Currently, power equipment defect detection predominantly relies on single-modal models, which exhibit significant limitations in capturing comprehensive device features and coping with environmental variations and noise interference. To address these issues, we propose a defect detection model that combines the features of visible light images (VIS) and thermal infrared images (TIR). Specifically, we first initialize the Differential Evolution (DE) population using a chaos-enhanced opposition learning strategy and introduce Gaussian variation, then the enhanced DE is used to select features extracted from various layers of the backbone network, and search for suitable fusion operations to integrate the features from different modalities. By fully leveraging the complementary advantages of multi-modal information, our method demonstrates significant improvements in noise suppression and detection feature capture, thereby markedly enhancing the accuracy and robustness of power equipment defect detection. We validate the performance on a self-constructed multi-modal insulator string defect detection data set and compared it with other multi-modal models. Experimental results indicate that the proposed algorithm outperforms its counterparts.