Semantic Data Governance and Collaborative Architecture for LLM-Based Natural Language Analytics in Urban Rail Transit

Abstract

Urban rail transit data are distributed across heterogeneous operational systems, where inconsistent semantics and fragmented metric definitions complicate the controlled use of large language models (LLMs) for natural language analytics. This study proposes a semantic data governance and collaborative analytics architecture for this setting. A lakehouse-based data foundation, a semantic middle layer, and a unified metric store are used to organize physical schemas, business terminology, and statistical definitions. Domain rules, network topology, and a controlled execution workflow are then introduced to constrain query generation and support cross-domain analysis. Two representative scenarios—power-supply fault assistance and passenger-flow–traction-energy analysis—are used to illustrate how data governance, knowledge augmentation, and execution validation work together. The contribution is an engineering architecture and its semantic governance mechanisms rather than a comparative evaluation of general-purpose LLM algorithms. The proposed approach provides a practical reference for integrating LLM-based analytics with existing urban rail transit data environments.