Towards cognitive manufacturing: integrating ontologies, digital twins, and large language models for industrial systems

The digital transformation of manufacturing is evolving industrial domains into highly interconnected, data-driven, and intelligent ecosystems. While these advancements offer unprecedented opportunities for operational optimization and informed decision-making, they also pose enduring challenges, including the unification of heterogeneous data, seamless integration of data and industrial knowledge, and the facilitation of flexible, human-centric interaction. To address these challenges, this thesis proposes a unified framework that integrates ontologies, digital twins, and Large Language Models (LLMs), thereby advancing the development of next-generation manufacturing systems that deliver context-enriched, explainable, and actionable insights. The research unfolds in three main stages. First, the Industrial Production workflow (InPro) ontology is developed as a formalized and standardized semantic framework for representing production workflows. This ontology ensures semantic interoperability and seamless data integration across heterogeneous manufacturing data sources. Second, a digital twin-driven industrial context-aware system is proposed. By coupling real-time data streams with structured domain knowledge, the system transforms raw data into high-level insights to support operator decision-making. This framework combines digital twin–based reflections with an ontology-driven semantic layer across external, user, and interface contexts. It is operationalized through an Augmented Reality (AR) interface that delivers personalized, situationally relevant information. This approach offers an endto-end solution that spans perception, integration, reasoning, and visualization within complex manufacturing environments. Third, a domain-specific Cypher query generation pipeline is introduced. It integrates schema-compliant synthetic training data generation, fine-tuning augmented with preference learning, and a structured inference process. This enables accurate and context-aware access to domain knowledge via natural language interactions. By integrating these three strands, the thesis advances industrial systems towards autonomous contextual understanding and decision support. The resulting framework strengthens interoperability, adaptability, and usability, thereby contributing to more cognitive manufacturing operations.