AI活用型海上オペレーション：リアルタイム船隊インテリジェンスによる安全・効率・脱炭素化の変革

Unofficial AI-generated summary based on the public title and abstract. Not an official translation.

🔬研究者:RAGを排出・コンプライアンス意思決定に適用する際の評価指標（検索精度・引用カバレッジ・修正率）の設計例として参考になる。

🏢実務担当者:船舶単位の燃料・活動・HSEデータを横断検索可能にし、Scope 3や規制対応の証跡整備と調査工数削減に活用できる。

🏛政策担当者:AI支援による排出・安全データの追跡可能性向上は、海運のMRVや開示品質の底上げ策として注目に値する。

Abstract This paper presents a Retrieval-Augmented Generation (RAG) framework for improving safety, efficiency, compliance visibility, and decarbonization decision support in offshore maritime operations. Vessel daily reports, fuel and lube oil records, machinery hours, HSE observations, defects, inspections, certificates, crew and training data, and emissions indicators are often dispersed across applications, spreadsheets, emails, and document stores. The resulting problem is not lack of data, but difficulty retrieving and interpreting the right evidence when a decision is required. The framework converts verified vessel and company records into an access-controlled maritime knowledge base. For each natural-language question, the system applies validation, metadata filters, hybrid retrieval, reranking, deterministic calculations, grounded generation, source citation, and human verification. Representative questions include abnormal consumption, machinery hours inconsistent with activity, recurring defects, overdue or unverified compliance items, HSE themes, training gaps, and emissions variance. The demonstration indicates that RAG is valuable when it connects records that are normally reviewed separately. Consumption becomes meaningful when combined with activity, weather, machinery configuration, and comparable periods; safety observations become more useful when linked to actions, inspections, and training; and compliance queries can distinguish expiry from an approaching due date or superseded document. The process also reveals inconsistent units, duplicate entries, missing activity codes, and stale records. RAG can therefore provide traceable fleet decision support when it uses approved data, preserves lineage, enforces vessel- and role-based access, and permits a no-answer response where evidence is insufficient. It should reduce manual searching and support earlier intervention, while operational authority remains with competent maritime personnel. The evaluation approach therefore specifies quantitative measures—retrieval precision, citation coverage, calculation accuracy, end-to-end investigation and verification time, manual workflow steps, source-system switches, human correction rate, and user adoption—benchmarked against the same questions completed through existing reports and dashboards. Deployment success also depends on governed source quality, scalable indexing, integration controls, and maintained maritime metadata and taxonomies. Cross-industry controlled and field studies report AI-related productivity or task-time improvements ranging from approximately 15% to 55.8% [11–14]. Given the higher consequence of maritime decisions and mandatory human verification, this paper proposes benchmark-informed initial targets of 25-35% lower end-to-end investigation time and 25-30% lower manual workflow effort, with at least 90% retrieval precision and at least 95% citation coverage. These are acceptance targets for paired trials, not measured fleet outcomes.

• openalex https://doi.org/10.2118/236375-msfirst seen 2026-11-03 05:00:02 · last seen 2026-11-06 05:00:24