Measurement uncertainty in machine learning - uncertainty propagation and influence on performance

Aug 3, 2026

Uploaded by Finn

Industry 4.0 relies on intelligent networking of machines and processes, making dependable sensor measurements essential for competitiveness and data reliability. Metrology defines internationally accepted measurement units and standards, while the Guide to the Expression of Uncertainty in Measurement (GUM) enables evaluation and interpretation of measurement uncertainty for cross-system comparability. Yet, GUM-consistent uncertainty handling has largely been neglected in machine learning or approximated via cross-validation. This dissertation combines GUM-based uncertainty principles with machine learning.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages180

Reading Time7 h 34 min

AccessFree to download

Abstract

Zusammenfassung

Appended Papers

Paper 1

Paper 2

Paper 3

Paper 4

Measurement uncertainty in machine learning – uncertainty propagation and influence on performance Dissertation zur Erlangung des Grades der Doktorin der Ingenieurswissenschaften der Naturwissenschaftlich-Technischen Fakult¨at der Universit¨at des Saarlandes von Tanja Dorst Saarbr¨ucken 2023 Tag des Kolloquiums: 14 . Juli 2023 Dekan: Prof. Dr. Ludger Santen Berichterstatter: Prof. Dr. Andreas Sch¨utze Prof. Dr.-Ing. Rainer Tutsch Vorsitz: Prof. Dr. Romanus Dyczij-Edlinger Akad. Mitarbeiter: Dr.-Ing. Amine Othmane “But in my opinion, all things in nature occur mathematically.” Ren Descartes Abstract Industry 4.0 is based on the intelligent networking of machines and processes in industry and makes a decisive contribution to increasing competitiveness. For this, reliable measurements of used sensors and sensor systems are essential. Metrology deals with the definition of internationally accepted measurement units and standards. In order to internationally compare measurement results, the Guide to the Expression of Uncertainty in Measurement (GUM) provides the basis for evaluating and interpreting measurement uncertainty. At the same time, measurement uncertainty also provides data quality information, which is important when machine learning is applied in the digitalized factory. However, measurement uncertainty in line with the GUM has been mostly neglected in machine learning or only estimated by cross-validation. Therefore, this dissertation aims to combine measurement uncertainty based on the principles of the GUM and machine learning. For performing machine learning, a data pipeline that fuses raw data from different measurement systems and determines measurement uncertainties from dynamic calibration information is presented. Furthermore, a previously published automated toolbox for machine learning is extended to include uncertainty propagation based on the GUM and its supplements. Using this uncertainty-aware toolbox, the influence of measurement uncertainty on machine learning results is investigated, and approaches to improve these results are discussed. Zusammenfassung Industrie 4.0 basiert auf der intelligenten Vernetzung von Maschinen und Prozessen und trgt zur Steigerung der Wettbewerbsfhigkeit entscheidend bei. Zuverlssige Messungender eingesetzten Sensoren und Sensorsysteme sind dabei unerlsslich. Die Metrologie befasst sich mit der Festlegung international anerkannter Maßeinheiten und Standards. Um Messergebnisse international zu vergleichen, stellt der Guide to the Expression of Uncertainty in Measurement (GUM) die Basis zur Bewertung von Messunsicherheit bereit. Gleichzeitig liefert die Messunsicherheit auch Informationen zur Datenqualitt, welche wiederum wichtig ist, wenn maschinelles Lernen in der digitalisierten Fabrik zur Anwendung kommt. Bisher wurde die Messunsicherheit im Bereich des maschinellen Lernens jedoch meist vernachlssigt oder nur mittels Kreuzvalidierung geschtzt. Ziel dieser Dissertation ist es daher, Messunsicherheit basierend auf dem GUM und maschinelles Lernen zu vereinen. Zur Durchfhrung des maschinellen Lernens wird eine Datenpipeline vorgestellt, welche Rohdaten verschiedener Messsysteme fusioniert und Messunsicherheiten aus dynamischen Kalibrierinformationen bestimmt. Des Weiteren wird eine bereits publizierte automatisierte Toolbox fr maschinelles Lernen um Unsicherheitsfortpflanzungen nach dem GUM erweitert. Unter Verwendung dieser Toolbox werden der Einfluss der Messunsicherheit auf die Ergebnisse des maschinellen Lernensuntersucht und Anstze zur Verbesserung dieser Ergebnisse aufgezeigt. Appended Papers Paper 1 T. Dorst, M. Gruber, B. Seeger, A. P. Vedurmudi, T. Schneider, S. Eichst¨adt, and A. Sch¨utze: Uncertainty-aware data pipeline of calibrated MEMS sensors used for machine learning, Measurement: Sensors (2022) Paper 2 T. Dorst, Y. Robin, S. Eichst¨adt, A. Sch¨utze, and T. Schneider: Influence of synchronization within a sensor netwo

FAQs

Why is measurement uncertainty important for machine learning in Industry 4.0?

Reliable sensor and sensor-system measurements are crucial for trustworthy learning outcomes. Measurement uncertainty also serves as data-quality information when applying machine learning in the digitalized factory.

How does the dissertation incorporate GUM into a machine-learning workflow?

It presents a data pipeline that fuses raw data from different measurement systems and determines measurement uncertainties from dynamic calibration information. It also extends an automated machine-learning toolbox to support GUM-based uncertainty propagation.

What is investigated using the uncertainty-aware toolbox?

The dissertation studies how measurement uncertainty influences machine learning results and discusses approaches to improve those results.

Aug 3, 2026

Uploaded by Finn