Saliency-aware heterogeneous independence decoupling network for mechanical transfer fault diagnosis under unseen target conditions.

To mitigate the degradation in transfer diagnostic performance caused by unseen target domains in complex industrial scenarios, a novel saliency-aware heterogeneous independence decoupling network (SA-HIDN) is proposed for mechanical transfer fault diagnosis. Initially, a heterogeneous decoupled feature extractor is constructed to align the network capacity with the distinct physical attributes of signal components. It employs a deep branch with deep resonance shrinkage to perform soft-thresholding denoising for domain-invariant fault impulses, and a shallow branch with a statistical feature layer to explicitly capture global statistical moments as domain-specific condition features. Subsequently, a statistical independence constraint based on the Hilbert-Schmidt Independence Criterion (HSIC) is introduced to mathematically force the fault and condition subspaces to be independent in the reproducing kernel Hilbert space, effectively suppressing high-order information leakage between decoupled branches. Finally, a saliency-aware adaptive fusion (SAAF) module is developed to proactively protect subtle fault signatures and achieve high-fidelity feature reconstruction through a saliency-masking compensation mechanism. Extensive experiments on planetary transmission and train transmission system datasets demonstrate that the proposed SA-HIDN achieves transfer diagnostic accuracies of 98.13% and 94.48% on two datasets, respectively. These quantitative results and mechanistic analyses confirm that the proposed method significantly outperforms state-of-the-art methods while maintaining a favorable balance between diagnostic precision and computational cost.