Few-Shot Bearing Fault Diagnosis Based on Multi-Metric Fusion Learning

Deep learning-based bearing fault diagnosis methods have achieved good results in recent years. However, most of them depend heavily on massive labeled data and are not available for the situations under limited labeled samples. To alleviate the problem, this paper proposes a few-shot bearing fault diagnosis model based on multi-metric fusion learning, which is suitable for the inadequate data and variable conditions. The proposed model introduces Manhattan distance (Ll metric) and Euclidean distance (L2 metric) to measure the feature similarities of the support and test signals which are extracted by a Siamese network architecture. By adopting the maximum fusion strategy, these two measurement representations are fused together and the fault diagnosis is performed accordingly. Various experiments on the two publicly available datasets, CWRU and JNU, demonstrate that the proposed method not only achieves better diagnostic performance with limited labeled samples, but also has good transfer capability between different conditions.