Filtern
Erscheinungsjahr
- 2025 (3)
Dokumenttyp
- Vortrag (3)
Referierte Publikation
- nein (3)
Organisationseinheit der BAM
Eingeladener Vortrag (wissenschaftliche Konferenzen)
- nein (3)
Operation and maintenance tasks in an industrial context are complex and often time- consuming processes. The manufacturer-provided manuals form the foundation for these processes. However, finding the required information from manuals remains an arduous task. During a maintenance task, the relevant information from plant documentation should be directly available. Along the individual process steps, questions may arise that require information from documentation. To make plant documentation queryable, a Retrieval Augmented Generation (RAG) system was developed. This technique is ideal for cases with little training data, in which fine-tuning of large language models (LLM) is ill-advised. The plant documentation serves as an external knowledge base for answering incoming questions with links to the original documents. Before the RAG development, comprehensive pre-processing of documents takes place, which here is divided into three main blocks: (1) Bill of materials are recognized, page references are stored, and tables are extracted. (2) Technical drawings are also recognized, their page reference is saved, and then passed to a multi-modal LLM (MMLLM), which extracts title and drawing number. (3) The manual is divided into chapters so that each query can be thematically assigned to one or more chapters where appropriate answers are sought. To provide further context, the operation and maintenance tasks are modeled using Business Process Model and Notation (BPMN), an international standard for representation of business processes. As a case study, the maintenance of an industrial pump is used. Along the maintenance process, questions are passed to LLM and RAG to provide concise answers.
Operation and maintenance tasks in an industrial context are complex and often time- consuming processes. The manufacturer-provided manuals form the foundation for these processes. However, finding the required information from manuals remains an arduous task. During a maintenance task, the relevant information from plant documentation should be directly available. Along the individual process steps, questions may arise that require information from documentation. To make plant documentation queryable, a Retrieval Augmented Generation (RAG) system was developed. This technique is ideal for cases with little training data, in which fine-tuning of large language models (LLM) is ill-advised. The plant documentation serves as an external knowledge base for answering incoming questions with links to the original documents. Before the RAG development, comprehensive pre-processing of documents takes place, which here is divided into three main blocks: (1) Bill of materials are recognized, page references are stored, and tables are extracted. (2) Technical drawings are also recognized, their page reference is saved, and then passed to a multi-modal LLM (MMLLM), which extracts title and drawing number. (3) The manual is divided into chapters so that each query can be thematically assigned to one or more chapters where appropriate answers are sought. To provide further context, the operation and maintenance tasks are modeled using Business Process Model and Notation (BPMN), an international standard for representation of business processes. As a case study, the maintenance of an industrial pump is used. Along the maintenance process, questions are passed to LLM and RAG to provide concise answers.
Der Vortrag stellt das Projekt „Handbuch2AR“ vor, das einen KI-gestützten, kontextbasierten Zugriff auf technische Dokumentation für Betriebs-, Wartungs- und Instandsetzungsaufgaben adressiert. Vorgestellt wird ein prototypischer Ansatz auf Basis von Retrieval-Augmented Generation (RAG), der Prozessmodelle der Wartungsaufgabe mit relevanten Inhalten aus Handbüchern verknüpft, um situationsabhängige Informationen bereitzustellen. Anhand mehrerer industrieller Anwendungsfälle wird der aktuelle Entwicklungsstand sowie das geplante Vorgehen zur qualitativen Evaluation der Ergebnisse erläutert.