On the Treatment of Measurement Errors for Magnetic Angular Sensors With Neural Networks

The wide-ranging digitization in many industrial sectors is a challenge for established sensor concepts, as new boundary conditions and requirements have to be taken into account. Often, the physical aspects are well-researched and developed, leaving little room for adaptation. Instead, the processing of the sensor signals is often extended, which provides more flexibility due to the large number of potential methods. In particular, the increased number of sensors in combination with machine-learning (ML) methods has proven to be beneficial in many ways. The use of small neural networks working directly with sensory data is a promising approach that can mitigate many unwanted effects. This can be seen as an extended calibration technique, as errors are corrected using knowledge from previous measurements. This work proves these abilities for a small two-layer neural network applied to angular sensing systems. The results are verified by simulations and measurements. Furthermore, the mathematical analysis is used to optimize the training dataset and improve the applicability of neural networks in a sensory context.

I. Introduction

Due to the ongoing digital transformation of many areas, the demand for reliable measurement systems in rough environments is growing rapidly. In particular, topics such as Industry 4.0 [1] or Healthcare 4.0 [2] introduce new measurement problems that extend the requirements for established sensing technologies. Due to unusual or new environments, additional sources of errors that were previously negligible must now be considered [3]. Electrically driven vehicles are a prime example, as electromagnetic distortions have to be reevaluated for many sensing concepts due to the large currents required for electrical drives [4].