Session: 03-04-02: Artificial Intelligent Applications in Manufacturing II

Paper Number: 167119

Generative AI-Driven Disruption Management in Matrix Production: A Digital Twin-Based Framework for Resilient Operations Towards Industry 5.0 

Manufacturing disruptions—including unexpected machine failures, human errors, and supply chain interruptions—significantly impact production efficiency, leading to unplanned downtime, reduced throughput, and substantial financial losses. A 2024 industry report by Siemens estimates that unscheduled downtime among the world’s 500 largest companies resulted in nearly $1.4 trillion in lost revenue, underscoring the urgent need for more adaptive and resilient manufacturing systems. However, traditional disruption management primarily relies on human expertise, where operators diagnose and resolve issues based on prior experience and domain knowledge. This manual approach requires processing vast amounts of data to formulate recovery strategies, a task that is not only time-consuming but also prone to suboptimal decision-making. As a result, human-driven methods often lead to inefficiencies, extended recovery times, and reduced manufacturing resilience.

This research explores the integration of Generative AI (GenAI)—particularly Large Language Models (LLMs) such as OpenAI’s GPT and Google’s Gemini—with the flexibility of matrix production systems to enhance manufacturing resilience. Unlike traditional linear production lines, where workflows follow a fixed sequence and are vulnerable to single points of failure, matrix production adopts a modular structure with interconnected workstations, allowing for real-time job reallocation, operation rescheduling, and parallel processing. This setup inherently improves system adaptability by redistributing workloads and maintaining production flow even when disruptions occur.

Additionally, Automated Guided Vehicles (AGVs) play a crucial role in matrix production by facilitating dynamic material transport between machines based on real-time production conditions. Unlike conventional conveyor-based systems, AGVs can reroute adaptively, minimizing delays and enhancing production efficiency in disruption scenarios. However, while AGVs provide flexibility, they also introduce significant decision complexity, as their routing is tightly coupled with machine scheduling. Poor coordination between AGV routing and production scheduling can lead to bottlenecks, increased waiting time, and suboptimal system performance. Thus, a key challenge is optimizing AGV dispatching and machine scheduling simultaneously to maximize system responsiveness to disruptions and overall resilience.

Furthermore, although autonomy is growingly adopted in modern manufacturing, human supervision remains essential to ensure decision transparency, reliability, and alignment with operational goals. Integrating human-in-the-loop mechanisms within GenAI-powered systems is critical for harmonize the machine autonomy with human expertise, fostering trust, and enabling effective collaboration. This research addresses the challenge of aligning AI-driven recommendations with human decision-making preferences, ensuring practical implementation in real-world manufacturing environments.

By integrating GenAI-enabled AGV routing and operation scheduling, the proposed framework aims to enhance adaptive logistics, mitigate disruptions, and maintain workflow continuity. This decentralized approach distributes workloads across multiple pathways, enabling production to continue despite machine failures or resource shortages through swift, explainable reallocation of jobs and materials. Unlike rule-based AI models, which rely on predefined conditions, GenAI synthesizes insights from historical data, recognizes disruption patterns, and generates adaptive recovery strategies in real time. The proposed framework provides real-time decision support communicated in a natural way, equipping workers with AI-generated recommendations to optimize response strategies in disruption scenarios.

To validate the effectiveness of this approach, a mathematical model is developed to simulate various disruption scenarios for a semiconductor manufacturer, capturing key performance metrics such as overall equipment effectiveness (OEE), system downtime, and recovery time. The structured knowledge base, built from historical disruption cases and expert decision strategies, is then used to train a fine-tuned LLM, enabling GenAI-driven decision recommendations for disruption management. The model provides context-aware, intuitive decision suggestions to operators, reducing reliance on trial-and-error troubleshooting and significantly improving response efficiency. Our preliminary results demonstrate that GenAI-generated recommendations can effectively assist workers in selecting optimal response strategies based on real-time disruption information and human inquiries.

Presenting Author: Yanchao Tan Purdue University

Presenting Author Biography: Yanchao Tan is a Ph.D. student in the School of Engineering Technology at Purdue University, West Lafayette. She holds two master's degrees: one in Computer Science from Wayne State University, Detroit, and another in Systems Engineering and Design from the University of Michigan, Ann Arbor.

Her research focuses on the development of advanced data-driven and machine learning models, with an emphasis on manufacturing digitalization, human-machine interaction, intelligent decision-making, lifecycle analysis, and the societal impacts of technology.

Authors:

Yanchao Tan Purdue University
Baicun Wang Zhejiang University
Wei Yuan Hitachi America, Ltd.
Quan Zhou Hitachi America, Ltd.
Yingxue He Zhejiang University
Wei Ye Purdue University
Marvin May Massachusetts Institute of Technology
Xingyu Li Purdue University
Sang-Gook Kim Massachusetts Institute of Technology
Ragu Athinarayanan Purdue University