Defect detection in laser-welded battery tabs using deep learning with imbalanced class datasets.

Abstract

As secondary batteries become more popular, it becomes increasingly important to ensure their safety. Many welding techniques are used in secondary batteries, and the welding process is a key part of battery safety. This high production volume requires welding vision technology to keep up. Traditional machine vision has difficulty detecting novel and atypical defects, and as production volume increases, human visual inspection becomes increasingly limited. Chapter 1 provides a description of the production process for secondary cells with laser welding on a mass production line. And in the secondary battery electrode foil joining process, the difficulty of laser welding technology is explained. Machine vision was used to monitor the process to overcome the difficulties, but its limitations were clear. Therefore, deep learning technology is being explored. This article will introduce the contents of this exploration. In addition, an object detection model was used to detect defects. The overall details of this object detection model will also be introduced. In Chapter 2, A deep learning-based detection method for irregular defects is proposed. The data set used vision images from a real production line. In terms of collecting datasets for deep learning training, it is difficult to get enough data because atypical defects are rare in mass production. Therefore, the dataset used in this paper is characterized by a small number of data and a non-uniform class distribution, which is a difficult data configuration for a deep learning model to learn. However, to overcome this difficulty, various methods to improve accuracy using a combination of data augmentation techniques were proposed. To detect defects, The YOLOv7-tiny object detection model achieved a test mAP@0.5 of 0.977, demonstrating its high accuracy. Its impressive speed shines through with 3.5 ms inference, 2.3 ms NMS, and 5.9 ms total processing per 640x640 image on a batch size of 10 (62 images). In conclusion, atypical defects were detected in real time with sufficient accuracy. In addition, the developed method successfully detected defect types that occur less frequently and effectively detected atypical defects that are difficult to detect with machine vision methods.