Knowledge graph-enhanced large language models toward reliable elevator fault diagnosis

The operational reliability of elevator systems is critical for the safety and efficiency of modern urban infrastructure. While large language models have demonstrated exceptional reasoning capabilities, their application to specialized industrial maintenance tasks is constrained by domain knowledge deficits. This dependency on generic pre-training corpora frequently leads to hallucinated or impractical strategies when confronting complex electromechanical fault logic. To address this bottleneck, we propose a structured knowledge-enhanced framework tailored for intelligent elevator fault diagnosis. First, we construct a curated, schema-constrained elevator knowledge graph derived from authentic industrial records. To reduce structural sparsity, we introduce a neuro-symbolic collaborative KG completion strategy, which combines deterministic meta-path logical rules with an LLM-driven retrieve-then-verify mechanism. Furthermore, we formulate a graph-based retrieval-augmented generation paradigm to inject these densified graph priors into the LLM’s inference pipeline. Comprehensive evaluations on state-of-the-art models show that, on our constructed elevator-maintenance question-answering dataset, the proposed framework consistently improves Key Entity Recall and BERTScore as well as preference under an LLM-as-a-Judge protocol, indicating better coverage of domain entities and closer agreement with record-derived reference answers. These results suggest that grounding generation in a schema-constrained and automatically screened domain knowledge graph is a promising direction for reducing hallucination and supporting more reliable LLM-based elevator fault diagnosis, while further validation with direct factuality and safety measures is needed before deployment in safety-critical settings.