A unified framework for interpretable elevator fault diagnosis and predictive maintenance via style-aware CoT fine-tuning.

Abstract

Although Large Language Models (LLMs) have shown potential in industrial applications, they encounter significant hurdles in vertical scenarios like elevator maintenance, including hallucinations, lack of domain specificity, and an inability to interpret numerical physical states. To bridge this semantic-physical gap, this paper proposes a Unified Style-Aware Chain-of-Thought (SA-CoT) framework tailored for Small Language Models (SLMs). The novelty of our approach lies in two aspects: first, we construct a robust instruction dataset using a style-aware augmentation strategy to simulate diverse real-world user behaviors and noise; second, we innovate by textualizing raw sensor data, enabling the fine-tuned 4B-parameter SLM to generate high-dimensional embeddings for downstream numerical analysis. Experiments demonstrate a dual breakthrough: in generative diagnosis, the SA-CoT framework consistently outperforms general models, achieving a 5.6-fold improvement in BLEU-4 scores compared to GPT-4o. Furthermore, its embeddings capture physical features more effectively than traditional baselines, yielding highly competitive accuracy in Alarm Type Classification and Vibration Magnitude Regression. These results suggest that domain-aligned SLMs offer a robust and cost-effective framework for autonomous predictive maintenance, indicating that knowledge density plays a more critical role than parameter scale in specialized industrial applications.