[Objective] Knowledge pertaining to power station equipment failures is typically dispersed across various reports

maintenance records

and expert experience. This fragmentation

coupled with inconsistencies in semantic description

severely hinders effective knowledge sharing and reuse. Conventional retrieval methods

which rely primarily on keyword matching

often fail to capture deeper semantic relationships

leading to inefficient queries

so knowledge graph and retrieval-augmented generation(KG-RAG) techniques are proposed. The method aims to consolidate fault knowledge

unify its semantics

and enable efficient yet precise querying

with the goal of enhancing the intelligence of power station operation and maintenance. [Methods] Using the failure mode and effects analysis tables as the core knowledge source

a formal domain ontology was first constructed through modeling. This ontology structurally represents key entities

such as equipment

failure modes

symptoms

and causes

along with their attributes and relationships. Based on this framework

unstructured text was converted into structured triples that comply with ontological constraints

progressively building a fault knowledge graph capable of semantic reasoning. Simultaneously

embedding models were employed to generate vector representations for nodes in the graph and their associated textual descriptions. A dual knowledge base was thus established

indexed for both graph structure and semantic vectors

effectively combining “factual relationships” with “semantic embeddings”. During the question-answering phase

structured retrieval is first performed using the knowledge graph to accurately locate relevant equipment and fault chains. If the graph-based results are insufficient

a semantic retrieval channel is activated

leveraging vector similarity to recall related text passages. The outputs from both retrieval channels are then deduplicated

ranked

and fused into a coherent

information-rich context. This enhanced context is fed into a large language model

which reasons over it to generate final answers. The responses not only present factual information but also incorporate explanations and relevant suggestions

yielding scientifically grounded and comprehensive results. [Results] To evaluate the proposed method

an experiment was conducted using 20 typical fault-related questions spanning multiple scenarios

including factual queries

cause inference

and handling recommendations. The performance of four approaches was compared: a standalone large language model

full-table text injection

vector-only retrieval augmentation

and the KG-RAG method. The evaluation metrics covered information coverage

semantic relevance

and response time. Experimental results indicated that the proposed method outperformed the others in both information coverage and semantic relevance

achieving an overall score of 91.2. This demonstrates that the hybrid mechanism

which integrates knowledge graph retrieval with semantic vector search

provides more comprehensive and accurate support for knowledge-intensive questionanswering tasks. [Conclusion] The KG-RAG-based question-answering method presented in this study effectively combines the precise relational reasoning of knowledge graphs with the semantic generalization capability of vector retrieval. Low response latency is maintained while accuracy

coverage

and explainability in the generated answers are improved. A practical and effective solution for managing and applying fault knowledge in power station contexts is thus offered. Future work could focus on dynamic knowledge updating

multi-modal fault knowledge integration

and cross-domain knowledge transfer

all of which would further enhance the method's practicality and generalizability.