LMKG: Application of a Retrieval-Augmented Reasoning Framework with Graph Topology Constraints in Rail Transit Locomotive Maintenance

This article introduces the LMKG (Locomotive Maintenance Knowledge Graph) framework, an innovative method combining graph topology constraints with Retrieval-Augmented Generation (RAG). It aims to address issues such as knowledge hallucination, domain knowledge gaps, and broken reasoning chains in large language models (LLMs) applied to the rail transit locomotive maintenance field, thereby enhancing the professional reasoning capabilities of LLMs. Through structured storage of knowledge graphs and reasoning guided by graph topology constraints, combined with hierarchical retrieval and multi-dimensional reasoning enhancement mechanisms, this framework has significant application value in scenarios like fault diagnosis, maintenance plan generation, and knowledge inheritance.

Rail transit locomotive maintenance involves massive amounts of information such as technical documents and fault records. The traditional approach relying on experienced engineers struggles to cope with the increasing system complexity. General-purpose LLMs have three major issues when applied in this field: 1. Knowledge hallucination: generating incorrect technical advice; 2. Domain knowledge gaps: lack of rail transit professional knowledge; 3. Broken reasoning chains: difficulty capturing causal relationships of complex faults.

Core Ideas: Integrate the structured characteristics of knowledge graphs with the generation capabilities of LLMs. The knowledge graph stores entities (components, fault types, etc.) and relationships (causal, assembly, etc.); graph topology constraints guide reasoning through paths, neighbors, and consistency constraints.

Technical Architecture: The retrieval enhancement module adopts a two-stage strategy (graph retrieval to narrow the scope → document retrieval to obtain precise information); the reasoning enhancement mechanism supports multi-hop reasoning (e.g., fault root cause tracing), temporal reasoning (remaining life prediction), and uncertainty handling (quantifying reasoning risks).

1. Fault Diagnosis Assistance: Analyze symptoms → retrieve fault paths → generate diagnostic reports with confidence levels and maintenance recommendations; 2. Maintenance Plan Generation: List tools and spare parts, step-by-step guides, safety precautions, and cost/time estimates; 3. Knowledge Inheritance and Training: Provide technical Q&A, case exercises, and fault simulation drills for new engineers.

Technical Implementation Highlights: Data samples include structured triples, unstructured text, and optional multi-modal data; supports integration with mainstream LLMs (prompt optimization, fine-tuning, RAG injection).

Industry Value: Improve maintenance efficiency, lower knowledge thresholds, promote knowledge accumulation, ensure operational safety, and enhance the safety and reliability of rail transit infrastructure.

Future Directions: Cross-domain migration (aviation, shipping, etc.), real-time data processing (IoT-integrated predictive maintenance), multi-modal fusion (image/video fault detection), edge deployment optimization.

Conclusion: LMKG demonstrates the prospects of combining knowledge engineering with LLMs, providing an industrial AI application solution that balances intelligence and controllability, which is worthy of industry attention and reference.