Aviation standards are characterized by complex data structures， rigorous semantics， and frequent cross-document references， which pose significant challenges for achieving efficient and accurate knowledge acquisition and question-answering applications. In response， an iterative retrieval-augmented generation （RAG） method for aviation standards based on large language models （LLMs） is proposed. A structure-aware vectorized knowledge base construction and retrieval mechanism is designed， which leverages the hierarchical structure and title chains of standard documents to support context-traceable knowledge representation. A hybrid retrieval strategy combining keyword matching and semantic similarity is further introduced. On this basis， an LLM-driven automatic iterative retrieval and generation mechanism is developed， enabling the model to autonomously determine whether sub-question decomposition and deep intent recognition are required. By integrating multi-turn retrieval with dynamic scheduling， the proposed framework forms a closed-loop process of problem decomposition， information acquisition， self-assessment， and content generation， effectively improving generation quality and semantic coverage for complex queries such as multi-point aggregation and semantic progression. Experiments are conducted on a knowledge base constructed from 7 459 aviation standard documents and a benchmark set of 500 expert-annotated question-answer pairs， evaluated across four mainstream open-source LLMs with varying parameter scales， model types， and bilingual capabilities. The results indicate that， for models with medium or large scales， the proposed method significantly outperforms traditional retrieval-augmented generation approaches in answer accuracy， coverage， and expression quality. Specifically， on DeepSeek-R1-70B， the bilingual evaluation understudy（BLEU） score improved by 27.97% and the simulated human preference score increased by 7.99%； on Qwen-2.5-32B， BLEU improved by 54.67% and the simulated score increased by 8.58%. The proposed method is applicable not only to aviation standards， but also to other structured document scenarios in the aviation field， including airworthiness regulations and maintenance manuals， as well as to domains such as law and healthcare that require high answer quality， reliability， and traceability. It offers a versatile technical framework and implementation approach for the construction of domain-specific question-answering systems.