A reference architecture for the operationalization of machine learning models in manufacturing
Tim Raffin, Tobias Reichenstein, Jonas Werner, Alexander Kühl, Jörg Franke
Abstract
Inherent characteristics of machine learning algorithms such as their probabilistic nature, their reliance on large datasets as well as their need for constant retraining pose major challenges to the operationalization of machine learning models (MLOps) in the manufacturing domain. As such systems are known to quickly accumulate technical debt due to system-level interdependencies of code, data, and models, clear abstractions boundaries are mandatory. Therefore, this publication derives a systematic functional decomposition of an MLOps system tailored to the manufacturing industry into specific domains and contexts. Moreover, a concrete deployment view is provided, and possible future research directions are discussed.