【摘要】 常规基于图像和深度学习的扁线绕组焊点缺陷检测方法需要大量缺陷焊点数据用于训练模型，而实际生产线上扁线绕组缺陷焊点样本十分匮乏，存在小样本和样本类别不平衡的问题。该文基于特征比对提出一种扁线绕组焊点缺陷检测方法，首先使用目标检测模型检测绕组端面上的所有焊点位置，再复用模型的骨干网络提取每个焊点区域的特征，并与训练过程中通过提取正常焊点的中间层特征构建的特征库中的特征作相似性比对，进而检测缺陷并定位缺陷位置。采用产线实拍照片制作了扁线绕组焊点的数据集，并在数据集上完成了实验与对比分析。结果表明，提出的方法在不依赖缺陷样本的情况下，能够准确实现产线实拍图像上焊点的缺陷检测，分类准确率达到98.4%，对缺陷样本的检测精度达97.0%，召回率达100%，算法的受试者工作特征曲线下的面积（AUROC）指标在图像级和像素级上分别达到97.4%、98.0%，满足工业需求。更多还原

【Abstract】 Conventional defect detection methods for hairpin winding weld joints based on image and deep learning require amount of defective weld joints data for model training, while there is a severe scarcity of such samples in actual production, resulting in challenges related to small sample sizes and imbalanced sample classes.Some existing studies have attempted to artificially create defective weld joints by controlling welding conditions,but this approach not only incurs additional costs but also fails to fully represent the types of defects that occur in real production. When the model encounters unknown defects, it cannot effectively detect or identify them.Therefore, this paper proposes a defect detection method for hairpin winding weld joints independent of defect data. The core of the method is to accurately identify defective weld joints by comparing the feature matching degree between weld joints to be detected on the production line and normal weld joints. The proposed method consists of two parts during the training phase. On one hand, it uses images of hairpin windings with annotated weld joints locations to train an object detection model. This process involves freezing the backbone network Resnet50, which is pretrained on ImageNet, and only training the detection head responsible for outputting location parameters. On the other hand, the backbone network of the model is used to extract intermediate layer features of a large number of normal weld joints, and a feature bank is constructed based on these features. In the detection phase, the trained object detection model is first used to detect the positions of weld joints in the images to be detected, but it does not perform defect detection on the potentially defective weld joints. Then, the backbone network of the model is reused to extract intermediate layer features from each weld joint region, which are compared with the features in the feature bank constructed during training to detect defects and locate the defective regions. The experimental data for this paper were collected from images taken by cameras on the automatic welding production line for hairpin windings, consisting of 111 images, with approximately 20 weld joints in each image.Based on whether the morphology of the weld joints met the standard, the weld joints were classified into two categories: normal and abnormal, and the positions and categories of the weld joints were annotated on the images.To evaluate the effectiveness of the proposed defect detection method, abnormal labels were cropped from the original images, and defect segmentation masks were further annotated. Comparative experimental results between the conventional classification network and the proposed method show that the accuracy, precision, and recall of the conventional classification network for defect detection are 89.1%, 87.9%, and 90.6%, respectively. In contrast,the proposed method improves these metrics to 98.4%, 97.0%, and 100%, respectively, and is also able to segment defect regions, achieving a pixel-level AUROC of 98.0%. Additionally, ablation experiments were conducted on the proposed method to explore the impact of different levels and sampling rates used in the feature bank construction process. This paper addresses the issue of the scarcity of defective hairpin winding weld joint samples in real industrial production, which makes it difficult to directly predict defect categories using object detection algorithms. By proposing a feature comparison-based method for defect detection that does not rely on defect data, the experimental results demonstrate that the method meets industrial requirements and solves the problem of limited defect samples in industrial production, which provides a new approach to the task of defect detection for hairpin winding weld joints in industrial applications.更多还原