Physics-guided deep learning-based domain-shift robust alignment network for bearing fault diagnosis

Bearing fault diagnosis under varying severity and operating conditions remains a critical challenge for safe industrial deployment. Traditional domain adaptation methods often rely on large over-parameterised backbones or adversarial training, suffering from negative transfer and poor interpretability. This paper proposes DS-RAN, a lightweight physics-constrained domain-shift robust alignment network that achieves state-of-the-art cross-severity transfer. The core contribution is a theoretically grounded combination of a two-parameter affine physics-constrained adapter that explicitly models dominant fault progression effects (amplitude scaling and DC offset) and a spectral-normalised convolutional blocks ensuring Lipschitz-continuous feature extraction with a class-conditional maximum mean discrepancy for fine-grained distribution alignment. Extensive experiments on CWRU and the non-stationary University of Ottawa datasets demonstrate 99.6% accuracy on severe cross-severity tasks (0.007” → 0.021”) and 88.0% under extreme speed inversion. The proposed method outperformed ResNet-50, EfficientNet and other SOTA DA methods by 8–10% while being significantly smaller and lightweight. Rigorous ablation and statistical analysis confirm the indispensable roles of each component.