Dynamic MAML with Efficient Multi-Scale Attention for Cross-Load Few-Shot Bearing Fault Diagnosis

Area of Science:

• Mechanical Engineering
• Artificial Intelligence
• Machine Learning

Background:

• Bearing fault diagnosis faces challenges due to limited labeled data and domain mismatch across operational conditions.
• Existing methods struggle with generalization in small-sample fault diagnosis scenarios under varying loads.

Purpose of the Study:

• To propose an adaptive meta-learning framework (AdaMETA) for enhanced generalization and diagnostic performance in small-sample bearing fault diagnosis.
• To address domain mismatches and improve feature extraction from vibration signals under constrained data conditions.

Main Methods:

• Developed an adaptive meta-learning framework (AdaMETA) combining dynamic task-aware model-independent meta-learning (DT-MAML) and efficient multi-scale attention (EMA) modules.
• Introduced a hierarchical encoder with C-EMA for multi-scale fault feature extraction from vibration signals.
• DT-MAML dynamically adjusts inner-loop learning rates based on task complexity to mitigate domain bias.

Main Results:

• AdaMETA achieved superior diagnostic accuracy (up to 99.26%) and robustness on the CWRU bearing dataset under cross-domain scenarios.
• Outperformed traditional meta-learning and classical diagnostic methods in small-sample bearing fault diagnosis.
• Ablation studies confirmed the significant contributions of the EMA module and dynamic learning rate components.

Conclusions:

• The proposed AdaMETA framework effectively enhances generalization and diagnostic accuracy for bearing faults with limited data.
• AdaMETA offers a robust solution for cross-domain fault diagnosis by mitigating domain bias and improving feature extraction.
• The framework demonstrates significant potential for practical applications in industrial equipment monitoring and maintenance.