Physical-Structure Evidence Learning for Robust Recognition of Overhead Ground Wire Damage Under Variable Imaging Conditions

Overhead ground wires exhibit a regularly repeated stranded configuration, providing a symmetry-related perspective on their local structural organization. Local damage disrupts this regularity, introducing deviations in strand continuity, contour organization, or surface morphology. These damage-associated structural deviations provide important cues for distinguishing different damage types. Reliable recognition is challenging because different damage types may present similar local appearances, while motion blur, haze, illumination variation, and other imaging degradations can further obscure damage-related cues. To improve recognition robustness under variable imaging conditions, this paper proposes a method based on physical-structure evidence learning. Here, physical-structure evidence refers to a structure-related representation learned under hierarchical supervision and structural constraints, rather than an independently validated damage-localization map. The structural differences among dispersed broken strands, non-dispersed broken strands, and conductor wear are represented through two hierarchical states: broken-strand presence and conditional spatial dispersion. A shared broken-strand evidence field is constructed from shallow visual features to encode localized damage-associated structural evidence, and the Broken-Strand Evidence Dispersion Index (BEDI) is introduced to characterize its spatial distribution. Physical-structure constraints and local evidence intervention are further incorporated to encourage structure-related representations. A two-stage training strategy is then employed, consisting of structural-evidence learning on clean samples followed by environment-adaptation training while preserving the learned structural representation. Experiments on a constructed overhead ground wire dataset show that the resulting CBEF-Env model achieves a Clean Acc of 93.18%, a Clean F1 of 93.37%, an OOD Avg of 92.79%, and an OOD Worst of 91.91%. Compared with the ResNet18 baseline, these metrics are improved by 2.76, 2.82, 2.71, and 3.10 percentage points, respectively. Compared with AugMix, RIDG, and RDM, the OOD Avg is improved by 1.96, 3.00, and 2.91 percentage points, while the OOD Worst is improved by 2.05, 3.38, and 3.74 percentage points, respectively. These results indicate improved robustness across the controlled synthetic degradations evaluated in this study, providing a basis for further validation on independently acquired field data.