| 陈佳琦,盛爽,林嘉曦,李勇乐,彭艳开,许冲.基于知识增强大语言模型的零样本电力设备本体缺陷等级识别方法[J].高技术通讯(中文),2025,35(4):429~439 |
| 基于知识增强大语言模型的零样本电力设备本体缺陷等级识别方法 |
| Zero sample power grid equipment ontology defect grade idetification method based on knowledge enhanced large language model |
| |
| DOI： |
| 中文关键词: 检索增强生成； 大语言模型； 电网设备缺陷检测； 知识库构建； 多阶段推理 |
| 英文关键词: retrieval-augmented generation, large language model(LLM), power grid equipment defect detection, knowledge base construction, multi-stage reasoning |
| 基金项目: |
| 作者 | 单位 | | 陈佳琦 | （国家电网有限公司大数据中心北京 100052） | | 盛爽 | | | 林嘉曦 | | | 李勇乐 | | | 彭艳开 | | | 许冲 | |
|
| 摘要点击次数: 6 |
| 全文下载次数: 4 |
| 中文摘要: |
| 知识检索增强技术通过引入外部知识库有效缓解了大语言模型（large language models，LLM）的幻觉问题与知识滞后性，成为提升领域任务性能的关键范式。本文针对电网设备缺陷等级识别任务中标注样本缺乏、知识利用率低与可解释性不足的问题，提出了一种零样本的知识增强的大语言模型协同推理框架。构建了层次化树状知识库，设计语义最相关的二阶段检索算法提升知识获取效率，并创新性地融合大语言模型先验知识与检索知识进行多阶段推理验证。该方法在218例测试数据上取得54.17%的分类准确率，较无知识检索方法提升了14.26%，同时通过思维链提示生成可验证的解释文本。此外，该方法为零样本，不需要标注数据进行训练。实验结果表明，本方法有效发挥了专业领域知识与通用知识的协同作用，为电力设备缺陷自动检测提供了准确与可解释的解决方案。 |
| 英文摘要: |
| Knowledge retrieval-augmented generation technology effectively alleviates the hallucination and knowledge lag issues of large language models (LLM) by incorporating external knowledge bases, becoming a crucial paradigm for enhancing domain-specific task performance. This paper addresses the problems of lack of labeled samples, low knowledge utilization and insufficient explainability in the task of power grid equipment defect grade identification for LLM, and proposes a zero-shot knowledge-enhanced collaborative reasoning framework for LLM to solve these problems. A hierarchical tree-structured knowledge base is constructed. Then, a two-stage retrieval algorithm is designed. This algorithm focuses on the most semantically relevant information to improve the efficiency of knowledge acquisition. Additionally, the method innovatively integrates the prior knowledge of large language models with retrieved knowledge for multi-stage reasoning and verification. The proposed method achieves a classification accuracy of 54.17% on 218 test samples, which is a 14.26% improvement compared to methods without knowledge retrieval. It also generates verifiable explanatory texts through chain-of-thought prompts. Moreover, this method is zero-shot and does not require labeled data for training. Experimental results demonstrate that the proposed method effectively leverages the synergy between domain-specific and general knowledge, providing an accurate and interpretable solution for the automatic detection of power equipment defects. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
|
|
|
