Abstract: Machine inspection has gradually been widely used in power inspection scenarios. Insulator is an important component for maintaining safe, reliable and stable operation of power systems, and it is important to accurately and effectively detect defects in insulators. The identification of insulator broken defect is a significant task of insulator defect detection. To address the problems of small and imbalanced insulator defect data samples, poor model generalization ability and inaccurate data annotation, this paper proposes a framework for automatic broken defect sample generation based on domain randomization and sample image quality assessment method, which achieves good performance in the domain adaptation of insulator broken defect detection from virtual domain to real domain.The image annotation data and broken defect 3D models for open source use are generated by this method. The domain randomization data generation method proposed in this paper first generates a structurally adjustable insulator umbrella disk model based on procedural modeling, and a procedural texture model containing common color and texture information of ceramic insulators is based on texture noise model.Furthermore, an insulator broken defect cutting module is built based on mesh noise model. Then, the complete insulator structure, texture model, defect structure, background information, and scene objects are generated by domain randomization. In the process of image rendering and automatic annotation part, camera alignment and visibility are first used to automatically generate and adjust the shooting points and camera parameters. Then, the data labeling class determination method based on light projection method is proposed, the image rendering pass corresponding to the instance is set for image rendering. Finally, the batch data generation is completed. In addition, 3000 virtual data generated by domain randomization are used to train the model without modifying the structure and model parameters of YOLO V5 network, and the model is tested on 300 real insulator defect images. The identification accurancy of normal insulators can reach 97.8% and the recall rate reaches 92.1%; the identification accurancy of defect insulators can reach s 79.0% and the recall rate reaches 75.9%. The inference results of the detection model on each class are better than the model trained on 400 real images. The image quality assessment method proposed in this paper takes into account the similarity with the real domain data and the independence of the samples in the dataset, and the obtained evaluation results are substituted into the loss function weight calculation to further improve the inference results. The accuracy of defect insulator identification is 85.3%, and the recall rate is 77.8%.