Source-only cross-condition bearing fault diagnosis via a lightweight multi-scale CNN with frequency-adaptive feature fusion

Cross-condition rolling-bearing diagnosis remains challenging in source-only scenarios where target-domain unlabeled samples are unavailable during training. To address this practical deployment setting, this paper proposes a lightweight multi-scale one-dimensional CNN with frequency-adaptive feature fusion (FAFF), in which full-spectrum FFT magnitude is mapped to branch-aligned band descriptors and then fused with content descriptors for adaptive branch weighting. Unlike target-assisted domain adaptation methods, the proposed framework is trained using only source-domain labeled data while still introducing frequency-guided branch-level fusion for cross-condition transfer. Experiments follow a hardened dual-benchmark protocol including source-train subsampling, three random seeds, paired task-seed analysis, cross-dataset validation, and additive-noise robustness assessment. On CWRU, aggregate scores remain near ceiling after hardening, so the most informative evidence comes from the difficult A→B and A→C transfers, where the proposed model achieves 97.96% mean accuracy, 1.51 percentage points higher than ResNet-18. On Paderborn, the proposed model reaches 79.68% average accuracy, exceeding ResNet-18 by 4.00 percentage points with a substantially smaller model size, while still remaining below DANN-CNN. Under additive noise with SNR levels from 0 to 8 dB, performance gradually recovers as SNR increases and remains strong under medium/high-SNR conditions, whereas the sharp degradation at 0 dB indicates that the present robustness claim should not be extended to severe-noise deployment.