基于ARGAN表面阴影预处理与迁移学习风电机组叶片故障识别

ARGAN-based Surface Shadow Preprocessing and Transfer Learning Fault Identification of Wind Turbine Blade

叶片是风电机组获取风能的重要部件。风电机组运行环境恶劣,叶片表面易出现剥落、开裂等多种故障。叶片故障的及时识别能够保障机组安全稳定运行。然而无人机拍摄叶片图像数据常出现光照不均,导致故障被阴影遮挡的情况,阴影边缘很难与叶片边缘区分开从而干扰计算机视觉算法,导致故障图像识别准确率低。对此利用专注递归生成对抗网络(Attentive Recurrent Generative Adversarial Network,ARGAN)对原始图像的阴影进行处理,减少阴影区域对算法识别分类的干扰,完整保留叶片故障区域,再将处理后的图像送入卷积神经网络,利用迁移学习思想进行训练学习。通过与传统图像处理方法以及条件生成对抗网络、深度卷积对抗网络对比,可得出ARGAN效果最佳,可使处理后的故障图像识别准确率达到89%,同时降低了均方根误差,从而验证了这种方法的可行性。

Blade is key for wind turbine to generate wind energy.Due to the bad operation environment,the blade surface is prone to peeling,cracking and other faults.In order to ensure the safe and stable operation of the unit,it is very important to detect blade fault in time.However,the uneven illumination,which often occurs in UAV image data,covers the fault by shadow.The overlap between shadow edge and blade edge interrupts computer vision algorithm and lowers fault image recognition accuracy.This paper processes original image shadow by focused Attentive Recurrent Generative Adversarial Network(ARGAN)to reduce its interruption with algorithm recognition and classification and to detect the whole blade fault area.Then,we sent the processed image to the convolution neural network,and used the migration learning idea for training and learning.Compared with traditional image processing methods,conditional generation countermeasure network and deep convolution countermeasure network,the method of ARGAN has been verified to have the best effect,increasing the recognition accuracy of fault image to 89%and reducing the root mean square error.