KG-SR-LLM: Knowledge-Guided Semantic Representation and Large Language Model Framework for Cross-Domain Bearing Fault Diagnosis.

Bearing fault diagnosis is crucial for stable operation and safe manufacturing as industry intelligence becomes increasingly advanced. However, under complicated non-linear vibration modes and multiple operating conditions, most of the current diagnostic methods are limited in terms of cross-domain generalization. To address these issues, this study develops a generalized diagnostic framework leveraging Large Language Models (LLMs), integrating multiple enhancements to improve both accuracy and adaptability. Initially, a structured representation approach is designed to transform raw vibration time series into interpretable text sequences by extracting physically meaningful features in both time and frequency domains. This transformation bridges the gap between sequential sensor data and semantic understanding. Furthermore, to explicitly incorporate bearings' structural parameters and operating condition information, a knowledge-guided prompt tuning strategy based on Low-Rank Adaptation (LoRA-Prompt) is introduced. This mechanism enables the model to adapt more effectively to varying fault scenarios by embedding expert prior knowledge directly into the learning process. Finally, a generalized fault diagnosis method named Knowledge-Guided Semantic Representation and Large Language Model (KG-SR-LLM) is established. Large-scale experiments using 11 public datasets from industrial, aerospace, and energy fields are carried out to extensively evaluate its performance. Based on experiment analysis and a comparison of results, KG-SR-LLM is superior to classical deep learning models by 9.22%, reaching an average diagnostic accuracy of 98.36%. KG-SR-LLM is effective for handling few-shot transfer and cross-condition adaptation tasks. All these results illustrate the theoretical significance and application benefit of KG-SR-LLM for intelligent fault diagnosis of bearings.