摘要
为实现航空发动机涡轮叶片射线检测自动化、智能化,有效改善传统射线检测费时费力、效率低下等问题,开展基于无监督学习的涡轮叶片X-ray图像缺陷检测方法研究。基于无监督生成对抗网络,提出一种适用于航空发动机涡轮叶片X-ray图像的缺陷检测算法;构建由生成网络、判别网络和附加自编码网络组成的深度卷积生成对抗网络,设计重构损失、判别损失、编码损失及中间编码损失,并利用4种损失的加权之和构造目标函数;利用完好涡轮叶片X-ray图像进行模型训练,基于训练得到的生成网络建立航空发动机涡轮叶片X-ray图像缺陷检测模型。研究了输入图像大小、编码长度和重构损失对缺陷检测模型性能的影响。结果表明:模型在输入图片像素尺寸为128像素×128像素、编码长度为600、重构损失为L2的情况下检测性能最佳,area under curve(AUC)可达到0.911。该缺陷检测算法能够实现实际生产缺陷零漏检的严苛技术指标,但误检率(>62.1%)较大,作为辅助检测手段应用于实际生产可将人工检测效率提高1.6倍。
To achieve the automation and intelligence of the radiographic inspection of turbine blades in aeroengines,and to effectively address the time-consuming,labor-intensive,and inefficient problems in traditional radiographic inspection methods,a research initiative was undertaken to develop a defect detection method for X-ray images of turbine blade based on unsupervised learning.A defect inspection algorithm suitable for X-ray images of aeroengine turbine blades was proposed based on an unsupervised generative adversarial network.It consisted of a generator network,a discriminator network,and an extra encoder network.Reconstruction,discrimination,encoding,and intermediate encoding loss were designed,and the weighted sum of the four losses was used to construct the objective function.Using non-defective X-ray images for model training.A defect inspection model for X-ray images of aeroengine turbine blades was established based on the trained generator network.The effects of input image size,encoding size,and type of reconstruction loss on the performance of the defect detection model were studied.Results showed that the proposed model with an input image size of 128 pixel×128 pixel,600 encoding size,and L2 reconstruction loss can achieve an area under curve(AUC)of 0.911.The defect inspection algorithm can realize strict technical indicators of zero missing rate for actual production,but the false detection rate(>62.1%)was relatively high.As an auxiliary detection method applied in actual production,it can improve the manual detection efficiency by 1.6 times.
作者
王栋欢
于艾洋
肖洪
WANG Donghuan;YU Aiyang;XIAO Hong(School of Power and Energy,Northwestern Polytechnical University,Xi’an 710072,China;Shenyang Engine Research Institute,Aero Engine Corporation of China,Shenyang 110015,China)
出处
《航空动力学报》
北大核心
2025年第6期247-258,共12页
Journal of Aerospace Power
基金
中国航空发动机集团产学研合作项目(HFZL2019CXY008-1)。
关键词
无监督学习
缺陷检测
射线检测
涡轮叶片
生成对抗
卷积自编码
unsupervised learning
defect detection
radiographic testing
turbine blades
generative adversarial
convolutional autoencoder
作者简介
王栋欢(1993-),男,工程师,博士,主要从事智能检测、航空发动机数字化智能化技术等方面研究。E-mail:wangdonghuan66@mail.nwpu.edu.cn。
