Scholarlink

<p>In smart manufacturing, rapidly and accurately retrieving knowledge from technical documents with high interpretability is critical for equipment maintenance and production decision-making. However, given the strict demands for rigor and interpretability in equipment maintenance question-answering (QA), conventional Retrieval Augmented Generation (RAG) approaches often suffer from limited transparency and hallucinations. To tackle this issue, we propose ERGR: a Knowledge-Enhanced LLM for Explainable Reinforced Graph Reasoning in equipment maintenance QA, designed to support equipment maintenance. In terms of model design, we first construct an equipment maintenance ontology based on patents, and employ a LLM to extract entity–relation–attribute and generate an equipment maintenance knowledge graph. Concurrently, we utilize LLM to generate question-level semantic understanding and reward signals. We then utilize Graph Attention Networks (GAT) to capture the local structural semantics of entity nodes. By integrating question intent, entity attributes, neighborhood context, and LLM-guided semantic alignment into the attention mechanism, nodes with semantically unreasonable meanings are effectively filtered out. In addition, we adopt a gated residual strategy to dynamically balance the original entity semantics and aggregated neighborhood information, thereby enhancing the robustness of the representation. Furthermore, we develop an Actor–Critic policy network that incorporates structural embeddings, query semantics, and LLM-derived semantic scores to enable controllable and interpretable reasoning for equipment maintenance QA. Overall, ERGR proposes a semantic alignment-driven reinforced graph reasoning paradigm, which unifies the modeling of question semantics, relation type constraints, and graph structure representation. It achieves dynamic optimization and interpretable path control of the multi-hop reasoning process through reinforcement learning. Finally, experimental results demonstrate that ERGR significantly outperforms various baseline methods on a real-world equipment maintenance question-answering dataset. In the comparison with RAG-enhanced large language models, compared with Qwen3-235B, H@1 is improved from 67.50% to 83.33%. In the comparison with graph neural network methods, compared with GraphSAGE, H@1 is increased by 7.69%.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

An explainable question answering system for equipment maintenance with knowledge-enhanced LLM

  • 期刊论文

  • 发表日期：

    2026年6月17日

查看全文
  • Yuexiang Yang,
  • Wenzhe Mu,
  • Yubo Chen,
  • Xiantao Fang

摘要

In smart manufacturing, rapidly and accurately retrieving knowledge from technical documents with high interpretability is critical for equipment maintenance and production decision-making. However, given the strict demands for rigor and interpretability in equipment maintenance question-answering (QA), conventional Retrieval Augmented Generation (RAG) approaches often suffer from limited transparency and hallucinations. To tackle this issue, we propose ERGR: a Knowledge-Enhanced LLM for Explainable Reinforced Graph Reasoning in equipment maintenance QA, designed to support equipment maintenance. In terms of model design, we first construct an equipment maintenance ontology based on patents, and employ a LLM to extract entity–relation–attribute and generate an equipment maintenance knowledge graph. Concurrently, we utilize LLM to generate question-level semantic understanding and reward signals. We then utilize Graph Attention Networks (GAT) to capture the local structural semantics of entity nodes. By integrating question intent, entity attributes, neighborhood context, and LLM-guided semantic alignment into the attention mechanism, nodes with semantically unreasonable meanings are effectively filtered out. In addition, we adopt a gated residual strategy to dynamically balance the original entity semantics and aggregated neighborhood information, thereby enhancing the robustness of the representation. Furthermore, we develop an Actor–Critic policy network that incorporates structural embeddings, query semantics, and LLM-derived semantic scores to enable controllable and interpretable reasoning for equipment maintenance QA. Overall, ERGR proposes a semantic alignment-driven reinforced graph reasoning paradigm, which unifies the modeling of question semantics, relation type constraints, and graph structure representation. It achieves dynamic optimization and interpretable path control of the multi-hop reasoning process through reinforcement learning. Finally, experimental results demonstrate that ERGR significantly outperforms various baseline methods on a real-world equipment maintenance question-answering dataset. In the comparison with RAG-enhanced large language models, compared with Qwen3-235B, H@1 is improved from 67.50% to 83.33%. In the comparison with graph neural network methods, compared with GraphSAGE, H@1 is increased by 7.69%.