Virtual-physical mapping network-driven digital twin for intelligent fault diagnosis with imbalanced samples

In advanced manufacturing systems, intelligent fault diagnosis of rolling bearings serves as a pivotal technology for ensuring equipment reliability and improving operational efficiency. Nevertheless, acquiring high-quality monitoring data across multiple fault states in real industrial environments remains a major challenge. This limitation makes achieving high-accuracy diagnosis under imbalanced samples particularly difficult. Digital twin technology provides a promising paradigm to alleviate these constraints. This research introduces a digital twin-based intelligent diagnosis framework under imbalanced samples. First, a nonlinear 6-DOF dynamic model is formulated for digital twin construction, generating high-fidelity simulated vibration responses. Second, a novel attention-assisted operational CycleGAN architecture is developed for cross-domain information exchange between virtual and physical domains. This mapping network integrates a generator constructed using an enhanced feature pyramid network and a discriminator based on the PatchGAN architecture. By leveraging self-organized operational neural network and attention-enhanced skip connections, the network strengthens intra-layer feature extraction and inter-layer feature transmission, thereby facilitating effective information exchange between virtual and physical entities. Finally, experimental validation demonstrates that the Pearson correlation coefficient between the normalized twin and real signal exceeds 0.9. The twin signal enables deployment of diagnostic models, achieving reliable fault diagnosis in data scarcity scenarios.