Fault‐Sensitive Relational Representation and Heterogeneous Global–Local Learning for Oil‐Immersed Power Transformer Fault Diagnosis

Overlapping dissolved gas signatures, inadequate representation of fault‐relevant relationships, and limited interaction between complementary diagnostic patterns remain major challenges in DGA–based fault diagnosis of oil‐immersed power transformers. To address these issues, a diagnosis‐oriented framework is developed around fault‐sensitive relational representation and heterogeneous global–local learning. Fault‐sensitive weighted Gramian angular field incorporates training‐derived feature discriminability together with sample‐specific pairwise gas‐feature differences into GASF and GADF construction, so that diagnostically informative relationships receive greater emphasis during encoding. The two representations are subsequently processed by a heterogeneous dual‐branch mobileViT. Its transformer‐oriented branch models global correlations among gas features, whereas the multiscale depthwise‐separable convolutional branch extracts local differential patterns. Cross‐branch information is coordinated through a complementary gating mechanism, which derives adaptive fusion weights from branch responses, discrepancy, and feature interaction. Model configuration is handled by diagnosis‐constrained opposition‐based golden jackal optimization, with classification performance, feature separability, prediction consistency under physically consistent gas perturbations, model complexity, and parameter feasibility considered jointly. Experiments show that the proposed method achieves favorable recognition performance and outperforms the compared deep‐learning models. Independent cross‐dataset evaluation also provides evidence of generalization under changes in DGA data distribution. Overall, the framework offers a practical solution for intelligent DGA‐based fault diagnosis of oil‐immersed power transformers.