Synergistic Progressive Adversarial Network for Domain Adaptation in Rotating Machinery

Abstract

The application of dual classifiers in adversarial learning significantly improves the convergence of cross-domain distributions. Traditional methods, however, largely focus on maintaining prediction consistency between classifiers, often neglecting their certainty. Additionally, existing strategies tend to emphasize domain alignment without fully tapping into the rich potential of unlabeled target data. To address these limitations, a synergistic progressive adversarial network (SPAN) is proposed. This framework initially designs a synergistic nuclear-entropy optimization (SNEO) strategy that embeds cooperative maximization and minimization of nuclear entropy within the adversarial training of dual classifiers. This approach not only promotes implicit domain alignment at the class level but also prevents the erroneous convergence of ambiguous samples into dominant categories during batch training. Moreover, an ensemble progressive self-training (EPST) strategy is developed to refine decision boundaries. By ensembling predictions from two classifiers and flexibly adjusting category-specific thresholds, the EPST strategy facilitates the progressive integration of information-rich, unlabeled target samples into the training process. This approach effectively enhances the delineation of decision boundaries through low-density regions. Experiments conducted on two distinct datasets have confirmed the effectiveness and superiority of the proposed network.