Einsatz Maschinellen Lernens zur datengetriebenen Synthese und Analyse von Produkteigenschaften im Kontext verknüpfter Produktdaten

Description

This thesis proposes a data-driven approach for the analysis and synthesis of product characteristics, drawing upon the WEBER model of product characteristics and properties. In the context of change and adaptation design, product properties and characteristics are conceptualized as variables within a machine learning model. The generation of a machine learning model follows a systematic sequence of steps, beginning with data preparation, modeling, evaluation, and optimization, as developed from the CRISP-DM. This systematic approach enables the application of data-driven analysis and synthesis. Ensuring consistent data management throughout the entire modeling process is imperative. Tothisend, aconcept fordata-drivenanalysisand synthesis has been developed, which is presented in the context of the required data and categorized accordingly. The objective of this work is to enhance three areas: metamodeling, solution space generation, and data management. The integration of metamodeling with global and local component properties through deep learning (data-driven analysis) facilitates the generation of solution spaces for novel component designs in data-driven synthesis. The management of data and the integration of semantics through metadata are integral components of the respective modeling processes. In summary, it can be concluded that deep learning in particular facilitates the generation of novel metamodels that enable, among other capabilities, a spatially resolved description with sufficient accuracy. Furthermore, the presented approaches demonstrate novel methodologies for leveraging solution spaces to expedite the generation of applicable component designs. The principles and methodologies outlined in this thesis are directly applicable to the development process.