Deep learning assisted vision inspection of resistance spot welds. (February 2021)

• Record Type: Journal Article
• Authors: Dai, Wei; Li, Dayong; Tang, Ding; Jiang, Qin; Wang, Dong; Wang, Huamiao; Peng, Yinghong
• Is Part Of: Journal of manufacturing processes. Volume 62(2021)
• Page Start: 262
• Page End: 274
• Publication Date: 2021-02
• Subjects: Resistance spot welding -- Quality vision inspection -- Deep learning model -- YOLOv3 -- Small object detection; Production management -- Data processing -- Periodicals; Manufacturing processes -- Periodicals; Procestechnologie; Productietechniek; Production -- Gestion -- Informatique -- Périodiques; Fabrication -- Périodiques; Manufacturing processes; Production management -- Data processing; Periodicals; 670.5
• DOI: 10.1016/j.jmapro.2020.12.015
• Languages: English
• ISSNs: 1526-6125

Abstract

Abstract: Spot welds are extensively used on autobodies and also a key factor affecting the performance of automobiles. Automatic detection of spot welding based on machine vision provides an effective way for car body welding quality control. Considering the traditional image processing methods are greatly disturbed by the environment and have unsatisfying robustness, a network model for small object detection is proposed to detect the position and quality of the spot welding of the car body. Based on the existing You Only Look Once (YOLOv3) model, the proposed model has three novel improvements. Firstly, the lightweight network MobileNetV3 is introduced to replace the backbone network of YOLOv3 to ensure accuracy and real-time performance. Secondly, to improve the model's ability for small object detection, a new feature pyramid network (FPN) with efficient cross-scale connections is proposed, which allows easy and fast multiscale feature fusion. Finally, considering the shortcomings of intersection and union ratio (IoU) loss, complete IoU (CIoU) loss is used to improve convergence speed and regression accuracy. Moreover, novel data augmentation is used to enrich the dataset during the model training. Quantitative results on the spot welding dataset show that the proposed approach achieves successful results for resistance spot welding vision inspection.