LLM-Driven Knowledge Graph Construction for Automotive Parts Fault Diagnosis

Large quantities of fault maintenance documents are produced in automotive parts manufacturing. non-standardized expressions, intensive professional terminology and hidden semantic features make it difficult to extract high-quality structured knowledge. Existing discriminative methods remain constrained within the sequence labeling paradigm, limiting their capacity to capture complex relational semantics and implicit fault logic inherent in heterogeneous industrial texts. To address this, this study proposes GIE-AFD, a schema-constrained generative information extraction from maintenance documents by leveraging large language model (LLM) to construct knowledge graph for fault diagnosis in automotive parts manufacturing. Initially, a domain-specific schema-constrained ontology encompassing six entity types and seven relation types is established, transforming unstructured text into a controlled structured data generation task. Subsequently, the chain-of-thought-guided structured parsing mechanism (CoT-SP) is designed to direct the large language model to simulate domain expert reasoning through step-by-step symbolic inference prior to structured output generation, effectively suppressing semantic noise in entity and relation extraction. Furthermore, the ChatGLM model is fine-tuned via LoRA on a domain-specific instruction dataset, AFD-Instruct, to inject automotive fault knowledge into the model's parametric space. Concurrently, an entity alignment strategy incorporating source text anchoring is designed, establishing a traceable and precise correspondence between the extracted results and the original document at the character level. Additionally, a retrieval-augmented generation (RAG) framwork with large language models and domain fault knowledge graph is introduced to support evidence-grounded fault diagnosis query-response. Experimental results on constructed dataset from the automobile parts manufacturing demonstrate that the proposed method outperforms comparison methods, achieving an NER-F1 score of 87.07% and an RE-F1 score of 79%. Based on this, a prototype system, InduKG, for generative information extraction, knowledge graph construction and fault diagnosis is implemented, illustrating the effectiveness and adaptability of the proposed method in real-world production scenarios.