IEVDET: A Multi-Granular Visual Event Detection Framework for Power Equipment Based on Cross-Modal Alignment

Accurate interpretation of visual scenes in electric power inspection is essential yet challenging due to dense equipment layouts and fine-grained spatial semantics. General vision-language models often struggle to resolve ambiguous references and align visual elements with domain-specific descriptions. We propose IEVDET (Incremental and Event-centric Visual Detection for Electric power Tasks), a multi-granular framework that integrates cross-modal alignment and an incremental layout graph to capture semantic and spatial relations. It also employs structured QA template generation for interpretable supervision. Trained on over 13,000 annotated image-event pairs from 500kV and 330kV substations, IEVDET outperforms baselines like GPT-4o, Gemini Pro, and Claude 3 Opus, achieving BLEU-4 of 51.6, CIDEr of 88.5, Visual Accuracy of 89.2%, and Spatial Accuracy of 85.7%. Its implementation, SGGM (State Grid General Multimodal model), shows strong adaptability to real-world inspection tasks. These results demonstrate the effectiveness of domain-adaptive multimodal modeling for intelligent electric power systems.