Source-free Open-set Domain Adaptation Network for Emerging Fault Diagnosis of Planetary Gearbox

Transfer learning-based intelligent fault diagnosis (IFD) methods have found extensive applications in mechanical equipment due to their models exhibiting strong generalization performance. However, two challenges remain unresolved: intelligent diagnosis for emerging faults and data privacy protection for training data. To address these issues, a novel source-free open-set domain adaptation network (SODAN) is proposed. An uncertainty entropy-based self-supervised pseudo-label technology is designed to obtain high-confidence pseudo-labels for known unlabeled samples. Benefit from labeled high-confidence samples, emerging fault categories are identified using an adversarial training strategy. The proposed SODAN efficiently classifies known faults and intelligently detects emerging new faults in the target domain, with source domain samples participating only in the pre-training stage. Experiments on a laboratory dataset to validate the effectiveness of the proposed SODAN.