TinyGC-Net: An Extremely Tiny Network for Calibrating MEMS Gyroscopes
Cui Chao, Zhao Jiankang
TL;DR
This work tackles the challenge of calibrating and denoising MEMS gyroscopes in resource-constrained environments. It introduces TinyGC-Net, an ultra-lightweight CNN architecture that separates calibration and denoising into two small subnets, enabling real-time deployment on MCUs while training on GPUs. A physics-informed loss using attitude references over short windows enables training without dataset-specific ground truth, and the method demonstrates competitive orientation accuracy on EuRoC data alongside real-world turntable experiments. The approach achieves robust performance with only hundreds of parameters, does not require accelerometer inputs, and is well-suited for industrial IMU calibration and embedded navigation systems; future work includes temperature drift handling and array-scale deployment.
Abstract
This paper presents a learning-based method for calibrating and denoising microelectromechanical system (MEMS) gyroscopes, which is designed based on a convolutional network, and only contains hundreds of parameters, so the network can be trained on a graphics processing unit (GPU) before being deployed on a microcontroller unit (MCU) with limited computational resources. In this method, the neural network model takes only the raw measurements from the gyroscope as input values, and handles the calibration and noise reduction tasks separately to ensure interpretability. The proposed method is validated on public datasets and real-world experiments, without relying on a specific dataset for training in contrast to existing learning-based methods. The experimental results demonstrate the practicality and effectiveness of the proposed method, suggesting that this technique is a viable candidate for applications that require IMUs.