There are few defective sample images among patrol images in the power system, leading to an imbalance between normal samples and defective samples,and It is impossible to use algorithms such as deep learning to study the fault detection of transmis sion lines further.At present,various of image generation methods based on deep machine le arning have problems such as low resolution and unconspicuous defect features,which makes i t difficult for the generated sample images to meet the needs of researchers.This paper pro poses a PCA weighted average (diverse integration,DI ) generation method based on (ensemble learning,EL). Experiments were carried out with normal and defective transmission line insulator images. The experimental results show that the quality of the generated images is obviou s,which can be effectively used in the construction of a professional sample library for pow er systems, providing big data support for subsequent related research,and also proposed a new and feasible research idea for this field.