As the errors of microelectromechanical system (MEMS) gyroscopes are complex and nonlinear, the current calibration methods, which rely on linear models or networks with numerous parameters, are inadequate for low-cost embedded computing platforms to achieve both precision and real-time performance. In this paper, we introduce a extremely tiny network (TGC-Net) that characterizes the measurement model of MEMS gyroscopes. The network has a small number of parameters and can be trained on a central processing unit (CPU) before being deployed on a microcontroller unit (MCU). The TGC-Net leverage the robust data processing capabilities of deep learning to derive a nonlinear measurement model from fragmented gyroscope data. Subsequently, this model is used to regress errors on the gyroscope data. Moreover, we analyze the relationship between the compact network and the traditional linear model for MEMS gyroscopes, and emphasize the significance of the adequate angular motion stimulation for train the network. The experimental results, based on public datasets and real-world scenarios, demonstrate the practicality and effectiveness of the proposed method. These findings suggest that this technique is a viable candidate for applications that require MEMS gyroscopes.

翻译：微机电系统（MEMS）陀螺仪误差具有复杂非线性特征，现有基于线性模型或参数数量庞大的网络标定方法难以在低成本嵌入式计算平台上兼顾精度与实时性能。本文提出一种超小型网络（TGC-Net）用于表征MEMS陀螺仪的测量模型。该网络参数量小，可在中央处理器（CPU）上完成训练并部署至微控制器（MCU）。TGC-Net利用深度学习强大的数据处理能力，从碎片化陀螺仪数据中推导出非线性测量模型，进而实现对陀螺仪数据的误差回归。此外，我们分析了紧凑型网络与传统MEMS陀螺仪线性模型之间的关联，并阐明了足量角运动激励对网络训练的重要意义。基于公开数据集与真实场景的实验结果验证了所提方法的实用性与有效性，表明该技术可作为需要MEMS陀螺仪应用场景的可行候选方案。