Knowledge-guided multimodal reasoning framework for intelligent power grid maintenance

Power grid infrastructure maintenance requires reliable integration of heterogeneous data sources including historical fault records, technical documents, and visual inspection imagery. Existing approaches face challenges in ensuring output completeness, consistency verification, and systematic multimodal data integration-critical requirements for high-stakes operational decisions. This paper presents an integrated multimodal reasoning system designed to address these reliability requirements through systematic quality assurance mechanisms. The system architecture comprises three components: (1) a document processing pipeline constructing a structured knowledge base from 10,247 heterogeneous maintenance records (8,198 for training, remaining for validation and testing); (2) a retrieval-augmented reasoning module implementing a Recursive Augmented Thinking (RAT) mechanism with three-dimensional self-evaluation (slot completeness, confidence assessment, logical consistency verification) tailored to power grid fault diagnosis requirements; (3) an attention-enhanced visual detection module for identifying equipment anomalies. Validation on 120 real-world fault cases demonstrates that the RAT mechanism achieves 95% output completeness and 0.82 confidence score through iterative refinement (approximately 98% convergence within 3 iterations), compared to 65% completeness and 0.55 confidence for single-pass generation. Visual detection on 15,648 inspection images achieves 0.952 mAP@.5 with maintained robustness under challenging conditions (87.4% performance retention under occlusion, distance, and adverse weather). These results indicate practical deployment potential within the validated scope of operational power grid maintenance.