Task Adaptation Meta Learning for Few-Shot Fault Diagnosis under Multiple Working Conditions

Abstract

Few-shot fault diagnosis is a challenging issue in manufacturing area, which rely on knowledge learned from historical data and limited data in new work condition. Nevertheless, the unbalanced distribution in historical working condition data and the distribution discrepancy between the finite small data and historical data lead to the poor generalization and low reliability of few-shot model. This study proposes a task adaptation meta learning framework. First, target domain is selected from historical working condition by relative entropy. Then, domain-adversarial training of neural networks is applied in historical samples for data distribution alignment to make tasks easy to learn. Finally, the fault diagnosis model trained with gradient based meta learning is adapted to new condition quickly with few data. On the Bearing Dataset under time-varying rotational speed conditions, the proposed framework has a good performance compared with the state-of-art method.