TinyGC-Net: an extremely tiny network for calibrating MEMS gyroscopes

Abstract

Abstract This paper introduces a learning-based calibration method tailored for microelectromechanical system (MEMS) gyroscopes. The proposed method integrates two linear networks, linked by a parametric rectified linear unit (PReLU), and boasts a compacted architecture with only 25 parameters. This simplicity allows for efficient training on a graphics processing unit (GPU) before deployment on resource-constrained microcontroller units (MCUs). The loss function has been carefully devised to strengthen the neural model by eliminating reliance on open-source datasets, and facilitates the swift collection of training data solely via a tri-axial manual rotation table. Furthermore, the proposed method has undergone rigorous validation through public datasets and real-world scenarios, which not only maintains its ultra-lightweight attributes but also outperforms other existing solutions in terms of accuracy. Experimental results demonstrate the method’s practicality and efficacy, indicating its suitability for applications requiring inertial measurement units (IMUs). And the open-source implementation of this method is accessible at: https://github.com/tsuibeyond/TinyGC-Net.git.