Learning semantic-specific visual representation for laser welding penetration status recognition

作     者：LIU TianYuan BAO JinSong ZHENG HangBin WANG JunLiang YANG ChangQi GU Jun LIU TianYuan;BAO JinSong;ZHENG HangBin;WANG JunLiang;YANG ChangQi;GU Jun

作者机构：Institute of Intelligent ManufacturingCollege of Mechanical EngineeringDonghua UniversityShanghai 201600China Shanghai Spaceflight Precision Machinery InstituteShanghai 201600China Shanghai Institute of Laser TechnologyShanghai 200235China 

出 版 物：《Science China(Technological Sciences)》 (中国科学（技术科学英文版）)

年 卷 期：2022年第65卷第2期

页      面：347-360页

核心收录：

学科分类：080503[工学-材料加工工程] 08[工学] 080203[工学-机械设计及理论] 0805[工学-材料科学与工程（可授工学、理学学位）] 0802[工学-机械工程] 080201[工学-机械制造及其自动化] 

基　　金：supported by the Fundamental Research Funds for the Central Universities and Graduate Student Innovation Fund of Donghua University(Grant No.CUSF-DH-D-2020053) 

主　　题：label semantic attention mechanism word embedding convolutional neural network laser welding penetration status 

摘      要：The degree of penetration can directly reflect the forming quality of laser welding. The fine-grained feature of the molten pool/keyhole image brings challenges to the vision-based laser welding penetration status recognition. In this paper, a novel knowledge-and-data-hybrid driven recognition model is proposed for solving the problem of difficult learning of discriminative visual features of molten pool/keyhole images. In addition, a label semantic attention mechanism(LSA) is designed with three modules: representation of image visual feature, representation of labels semantic feature, and generation of label semantic attention. For learning discriminative features in visual space, LSA uses discriminative information in label semantics to guide the convolutional neural network. The experimental results show that the proposed LSA method has faster convergence and higher accuracy than the traditional attention mechanism. Further comparative experiments reveal that LSA is less dependent on the amount of training data and model complexity. The results of visualization experiments show that the visual features learned by the proposed method are more discriminative.