BiCAN: Bidirectional Contrastive Adversarial Network for Unsupervised Domain Adaptation in Mechanical Fault Diagnosis

Supervised learning-based approaches have demonstrated effectiveness in mechanical fault diagnosis, yet their practical deployment is constrained by heavy reliance on large-scale labeled datasets and assumptions of consistent data distributions. To bridge these gaps, a bidirectional contrastive adversarial network with unsupervised domain adaptation is proposed for mechanical fault diagnosis. First, a momentum-enhanced siamese network architecture built upon Momentum Contrast (MoCo) principles is developed to learn robust high-dimensional representations from unlabeled vibration signals. Then, a symmetric contrastive learning mechanism integrated with a dynamic feature-swapping module is designed that amplifies self-supervised signals through augmented feature interactions, significantly improving feature discriminability. Furthermore, a domain-adversarial network aligns cross-domain feature distributions through adversarial optimization between a momentum-updated feature extractor and a domain discriminator, enabling robust knowledge transfer without target domain labels. Experimental validation on a multi-condition bearing fault dataset confirms the method’s ability to achieve precise fault diagnosis. The proposed method exhibits significant advantages in cross-domain knowledge transfer, outperforming state-of-the-art methods.