Journal article 68 views 11 downloads
An instruction fine-tuning and retrieval-augmented generation framework for intelligent defect diagnosis and maintenance decision support in high-speed railway turnouts
,
Xiaopei Cai,
Bin Cui,
Xueyang Tang,
Clare Wood ,
Yue Hou
Computer-Aided Civil and Infrastructure Engineering, Volume: 49, Start page: 100098
Swansea University Authors:
Clare Wood , Yue Hou
-
PDF | Version of Record
© 2026 The Authors. This is an open access article under the CC BY license.
Download (7.3MB)
DOI (Published version): 10.1016/j.cacaie.2026.100098
Abstract
High-speed railway (HSR) turnouts are among the most mechanically demanding components in the railway infrastructure, yet current operation and maintenance (O&M) practices remain largely reactive, experience-dependent, and disconnected from automated decision support. This paper presents a large...
| Published in: | Computer-Aided Civil and Infrastructure Engineering |
|---|---|
| ISSN: | 1093-9687 |
| Published: |
Elsevier BV
2026
|
| Online Access: |
Check full text
|
| URI: | https://cronfa.swan.ac.uk/Record/cronfa71989 |
| Abstract: |
High-speed railway (HSR) turnouts are among the most mechanically demanding components in the railway infrastructure, yet current operation and maintenance (O&M) practices remain largely reactive, experience-dependent, and disconnected from automated decision support. This paper presents a large language model (LLM)-driven expert system that bridges the gap between raw sensor data and actionable maintenance decisions for turnout defect diagnosis. Three core contributions are made. First, a data textualization strategy is developed to convert train body acceleration signals into structured text sequences comprehensible to LLMs, enabling domain-specific diagnosis without architectural modification of the base model. Second, an enhanced instruction fine-tuning scheme is proposed, incorporating a contrastive loss function that tightens intra-class feature clusters and widens inter-class margins, alongside a hierarchical evaluation method that reliably extracts categorical intent from free-form model outputs. Third, a retrieval-augmented generation (RAG) module is integrated with the fine-tuned model, enabling the system to generate standards-compliant maintenance recommendations directly from diagnostic results. Controlled experiments across four pre-trained models and 26 experimental groups demonstrate that the proposed system reaches a peak diagnostic accuracy of 89.6%, while preserving the natural language generation capabilities essential for report production. The framework is evaluated on a physically representative dataset generated by a validated stochastic vehicle–turnout dynamics model. The resulting integrated pipeline, from extracted signal features to maintenance decision output, offers a practical and scalable solution for intelligent O&M of complex railway turnout infrastructure and beyond. |
|---|---|
| Keywords: |
High-speed railway; Turnout; Large language model; Defect diagnosis; Intelligent operation and maintenance |
| College: |
Faculty of Science and Engineering |
| Funders: |
The work was supported by the State Key Laboratory of Advanced Rail Autonomous Operation (RAO2025ZT002), Beijing Jiaotong University; Tianjin Key R&D Programme for Beijing–Tianjin–Hebei Collaborative Innovation (25YFXTHZ00260); the Natural Science Foundation of Beijing, China (L251029); the Fundamental Research Funds for the Central Universities (2025QYBS007); and the Key Research Project of China Railway Design Corporation (2024A0253805). |
| Start Page: |
100098 |