Use of AI techniques to extract data from CAD drawings: To create a COW (Crude Oil Washing) manual

Abstract

Crude oil washing (COW) manuals require vessel-specific technical information from drawings, questionnaires and system documentation. Much of this information is available in documents provided by the shipyard, but it is distributed across drawing pages, tables, labels and diagram regions, making manual extraction time-consuming and sensitive to missing or inconsistent source material. This thesis asks to what extent AI-based techniques can support the production of COW manuals from CAD derived drawings and related technical documents with varying formats, layouts and information content. Confidential source documents, complete page images, full OCR output and system-connectivity information were processed locally. Only selected metadata and bounded excerpts assessed as insufficient to reconstruct a drawing or its connection logic were made available to a cloud-based large language model (LLM). The work implemented a multi-stage pipeline for page categorization, text extrac tion, token-level semantic labelling, structural page analysis, target-region detection, retrieval and LLM-assisted draft generation. The source material described in the dataset appendix contained 32 project entries from anonymized Chinese and Ko rean shipyards. The pipeline combined direct PDF text extraction, OCR, TF-IDF page-category scoring, rules, dictionary matching, LayoutLMv3 token classification, YOLO-based page-content and target-region detection, YAML-guided search and an evidence-controlled LLM workflow with local preprocessing. The final page-category model classified 64 of 67 validation pages correctly, corre sponding to 95.5 % accuracy. The final directed-search retrieval reached 96.5 % validation box recall, 94.4 % page recall and 66.5 % candidate precision. In a one project proof-of-concept case study, the workflow generated a draft of the cargo oil system section and an LLM-generated questionnaire. Of the 24 questionnaire fields, 11 were supported by the available evidence and 13 remained unresolved. The draft included the system family, tank arrangement, main cargo oil pump data, major pipe dimensions and stripping-system information, while control, monitoring, heating and several operational fields were omitted or marked for review. The case study demonstrates workflow feasibility but does not estimate generalization to new projects. The results show that AI methods can support parts of COW manual production, particularly evidence location, structuring and preparation of a traceable first draft on the evaluated material. The system is not suitable as an autonomous manual generation workflow, and no conclusion can be drawn about performance on new projects from the available LLM case study. Human engineering review and addi tional structured source material remain necessary for completeness, approval qual ity and operational safety.