A Few-shot Fault Diagnosis Method under Target Domain Data Based on Improved Metric Space SVM

The intelligent bearing diagnosis with big data has been widely researched. Despite high performance diagnosis models are constantly being proposed, these models are difficult to apply in some practical application with limited collectable samples. In this paper, an Improved Metric Space Support Vector Machine with a two-stage learning strategy is proposed to overcome the challenge of few-shot scenarios under target domain samples. First, an improved triplet loss is designed in the metric network to select features by utilizing the similarity of samples. Second, kernel method is used to model in few-shot scenarios by maximizing the inter-class interval. Experiments on CWRU dataset verify the superiority and robustness of IMSSVM comparing with popular baseline methods. Furthermore, the effectiveness of IMSSVM is verified with visualizing the feature embeddings by t-distributed stochastic neighbor embedding.