摘要
太阳能在可再生能源中扮演越来越重要的角色,但是积灰和鸟粪等影响因素会降低太阳能电池板的发电效率,针对太阳能电池板的缺陷识别十分重要,为此基于改进的Faster R-CNN(faster region-based convolutional neural network)算法对太阳能电池板的缺陷识别进行研究。网络的改进内容如下:实验样本量较少,首先采取了色域转换、旋转等操作以及mosaic数据增强方法,将主干网络替换为效果更好的RestNeSt-50网络;由于检测目标中积灰和鸟粪目标尺寸相差较大,采取了目标尺寸均衡策略;为了使分类和回归任务独立学习,采用了参数不共享双分支策略;并采用了Cosine学习率策略避免网络陷入局部最小值。以上改进方法使得评价指标平均准确率(mean average precision,mAP)值从基准模型的78.91%提升至94.05%。最后成功将单个太阳能电池板从无人机拍摄的图像中提取并修正角度,同时利用改进的Faster R-CNN算法进行缺陷识别,结果表明所提方法可以更准确地识别出积灰和鸟粪等缺陷。
Solar energy is playing a more important role in renewable energy.However,the power generation efficiency of solar panels will be greatly reduced by the dust and bird droppings on the panels.Therefore,it is essential to identify the defects of solar panels.This paper investigates the defect identification of solar panels using the improved Faster R-CNN.The improvements on the model are given as follows:Due to insufficient samples,the operations such as color space conversion,rotation and mosaic data enhancement are adopted.The backbone network is replaced by a better ResNeSt-50 network.Because the sizes of dust and bird droppings are quite different,a target size equalization strategy is applied.In order to independently learn the classification and regression tasks,the task-aware disentanglement is used.Moreover,the Cosine learning rate is adopted to avoid the network falling into the local minimum.All these improvements increase the mAP value from the baseline of 78.91%to 94.05%.Finally,the single solar panel is extracted from the UAV images and the angle correction is also implemented.In addition,its defect identification is conducted using the improved Faster R-CNN.The results prove that the defects of dust and bird droppings can be more accurately identified.
作者
张文彪
马永华
白晓静
谈元鹏
皮宇啸
ZHANG Wenbiao;MA Yonghua;BAI Xiaojing;TAN Yuanpeng;PI Yuxiao(School of Control and Computer Engineering,North China Electric Power University,Changping District,Beijing 102206,China;China Electric Power Research Institute,Haidian District,Beijing 100192,China)
出处
《电网技术》
EI
CSCD
北大核心
2022年第7期2593-2600,共8页
Power System Technology
基金
中央高校基本科研业务费专项资金项目(2021MS016)。
