Adversarial Contrastive Learning-Enabled Domain Extension Method for Time-Varying Cross-Domain Fault Diagnosis

Cross-domain fault diagnosis is useful in handling cross-working condition diagnosis scenarios. Most of the existing approaches mine domain knowledge from multiple working conditions. However, for time-varying working conditions, collecting valuable sensor signal samples from polytropic operating conditions can indeed be impossible; the available sensor data is commonly from a single working condition. Therefore, an adversarial contrastive learning-enabled domain extension method is proposed for time-varying cross-domain fault diagnosis under a single working condition. Specifically, to improve the generalization capability, weak and strong augmentations are proposed to enrich sample diversity at the sample level. The domain extension module with adversarial contrastive learning is designed for effective and reliable sample generation to overcome variable working conditions at the feature level. Moreover, improved supervised contrastive learning is leveraged in the discriminator module to learn the invariant representation and perform fault recognition. Meanwhile, the adversarial contrastive training between generator and discriminator is used to balance domain-specificity and domain-invariance and achieve cross- and within-comparison, improving the time-varying cross-domain diagnosis performance. Extensive experiments are conducted and show the effectiveness of the proposed framework, improving diagnosis accuracy by at least 5%~10%.