基于光伏组件图像特征的故障检测方法  

FAULT DETECTION METHOD BASED ON IMAGE FEATURES OF PHOTOVOLTAIC MODULES

作　　者：殷孝雎 于金池 郝志鹏 潘雪 Yin Xiaoju;Yu Jinchi;Hao Zhipeng;Pan Xue(School of New Energy,Shenyang Institute of Engineering,Shenyang 110136,China;Liaoning Solar Energy Research and Application Co.,Ltd.,Shenyang 110135,China)

机构地区：[1]沈阳工程学院新能源学院,沈阳110136 [2]辽宁太阳能研究应用有限公司,沈阳110135

出　　处：《太阳能学报》2025年第4期273-279,共7页Acta Energiae Solaris Sinica

基　　金：辽宁省应用基础研究计划(2022JH2/101300251);辽宁省教育厅科学研究项目(LJKZ1101)。

摘　　要：针对集中式光伏电站地形复杂、面积广、光伏组件故障识别较困难的情况,提出基于YOLOv8模型改进的光伏组件故障识别检测方法。基于Backbone结构采用渐进特征金字塔(AFPN)融合不同层级的图像提取多尺度信息,增强上下文信息的融合。通过在Neck结构添加无参数注意力机制(SimAM),由能量函数推断出特征图中的三维注意力权重,轻量化地提高模型表征能力。取代每个池化层和每个跨步卷积层而建立SPD-Conv卷集神经网络,提高光伏组件图像中出现热斑、黑边和划痕等小目标特征的故障识别能力。实验结果表明,改进模型召回率和精确率分别达到78.7%和84.9%,平均精度mAP50和mAP50-95分别达到86%和57.9%,实现对光伏组件故障的识别与定位,验证该模型的正确性和有效性。Centralized photovoltaic power stations have complex terrain and wide areas,making it difficult to identify photovoltaic modules faults.An improved photovoltaic modules fault identification and detection method based on the YOLOv8 model is proposed.Based on the Backbone structure,the asymptotic feature pyramid network(AFPN)is used to fuse images at different levels to extract multi-scale information and enhance the fusion of contextual information.By adding a parameterless attention mechanism(SimAM)to the Neck structure,the three-dimensional attention weight in the feature map is inferred from the energy function,which lightweightly improves the model′s representation ability.Instead of each pooling layer and each strided convolution layer,the SPD-Conv convolutional neural network is established to improve the fault identification ability of small target features such as hot spots,black edges and scratches in photovoltaic modules images.Experimental results show that the recall rate and accuracy rate of the improved model reach 78.7%and 84.9%respectively,and the average precision mAP50 and mAP50-95 reach 86%and 57.9%respectively.The identification and location of photovoltaic module faults are achieved and the correctness and effectiveness of the model is verified.

关 键 词：光伏组件 深度学习 目标检测 卷积神经网络 改进YOLOv8 注意力机制 

分 类 号：TP391.413[自动化与计算机技术—计算机应用技术]