基于改进DETR的机器人铆接缺陷检测方法研究

Research on robot riveting defect detection method based on improved DETR

铆接作为铁道车辆结构件的主要连接方式,合格的铆接质量是车辆安全稳定运行的重要保证。针对现有铆接缺陷检测方法存在检测精度低、检测点位少、检测智能化水平不高等问题,提出一种基于改进DETR的机器人铆接缺陷检测方法。首先,搭建铆接缺陷检测系统,依次采集工件尺寸大、铆钉尺寸小工况下的铆接缺陷图像。其次,为了增强DETR模型在小目标中的图像特征提取能力和检测性能,以EfficientNet作为DETR中的主干特征提取网络,并将3-D权重注意力机制SimAM引入EfficientNet网络,从而有效保留图像特征层的镦头形态信息和铆点区域的空间信息。然后,在颈部网络中引入加权双向特征金字塔模块,以EfficientNet网络的输出作为特征融合模块的输入对各尺度特征信息进行聚合,增大不同铆接缺陷的类间差异。最后,利用Smooth L1和DIoU的线性组合改进原模型预测网络的回归损失函数,提高模型的检测精度和收敛速度。结果表明,改进模型表现出较高的检测性能,对于铆接缺陷的平均检测精度mAP为97.12%,检测速度FPS为25.4帧/s,与Faster RCNN、YOLOX等其他主流检测模型相比,在检测精度和检测速度方面均具有较大优势。研究结果能够满足实际工况中大型铆接件的小尺寸铆钉铆接缺陷实时在线检测的需求,为视觉检测技术在铆接工艺中的应用提供一定的参考价值。

Riveting is the main connection method for structural components of railway vehicles,and qualified riveting quality is an important guarantee for the safe and stable operation of vehicles.Aiming at the problems of low detection accuracy,few detection points,and low level of intelligent detection in existing riveting defect detection methods,a robot riveting defect detection method based on improved DETR was proposed.First,a riveting defect detection system was established,which sequentially collected riveting defect images under working conditions of large workpiece size and small rivet size.Second,in order to enhance the image feature extraction ability and detection performance of the DETR model in small targets,EfficientNet was used as the backbone feature extraction network in DETR.The 3D weighted attention mechanism SimAM was introduced into the EfficientNet network,effectively preserving the header shape information of the image feature layer and the spatial information of the rivet point area.Then,a weighted bidirectional feature pyramid module was introduced into the neck network.The output of the EfficientNet network was used as the input of the feature fusion module to aggregate the feature information at each scale,which increased the variation of different riveting defects.Finally,the regression loss function of the original model prediction network was improved by using the linear combination of Smooth L1 and DIoU,which improved the detection accuracy and convergence speed of the DETR model for defect types.The experimental results show that the improved model exhibits high detection performance,with an average detection accuracy mAP of 97.12%and a detection speed FPS of 25.4 f/s for riveting defects.Compared with other mainstream detection models such as Faster RCNN and YOLOX,the improved model has significant advantages in detection accuracy and detection speed.The research results can meet the demand for real-time online detection of small-scale rivet riveting defects in large riveted parts under actual working conditions,and provide certain reference value for the application of vision detection technology in riveting processes.