的姿态估计算法。该算法利用空洞卷积神经网络学习MIMU提供的加速度和角速度数据来补偿陀螺仪的误差,使用基于四元数的四阶龙格库塔法求解误差补偿后陀螺仪的角度。将修正后的陀螺仪数据作为预测量;加速度计和磁力计数据作为观测量,通过扩展卡尔曼滤波实现多传感器融合准确估计姿态。利用MPU9250模块进行算法对比实验,结果表明:本算法可有效降低陀螺仪随机漂移,稳定输出高精度姿态。

Aiming at the problem of micro-electro-mechanical system(MEMS)gyroscope drift and error divergence in attitude estimation,an attitude estimation algorithm based on cavity convolution neural network(CNN)is proposed to compensate the error of micro inertial measurement unit(MIMU).The algorithm uses the hole CNN to learn the acceleration and angular velocity data provided by MIMU to compensate the gyroscope error,and uses the four elements four-order Runge-Kutta method to solve the angle of gyroscope after error compensation.The modified gyroscope data is used as the predictor,the accelerometer and magnetometer data are used as the observation measurements,and the multi-sensor fusion is realized through extended Kalman filtering(EKF)to accurately estimate the attitude.MPU9250 module is used for algorithm comparison experiment.The results show that the algorithm can effectively reduce the random drift of the gyroscope and stabilize the high precision attitude.