Journal Article

Inspection of Welding Defect Based on Multi-feature Fusion and a Convolutional Network

Fan, Junfeng, Liu, Yanhong, Yang, Lei, Huo, Benyan

2021

Overview

Robot welding is a basic but indispensable technology for many industries in modern manufacturing. However, many welding parameters affect welding quality. During the real welding process, welding defects are inevitably generated that affect the structural strengths and comprehensive performances of different welding products. Therefore, an accurate welding defect recognition algorithm is necessary for automatic robot welding to assess the effects of defects on structural properties and system maintenance. Much work has been devoted to welding defect recognition. It can be mainly divided into two categories: feature-based and deep learning-based methods. The detection performances of feature-based methods rely on effective image features and strong classifiers. However, faced with weak-textured and weak-contrast welding images, the realization of strong image feature expression still faces a certain challenge. Deep learning-based methods can provide end-to-end detection schemes for welding robots. Nevertheless, an effective deep network model relies on much training data that are not easily collected during real manufacturing. To address the above issues regarding defect detection, a novel welding defect recognition algorithm is proposed based on multi-feature fusion for accurate defect detection based on X-ray images. To improve network training, an effective data augmentation process is proposed to construct the dataset. Combined with transfer learning, the multi-scale features of welding images are acquired for effective feature expression with the pre-trained AlexNet network. On this basis, based on multi-feature fusion, a welding defect recognition algorithm fused to a support vector machine with Dempster–Shafer evidence theory is proposed for multi-scale defect detection. Experiments show that the proposed method achieves a better recognition performance in terms of detecting welding defects than those of other related recognition algorithms.

• Publisher: Springer US,Springer Nature B.V
• Subject: Algorithms / Automatic welding / Characterization and Evaluation of Materials / Classical Mechanics / Control / Deep learning / Defects / Dynamical Systems / Engineering / Image acquisition / Image contrast / Inspection / Machine learning / Manufacturing / Robots / Solid Mechanics / Support vector machines / Training / Vibration / Weld defects / Welding parameters