Accurate Calibration for Magnetic Measurements Using Deep Learning

Abstract

Accurate magnetic measurement is necessary in various applications. However, due to the interferences existing in the magnetic field, the calibration is typically required to correct magnetic measurements. This study introduces a deep learning method, namely, magnetic measurement calibration network (MagMCNet), to calibrate the raw measurements of magnetic sensors. Unlike the conventional methods that are reliant on the predefined measurement error models, our approach adopts deep networks to learn rich calibration parameters to effectively address the nonlinear measurement errors that the existing methods cannot resolve. Given the limited computational resources in practical applications, we integrate two networks in MagMCNet. Specifically, a complex network transfers its prediction capability to a lightweight network through a hybrid regression loss, enabling the real-time calibration. In addition, two new evaluation metrics are introduced for the direct assessment of calibration performance. The proposed approach is rigorously evaluated across simulated, laboratory, and practical application settings. MagMCNet achieves the calibration error of $0.04~\pm ~0.56$ mG in the evaluation with ferromagnetic interferences. Experimental results show that MagMCNet performs consistently better than related methods, suggesting the state-of-the-art performance in complex applications.