航空学报 > 2022, Vol. 43 Issue (4): 524961-524961   doi: 10.7527/S1000-6893.2021.24961

面向激光焊接缺陷识别的可解释性深度学习方法

刘天元1, 郑杭彬1, 杨长祺2, 鲍劲松1, 汪俊亮1, 顾俊3   

  1. 1. 东华大学 机械工程学院 智能制造研究所, 上海 201600;
    2. 上海航天精密机械研究所, 上海 201600;
    3. 上海市激光技术研究所, 上海 200235
  • 收稿日期:2020-11-11 修回日期:2020-11-30 发布日期:2021-02-24
  • 通讯作者: 鲍劲松 E-mail:bao@dhu.edu.cn
  • 基金资助:
    国家自然科学基金(51905091);中央高校基本科研业务费专项资金-东华大学研究生创新基金(CUSF-DH-D-2020053)

Explainable deep learning method for laser welding defect detection

LIU Tianyuan1, ZHENG Hangbin1, YANG Changqi2, BAO Jinsong1, WANG Junliang1, GU Jun3   

  1. 1. Institute of Intelligent Manufacturing, College of Mechanical Engineering, Donghua University, Shanghai 201600, China;
    2. Shanghai Spaceflight Precision Machinery Institute, Shanghai 201600, China;
    3. Shanghai Institute of Laser Technology, Shanghai 200235, China
  • Received:2020-11-11 Revised:2020-11-30 Published:2021-02-24
  • Supported by:
    National Natural Science Foundation of China (51905091); Fundamental Research Funds for the Central Universities and Graduate Student Innovation Fund of Donghua University (CUSF-DH-D-2020053)

摘要: 激光焊接在航空领域具有广泛的应用场景,基于视觉的激光焊接缺陷识别对于产品质量的提高至关重要。针对当前基于深度学习的激光焊接缺陷识别方法存在可解释性差的问题,提出了一种融合多尺度特征的类激活映射(MSF-CAM)方法。在训练阶段,以VGG16为骨架模型并将监督信息施加于多个尺度以促进模型对多尺度特征的学习。在测试阶段,对输出类别在多个尺度上的激活图进行叠加,并以此作为模型的判断依据。多尺度特征的融入不但增强了模型的可解释性,而且还提高了激光焊接缺陷识别的准确性。试验结果表明:MSF-CAM在测试集上的准确率为98.12%,识别单幅图像耗时8.28 ms。此外,MSF-CAM可以从边缘、轮廓这种初级特征的角度对模型的决策依据提供人类更容易理解的解释。

关键词: 激光焊接, 缺陷识别, 深度学习, 可解释性, 多尺度

Abstract: Laser welding has a wide range of applications in the aerospace industry. Vision-based laser welding defect detection is crucial to the improvement of product quality. To overcome the problem of poor explainability in current deep learning-based laser welding defect detection methods, a method of Class Activation Mapping by Incorporating Multiscale Features (MSF-CAM) is proposed. In the training phase, VGG16 is used as the backbone network, and supervisory information is applied to multiple scales to facilitate learning of multiscale features. In the testing phase, the activation maps of the output categories on multiple scales are superimposed and used as the basis for the judgment by the model. Integration of multiscale features not only enhances explainability of the model, but also improves accuracy of laser welding defect detection. The test results show that the accuracy of MSF-CAM on the test set is 98.12% and the time used by the method to identify a single image is 8.28 ms. In addition, MSF-CAM can provide more understandable explanation for the decision-making basis of the model from the perspective of such primary features as edge and contour.

Key words: laser welding, defect recognition, deep learning, explainability, multiscale

中图分类号: