融合迁移学习的AlexNet神经网络不锈钢焊缝缺陷分类

PDF下载HTML

Title:

Welding defect classification of stainless steel based on AlexNet neural network combined with transfer learning

作者:

陈立潮¹, 闫耀东¹, 张睿¹, 傅留虎², 曹建芳^1,3

1. 太原科技大学 计算机科学与技术学院，山西 太原 030024;
2. 山西省机电设计研究院 机械产品质量监督检验站，山西 太原 030009;
3. 忻州师范学院 计算机科学与技术系，山西 忻州 034000

Author(s):

CHEN Lichao¹, YAN Yaodong¹, ZHANG Rui¹, FU Liuhu², CAO Jianfang^1,3

1. School of Computer Science and Technology, Taiyuan University of Science and Technology, Taiyuan 030024, China;
2. Mechanical Product Quality Supervision and Inspection Station, Shanxi Mechanical and Electrical Design & Research Institute, Taiyuan 030009, China;
3. Department of Computer Science and Technology, Xinzhou Teachers University, Xinzhou 034000, China

关键词:

不锈钢焊缝缺陷分类; 卷积神经网络; 图像预处理; AlexNet模型; 迁移学习; 数据增强; 焊缝数据集; 深度学习

Keywords:

classification of weld defects in stainless steel; convolutional neural network; image preprocessing; AlexNet model; the migration study; data enhancement; weld data set; deep learning

分类号:

TP391.4

DOI:

10.11992/tis.202005013

摘要:

针对不锈钢焊缝缺陷特征提取存在主观单一性和客观不充分性等问题，提出一种融合迁移学习的AlexNet卷积神经网络模型，用于不锈钢焊缝缺陷的自动分类。首先，由于不锈钢焊缝缺陷数据较为缺乏，通过采用迁移学习对网络前3层冻结，减少网络对输入数据量的要求；对后2层卷积层提取的特征信息批量归一化(batch normalization, BN)，以加快网络的收敛速度；并使用带泄露线性整流(leaky rectified linear unit, LeakyReLU)函数对抑制神经元进行激活，从而提高模型的鲁棒性和特征提取能力。结果表明，该模型最终达到了95.12%的准确率, 相比原结构识别精度提高了9.8%。验证了改进后方法能够对裂纹、气孔、夹渣、未熔合和未焊透5类不锈钢焊缝缺陷实现高精度分类。相比现有方法，其识别面更广，精度更高，具有一定的工程实践意义。

Abstract:

In order to solve the problems of subjectivity and objectivity in feature extraction of stainless steel weld defects, an AlexNet convolutional neural network model based on transfer learning is proposed for automatic classification of stainless steel weld defects. First, due to the lack of stainless steel weld defect data, the first three layers of the network are frozen by transfer learning, which reduces the requirement of the network on the input data. In order to speed up the convergence of the network, the batch normalization (BN) of the feature information extracted from the two latter layers of convolution is carried out. The LeakyReLU function is used to activate the features in the negative interval so as to improve the robustness of the model and the ability of feature extraction. The results show that the accuracy of the model is 95.12%, and the recognition accuracy is 9.8% higher than that of the original structure. It has been verified that the improved method can classify five kinds of stainless steel weld defects such as crack, blowhole, slag inclusion, incomplete fusion, and incomplete penetration with high precision. Compared to the existing methods, this method has a wider recognition area, higher accuracy, and certain engineering significance.