Human-Machine Adaptation in Lean Waste Elimination for Customized Building Manufacturing Systems

 The growing demand for customization in the building manufacturing industry has led to more frequent changeover operations, necessitating precise and effective adjustments to numerous control parameters to ensure production quality and efficiency. These adjustments often result in significant production waste, such as product scraps and stoppages. Currently, highly trained engineers manually develop these processes, seeking optimal parameter combinations to achieve satisfactory outcomes on production lines. On the other hand, the limited availability of experimental data and the high-dimensional search space present challenges for existing computer algorithms, impeding the development of accurate and robust learning models. To address these challenges, this study explores human-machine collaboration strategies to reduce the costs associated with complex manufacturing tuning processes. We designed a controlled virtual process game to systematically benchmark the performance of human engineers and computer algorithms, including Reinforcement Learning and Bayesian Optimization, in process control. Our platform captured and analyzed human-system dynamics across various manufacturing scenarios for operators with different backgrounds, providing insights into operator effectiveness and variability. We implemented and compared the performance of human operators and state-of-the-art algorithms, highlighting their strengths and improvement areas. Additionally, this study establishes an end-to-end framework for decoding cognitive behaviors and proposes a context-aware decision-making support method to improve decision quality and operational efficiency. We proposed and evaluated two distinct human-machine collaboration strategies—explicit and implicit—demonstrating the transformative potential of these approaches. Our findings suggest that a hybrid strategy, integrating human expertise with algorithmic capabilities, can improve the performance of manufacturing process control compared to using algorithms alone.