Abstract: Abstract:To address the concealment, power-off requirement, and low accuracy of traditional corrosion detection for substation grounding grids, this paper proposes an online evaluation system. First, an electrochemical corrosion kinetics and equivalent circuit model for non-uniform strata is established to reveal degradation patterns. Second, a multi-frequency weak electrical signal injection topology is designed, and Hilbert-Huang transform is applied to extract anti-interference corrosion fingerprints. Finally, a lightweight graph convolutional network (GCN) is constructed for quantitative inversion decision-making. Experiments confirm that the method requires no power outage, achieves 96.88% cross-section loss detection accuracy, and locates defects within 0.46 m, enabling transparent dynamic perception of hidden infrastructure.