首页 | 本学科首页   官方微博 | 高级检索  
     

基于 EfficientDet 的风机叶片缺陷检测方法
引用本文:辛 彦,吴国新,左云波.基于 EfficientDet 的风机叶片缺陷检测方法[J].电子测量技术,2022,45(5):124-131.
作者姓名:辛 彦  吴国新  左云波
作者单位:1.北京信息科技大学现代测控技术教育部重点实验室100192;
基金项目:国家重点研发计划(2020YFB1713200);北京市教委科研计划(KM202011232001)项目资助。
摘    要:受工作环境恶劣等原因影响,风机叶片常会出现裂纹、凹坑等缺陷。针对当前常用目标检测算法对风机叶片小尺寸缺陷检测准确率低的问题,提出一种基于EfficientDet算法的风机叶片缺陷检测方法。首先采集图像数据并建立Pascal VOC格式的风机叶片缺陷图像数据集,然后对EfficientDet算法中的主干特征提取网络进行改进,减少向下采样次数并调整有效特征层从而增强主干特征提取网络对小尺寸缺陷的检测能力,同时为特征融合网络增加融合路径提升算法的多尺度特征融合能力,选用FReLU作为激活函数实现像素级空间信息建模,并通过Mosaic数据增强和Focal Loss损失函数增加小尺寸缺陷样本对于检测器的贡献。在建立的风机叶片缺陷图像数据集上的测试结果表明改进后的算法模型平均类别精度达到了96.15%,相较于原版的EfficientDet提升了3.77%,对小目标的检测性能有明显提升。

关 键 词:深度学习  目标检测  EfficientDet  图像处理  风机叶片

Defect detection method of wind turbine blade based on EfficientDet
Xin Yan,Wu Guoxin,Zuo Yunbo.Defect detection method of wind turbine blade based on EfficientDet[J].Electronic Measurement Technology,2022,45(5):124-131.
Authors:Xin Yan  Wu Guoxin  Zuo Yunbo
Affiliation:The Ministry of Education Key Laboratory of Modern Measurement and Control Technology, Beijing Information Science & Technology University, Beijing 100192, China
Abstract:Affected by the poor working environment and other reasons, the fan blades often have defects such as cracks and pits. Aiming at the low accuracy of the current common target detection algorithms for the detection of small-size defects of the fan blades, a fan blade defect detection method based on the EfficientDet algorithm is proposed. . First collect image data and establish a wind turbine blade defect image data set in Pascal VOC format, and then improve the backbone feature extraction network in the EfficientDet algorithm to reduce the number of downsampling and adjust the effective feature layer to enhance the backbone feature extraction network for small-size defects Detection capability; At the same time, the multi-scale feature fusion capability of the fusion path enhancement algorithm is added to the feature fusion network. The algorithm uses FReLU as the activation function to achieve pixel-level spatial information modeling, and uses Mosaic data enhancement and Focal Loss loss function to increase small-size defect samples for Contribution of the detector. The test results on the established defect image data set of fan blades show that the improved algorithm model has an average category accuracy of 96.15%, which is 3.77% higher than the original EfficientDet, and the detection performance of small targets has been significantly improved.
Keywords:deep learning  object detection  EfficientDet  image processing  wind turbine blade
本文献已被 维普 等数据库收录!
点击此处可从《电子测量技术》浏览原始摘要信息
点击此处可从《电子测量技术》下载全文
设为首页 | 免责声明 | 关于勤云 | 加入收藏

Copyright©北京勤云科技发展有限公司  京ICP备09084417号