Multimodal large-model-driven intelligent inspection, maintenance decision, and closed-loop control for power transmission and substation equipment

Abstract

Reliable inspection and maintenance of power transmission and substation equipment require degraded optical observations and heterogeneous monitoring data to be converted into timely, verifiable operating actions. This study aims to develop an edge-deployable framework that integrates multimodal perception, health reasoning, maintenance decision, and safety-supervised closed-loop control. A residual dual-attention network reconstructs and fuses visible-light and infrared images, while an entropy-gated cascade invokes a large vision-language model only for uncertain regions. Operating signals, dissolved-gas indicators, acoustic spectra, images, and maintenance records are aligned through a multimodal transformer and causal event graph; adaptive-temperature knowledge distillation then transfers structured maintenance and control reasoning to a compact student model. In controlled prototype tests, the framework improved defect-recognition accuracy by 8.7 percentage points over the lightweight baseline and reduced computation by 35.1% relative to always-on large-model inference. Across 240 held-out hardware-in-the-loop episodes, command success, feedback completion, and safety-constraint satisfaction reached 96.3%, 97.1%, and 99.5%, respectively, while normalized target risk decreased by 28.0%. The proposed framework establishes a practical and auditable pathway from multimodal inspection evidence to bounded automated control.