Diffusion-Enhanced Anomaly Perception for IoT-Enabled Fault Prediction in Multi-Center Power Dispatch Platforms
Abstract
The high-availability operation of IoT-enabled geographically distributed multi-center power dispatch platforms demands reliable early fault perception from heterogeneous sensing and surveillance data to support proactive failover preparation and resilience management. However, two bottlenecks hinder this goal: visual fault indicators in surveillance footage are extremely compact (as small as 15 × 10 pixels), and severe sample scarcity under rare cross-center failover and extreme weather conditions limits model robustness. To address these challenges, we propose a diffusion-enhanced anomaly perception framework integrating Edge-conditioned Diffusion Preprocessing (EDP) and Scenario-Adaptive Generation (SAG), coupled with a YOLO-based detector. EDP recovers micro-scale fault cues via edge-regularized diffusion reconstruction, while SAG synthesizes physically consistent fault scenarios under rare dispatch conditions through dual-condition control and low-rank adaptive fine-tuning. Experiments demonstrate that EDP reduces LPIPS by 31.2%, and the unified EDP–SAG–YOLO framework achieves 67.9% fault indicator recall (+15.2% over baseline) with performance degradation under adverse conditions limited to 8.3%, providing a reliable technical foundation for intelligent fault prediction in multi-center power dispatch platforms.