CGML++: Coarse-Grained Meta-Learning for Few-Shot Bearing Fault Detection Under Multiple Domain Shifts

Early detection of bearing defects in rotating machinery is crucial for preventing failures and reducing operational costs. However, deploying neural networks for vibration-based diagnostics is often hindered by a scarcity of labeled data under novel operating conditions. Existing adaptation techniques, including transfer learning and classical meta-learning algorithms, either suffer from negative transfer or fail to maintain sufficient accuracy under severe domain shifts. To address this, we propose CGML++ (Coarse-Grained Meta-Learning++), a novel few-shot method that aggregates MAML-trained domain experts through a tunable parameter γ combined with a transfer-learned model initialization, adapting to shift severity. We evaluated CGML++ on the CWRU bearing dataset under two domain shift scenarios: cross-sampling-rate transfer (12 kHz → 48 kHz) and cross-load transfer. In the 1-shot setting, CGML++ achieved a Macro F1-score of 0.88 and a Balanced Accuracy of 0.91 under cross-sampling-rate transfer, outperforming standard transfer learning by 0.18 and 0.12, respectively. Under cross-load transfer, CGML++ achieved a Balanced Accuracy of 0.96, surpassing all baselines by 0.04–0.09. The method’s versatility was validated across four architectures: ResNet1D, LightCNN, Transformer1D, and Informer1D. A key empirical finding is the identification of a phase transition at γ·N ≈ 2.0–3.5, enabling a scaling rule for adaptive hyperparameter tuning.