In order to meet the demand of fault prediction of power distribution network equipment, this paper proposes an intelligent inspection image recognition algorithm framework which combines visible light and infrared images to realize the transformation from "passive emergency repair" to "active prevention". In this method, the spatial-temporal joint feature representation is constructed, the dual-mode image is fused by using dual-stream CNN (Convolutional Neural Network), and the channel attention mechanism is introduced to adaptively weight the key features. Aiming at the scarcity of fault samples, a data enhancement strategy based on diffusion model is designed to effectively alleviate the data imbalance, A dynamic channel pruning algorithm is proposed, which can reduce the model parameters by 80% while maintaining 95% accuracy, and meet the requirements of edge deployment. Integrating LSTM (Long Short Term Memory) module to achieve time series prediction of fault probability, generating warning information 3-7 days in advance and integrating with maintenance scheduling system. The experiment is based on a self built multimodal dataset covering five types of distribution network equipment. The results show that the proposed method has an accuracy rate of 98.5% and an F1 score of 94.0%, which is significantly better than the single modal baseline model and has good practical promotion prospects.
Zhou et al. (Sun,) studied this question.