This study applies a YOLOv8-based object detection framework to insulator defect detection. A three-class dataset—intact insulators, broken insulators, and pollution-flashover defects—was used, with separate training and test splits for unbiased evaluation. The model converges quickly and attains mAP@50 = 0.9507, precision = 0.9268, recall = 0.9379, and F1-score = 0.9323 on the test set. Confidence analysis shows that intact and broken insulators are detected with high reliability, whereas pollution-flashover exhibits greater variability and lower confidence. Detection visualizations and activation heatmaps corroborate the quantitative metrics, indicating accurate localization and interpretable defect-relevant regions. These findings confirm that YOLOv8 provides a robust and efficient solution for automated insulator defect detection, supporting future applications in power-grid monitoring and maintenance.