基于ResNet的航空发动机制件表面缺陷分类研究

Research on Surface Defects Classification of Aeroengine Parts Based on ResNet

摘要: 针对现有航空发动机制件缺陷分类所存在的检测效率低、适用范围有限等缺陷，提出了一种基于ResNet-18算法的缺陷分类方法。该算法使用深度残差网络提取缺陷特征，并通过修改网络结构适应于不同的缺陷种类。在实验过程中，首先对原始的钢带表面图像进行预处理，使用裁剪、旋转角度等方法扩增数据集。然后使用PyTorch深度学习框架搭建卷积神经网络模型，并将增强后的图片数据输入到模型中进行训练，实现对缺陷的分类。最后，使用东北大学钢带表面缺陷公共数据集进行训练与评估。本文算法在东北大学钢带表面缺陷公共训练集上的分类准确率为97.33%，在测试集上的准确率达到95.36%，为真实工业场景下缺陷的分类提供了可能。

Abstract: A defect classification method based on the ResNet-18 algorithm is proposed to address the defects of the existing defect classification for aeroengine parts with low detection efficiency and limited applicability. The algorithm uses a deep residual network to extract defect features and adapts to different defect types by modifying the network structure. In the experimental process, the original steel strip surface images are first preprocessed and the dataset is augmented using cropping and rotation angles. Then a convolutional neural network model was built using the PyTorch deep learning framework, and the enhanced image data were input into the model for training to achieve classification of defects. Finally, a public dataset of steel strip surface defects at Tohoku University is used for training and evaluation. The classification accuracy of the algorithm in this paper is 97.33% on the Northeastern University public training set of steel strip surface defects and 95.36% on the test set, which provides the possibility of classifying defects in real industrial scenarios.