Revamping Manufacturing Systems: Utilization of Data Driven Models, Interpretable Machine Learning, and Data-Product Stakeholder Flow Analysis

Abstract

In the manufacturing environment, high volume of data can be easily generated. However, to provide valuable insight, the right tools, medium, and communication flow within stakeholders are crucial. This thesis presents a comprehensive exploration of developing data products in the manufacturing sector. It includes modeling industrial coffee roaster systems, improving the interpretability of machine learning models, and analyzing stakeholder flow to develop effective manufacturing data products.

The first study involves modeling an industrial coffee roaster system. Using production data collected during the roasting process and multiple experiments, an 11-stacked long short-term memory (LSTM) neural network was developed and trained to model the dynamics of the industrial coffee roaster plant. The model was validated, and an initial closed-loop system was developed in MATLAB to further validate the model.

The second study focused on improving the interpretability of machine learning models in the Semiconductor Fab. The SHapley Additive exPlanations (SHAP) methodology was applied to generate beeswarm and bar plots for the SHAP results, which identified the most important features to improve the throughput prediction. The study showed that Machine E utilization has a significant influence on the throughput prediction.

Finally, the third study involved conducting a qualitative analysis of stakeholder flow in developing data products for manufacturing cases using interview methods and design structure matrix (DSM). The study found that the stakeholder flow varied based on the resources and stage of an organization, with the interaction with end-users being the main driver of the flow. The study also highlighted the importance of identifying and managing stakeholders with the highest coupled interaction for the development of data products.