An Automatic Detection and Identification Method of Welded Joints Based on Deep Neural Network

Lei Yang,Yanhong Liu,Jinzhu Peng

doi:10.1109/access.2019.2953313

Lei Yang, Yanhong Liu + Show 1 more

Open Access

https://doi.org/10.1109/access.2019.2953313

Copy DOI

Abstract

Welding quality is an important factor to affect the performance, quality and strength of different products, and it will affect the safe production. Therefore, welding quality detection is a key process of industrial production. And the detection and identification of welded joints are the premise of welding quality detection, which could reduce the quality detection range and improve the detection precision. Welded joint identification is also important for providing information for automatic control of welding process. Faced with the complex characteristics of industrial environment, such as weak texture, weak contrast and corrosion, we propose a detection and identification method of welded joints based on deep neural network. Firstly, aimed at the problem of insufficient training samples, combined with image processing and Generative Adversarial Network (GAN), the high-quality training samples are generated. Secondly, the updating mechanism of training samples is established to guarantee that the deep neural network model could cover all samples. Finally, the detection and identification of welded joints are realized by the deep neural network which could avoid the handcrafted features of conventional machine learning methods. Experiments show that the proposed method could quickly and efficiently finish the detection and identification task of welded joints. Meanwhile, the proposed method could well solve the detection and identification problems of complex industrial environment.

Highlights

Nowadays, welding automation technology has been widely applied into much production scenes, such as automobile processing, shipbuilding and aerospace
Combined with the CycleGAN network and traditional data enhancement methods, the data set of welded joints could be expanded to train the deep neural network model
Faced with the detection and identification of welded joints, an automatic detection and identification method of welded joints is proposed based on deep convolution neural network (DCNN) models

Summary

INTRODUCTION

Nowadays, welding automation technology has been widely applied into much production scenes, such as automobile processing, shipbuilding and aerospace. L. Yang et al.: Automatic Detection and Identification Method of Welded Joints Based on Deep Neural Network main non-contact assessment methods. To detect the welded joints with different welding quality, Yang et al designed the curvature features for different weld images On this foundation, a detection method based on SVM model was proposed to detect different welded joints [23]. It has been widely applied into object classification [28], behavior prediction [29], autonomous vehicle [30] and other much application scenarios Through model training, it could well extract the high-level image features and do the classification or detection tasks. To better represent the laser welding images, Günther et al [36] proposed a deep auto-encoding neural network to realize automatic feature extraction. Two network models are trained to realize automatic welded joint assessment.

SYSTEM FRAMEWORK

Findings

CONCLUSION