Moritz Berger

Sensirion AG, Stäfa, Zürich, Schweiz

Wednesday, January 31, 2024, 9:15 - 9:55

Building 101, HS 036

This talk introduces a probabilistic approach to calibrate sensors – the Bayesian sensor calibration.

Sensors are designed to measure a physical quantity and to provide a sensor signal allowing to infer it. However, sensor signals often possess so-called cross-sensitivities to other physical quantities than the measurand. Therefore, a sensor needs to be calibrated by measuring the sensor data in an experiment, modeling the sensor’s input-output relation and determine the model parameters by some sort of regression.

Bayesian sensor calibration takes advantage of prior information about a sensor, combines it with new evidence, and infers the measurand from the sensor signals using updated posterior knowledge. One of the key features of the Bayesian method is the possibility to calibrate a sensor with fewer measurements than model parameters while still ensuring a satisfying sensor accuracy. This might allow to save costs and time in the calibration process, especially in the context of large sensor production volumes. Furthermore, Bayesian sensor calibration enables to quantify the uncertainty of a sensor by its confidence interval (CI), which can be interpreted as a measure of trust.