Digital Twin Based Bearing Fault Diagnosis with Convergent Attention

Rolling bearings are crucial in various rotating machinery, playing a key role in maintaining safety and reliability in industrial operations. Deep learning-based fault diagnosis approaches typically depend on training datasets containing samples for all possible fault scenarios, which may not be readily accessible in certain industrial settings. To overcome this limitation, this paper proposes a visual perception model for diagnosing bearing faults using aggregated attention and digital twins. Initially, a digital twin model of the bearing is developed to simulate its vibration characteristics and produce vibration data representing different health conditions. Second, the vibration signals of the bearing are transformed into time-frequency diagrams, allowing the model to automatically learn important features, as well as local and global features, across different time and frequency scales through the aggregation of attention. These features are then input into the network for classification. Finally, after reducing the distribution gap between virtual and real data within the feature space, real data is then utilized to validate the effectiveness of the training model based on virtual data. Experimental results show that the digital twin-based approach provides high accuracy in diagnosing bearing faults and can effectively simulate and identify various fault types.