LLM-Based Framework Boosts Bearing Fault Diagnosis with Limited Data

Bearing fault diagnosis in industrial settings faces significant hurdles due to dataset heterogeneity, varying operating conditions, and a scarcity of labeled data. Current diagnostic approaches often tackle these issues in isolation and rely on implicit feature alignment, which limits their effectiveness when multiple challenges occur simultaneously. This paper introduces a novel knowledge-guided two-stage transfer learning framework designed to address these complex problems. The framework employs a lightweight GPT-2-style Transformer, featuring causal self-attention, for hierarchical feature extraction from vibration signals. It establishes explicit knowledge transfer pathways where pre-trained encoder weights and fault prototype embeddings carry information from multi-source pre-training to target adaptation. The proposed framework effectively handles the dual-shift challenge through multi-source learning for generalizable representations, prototype-based knowledge modulation for adapting to target domains, and taxonomy-adaptive classification for seamless transfer across diverse fault categories. Experimental validation across four real-world datasets demonstrated an impressive 92.61% average accuracy using only 10% labeled target data, surpassing state-of-the-art methods by 17.24 percentage points. This breakthrough offers a practical and cost-effective solution for predictive maintenance in Industry 4.0 applications.