Adversarial Domain Adaptation for Gear Crack Level Classification Under Variable Load

Traditional intelligent fault diagnosis assumes that the training and testing samples are drawn from the same distribution. This assumption does not hold when working condition changes, as variable working condition can make the training and the testing datasets have different distributions. A novel working condition might be encountered in the testing stage, and there will be no label available under that novel working condition. This paper studies domain adaptation for gear crack level diagnosis under variable loading. A new two-stage fault diagnostic method for variable load condition is developed based on adversarial training strategy and gradient reversal layer. Both labeled and unlabeled data are utilized to learn best model for the novel load condition. An experimental case study is carried out to demonstrate the effectiveness of the proposed method.