航空发动机鼓风机机匣表面缺陷检测

Surface Defect Detection of Aero Engine Blower Casing

本文针对航空发动机鼓风机机匣表面缺陷,设计了一种新型深度卷积神经网络及优化算法缺陷检测。新型深度卷积神经网络(Convolution Neural Network,CNN)由三部分组成:骨干卷积神经网络生成不同尺度的特征图;多尺度特征融合网络(Multilevel-Feature Fusion Network,MFN)将多个尺度特征进行融合,以便包含更多的定位细节;区域建议网络(Region Proposal Network,RPN)生成感兴趣区域,每个感兴趣区域由一个分类器和一个锚框回归器组成的检测器产生最终的检测结果。然后分析了鼓风机机匣表面缺陷的特点,设计了新的检测损失函数。最后利用公司采集的包含两种典型缺陷(划痕和污秽)的数据集,利用单张GTX 1080ti显卡对网络进行训练和测试,测试结果表明两种典型缺陷检测平均可达到68.9%m AP、87.6%的精准率和35FPS的检测速度。

A new deep convolutional neural network and its optimization algorithm are designed to detect the surface defects of the aero-engine blower casing.A new deep convolutional neural network consists of three parts:the backbone convolution neural network generates feature graphs of different scales;Multi-scale feature fusion network(MFN)fuses multiple scale features to include more localization details.Region Proposal Network(RPN)generates regions of interest(ROI),and each ROI is composed of a detector including a classifier and an anchor box regression to generate the final detection results.Then,the characteristics of the surface defected of the blower casing are analyzed,and a new detection loss function is designed.Finally,the data set collected by the company is used to train and test the network on a single GTX 1080TI graphics card,the deep learning model can detect the two typical surface defects(scratches and filth)of the blower casing with an average of 68.9%mAP,87.6%Precision and 35 FPS detection speed.