Supervised Learning Regression for Sensor Calibration

Sensor calibration is of paramount importance for ensuring the accuracy and reliability of multi-sensor systems. This paper focuses on Inertial Measurement Unit (IMU) sensors and investigates the feasibility of framing the calibration problem as a supervised learning task. In a work in process framework, the calibration process is explored, specifically for an acceleration sensor utilizing the gravity constant as a reference. Practical experiments are conducted, comparing traditional calibration using the Gauss-Markov Equation with a supervised learning method based on Descent optimization. By juxtaposing these approaches, we examine the potential advantages of employing machine learning to address data recording errors. Additionally, we share the dataset, encompassing both acceleration and gyroscope sensor data, to facilitate further research in this domain. Our initial findings indicate that learning approaches may not be ideal for intermediate parameter calculation steps compared to traditional methods. However, they exhibit remarkable improvements at the results level, making them suitable for problems directly reliant on the calibrated data. One such application, data-driven inertial localization, is discussed in detail. We conclude that, especially for sensors with unstable calibration parameters, the calibration process is inherently integrated into the learning regression approach.