Clustering Insights in Highly Customized Manufacturing: A Digital Twin Approach with Sensor Fusion and Machine Learning

A digital twin serves as a real-time virtual representation of a physical system, achieved by identifying and capturing critical process parameters through sensor fusion. In a conventional production line, efficiency is typically calculated by comparing actual production data with the production plan. However, in the case of highly customized products, production processes may significantly vary based on product variations, rendering it challenging to predict system behavior during the planning stage. As a result, the actual capacity of the production system will vary with each product variation. A proper product architecture is essential for a digital twin to distinguish individual products. This architecture assigns a unique identity to each product based on specific product variables. This study focuses on sensor fusion in a highly customized windows manufacturing facility. The real-time process variables, captured through sensors, facilitate the creation of a digital value stream map (DVSM) for the assembly line. These process variables, combined with the product architecture, reveal the actual variation in the process corresponding to product differences. Drawing insights from historical data, machine learning techniques will establish connections between process variables captured through sensors and the product architecture. Product clusters will be developed using partitional (k-means) and density-based (Gaussian mixture model) clustering techniques based on product variables to achieve this. The machine learning algorithm should adeptly predict the process variation for any newly introduced product and anticipate and plan for future changes and their impact on the production line.