Interchangeability of neural networks in different industry settings: an approach of deriving a general framework for the implementation of predictive machine actions.

Abstract

The aim of this master thesis is to propose a general framework that facilitates the implementation of neural networks in the manufacturing industry and beyond. The process steps described in this framework provide a structured approach and serve as a guide if neural networks are considered to be implemented in an industrial setting. Furthermore, having and following this approach not only saves time as there is a risk of overlooking certain steps, which then need to be reconsidered, but can also increase the accuracy of the model itself. Moreover, a key point of this thesis is the question, when and for which problems state-of-theart artificial neural networks can be applied in manufacturing. To answer this, first an extensive review of recent academic literature was conducted. Secondly, an analysis was performed, which includes the qualitative assessment of two recent use-cases of neural networks used in a manufacturing context. Based on the findings of the literature review as well as the qualitative assessment a general framework was derived. In order to prove the usability of this framework was used to self-develop a neural network application solving an existent problem in supply chain management and procurement, specifically the prediction of whether a delivery of goods will be on time or late. Subsequently, the performance of the network and its prediction accuracy were validated. Indeed, it was observed that neural networks are most suitable for solving regression and classification problems, whereas for manufacturing the focus is on classification problems. In addition, a general framework was derived which was also used to self-develop a neural network. This neural network was not only able to solve the underlying problem of predicting and classifying on time or late deliveries of goods, but also achieved a performance measured by binary accuracy of 0.82, a validation accuracy for on time deliveries of 0.82 and a validation accuracy for late deliveries of 0.72.