| 549 | 1 | 98 |
| 下载次数 | 被引频次 | 阅读次数 |
为满足爆破领域知识问答对准确性、时效性与安全合规性的三重需求，提出融合模型微调与安全增强检索的BlastRAG系统。该系统通过自动化知识结构化处理，将《爆破安全规程》等文献转化为高质量数据集与向量知识库；进而采用低秩适配技术对Qwen模型进行领域知识注入。在检索层面，设计了结合混合检索与安全关键度重排器的增强架构，优先提升官方安全规范权重，从根本上保障答案的合规性。实验表明，BlastRAG在检索准确性与合规性上表现优异，其安全增强检索模块准确率达91.8%，综合F1值达89.9%，生成答案的ROUGE指标也有明显提升。模型回答案例和专家评估证实，系统能够生成准确、专业且严格合规的答案，为爆破相关高危垂直领域知识问答提供了可靠的智能化解决方案。
Abstract：To address the triple demands of accuracy,timeliness,and safety compliance in knowledge question-answering for the blasting domain,this paper proposes the BlastRAG system,which integrates model fine-tuning with safety-enhanced retrieval. The system performs automated knowledge structuring to convert literature,such as the Blasting Safety Regulations,into high-quality datasets and a vector knowledge base. Furthermore,LowRank Adaptation(LoRA)technology is employed to inject domain-specific knowledge into the Qwen model through fine-tuning. At the retrieval stage,an enhanced architecture combining hybrid retrieval and a Safety Criticality Re-ranker(SCR) is designed. This mechanism prioritizes official safety regulations by assigning them higher weights,thereby fundamentally ensuring the compliance of generated answers. Experimental results demonstrate that BlastRAG performs effectively in terms of retrieval accuracy and compliance. The safety-enhanced retrieval module achieves an accuracy of 91. 8% and an overall F1 score of 89. 9%,and the ROUGE metrics of the generated answers are also significantly improved. Examples of model responses and expert evaluations further confirm that the system can generate accurate, professional, and strictly compliant answers,providing a reliable intelligent solution for blasting-related high-risk vertical domains.
[1]中华人民共和国国家质量监督检验检疫总局，中国国家标准化管理委员会.爆破安全规程：GB 6722—2014[S].北京：中国标准出版社，2014.
[2]DHAR B K,RAHMAN L,RAMULU M,et al. Blasting for urban construction:a case study from bengaluru metro project[J]. The Indian Mining&Engineering Journal,2022,61(2):5-11.
[3]CHEN L,FISSHA Y,HASANIPANAH M,et al. Accurate prediction of blast-induced ground vibration intensity using optimized machine learning models[J]. Defence Technology,2025,52:32-46.
[4]LI Q,WANG Y,SHAO Y,et al. A comparative study on the most effective machine learning model for blast loading prediction:From GBDT to transformer[J]. Engineering Structures,2023,276:115310.
[5]TAN J. Knowledge graph-based Q&A system for blasting safety management in open-pit mines[J]. Blasting,2025,42(2):188-201.
[6]PAN L,AL-KAMEL A,WANG Y,et al. Unifying large language models and knowledge graphs:a roadmap[EB/OL].(2024-01-30)[2025-08-22]. https：∥arxiv.org/abs/2401.10309.
[7]KANG M,KWAK J M,BAEK J,et al. Knowledge graph-augmented language models for knowledgegrounded dialogue generation[J]. arXiv preprint arXiv:2305,2023:18846.
[8]HOGAN A,BLOMQVIST E,COCHEZ M,et al. Knowledge graphs[J]. ACM Computing Surveys(CSUR),2021,54(4):1-37.
[9]王海峰，李纪为，HUA W，等.预训练语言模型及其应用（英文）[J]. Engineering,2023,25(6):51-65.
[10]HU E J,SHEN Y,WALLIS P,et al. LoRA:Low-rank adaptation of large language models[J]. International Conference on Learning Representations(ICLR),2022,1(2):3.
[11]DING N,XIE Y,ZHANG Y,et al. Parameter-efficient fine-tuning of large-scale pre-trained language models[J]. Nature Machine Intelligence,2023,5(3):220-235.
[12]ZHENG Y,ZHANG R,ZHANG J,et al. Llamafactory:Unified efficient fine-tuning of 100+language models[EB/OL].(2024-03-21)[2025-08-22]. https：∥arxiv.org/abs/2403.13372.
[13]LEWIS P,PÉREZ E,PETRONI A,et al. Retrieval-augmented generation for knowledge-intensive NLP tasks[C]∥Advances in Neural Information Processing Systems 33. 2020:9459-9474.
[14]HUANG L,YU W,MA W,et al. A survey on hallucination in large language models:Principles,taxonomy,challenges,and open questions[J]. ACM Transactions on Information Systems,2024,43(2):1-55.
[15]MAYNEZ J,NARAYAN S,BOHNET B,et al. On faithfulness and factuality in abstractive summarization[EB/OL].(2020-05-01)[2025-08-22]. https：∥arxiv.org/abs/2005.00661.
[16]KARPUKHIN V,OGUZ B,MIN S,et al. Dense passage retrieval for open-domain question answering[C]∥Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing(EMNLP).2020:6769-6781.
[17]AGARWAL L, NASIM A. Comparison and analysis of Large Language Models(LLMs)[EB/OL].(2024-09-02)[2025-08-22]. https：∥ssrn.com/abstract=4939534.
[18]YANG A,YANG B S,HUI B Y,et al. Qwen2 technical report[EB/OL].(2024-06-06)[2025-08-21]. axXiv:2407. 10671v4[cs. CL],2024.
[19]GU X,LIU J,ZHANG Y,et al. A comprehensive approach to instruction tuning for Qwen2.5:Data selection,domain interaction,and training protocols[J]. Computers,2025,14(7):264.
[20]THAKUR N,REIMERS N,RÜCKLÉA,et al. BEIR:A heterogeneous benchmark for zero-shot evaluation of information retrieval models[EB/OL].(2021-10-04)[2025-8-22]. https：∥arxiv. org/abs/2104.08663.
[21]DRORI I,TE’ENI D. Human-in-the-loop AI reviewing:feasibility,opportunities,and risks[J]. Journal of the Association for Information Systems,2024,25(1):98-109.
基本信息:
DOI：10.16389/j.cnki.cn42-1737/n.2026.03.001
中图分类号:TP18;TP391.1;TB41
引用信息:
[1]吕浩源,兰慧,徐子璎,等.模型微调与RAG的爆破领域智能问答系统研究[J].江汉大学学报(自然科学版),2026,54(03):5-14.DOI:10.16389/j.cnki.cn42-1737/n.2026.03.001.
基金信息:
湖北省自然科学基金面上项目（2024AFB1008）; 武汉市科技计划项目（2024050802030155）; 2024年楚天英才计划科技创新团队项目
2026-03-27
2026-03-27
2026-03-27