Research on CycleGAN sample generation method for dropper fault detection
Aiming at the problems of unbalanced data samples,few defect samples and difficulty in accurately reflecting the effectiveness of the identification algorithm when the depth learning algorithm is used to identify the dropper defects of the catenary,a method for generating dropper defect samples based on the improved cycle generative adversarial network(CycleGAN)model is proposed.Firstly,the residual blocks are replaced with dense convolutional blocks in the generator,so that the generatived model behaves more stably and converges faster.Then,a coordinate attention mechanism is added after the generator convolutional layer and the dense convolutional block to make the generated dropper defect sample more clearer.Finally,the common defect dropper data is migrated to the normal dropper data to generate a sample of the dropper defect.The simulation results show that the proposed method generates clearer images than deep convolutional GAN (DCGAN )algorithm and CycleGAN algorithm,and the finally generated samples can replace the real samples.
万方数据
维普
