Abstract: In the field of intelligent question answering for equipment fault diagnosis, traditional entity linking and intent recognition methods exhibit low accuracy, difficulties in disambiguation, and imprecise intent identification when processing long texts. Meanwhile, large-model-based question-answering systems in this domain suffer from limited answer precision and high computational costs, making it challenging to meet the demands for high accuracy and real-time performance. To address these issues, this paper proposes a retrieval-based question-answering framework that enhances retrieval performance by jointly optimizing the entity linking and intent recognition modules. The entity linking module employs a multi-feature fusion approach that integrates textual and semantic features. The intent recognition module incorporates a deep learning model based on global-local semantic collaborative modeling to achieve fine-grained intent analysis. Additionally, an interactive mechanism is designed between the entity linking and intent recognition modules: on one hand, candidate entity disambiguation is performed based on intent semantic features; on the other hand, entity linking feedback is leveraged to refine query intent representation. A question-answering dataset is constructed for multiple fault types and diagnostic scenarios of a publicly available device, followed by extensive comparative and ablation experiments. Experimental results demonstrate that the proposed framework achieves a question-answering accuracy of 0.9607 and an F1 score of 0.9610, outperforming baseline methods. Furthermore, comparison experiments with large-model-based question-answering systems show that the proposed method achieves a BLEU score of 8.7894, a ROUGE-L score of 0.9530, and an average response time of 0.3716 s, verifying its superiority over both foundational large models and large-model-based GraphRAG question-answering systems in terms of accuracy and real-time performance.

Key words: equipment fault diagnosis, retrieval-based question answering, intent recognition, entity linking

摘要： 在设备故障诊断智能问答领域，传统的实体链接和意图识别方法在长文本处理中存在精度低、消歧困难、意图识别不准确等问题。同时，大模型问答系统在该领域存在答案精确度有限、计算开销较高的问题，难以满足高精度、高实时性的应用需求。为此，提出了一种检索式问答算法框架，该框架通过协同优化实体链接与意图识别模块提升检索性能。实体链接模块采用多特征融合方法，整合文本特征与语义特征信息；意图识别模块构建基于全局与局部语义协同建模的深度学习模型，实现细粒度意图解析。设计实体链接与意图识别双模块交互机制：一方面基于意图语义特征实现候选实体消歧，另一方面利用实体链接反馈优化问句意图表征。针对某公开设备的多种故障类型与诊断情境，构建问答数据集，进行了多组对比实验与消融实验。实验表明，该框架的问答精确率为0.960 7，F1值为0.961 0，均优于基准方法。与大模型问答系统的对比实验表明，BLEU得分8.789 4，ROUGE-L得分0.953 0，平均响应时间0.371 6 s，验证了该方法在精确性和实时性方面均优于基座大模型及大模型GraphRAG问答系统。

关键词: 设备故障诊断, 检索式问答, 意图识别, 实体链接