Data Science Approach to Process Control
42 Pages Posted: 14 Jun 2024
Date Written: May 30, 2024
Abstract
Advancements in manufacturing have led to increased complexity in production processes, rendering many existing Statistical Process Control (SPC) techniques ineffective due to nonstationary and autocorrelated behavior. Researchers have turned to machine learning to address these challenges, but manufacturing data poses obstacles such as concept drift, weak correlations, substantial noise, and endogeneity, which have not been adequately addressed in process control literature. In this paper, we worked with a manufacturing company and developed a simple aggregation technique to mitigate the effects of noise and provide a solution to address endogeneity in their data. Through a bias-variance decomposition, we demonstrate that our aggregation technique trades off latency and granularity in predictions for a reduction in noise. We also empirically validated the presence of endogeneity in the manufacturer's data, supporting our solution for endogeneity. By employing a blend of established Data Science tools and our novel solutions, the manufacturer obtained over 34.5% in cost reductions. To bridge the gaps between the Data Science approach and traditional Process Control literature, we summarize the key points of our process control solution into what we call the ``Predict, Explain, and Act'' (PEA) framework. Our framework helps engineers to adopt, develop and communicate the key ideas in data science for process control solutions that will enable the manufacturer to anticipate, diagnose, and implement targeted pre-emptive remedies to resolve process issues even before the products are assembled.
Keywords: manufacturing, statistical process control, bias-variance, endogeneity, data science, machine learning
Suggested Citation: Suggested Citation