基于轻量神经网络的MEMS陀螺仪降噪与标定方法

Lightweight Neural Network-based Denoising and Calibration Method for MEMS Gyroscopes

针对MEMS陀螺仪测量模型中时变、非线性误差和高频噪声引起的姿态估计精度低及易发散的问题,提出一种基于深度学习的陀螺仪降噪与标定方法。对陀螺仪测量误差进行建模,采用卷积神经网络(CNN)从陀螺仪历史数据中提取误差模型特征,实现对陀螺仪数据实时降噪与标定,获得高精度姿态估计结果。原始陀螺仪数据经过网络降噪和标定后进行姿态估计,并将结果与参考姿态真值构建损失函数训练网络。在EuRoC导航数据集上的实验结果表明:与基于循环神经网络的方法和直接使用原始陀螺仪数据进行的姿态估计相比,基于CNN的方法误差分别降低了55.9%和96.4%,有效降低陀螺仪误差与噪声并提高姿态估计精度。网络轻量,参数仅有180个,适合嵌入式系统的应用。

To address the issues of low accuracy and divergence in attitude estimation caused by time-varying,nonlinear errors,and high-frequency noise in MEMS gyroscope measurement models,a gyroscope denoising and calibration method based on deep learning was proposed.A model for gyroscope measurement errors was developed,utilizing a convolutional neural network(CNN)to extract error model features from historical gyroscope data,thereby achieving real-time denoising and calibration of gyroscope data for high-precision attitude estimation results.After the raw gyroscope data was denoised and calibrated by the network,attitude estimation was performed.The outcomes,along with the true reference attitude,were used to construct a loss function for training the network.Experimental results on the EuRoC navigation dataset indicate that the CNN-based method achieves error reductions of 55.9%and 96.4%compared to methods based on recurrent neural networks and direct use of raw gyroscope data,respectively.The CNN network effectively reduces gyroscope errors and noise,thus enhancing the precision of attitude estimation.The network is lightweight and has only 180 parameters,making it suitable for embedded system applications.