基于卷积神经网络的焊缝缺陷超声波识别研究

Research on Ultrasonic Identification of Weld Defects Based on Convolutional Neural Network

摘要: 桥梁钢结构焊接构件中，焊缝质量是影响构件力学性能的主要因素之一。针对超声检测回波信号具有步长大、多模态、多峰分布等特点和焊缝缺陷图像进行深度学习识别已经取得了很好的效果。本文用超声探伤仪对70个带焊缝缺陷的T型管节点和18个对接钢板试件、25个对接钢管试件进行超声检测。建立包含8800张超声检测回波信号图像的数据集作为网络训练、验证和测试的对象。调整卷积神经网络MobileNet-v2的结构层、参数和权重，利用卷积神经网络MobileNet-v2进行训练和测试，得到T型管节点超声检测回波信号识别的卷积神经网络的平均识别准确率分别为91%。计算识别精确率和召回率可作为调整后卷积神经网络识别性能的评价指标。

Abstract: In the welded components of bridge steel structure, the quality of the weld is one of the main factors affecting the mechanical properties of the components. Good results have been achieved in the deep learning recognition of ultrasonic testing echo signals with the characteristics of large step length, multi-modality, multi-peak distribution and weld defect images. In this paper, 70 T-pipe joints, 18 butt steel plate specimens, and 25 butt steel pipe specimens with weld defects were tested by ultrasonic flaw detector. A data set of 8800 ultrasonic echo signal images was established as the object of network training, validation and testing. The average recognition accuracy of the convolutional neural network MobileNet-v2 is 91% after adjusting the structural layer, parameters and weights of the convolutional neural network MobileNet-v2 for training and testing. The calculation accuracy and recall rate can be used as evaluation indicators for the recognition performance of the adjusted convolutional neural network.