An intelligent Digital Twin based on machine learning for interpretable decision-making in manufacturing
本文通过行动研究，在Adige Spa公司开发了一个智能数字孪生框架，结合可解释机器学习（进化学习生成决策树）和大语言模型，为生产调度提供透明、可解释的决策支持，并允许用户表达偏好以引导学习过程。
In the context of Industry 4.0, several technologies converge to orchestrate improvements in business performance. Among these, Artificial Intelligence and Digital Twins stand out as some of the most promising. These two technologies are connected through the concept of intelligent Digital Twins (iDTs), which enhance standard Digital Twins with intelligent capabilities while keeping humans at the core of the process. One of the main obstacles to the broad adoption of iDTs in operations and supply chain management is the reliance on opaque AI models, which often limit trust and acceptability among operations experts and managers. To address this, it is critical to design iDTs that not only leverage the advanced capabilities of AI but also provide interpretable and actionable insights to stakeholders. In this paper, we present an action research in Adige Spa to develop an iDT framework for production scheduling. Our framework integrates interpretable machine learning techniques, employing evolutionary learning to produce decision trees that are transparent by design. Additionally, we incorporate Large Language Models to explain decision tree policies in natural language, enhancing user understanding. The framework also facilitates human interaction, allowing users to express preferences and guide the tree learning process. Results in a hybrid flow shop setting demonstrate that the proposed iDT framework delivers interpretable and effective decision-support policies while empowering users to influence and refine its outcomes, hence bridging the gap between AI-driven insights and real-world applicability. • We propose an intelligent Digital Twin for decision-making in production scheduling. • The tool automatically induces decision trees for interpretable decision-making. • Trees are explained in natural language using four different Large Language Models. • Users express preferences on the tree-based policies to guide the learning process.