Potenzialanalyse zum Einsatz minimalinvasiver Datenerfassung und maschineller Lernverfahren für die Identifizierung von SMC-Prozessschritten

• Potential analysis for the application of minimally invasive data acquisition and machine learning methods for the identification of SMC process steps

Breiing, Markus; Hopmann, Christian (Thesis advisor); Brecher, Christian (Thesis advisor)

Aachen : RWTH Aachen University (2023)
Dissertation / PhD Thesis

Dissertation, Rheinisch-Westfälische Technische Hochschule Aachen, 2023

Abstract

Production digitalization is an important factor in maintaining and increasing the competitiveness of manufacturing companies. In addition, the use of artificial intelligence in production offers further potential for increasing productivity. However, for small and medium-sized enterprises in particular there are many financial, structural and organizational obstacles that impede and delay the use of new, digital technologies. This also affects many companies in the FRP sector, where complex processes are moreover often characterized by manual activities. Thus, there is a high demand for low-cost and low-effort solutions for the acquisition and evaluation of productivity-relevant production data. This can be addressed by minimally invasive concepts, whose benefits have already been demonstrated in different fields of application. Therefore, the aim of this work is the potential analysis for the use of minimally invasive data acquisition as well as low-effort data evaluation by means of machine learning methods for the classification of process steps as the basis for productivity evaluation of the SMC process as a representative example for FRP manufacturing processes. For this purpose, a system for minimally invasive data acquisition by using wireless sensor modules is developed and applied to the SMC manufacturing use case. The minimally invasive production data is pre-processed and then evaluated using different selected models, algorithms and methods of supervised and unsupervised machine learning. The classification of process steps relevant for the productivity of the SMC process is performed using true state labels, and the influence of a basic or more detailed process step definition is also investigated. In addition, the classification is also implemented and analyzed using manually acquired state labels as a low-effort alternative. Overall, the goal of 90% classification performance for unknown transfer data, as declared in the thesis, is achieved for several models. Finally, model variations for the extended evaluation of the potential for the most promising model types in the previous investigations are carried out. For this purpose, selected hyperparameters are adapted exemplarily.