A joint weighted transfer model for open-set adaptation fault diagnosis of rotating machinery

Abstract Recently, transfer learning has been widely used for domain adaptation (DA) fault diagnosis in rotating machinery. Since machines may incur new unknown faults in testing scenarios, it is necessary to address the open-set DA (ODA) problem, where the target domain contains private unknown-class samples not shared by the source domain. Existing ODA diagnostic methods mainly adopt entropy uncertainty and domain similarity as weighting criteria to distinguish known-class and unknown-class samples. However, entropy uncertainty exhibits unreasonable quantification results on approximately fuzzy and non-sharp class distributions, and these two weighting criteria are not reliable enough due to their susceptibility to prediction errors. To address the issue, the paper proposes a joint weighted transfer network. By analyzing the complementarity between entropy and relative confidence, a dual uncertainty weight is devised to achieve reasonable classification uncertainty quantification for the target samples. Furthermore, by combining classification uncertainty, domain similarity, and feature distance, a joint weighting (JW) criterion is developed to reliably reflect the probability of each target sample in unknown classes. During model training, the JW criterion is introduced into both domain adversarial learning and classification learning to separate unknown-class samples from known-class samples. The experimental results demonstrate that the proposed model can achieve ODA fault diagnosis in rotating machinery with high diagnostic accuracy, and its performance is superior to existing methods.