A domain adaptation method based on interpolation and centroid representation for open-set fault diagnosis of bearing

Fault diagnosis in open-set world refers to the condition that source domain and target domain do not share a uniform health state space. Unknown fault categories are extremely likely to occur in the target domain, which is more in line with the actual engineering scenarios, simultaneously exposing a great challenge to find a decision boundary for the unknown. However, in existing open-set fault diagnosis studies, most studies neglect the influence of spurious correlations between label and domain. Here we propose a domain adaptation method based on interpolation and centroid representation for this problem. Specifically, an interpolation mechanism is implemented to enhance the predictor’s ability of reducing the inference of spurious correlations and recognizing the unknown fault categories. For better alignment with the target domain, relative centroid distance minimization method is designed to make the source domain discriminative. Centroid representation alignment and decision boundary for rejecting the unknown are executed from both category centroids and specific samples. Comprehensive experiments are conducted to demonstrate that the proposed method achieves a promising performance for various open degrees.