基于改进EfficientNet的海上风机叶片早期缺陷检测及分类  

Early defect detection and classification of offshore wind turbine blades based on improved EfficientNet

作　　者：付军军 陈林[2] 童华敏 舒征宇 王灿 Fu Junjun;Chen Lin;Tong Huamin;Shu Zhengyu;Wang Can(College of Electrical Engineering&New Energy,China Three Gorges University,Yichang 443002,China;State Grid Yichang Power Supply Company,Yichang 443000,China)

机构地区：[1]三峡大学电气与新能源学院,宜昌443002 [2]国网宜昌供电公司,宜昌443000

出　　处：《电子测量技术》2024年第16期110-119,共10页Electronic Measurement Technology

基　　金：国家自然科学基金(52107108)项目资助。

摘　　要：针对海上风机叶片小尺寸缺陷检测准确率低、分类效果较差的问题,提出一种基于EfficientNet的改进海上风机叶片表面早期缺陷检测模型。首先,在EfficientNet特征提取网络中引入非对称卷积替换普通3×3卷积,增强了卷积核骨架信息,提高网络提取缺陷信息的能力;其次提出一种混合空间通道注意力模块聚焦空间和通道信息,结合BiFPN特征融合模块对不同深度的语义信息进行特征融合,提升算法多尺度特征融合能力;最后引入Focal-EIOU和Focal Loss损失函数计算位置损失和分类损失,提高定位精度,解决模型训练过程中正、负图像样本的比例失衡的问题。实验结果表明,本文所提算法模型平均精度均值为97.6%,对风机叶片表面早期缺陷的检测性能有明显提升。Aiming at the problem of low accuracy and poor classification effect of small size defect detection of offshore wind turbine blades,an improved early defect detection model of offshore wind turbine blade surface based on EfficientNet is proposed.Firstly,the asymmetric convolution is introduced into the EfficientNet feature extraction network to replace the ordinary 3×3 convolution,which enhances the convolution kernel skeleton information and improves the ability of the network to extract defect information.Secondly,a hybrid spatial channel attention module is proposed to focus on space and channel information,and the BiFPN feature fusion module is used to fuse the semantic information of different depths to improve the multi-scale feature fusion ability of the algorithm.Finally,Focal-EIOU and Focal Loss functions are introduced to calculate the position loss and classification loss,so as to improve the positioning accuracy and solve the problem of imbalance between positive and negative image samples in the model training process.The experimental results show that the average accuracy of the proposed algorithm model is 97.6%,and the detection performance of early defects on the surface of wind turbine blades is significantly improved.

关 键 词：缺陷检测 深度学习 EfficientNet 海上风机叶片 注意力机制 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术] TN957.52[自动化与计算机技术—计算机科学与技术]