基于改进Faster R-CNN的焊缝缺陷图谱识别

The Weld Defect Maps Identification Based on Improved Faster R-CNN

焊缝缺陷是锅炉、压力容器等承压特种设备使用安全的严重威胁,相控阵超声缺陷检测方法以其快速、无损等优点将逐渐成为此类设备的原位全体积焊缝缺陷检测的主要技术。然而,由于检测数据量大,缺陷形状和大小各异,缺陷识别是一项具有挑战性的任务。在本文中,提出了基于深度学习的方法来识别相控阵超声检测图谱中的缺陷,基于Faster R-CNN网络框架,根据焊缝缺陷的特点,使用带有特征金字塔网络的ResNet50作为特征提取的主干,以更好的检测缺陷图谱中的小缺陷。实验结果表明,与传统的Faster R-CNN相比,改进后的模型可以获得更好的性能。

Weld defects were serious threat to the safety of pressure-bearing special equipment such as boilers and pressure vessels.Phased-array ultrasonic flaw detection method will gradually become the main technology of in-situ full-volume weld defect detection in this kind of equipment field because of its advantages of rapidity and non-damage.However,due to a large number of testing data and the diversity of defect shapes and sizes,defect identification was a challenging task.In this paper,a deep learning-based method was proposed to identify defects in phased array ultrasonic inspection maps.Based on the Faster R-CNN network framework,according to the characteristics of weld defects,ResNet50 with feature pyramid network was used as the backbone of feature extraction,which can better detect small defects in the defect map.Experimental results show that the improved model can achieve better performance than the traditional Faster R-CNN.