Lightweight Neural Network-based Denoising and Calibration Method for MEMS Gyroscopes
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 gy-roscope data for high-precision attitude estimation results.After the raw gyroscope data was denoised and calibrated by the net-work,attitude estimation was performed.The outcomes,along with the true reference attitude,were used to construct a loss func-tion 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 esti-mation.The network is lightweight and has only 180 parameters,making it suitable for embedded system applications.
MEMS gyroscopedeep learningattitude estimationdenoising and calibrationconvolutional neural network
NETL
NSTL
万方数据
