A Cognitive Digital Twin for Industry 5.0 Based on a Large Language Model Agent

While digital twins have made significant strides in creating digital replicas of physical manufacturing systems, their cognitive capabilities remain inadequate to address the dynamic complexities in manufacturing, including process variations, environmental fluctuations, and human interactions. These limitations hinder their applicability in Industry 5.0, which emphasizes Self-X cognitive capabilities. This work proposes an improved five-dimensional framework for developing a cognitive digital twin (CDT) that adopts a Large Language Model (LLM) agent at its core. The LLM agent enhances domain-specific tasksolving capabilities through retrieval-augmented generation (RAG) and in-context learning. RAG compensates for the general LLM's limitations by utilizing external tool libraries and industrial knowledge graphs to establish context awareness, retrieve domain-specific knowledge, and convert human commands into sequential task plans via function calls. Incontext learning further enables the LLM agent to learn specific tasks based on contextual examples without retraining. It empowers CDT to address domain-specific challenges with efficiency, flexibility, and cost-effectiveness. The effectiveness of the proposed CDT is demonstrated in a lab-scale manufacturing unit, highlighting its ability to perform valid task planning and handle dynamic incidents, paving the way for more resilient manufacturing systems aligned with Industry 5.0 objectives.