Self-Supervised Multimodal Data Fusion for Knowledge-Enhanced Feature Representation in Power Transformer Defect Risk Assessment

Résumé fourni par la source

Power transformers are critical components in electrical power systems, where accurate defect risk assessment is essential for maintaining grid stability. Traditional approaches rely on single-modal data or simple feature-level fusion, limiting their ability to capture complex defect patterns and provide interpretable risk assessments. This paper proposes a novel self-supervised multimodal data fusion framework that integrates visual information from defect images with semantic information derived from risk knowledge texts. This approach employs multitask collaborative self-supervised pretraining to deeply explore cross-modal correlations through concept-guided prototype contrastive learning and mutual information maximization strategies. A knowledge factor extraction module is developed to identify and incorporate critical domain knowledge, thereby guiding intra-modal feature refinement. Subsequently, a hierarchical cross-modal interaction fusion network with knowledge-enhanced attention mechanisms enables deep feature interaction. Finally, a knowledge-guided multi-perspective pooling mechanism is implemented to generate highly discriminative and interpretable fused feature representations. Experimental results demonstrate that the method effectively mines deep inter-modal associations, fully integrates domain knowledge, and significantly improves the accuracy and interpretability of risk identification without requiring extensive manual annotations.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.