Sensor Individual Non-Orthogonality Correction in Low-Cost MEMS Gyroscopes Using Neural Networks

Abstract

The research presented in this work compensates non-orthogonality over temperature stress effects in low-cost open-loop MEMS gyroscopes using neural networks (NN) for a sensor individual compensation to improve the sensor performance. The non-orthogonality is included in the sensor cross-axis sensitivity (CAS) of MEMS gyroscopes. Using the model-agnostic meta-learning algorithm (MAML) as a self-calibration algorithm and one initial measurement after soldering, an individual compensation model is generated for each sensor that predicts the non-orthogonality using the MEMS gyroscope's quadrature value as an input. It will be shown that a sensor-individual model outperforms a compensation model that should fit for all sensors at once like linear regression or classic NN and improves the non-orthogonality by 82.7 %, 7.5 % and 70 % for yx-, zx-, and zy-non-orthogonality,