26 March 2026 in IEEE Transactions on Dielectrics and Electrical Insulation
With the widespread adoption of power cables in urban distribution networks, accurate aging assessment remains challenging due to non-uniform aging patterns and severe class imbalance across aging levels. Traditional models, relying on identical training-test distribution assumptions, often fail when encountering unseen operational data. To overcome these limitations, this paper proposes a novel cable aging assessment framework applying feature map-based knowledge distillation (FMKD) and adaptive layer regeneration (LR). The key innovation lies in a dual-phase learning mechanism:(1) The knowledge distillation (KD) module extracts latent cross-category correlations between minority and majority classes, effectively mitigating catastrophic forgetting during updates, (2) A parameter transfer-based LR strategy enables rapid adaptation to new-task data (minority samples) while preserving old-task knowledge. Comprehensive experimental results demonstrate that the model simultaneously optimizes new-task identification (unseen samples) and old-task retention (known samples), achieving satisfactory accuracy across varying data ratios.