To improve the identification efficiency of quality issues and the level of intelligent decision-making in the supervision process of urban rail vehicle manufacturing, a Sparse Retrieval-Augmented Generation (SRAG) method dedicated to the manufacturing supervision domain is proposed. Firstly, 5 836 technical documents and specifications were screened to construct a knowledge base consisting of 412 authoritative documents and a corresponding knowledge graph. Secondly, a total of 12 248 supervision records from 2019 to 2026 were collected, and professional lexicons combined with word segmentation techniques were employed to reveal the temporal patterns of faults, high-frequency failure modes of components, and the coupling relationships between locations and fault types. Finally, the manufacturing supervision corpus and knowledge graph were embedded into Large Language Models (LLMs) to build a sparse-driven knowledge-augmented generation architecture, enabling accurate knowledge invocation and semantically consistent decision-making in complex contexts. The results show that in terms of the semantic similarity metric, the traditional Back Propagation Neural Network (BPNN), Convolutional Neural Network (CNN), and Recurrent Neural Network (RNN) all achieve scores below 0.50, while the T5 model reaches 0.73. After introducing knowledge augmentation, all mainstream LLMs achieve semantic similarity scores exceeding 0.85. Specifically, DeepSeek R1 is improved from 0.92 to 0.96, and ChatGPT-4o is enhanced from 0.88 to 0.97. Compared with vanilla RAG, SRAG also achieves significant improvements in semantic coherence and structural consistency. This method systematically verifies the effectiveness and engineering promotion potential of the sparse retrieval strategy in the context of industrial manufacturing supervision. It helps promote the intelligent, precise and sustainable development of urban rail vehicle manufacturing supervision.
近年来,大语言模型(Large Language Model,LLM)作为1种大数据时代的标志性技术已不断向各个领域发展[3-7]。LLM通常通过大规模无监督预训练、有监督微调以及基于人类价值对齐的强化学习等阶段构建而成,在信息抽取、机器翻译、智能问答等任务中展现出卓越的性能。然而,由于LLM的训练语料主要来源于互联网的公开数据集,对于诸如地铁车辆监造等企业内部的专有知识,其建模能力和泛化效果仍显不足,直接应用往往难以满足工业场景的精确性与可靠性要求[8-10]。基于此,Lewis等[11]提出检索增强生成(Retrieval-Augmented Generation,RAG)方法,通过集成外部知识检索机制,能够在生成过程中动态获取并更新所缺失或遗忘的信息,从而同时提升输出内容的准确性与时效性。与仅依赖内部参数存储知识的传统语言模型相比,RAG具备更强的可控性、解释性以及灵活轻量的部署特性,因此已成为增强语言模型知识能力与生成质量的关键研究方向之一[12]。然而,朴素RAG[13]以固定长度文本块进行向量化存储[14],导致其难以保留监造数据中隐含的语义关联,例如“风机接线盒翻边有毛刺”等口语化质量记录往往与工序、部件编号及处理措施等信息紧密关联,但朴素RAG仅依靠相似度检索难以捕捉此类上下文关系,进而导致生成结果不准确、解释性不足。此外,随着监造数据量激增,固定分块机制还易造成检索延迟与信息丢失,限制了其在复杂工业知识场景下的应用效果[15]。
ZHANGGuizhong, DUYanliang, GAOYang, et al. Development of Urban Service Function of Multi-Level Rail Transit in Cities (Clusters) [J]. Strategic Study of CAE, 2025, 27 (4): 199-209. in Chinese
Urban Passenger Transport. GB/T 38707—2020 Regulation for Operation Technology of Urban Rail Transit [S]. Beijing: Standards Press of China, 2020. in Chinese )
[5]
WANGY T, PANY H, YANM, et al. A Survey on ChatGPT: AI-Generated Contents, Challenges, and Solutions [J]. IEEE Open Journal of the Computer Society, 2023, 4: 280-302.
LINWenguang, SHENFan, XIAORenbin. Research on the Construction Method of a Multi-Modal Patent Data-Driven Professional Large Language Model for Product Innovation Design [J/OL]. Journal of Mechanical Engineering, 2026 (2025-11-26) [2026-04-20]. in Chinese
WANGXiao, ZHANGXiangyu, ZHOURui, et al. An Intelligent Architecture for Cognitive Autonomous Driving Based on Parallel Testing [J]. Acta Automatica Sinica, 2024, 50 (2): 356-371. in Chinese
[10]
FANL L, GUOC, TIANY L, et al. Sora for Foundation Robots with Parallel Intelligence: Three World Models, Three Robotic Systems [J]. Frontiers of Information Technology & Electronic Engineering, 2024, 25 (7): 917-923.
[11]
JINH, LEES, SHINH, et al. Teach AI How to Code: Using Large Language Models as Teachable Agents for Programming Education [C]// Proceedings of the 2024 CHI Conference on Human Factors in Computing Systems. USA: Association for Computing Machinery, 2024: 1-28.
[12]
KANDPALN, DENGH, ROBERTSA, et al. Large Language Models Struggle to Learn Long-Tail Knowledge [C]// Proceedings of the 40th International Conference on Machine Learning. USA: Association for Computing Machinery, 2023: 15696-15707.
CHEWanxiang, DOUZhicheng, FENGYansong, et al. Towards a Comprehensive Understanding of the Impact of Large Language Models on Natural Language Processing: Challenges, Opportunities and Future Directions [J]. Scientia Sinica Informationis, 2023, 53 (9): 1645-1687. in Chinese
TIANYonglin, WANGYutong, WANGXingxia, et al. From Retrieval-Augmented Generation to SAGE: the State of the Art and Prospects [J]. Acta Automatica Sinica, 2025, 51 (6): 1145-1169. in Chinese
[17]
LEWISP, PEREZE, PIKTUSA, et al. Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks [C]// Proceedings of the 34th International Conference on Neural Information Processing Systems. USA: Association for Computing Machinery, 2020: 9459-9474.
YUANLe, LIUShaohua, WANGYu, et al. A Comprehensive Review of Optimization Techniques in Retrieval-Augmented Generation for Large Language Models [J]. Chinese Journal of Computers, 2026, 49 (2): 383-422. in Chinese
[20]
GAOY, XIONGY, GAOX, et al. Retrieval-Augmented Generation for Large Language Models: a Survey [J]. arXiv preprint arXiv:2023.
SONGYuxin, LIULin, WANGHailong, et al. A Survey of Retrieval Granularity in Retrieval-Augmented Generation [J]. Data Analysis and Knowledge Discovery, 2026, 10 (2): 23-40. in Chinese
[23]
吴璇,付涛.检索增强生成技术研究综述[J].计算机工程与应用,2025,61(20):19-35.
[24]
WUXuan, FUTao. Comprehensive Review of Retrieval-Augmented Generation [J]. Computer Engineering and Applications, 2025, 61 (20): 19-35. in Chinese
[25]
YANGL Y, CHENH Y, LIZ, et al. Give Us the Facts: Enhancing Large Language Models with Knowledge Graphs for Fact-Aware Language Modeling [J]. IEEE Transactions on Knowledge and Data Engineering, 2024, 36 (7): 3091-3110.
[26]
WEIJ, TAY Y, BOMMASANIR, et al. Emergent Abilities of Large Language Models [J]. arXiv preprint arXiv:2022.
[27]
XIAY J, CHENY H, ZHAOY X, et al. FCLLM-DT: Enpowering Federated Continual Learning with Large Language Models for Digital-Twin-Based Industrial IoT [J]. IEEE Internet of Things Journal, 2025, 12 (6): 6070-6081.
[28]
DEVLINJ, CHANGM W, LEEK, et al. Bert: Pre-Training of Deep Bidirectional Transformers for Language Understanding [C]// Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. Minnesota: Association for Computational Linguistics, 2019, 1: 4171-4186.
[29]
WELLAWATTEG P, GUOH X, LEDERBAUERM, et al. ChemLit-QA: a Human Evaluated Dataset for Chemistry RAG Tasks [J]. Machine Learning (Science and Technology), 2025, 6 (2): 020601.
[30]
CUCONASUF, TRAPPOLINIG, SICILIANOF, et al. The Power of Noise: Redefining Retrieval for RAG Systems [C]// Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval. USA: Association for Computing Machinery, 2024: 719-729.
[31]
PICHAPPANP. Adaptive and Enhanced Retrieval Augmented Generation (RAG) Systems: a Summarised Survey [J]. Journal of Digital Information Management, 2025, 23 (3): 190-195.
[32]
DAVOODIJAME, ALAMBARDAR MEYBODIM. Evaluation Metrics on Text Summarization: Comprehensive Survey [J]. Knowledge and Information Systems, 2024, 66 (12): 7717-7738.
[33]
DADAI D, AKINWALEA T, TUNDE-ADELEKET. A Structured Dataset for Automated Grading: from Raw Data to Processed Dataset [J]. Data, 2025, 10 (6): 87.
[34]
ZHANGT, KISHOREV, WUF, et al. Bertscore: Evaluating Text Generation with BERT [C]// ICLR 2020 Confernce and Blind Submission. USA: OpenReview, 2025.
LINC Y. ROUGE: a Package for Automatic Evaluation of Summaries [C]// Text Summarization Branches out. USA: Association for Computational Linguistics, 2004: 74-81.
[37]
LUANY D, LINS F. Research on Text Classification Based on CNN and LSTM [C]// 2019 IEEE International Conference on Artificial Intelligence and Computer Applications (ICAICA). New York: IEEE Press, 2019: 352-355.
[38]
RAFFELC, SHAZEERN, ROBERTSA, et al. Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer [J]. Jornal of Machine Learning Research, 2020, 21 (140): 1-67.
[39]
GUOD, YANGD, ZHANGH, et al. DeepSeek-R1: Incentivizing Reasoning Capability in LLMS via Reinforcement Learning [J]. arXiv preprint arXiv: 2025.
[40]
ISLAMR, MOUSHIO M. GPT-4o: the Cutting-Edge Advancement in Multimodal LLM [C]// Intelligent Computing-Proceedings of the Computing Conference. Cham: Springer, 2025: 47-60.
[41]
COMANICIG, BIEBERE, SCHAEKERMANNM, et al. Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities [J]. arXiv preprint arXiv:2025.