基于注意力机制和多路径YOLOv7的光伏组件故障诊断研究

Research on Photovoltaic Module Fault Diagnosis Based on Attention Mechanism And Multipath YOLOv7

中文关键词:  Conv_SE注意力机制  多路径优化  YOLOv7目标检测  光伏组件的故障诊断  ELAN结构模块

英文关键词:Conv_SE attention mechanism  multipath optimization  YOLOv7 target detection  fault diagnosis of photovoltaic modules  ELAN structural modules

中文摘要:

      针对无人机对光伏组件的故障（热斑和遮挡）诊断准确率较低和检测速度较慢的问题，提出了使用改进后的ELAN_MSE(Efficient Layer Aggregation Networks_Multipath Selective Enhancement)模块替换YOLOv7(You Only Look Once version 7)网络的ELAN模块，提高了有效特征学习的速率和准确率。首先，将YOLOv7的主干特征提取网络中的ELAN结构模块引入通道注意力机制SE(Squeeze-and-Excitation)，提高特征提取的精确率；其次,增加了多路径卷积支路，实现中间层的跳跃结构连接，能够对目标特征图进行不同尺度的特征学习，提升故障缺陷识别的精度和检测速度。通过对数据增强后的光伏组件缺陷数据集进行实验验证，改进的YOLOv7算法与传统的YOLOv7和单发多框检测(Single Shot MultiBox Detector,SSD)算法对比，F1 Score分别提高了3.96%和5.98%，检测速度分别提高了1.43 ms和2.25 ms，为光伏组件故障检测提供了更有效的算法。

英文摘要:

      In order to solve the problems of low accuracy and slow detection speed of photovoltaic module faults (hot spots and occlusion) by unmanned aerial vehicle(UAV),a method of replacing efficient layer aggregation networks（ELAN）module of you only look once verston 7（YOLOv7）network with improved efficient layer aggregation networks_multipath selective enhancement（ELAN-MSE）module is proposed,which improves the rate and accuracy of effective feature learning.Firstly,ELAN structure module in the backbone feature extraction network of YOLOv7 is introduced into channel attention mechanism squeeze-and excitation（SE）to improve the accuracy rate of feature extraction.Secondly,multipath convolutional branch is added to realize skip structure connection in the middle layer,and feature learning of different scales can be carried out on the target feature map to improve the accuracy and speed of fault defect detection.Through the experimental verification of the data enhanced photovoltaic module defect dataset,compared with the traditional YOLOv7 and single shot multiBox detector(SSD)algorithms,the improved YOLOv7 algorithm can increase F1 Score by 3.96% and 5.98%,and the detection speed can increase by 1.43 ms and 2.25 ms,respectively.It provides a more effective algorithm for photovoltaic module fault detection.