Defect Reconstruction for Class-imbalanced Power System Defect Recognition

Performing fault diagnosis is an important routine to keep power systems functioning properly. Since most facilities of power systems are located in the wild, unmanned aerial vehicles (UAV) are used to collect potentially damaged components of power systems by taking pictures. Those pictures are categorized into a certain type to take corresponding actions to repair the damaged components. It is vital to classify collected images accurately. However, the collected images distribute in a class-imbalanced style, which degrades the performance of the classifier if directly used for training. In this paper, we make use of the generative adversarial network (GAN) to generate extra images for classes that have fewer images. Our method achieves decent improvements on 4 different scenes, showing the effectiveness of GAN-generated images on the class-imbalanced power system defect classification task.