Abstract

Deep learning algorithms have shown strong classification performance in bearing fault diagnosis, but their performance can deteriorate significantly when vibration measurements are corrupted by noise. In complex practical environments, sensor locations, installation conditions, and surrounding machinery can differ across field settings, reducing the observability of fault-related patterns. To address this issue, this study proposes a fault characteristic frequency-aware multi-task learning framework for robust bearing fault diagnosis in noisy environments. The proposed method emphasizes fault-related repetitive structure in vibration signals and jointly learns fault classification with an auxiliary reconstruction task, encouraging the model to base its diagnosis on physically meaningful fault information even under noisy conditions. By leveraging the shared representation toward fault-related periodicity, the model is intended to reduce noise and improve diagnostic robustness in practical settings. Validation on the experimental dataset (Case Western Reserve University (CWRU)) and a laboratory test dataset (Seoul National University, SNU) bearing dataset under progressive test-time noise degradation showed average accuracies of 98.56 percent and 97.99 percent, with clear-to-low-source accuracy changes of −0.17 and 0.56 percentage points, respectively. These results demonstrate robust diagnosis under severe noise contamination, supporting physics-informed multi-task-learning in real-world settings.