YOLOv5框架下的飞机蒙皮铆接缺陷检测

Defect Detection Method for Aircraft Skin Riveting in Framework of YOLOv5

飞机蒙皮制孔及铆接质量的视觉辅助检测是飞机装配过程的关键环节,为提高其智能检测效率和精度,提出改进YOLOv5模型的飞机蒙皮铆接装配缺陷的检测方法。研究蒙皮制孔、铆接质量等,执行多种特征缺陷检测。为减少数据增强过程中计算损耗,避免训练过程中过拟合现象,引入随机数据增强策略作数据增广,采用强化学习方法确定随机搜索策略的超参数,可提升深度学习的数据驱动性能;更改损失函数为α-CIoU函数,并使用非极大抑制软计算方法(soft-NMS)进行预测框的筛选,以避免漏检情况,提高多种缺陷检测判别的准确性;最后将YOLOv5改进前、后算法进行对比实验。实验结果表明,在检测速度上单张图片检测时间可达到1.78 s,检测平均精度均值提升了16.4%,达到94.6%,精确率提升18.8%,达到90.7%,满足飞机批量化检测的精度要求和实时检测需求。

It is critical for the visual aided defection inspection of aircraft skin hole making and riveting in the aircraft assembling process.A multi-defect detection method is proposed for aircraft skin riveting using the improved YOLOv5 model to improve the efficiency and accuracy of intelligent detections.The multi-feature defects of skin hole making and riveting quality,including riveting hole aperture accuracy,riveting difference and rivet corrosion,are investigated.The random search strategy of data augmentation is implemented in order to reduce the calculation loss of the data augmentation process and avoid the training process overfitting.Reinforcement learning method is adopted to determine the hyperparameters of random search strategy which improves the data-driven performance of deep learning.The loss function is changed toα-CIoU function.Soft-NMS is adopted for screening the prediction box,so as to avoid missing detection and improve the accuracy of detection and discrimination of various defects.The improved YOLOv5 algorithm is compared and tested.The experimental results show that the cost time of single image detection is confined to 1.78 s.The mean average precision has increased by 16.4%,reaching 94.6%.The accuracy rate has increased by 18.8%,reaching 90.7%.It meets the requirements of mass detection accuracy and real-time detection for the aircraft skin hole making and riveting quality.