Synthetic Insulator Broken Defect Data Generation With Causal and Foggy Augmentation for Defect Detection

The insulator defect detection is essential for the reliability and safety of power transmission systems. The effective detection enables early defect identification, ensuring a stable power supply and protecting infrastructure. However, insufficient data, especially in new inspection scenarios, limits model generalization. In this work, instead of using the traditional image augmentation and 2-D image composition methods to improve the number of training data, we adopt a domain randomization generative model in virtual 3-D rendering software to create synthetic images. Rather than training with randomly generated synthetic data, we introduce causal and 3-D foggy augmentations to enhance model explainability and generalization performance. Overall, our methods achieve a 6.8% improvement over the real-world data training baseline, 6.1% over the model with the same foreground causal instances, and 2.9% over random generation, using much fewer foreground causal instances. In cross-domain validation, we achieve 42.3% and 8.2% higher $\text {mAP}_{{50}}$ than the baseline and random generation, respectively. The increased focus on causal elements and reduced focus on noncausal elements in the attention maps, along with detailed ablation studies, further validate the effectiveness of our approach.