Zeitschriftenartikel
Machine learning in sensor identification for industrial systems
Vorschaubild nicht verfügbar
Volltext URI
Dokumententyp
Text/Journal Article
Zusatzinformation
Datum
2023
Autor:innen
Zeitschriftentitel
ISSN der Zeitschrift
Bandtitel
Verlag
De Gruyter
Zusammenfassung
This paper explores the potential and limitations of machine learning for sensor signal identification in complex industrial systems. The objective is a tool to assist engineers in finding the correct inputs to digital twins and simulations from a set of unlabeled sensor signals. A naive end-to-end machine learning approach is usually not applicable to this task, as it would require many comparable industrial systems to learn from. We present a semi-structured approach that uses observations from the manual classification of time series and combines different algorithms to partition the set of signals into smaller groups of signals that share common characteristics. Using a real-world dataset from several power plants, we evaluate our solution for scaling-invariant measurement identification and functional relationship inference using change-point correlations.