Prädiktive Qualität durch Werkzeugmaschinensignale: Effekte der Datenvorbereitung auf Klassifikationsergebnisse maschineller Lernverfahren

Manufacturing enterprises operate within the scope of market pull and technology push. Traditionally, this is characterised by the optimisation triangle of cost -time -quality and the market demand for customised high quality products and short delivery times at a low price. One major technology trend is the introduction of artificial intelligence (AI) applications into social and professional life. While AI applications are perceived to have considerable potential in the manufacturing environment, organisations lack the capabilities to leverage this potential in a goal-oriented manner in practice.This thesis aims at contributing to resolving this challenge by approaching data integration and analysis as a means to an end rather than an end in itself in the context of industrial framework circumstances. The idea is to utilise existing systems and strengths to prevent waste. This requires an understanding of the interdependencies involved in the AI-based analysis process, which are the subject of this study.In the scope of this empirical work, the use case of predictive quality through the analysis of machine tool control data is explored. The objective is to establish the feasibility of predicting quality through machine tool control data and to quantify the required data processing steps to then identify the interaction effects on the model quality. Based on these results, the deliberate redesign of the data processing steps are discussed.For this purpose, the main steps of data preparation for the construction of machine learning models with a focus on time series data are first identified through a literature review. The use case of binary classification of bores based on shape attributes demonstrates the suitability of machine tool control data as explanatory variables for modeling. The interaction effects of the data processing steps on the classification result are identified and quantified by means of statistical analysis.These results are discussed in regard to expense and benefit, integration depth and requirements, along with practical recommendations for addressing the possibilities and constraints of employing machine learning technologies in industrial production.The main practical implication of these findings is the prospect of risk minimisation and strength utilisation through well-designed data processing procedures. Especially since the impacts of prolonged technical trial runs might be conditioned data-based and vice versa.As a result, the findings of this study can be used as a foundation for the development of individual target functions with appropriate prioritisation in accordance with a company's strategic orientation, as well as a procedural guideline for the identification of influencing factors in other fields of application.