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

Area of Science:

• Artificial Intelligence
• Industrial IoT
• Machine Learning

Background:

• Large Language Models (LLMs) show industrial promise but struggle with domain specificity and interpreting physical states.
• Challenges include hallucinations and lack of specialized knowledge in vertical applications like elevator maintenance.

Purpose of the Study:

• To bridge the semantic-physical gap in industrial AI by proposing a novel framework for Small Language Models (SLMs).
• To enhance predictive maintenance capabilities in specialized industrial scenarios.

Main Methods:

• Developed a Unified Style-Aware Chain-of-Thought (SA-CoT) framework for SLMs.
• Created a robust instruction dataset using style-aware augmentation to simulate real-world data and noise.
• Textualized raw sensor data to enable SLMs to generate high-dimensional embeddings for numerical analysis.

Main Results:

• SA-CoT framework significantly outperformed general models in generative diagnosis, achieving a 5.6-fold improvement in BLEU-4 scores over GPT-4o.
• Generated embeddings effectively captured physical features, showing competitive accuracy in Alarm Type Classification and Vibration Magnitude Regression.
• Demonstrated that domain-aligned SLMs offer a cost-effective solution for autonomous predictive maintenance.

Conclusions:

• Domain-aligned SLMs, particularly when utilizing frameworks like SA-CoT, provide a robust and efficient solution for industrial predictive maintenance.
• Knowledge density is more critical than parameter scale for specialized industrial AI applications.
• The proposed framework effectively addresses limitations of LLMs in vertical industrial settings.