Classification of material master data using a neural network (Part I)

Allgemeine Informationen

Veranstaltungsart: Studienprojekt
Semester: WiSe 2020/21
ECTS-Punkte: 12
Veranstaltungsnummer: 8.3458
Details: Link zur Veranstaltung in StudIP

Dozent*innen

Prof. Dr. Gordon Pipa

Beschreibung

Purpose – This study project intends to design and implement a neural network, which will carry out the classification of the material master data. Since the ZF Friedrichshafen AG has identified potential improvements in the area of material master data classification, the network will be applied to their processes for the reduction of manual efforts and minimization of wrongly classified master data.
Approach – The first step is an analysis of the existing master data and which parameters can be used for classification purposes. In this context it is particularly important to ensure that these parameters apply to all or almost all master data classes, as ZF procures various materials, which can be fundamentally differ from each other (e.g., electronic control units, functional plastics and castings). Afterwards the theoretical basics are considered in order to discuss exemplarily, which minimum amount of correctly classified master data per class is needed to train a network with an accuracy that must be defined. To conduct a first proof of concept, the prepared concept is validated with an initial trained network for a predefined material class and the required expert knowledge about the classification within an early project gate review. The next steps are the further development of the network regarding all needed material classes and the ongoing optimization of the classification accuracy.
Implementation – The implementation of the network is carried out in the environment of the existing IT landscape of ZF, therefore it is important to ensure that no conflicts arise during the roll-out. In addition to the purely technical implementation, the integration into the respective process landscape is also required. This also serves to prevent manual and potential incorrect classifications as it is currently being done.