基于深度学习的管道焊缝超声检测缺陷识别方法

超声无损检测是目前管道焊缝质量检测应用最为广泛的一种检测手段,但是迄今为止在很大程度上依赖于训练有素的人体检查员的专业知识和判断。目前在深度学习的方法已经很好的应用于基于图像数据的管道焊缝缺陷智能检测，但是对于深度学习辅助判断超声无损检测却进展缓慢。主要原因是超声无损检测数据的复杂性（步长大、多模态、多峰分布等），神经网络训练往往出现梯度消失或爆炸的问题，而且能用于训练的标准数据集也严重匮乏。为了克服这些困难，首先引入特殊标准化方法和全连接隐含层实现了一种超声无损检测数据增强方法FMC-GAN构建虚拟数据集，再根据改进的LSTM-FCN模型并引入门函数，以此彻底克服超声无损检测数据复杂性。最后实验表明LSTM-FCN网络识别真实检测数据中的缺陷漏检率为0，各缺陷综合正确识别率高于95.6%，达到工业检测的要求，为超声无损检测智能化发展提供重要研究基础。

Defect identification method of pipeline weld ultrasonic testing based on deep learning

The method of in-depth learning has been well applied to the intelligent detection of pipeline weld defects based on image data. Still, the progress of ultrasonic nondestructive testing assisted by in-depth learning is slow. The main reason is the complexity of ultrasonic nondestructive testing data (step length, multimodality, multimodal distribution, etc.), the gradient disappears or explodes in neural network training, and the standard data sets that can be used for training are also seriously scarce. In order to overcome these difficulties, firstly, a particular standardized method and fully connected hidden layer are introduced to realize an ultrasonic nondestructive testing data enhancement method FMC-GAN to construct a virtual data set, and then according to the improved LSTM-FCN model and the introduction function, to completely overcome the complexity of ultrasonic nondestructive testing data. Finally, the experiment shows that the missing detection rate of defects in the accurate detection data identified by the LSTM-FCN network is 0, and the comprehensive correct recognition rate of each defect is higher than 95.6%, which meets the requirements of industrial detection, and provides a vital research basis for the intellectual development of ultrasonic nondestructive testing.