A lightweight cross-bearing fault diagnosis network incorporating implicit high-dimensional feature interaction and multi-kernel maximum correlation alignment

To address the issues of large domain shift, high model complexity, and low transfer diagnosis accuracy in cross-equipment fault diagnosis across different rolling bearing models, this paper proposes a transfer learning fault diagnosis framework based on a lightweight domain adaptation network (LDAN). Firstly, a lightweight feature extraction structure composed of basic convolution blocks and deep residual blocks is constructed to reduce model parameters and achieve a lightweight network architecture. Meanwhile, an implicit feature interaction mechanism based on element-wise multiplication is embedded in the deep residual blocks to boost the network’s capacity for capturing high-order discriminative information. Secondly, to mitigate cross-domain distribution discrepancy in cross-bearing scenarios, a novel domain alignment strategy named Multi-Kernel Maximum Correlation (MMC) is proposed. It formulates a joint loss combining multi-kernel maximum mean discrepancy and correlation alignment, enabling simultaneous alignment of the source and target domains in both feature space and second-order covariance statistics. Finally, the proposed method is validated on cross-bearing transfer tasks using three datasets. Experimental results show that the average diagnostic accuracy reaches 99.48% under constant working conditions and 98.11% under varying working conditions, which verifies the superior transfer diagnosis performance of the proposed method.