Deep prototypical networks based domain adaptation for fault diagnosis

Due to the existence of domain shifts, the distributions of data acquired from different machines show significant discrepancies in industrial applications, which leads to performance degradation of traditional machine learning methods. In this paper, we propose a novel method that combines supervised domain adaptation with prototype learning for fault diagnosis. The proposed method consists of two modules, i.e., feature learning and condition recognition. The module of feature learning applies the Siamese architecture based on one-dimensional convolutional neural networks to learn a domain invariant subspace, which reduces the inter-domain discrepancy of distributions. The module of condition recognition applies a prototypical layer to learn the prototypes of each class. Then the classification task is simplified to find the nearest class prototype. Compared with existing intelligent fault diagnosis methods, this proposed method can be extended to address the problem when the classes from the source and target domains are partially overlapped. The model must generalize to unknown classes in the source domain, given only a few samples of each new target class. The effectiveness of the proposed method is verified using two bearing datasets. The model quickly converges with high classification accuracy using a few labeled target samples in training, even one per class can be effective.