A Multiscale Recalibration Attention Network With Adaptive Grouped Convolution for Bearing Fault Diagnosis via Transfer Learning

Addressing the critical challenges of limited labeled data and operational variability in bearing fault diagnosis, a multiscale recalibration attention network (MRAN) integrated with adaptive grouped convolution is proposed in this article. First, a statistical-aware recalibration module that dynamically fuses mean and standard deviation information to suppress noise and amplify discriminative features under data scarcity. Second, a kernel-adaptive grouped convolution mechanism that intelligently assigns computational resources based on kernel size to optimize efficiency. Third, a hierarchical fine-tuning strategy that selectively adapts deep layers to resolve cross-domain feature misalignment. Extensive experiments on BJTU-RAO and cross-device datasets demonstrate the superior performance of MRAN: it achieves an accuracy of 98.67% without transfer learning (TL), which is significantly better than the state-of-the-art convolutional neural networks (CNNs) and vision transformers. Crucially, using only 30% of the target domain data, MRAN has an accuracy of 94.57% in cross-conditional tasks and an accuracy of 92% in demanding cross-device diagnostics, while greatly reducing training time and computational burden. This work establishes MRAN as a groundbreaking, efficient, and highly accurate solution for practical industrial bearing fault diagnosis under data and operational constraints.