Research On Weld Defect Classification Based On Deep Convolutional Neural Network

Welding technology plays a crucial role in the production of pressure vessels,but failure to detect welding defects in a timely manner may lead to disasters.X-ray testing,as a practical non-destructive testing(NDT)method,has been widely used in industrial weld defect detection.The defect recognition technology based on X-ray images has also made rapid progress with the development of artificial intelligence.However,current research mainly focuses on feature recognition of weld defects,and few studies consider classifying weld defects based on industry standards.And in the process of identifying weld defects,there are also few scholars studying the problem of weld position positioning.Based on these issues,the specific research content of this article is as follows:(1)This article first proposes a welding seam location method based on Hough line detection and K-means clustering,which locates the welding seam area from the original X-ray film image.To facilitate the subsequent dataset cutting work and the recognition of the final entire weld seam image.(2)This article cuts the defects in weld seam images into small pixel blocks as samples for deep learning,and divides weld seam defects into four categories based on industry standards for weld seam quality evaluation and the shape characteristics of weld seam defects.Subsequently,based on the issue that a certain type of weld defect image cannot be enhanced using rotation methods for data,this paper proposes an improved generative adversarial network(GAN)based on residual network(Res Net)for data enhancement of a single defect,which has better data generation ability compared to the basic GAN model.This article constructs a new dataset of weld defects.(3)In this paper,two kinds of deep convolution neural network(DCNN)models are built through a convolution block including convolution layer,batch normalization layer and Re LU activation function layer.And based on the evaluation indicators of traditional classification models,a new evaluation indicator for the classification of weld defects is proposed,called the rate of weld defect misdetection.This article uses a five fold cross validation to compare the generalization ability of two models,and verifies the effectiveness of the new weld defect dataset in improving accuracy through these two models.(4)On the basis of the new dataset,this article adopts a global maximum pooling layer dimensionality reduction approach that preserves the feature information of each feature channel based on the optimal model among the two models.The channel attention mechanism is used to enhance the feature extraction of the front-end network of the optimal model,and thus the best model for weld defect classification in this article is obtained.The optimal model can achieve an average accuracy of 92.27% in the 5-fold cross validation,which is 3.16% and 3.73% higher than the traditional Alex Net classification model and VGG classification model,respectively.At the same time,the optimal model also achieved the optimal rate of weld defect misdetection.(5)Finally,this article uses the optimal model to traverse the entire weld seam negative image through multiple sliding windows of different sizes to obtain the recognition effect of the entire weld seam negative image defects.At the same time,the recognition results of the original weld seam image and the recognition results of the weld seam image that only retains the weld seam area are compared,which indirectly proves the necessity of the proposed weld seam position positioning algorithm.