基于卷积神经网络的齿轮表面缺陷检测方法  

Gear Surface Defect Detection Methods Based on Convolutional Neural Network

作　　者：Abdou Yahouza Maman Rabiou 闫娟 杨慧斌 刘向前 Abdou Yahouza Maman Rabiou;Yan Juan;Yang Huibin;Liu Xiangqian(School of Mechanical and Automotive Engineering,Shanghai University of Engineering Science,Shanghai 201620,China)

机构地区：[1]上海工程技术大学机械与汽车工程学院,上海市201620

出　　处：《农业装备与车辆工程》2022年第6期154-157,共4页Agricultural Equipment & Vehicle Engineering

摘　　要：为了准确、定量地检测齿轮的表面缺陷,提出了2种采用深度学习算法的卷积神经网络模型(CNN)检测齿轮的表面缺陷。该方法对卷积核和卷积层进行了优化,并使用最大池化代替了大步长卷积,以扩大接收场的大小并以高分辨率捕获齿轮的精细特征,改进了分类器模块。使用和不使用数据扩充的Alex Net和Res Net模型都涉及通过操纵原始数据创建新数据点的过程,此过程无需添加新照片即可增加深度学习(DL)中训练图像的数量,适用于数据集较小的情况;收集生产过程中齿轮的200个图像的自数据集,通过灰度处理、调整图像大小获得清晰的目标齿轮轮廓并识别齿轮特征点。实验结果表明,经过训练的数据增强模型对Res Net和Alex Net分别具有95.83%和97.94%的最佳效果。与目前仅基于机器视觉的齿轮表面缺陷检测技术相比,该方法具有很好的通用性,获得了最高的识别率。To accurately and quantitatively detect the surface defects of gear, this paper proposes two models of convolutional neural network(CNN) with Deep Learning algorithm for surface defect detection. In this proposed method, the convolution kernel and convolution layer are optimized, and the maximum pooling is used instead of the large step convolution to enlarge the size of the receptive field and capture the fine features of the gear at high resolution. The classifier module is improved. AlexNet and ResNet models with and without data augmentation involve the process of creating new data points by manipulating the originals data. This process increases the number of training images in Deep learning(DL) without the need to add news photos, applicable in case of small dataset. A self-dataset of 200 images of gears during production are collected then, which are processed by grayscale processing and image resizing to obtain a clear target gear contour and identify the gear feature points. Experimental results show that the trained models with data augmentation give the best results of 95.83% and 97.94% for ResNet and AlexNet respectively. The proposed method has good versatility on detection of gear surface defects and achieved the highest recognition rate than the current gear surface defect detection technologies based on machine vision only.

关 键 词：AlexNet 深度学习 齿轮 表面缺陷检测 数据扩充 ResNet 

分 类 号：TG86[金属学及工艺—公差测量技术]