Retrieval-augmented generative decision support for marine diesel engine maintenance: a domain-adaptive Reorganizer-Retriever-Reranker-Responder framework

Efficient marine diesel engine repair and maintenance (MDE-RM) require accurate, timely, and resource-efficient decision support. However, existing retrieval-augmented generation systems face fragmented knowledge, limited retrieval precision, and difficulties in balancing domain adaptability with deployment efficiency. Therefore, an end-to-end 4D-4R framework is proposed to address these challenges. The framework comprises four key methods: domain-perspective augmentation of the corpus Reorganizer, domain-semantic alignment of the Retriever based on contrastive learning, domain-related rating optimization of the Reranker, and domain-specific instruct tuning of the Responder. Experimental results show substantial improvements: the Retriever's top-1 precision increases from 83.07% to 93.58%. Concurrently, the mean average precision of the Reranker achieves a rise from 92.40% to 94.21%, while the generation quality of the Responder is enhanced, as reflected in the word overlap rate increasing from 58.79 to 61.10. The framework demonstrates high efficiency in resource-constrained environments, providing a reliable paradigm for intelligent decision support of maintenance.