改进YOLOv5的光伏组件热斑及遮挡物检测  

Detection of photovoltaic module hot spots and shelters on improved YOLOv5

作　　者：魏卓航 林培杰 陈志聪 吴丽君 卢箫扬 程树英 WEI Zhuohang;LIN Peijie;CHEN Zhicong;WU Lijun;LU Xiaoyang;CHENG Shuying(Institute of Micro-Nano Devices and Solar Cells,College of Physics and Information Engineering,Fuzhou University,Fuzhou,Fujian 350108,China)

机构地区：[1]福州大学物理与信息工程学院,微纳器件与太阳能电池研究所,福建福州350108

出　　处：《福州大学学报（自然科学版）》2023年第1期33-40,共8页Journal of Fuzhou University(Natural Science Edition)

基　　金：福建省自然科学基金资助项目(2021J01580);福建省工信厅基金资助项目(82318075);福州市科技计划资助项目(2021-P-030,2021-P-059)。

摘　　要：针对光伏组件热斑若未及时发现处理,会严重影响光伏组件及阵列正常运行的问题,为了有效检测光伏阵列热斑,提出一种基于YOLOv5框架的深度学习热斑检测方法.首先,采用像素加权平均法融合红外和可见光图像作为检测对象,实现同时对光伏组件热斑和遮挡物的检测,并初步分析热斑成因.其次,改进模型框架,在轻量级网络MobileNetV3-large的基础上,融合坐标注意力机制,设计更轻量、更高效的MobileNetCA作为特征提取网络.然后,针对训练中正负样本数量极不平衡的情况,更换损失函数为变焦距损失函数,达到训练中突出正例的效果.同时,改进模型anchor box目标框生成算法,使生成的目标框与实际标注框更一致.实验结果表明,改进后的模型mAP为88.9%,较原YOLOv5s模型提升了3.8%,且模型参数量仅为原模型的48.6%.Photovoltaic(PV)modules can produce hot spot effect due to various conditions,such as mismatch,cracks,short-circuit,partial coverage,which leads to the abnormal operation of the PV modules or PV array.In this paper,a deep learning(DL)based target detection method using YOLOv5 framework is proposed to the detect the hot spots on PV array.Firstly,the pixel weighted average method is used to fuse infrared and visible images as detection objects,so that the hot spots and coverings can be simultaneously detected and located,and then the cause of hot spots can also be preliminarily analyzed.Secondly,a more lightweight and efficient MobileNetCA(MobileNetV3-large with CoordAttention)by integrating the coordinate attention mechanism on the basis of lightweight network MobileNetV3-large is designed,which are applied to replace the feature extraction network of the proposed model framework.In addition,the loss function is replaced with Varifocal Loss function to enhance positive samples when positive and negative samples are unbalance in training phase.Furthermore,the model anchor box generation algorithm is also improved,so that the generated target box is more consistent with the actual labeled box.Experimental results show that the mAP of the proposed model is 88.9%,which is 3.8%higher than that of the original model YOLOv5s,and the parameter size of the proposed model is only 48.6%of that of the YOLOv5s.

关 键 词：光伏组件 热斑检测 红外可见光融合 YOLOv5 MobileNetV3-large 变焦距损失函数 坐标注意力机制 

分 类 号：TM91[电气工程—电力电子与电力传动]