Self-Training Reinforced Adversarial Adaptation for Machine Fault Diagnosis
Jinyang Jiao, Hao Li, Jing Lin
Abstract
To alleviate the predicament of data annotating and the need for collecting data from identical distribution, unsupervised domain adaptation technologies have been widely deployed in the field of machine fault diagnosis. Nevertheless, most of them focus only on domain alignment and fail to make full use of the unlabeled target data. Given this, a novel self-training reinforced adversarial adaptation (SRAA) diagnosis method is developed in this article. In SRAA, an adversarial adaptation mechanism based on a bi-classifier disparity measure is introduced to implicitly align two domains. Meanwhile, a bi-classifier-oriented self-training algorithm is further proposed to fully mine the connotation information of unlabeled target data and reinforce model performance, in which an adaptive sample selection strategy and soft cross-entropy loss are included to guide high-confidence self-training. Empirical evidence from extensive experiments demonstrates the efficacy of SRAA, and comprehensive comparisons against the existing approaches show the superiority of our method.