Dynamic adversarial joint domain adaptation for unsupervised cross-machine bearing fault diagnosis

Unsupervised domain adaptation (UDA) methodologies have demonstrated notable success in solving the distribution shifts between training and testing data for rotating machinery fault diagnosis. However, their effectiveness remains limited in cross-machine scenarios. To address this, a dynamic adversarial joint domain adaptation (DAJDA) framework is proposed to enhance diagnostic adaptability across heterogeneous machinery. The methodology incorporates three characteristics: First, an inception-enhanced wide-kernel deep convolutional neural network architecture is constructed for extracting the discriminative and domain-invariant features. Second, a hybrid discrepancy reduction strategy that integrates joint distribution adaptation and adversarial adaptation is proposed. At last, a dynamic balancing mechanism for multiloss interactions during joint explicit-implicit domain adaptation is employed. Experiments on six cross-machine transfer tasks involving single-row to double-row bearings and small-to-large size bearings confirm the effectiveness of DAJDA. Comparative analyses further demonstrate consistent superiority over existing UDA approaches in both diagnostic performance and domain transfer robustness.