Meta-learning for few-shot sensor self-calibration to increase stress robustness

Abstract

Inertial measurement units (IMU) are able to sense the acceleration and rotation rate of a system and are widely used in mass products like mobile phones and toy drones. These different use-cases have various requirements on the IMU in dependency of the occurring environmental influences. These influences are for example stress effects like temperature, humidity or diverse soldering processes which cause the performance of the sensors to suffer. The initial sensor calibration during the manufacturing of the IMUs is not sufficient anymore and the individual further processing and usage reduces this performance. In this work, an approach that increases the performance and improves the stress robustness against environmental influences is presented. This approach self-calibrates each sensor part individually by using a machine-learning technique called meta-learning. This allows to efficiently adapt a general meta-model to different sensor units to achieve an improved stress robustness with only one data sample. While the quadrature signal of IMUs gyroscope is used to demonstrate the performance and behavior of the chosen approach, this method is applicable for a wide range of various sensor systems to speed up the calibration process and to increase the performance for multiple use-cases.