Fault diagnosis of bearings under small sample and variable working conditions based on DG-FWC-KELM

Abstract To address the issues of data scarcity (small samples) and data distribution shift (domain generalization)DG caused by variable working conditions in bearing fault diagnosis within actual industrial scenarios, traditional deep learning methods often suffer from defects such as severe overfitting, poor DG ability, and time-consuming training. Therefore, this paper proposes a DG-oriented wide-field convolution (FWC) Kernel extreme learning machine (KELM) fault diagnosis method with strong generalization capabilities. First, a FWC module is constructed, utilizing large-size convolution kernels to simulate ‘wide-angle’ perception, which efficiently captures global noise-resistant features while reducing computational complexity. Second, a high-dimensional feature constraint mechanism is designed, introducing LeakyReLU combined with Tanh functions and L2 norms to impose physical constraints, effectively resolving numerical explosion and enhancing feature separability. Finally, the KELM is adopted to replace the traditional Softmax layer, leveraging its analytical solution characteristics to construct a robust classification hyperplane under small sample conditions, significantly improving training speed and generalization accuracy. Experimental results on cross-domain variable working conditions using Jiangnan University and Case Western Reserve University datasets demonstrate that the method can effectively identify fault states in unseen target domains using only a small number of source domain samples for training. Furthermore, to deeply examine the global feature capture capability of the proposed ‘FWC’ module under strong background noise and complex vibration coupling, targeted high-difficulty test data from Shanghai Maritime University is introduced for verification. The results confirm that the proposed DG-FWC-KELM model can effectively overcome interference under extremely harsh working conditions and maintains excellent diagnostic robustness in DG scenarios.