Physics-guided cross-domain adaptation: a hierarchical hybrid transformer framework with contrastive learning for...

The domain shift problem induced by variable working conditions severely constrains the cross-domain generalization capability of data-driven fault diagnosis models. Existing methods lack collaborative modeling of the multi-scale time-frequency characteristics of vibration signals at the feature extraction level. They also neglect the constraint guidance of fault physical mechanisms on the feature alignment process at the domain adaptation level. To address these deficiencies, this paper proposes a physics-guided cross-domain adaptation framework-a Hierarchical Hybrid Transformer network with Contrastive Learning (PgHHT-CL). The framework comprises three key designs. At the feature encoding level, a hierarchical hybrid Transformer encoder is constructed. It achieves collaborative extraction of transient impulse components and periodic modulation components in vibration signals through gated interactive fusion of local convolutional branches and global self-attention branches across multiple abstraction levels. At the domain adaptation level, a physics-guided cross-domain contrastive learning strategy leverages the order-invariance relationship between fault characteristic frequencies and rotational frequency from bearing dynamics prior knowledge to constrain the construction of cross-domain positive and negative sample pairs. The feature alignment process is thereby required to satisfy physical consistency beyond statistical distribution matching. At the training optimization level, a joint optimization objective integrates classification loss, cross-domain contrastive loss, and physical consistency loss, with a progressive weight adjustment strategy to ensure stable convergence of multi-task learning. Extensive cross-condition transfer experiments on two public bearing datasets from Case Western Reserve University and Paderborn University show that PgHHT-CL achieves average diagnostic accuracies of 94.94 ± 0.32% and 90.26 ± 0.50%, respectively, attaining the highest mean accuracy across all 18 transfer tasks among the representative state-of-the-art baselines selected for comparison. The framework also exhibits notable robustness under large domain shift and strong noise conditions. Ablation experiments and feature visualization analyses further validate the effectiveness and physical interpretability of the physics-guided strategy and hierarchical hybrid architecture.