Abstract: To address the limitations of traditional insulator fault detection methods: specifically the inadequate capture of details and poor performance in identifying small targets, a novel dual-stream attention-based approach is proposed. This method combines a deep supervised attention generative adversarial network (DSAGAN) with the YOLOv8 object detection algorithm. In the proposed DSAGAN, infrared and visible light images of insulators are fused using an attention mechanism embedded within the generator of the generative adversarial network (GAN) to enhance fusion quality. The generator and discriminator form an adversarial network, where the least squares (LS) loss function is employed instead of the conventional cross-entropy loss. This substitution helps preserve finer image details and improves the stability of the DSAGAN. The fused images are then subjected to fault detection using the YOLOv8 object detection algorithm. Experimental results demonstrate that the five evaluation indexes of the insulator images fused by DSAGAN outperform those of seven other fusion methods. YOLOv8 object detection algorithm achieves a mean average precision (mAP) of 0.917 and 0.639 for detecting insulator damages, flashovers, glass losses, and polymer contaminations at thresholds of 0.5 and 0.5 to 0.95, respectively, representing improvements of 0.026 and 0.08 compared to YOLOv5. Furthermore, the fault recognition rates of fused images for different types of insulator faults surpass those of single infrared or visible light images. The average recognition rate reaches 93%, marking improvements of 6.25% and 4.5% over infrared and visible light images, respectively.