Volume 68

Learn physics, then diagnose: physics-augmented curriculum learning for large language model-based fault diagnosis in thermal power plants Zeyu Zheng, Aiqing Wang, Jie Lu*, Yang Zhao

Abstract

Fault diagnosis is essential for safe thermal power system operation. Large language models (LLMs) can process complex industrial logs and provide diagnostic insights, making them a promising approach for fault diagnosis. However, conventional LLMs often produce physical hallucinations and logical inconsistencies, since physical knowledge and fault-specific patterns are injected together in a single training stage. This paper proposes a physics-augmented curriculum learning (PACL) method that uses a two-stage curriculum to Integrate physical laws into diagnostic reasoning. The method includes two steps: physics-field chain-of-thought-driven knowledge extraction and diagnostic alignment to fine-tune the model. Multi-source documents from Hanglian thermal power plant in Hangzhou, China are utilized, including fault cases, equipment regulations, and technical manuals. Evaluations on representative fault scenarios show that the proposed PACL achieves an average diagnostic score of 92.0, outperforming SFT (47.8), Mixed-FT (53.5), and R-PACL (60.8), demonstrating expert-level reasoning in complex scenarios.

Keywords thermal power systems, large language models, physics-augmented curriculum learning, fault diagnosis, physics injection