Dual-space multi-node collaborative transfer learning for cross-condition and cross-machine axlebox bearing fault diagnosis

Although transfer learning (TL) is a potential solution to the problem of data scarcity and model generalization, existing TL methods predominantly address single-source domains, and multi-source TL still suffers from negative migration and under-utilization of target domain-specific features. This study introduces a novel dual-space multi-node collaborative TL method for multi-source transfer learning in cross-condition and cross-machine axlebox bearing fault diagnosis. The proposed model employs a parallel shared-private dual-space framework, aimed at extracting domain-invariant features across various source-target domain pairs, while simultaneously capturing the unique features of the target domain. We design a general feature extractor (GFE) based on the Inception Transformer (IFormer) and a specialized feature extractor with a deformable convolution (DC) module for comprehensive feature extraction. Furthermore, two novel domain-adaptive loss functions—dynamic MK-MMD and attention LMMD—aimed at focusing on key feature information and are suitable for multi-source TL. On this basis, a multi-node loss function strategy collaboratively integrates five distinct types of loss functions, effectively aligning the global and local feature distributions of the multiple source and target domains. Extensive experiments are conducted on a publicly available bearing dataset and a self-constructed axlebox bearing dataset. This research proposes a TL-based axlebox bearing fault diagnosis method that can be effectively applied to cross-condition and cross-machine fault diagnosis tasks, exhibiting superior performance compared to state-of-the-art TL approaches.