A bearing fault diagnosis method for hydrodynamic transmissions integrating few-shot learning and transfer learning.

To address the insufficient generalization capability of bearing fault diagnosis models caused by scarce vibration data from high-power hydrodynamic transmission testbeds, this study proposes a diagnostic method integrating deep few-shot learning with transfer learning. First, a Siamese Wide Convolutional Neural Network (Siamese-WDCNN) is constructed based on public bearing datasets to extract essential features of vibration signals through few-shot contrastive learning. Second, we introduce a transfer learning strategy to address cross-condition generalization challenges. This approach adapts pre-trained model parameters from the CWRU dataset to real industrial hydrodynamic transmission data. We then fine-tune the model using limited target-domain samples to optimize performance. Experiments evaluating the generalization capability under variable operating conditions compare diagnostic performance across SVM, WDCNN, WDCNN + TL, FSL + TL, and FSL + TL + AM methods. Results demonstrate that FSL + TL achieves an accuracy of 85.30% under mixed operating conditions. Further optimization by incorporating an attention mechanism (FSL + TL + AM) elevates accuracy to 88.75%, effectively enhancing the generalization capability of the bearing fault diagnosis model. This validates the engineering practicality of the proposed method and explores a viable pathway for industrial equipment health monitoring.