Abstract
This work addresses fault diagnosis in power communication networks (PCNs) under three practical challenges-limited labels, evolving topology, and noisy alarms-by proposing GCFR, a graph-contrastive fault representation framework. GCFR learns robust node embeddings via lightweight, semantics-preserving multi-view augmentations (feature masking, edge dropping, and mild topology perturbations) and a mini-batch InfoNCE objective, followed by a small classification head for deployment-friendly inference. To withstand structural drift, we incorporate a time-consistency regularizer that stabilizes representations across successive snapshots. We further enhance robustness with domain-aware alarm-propagation features and a lightweight wavelet denoising module that mitigates timestamp jitter, missing events, and spurious alarms. For reliable decision-making in operations, we provide approximate uncertainty quantification and selective prediction, enabling explicit control of the coverage-risk trade-off. Experiments on two benchmark graphs (GridGraph and Texas2000) and an anonymized real-world slice show that GCFR maintains strong recall with scarce labels and consistently outperforms strong supervised and signal-processing baselines under topology perturbations and alarm noise. Overall, GCFR offers a low-overhead, deployment-oriented solution that couples label efficiency with calibrated, risk-aware predictions for PCN fault diagnosis.