Knowledge-Guided Two-Stage Transfer Learning Framework for Bearing Fault Diagnosis

In industrial environments, bearing fault diagnosis is hindered by three concurrent challenges: dataset heterogeneity (different sensors/geometries), operating condition variations (speed/load changes), and a scarcity of labeled data. Existing deep learning methods typically address these issues in isolation or rely on implicit feature alignment, which often fails under severe data scarcity. This paper investigates how to establish explicit knowledge pathways to enable effective transfer learning when these challenges occur simultaneously.

The authors propose a two-stage, knowledge-guided framework based on a lightweight, 6-layer GPT-2-style Transformer.

• Stage 1 (Cross-Dataset Prior Learning): The model is pre-trained on multiple source datasets to learn universal fault representations. This stage produces two explicit knowledge carriers: initialized encoder weights and fault prototype embeddings (class-specific characteristic patterns).
• Stage 2 (Knowledge-Guided Adaptation): The model adapts to a new target domain using only 10% of the available labeled data. It employs dual-level transfer: parameter-level initialization using Stage 1 weights and feature-level modulation where input signals are weighted by their similarity to the learned fault prototypes. The framework also uses LoRA (Low-Rank Adaptation) for parameter-efficient fine-tuning and a dynamic classification head to handle varying fault taxonomies across different datasets.

The framework was validated on four real-world bearing datasets (CWRU, MFPT, JNU, and PU). It achieved an average accuracy of 92.61% using only 10% labeled target data, outperforming state-of-the-art methods by 17.24 percentage points. The explicit use of fault prototypes and parameter initialization proved significantly more effective than relying on universal feature extraction alone, particularly in scenarios with complex load variations and significant distribution shifts.