Wire composition and shielding gas flow monitoring based on image and spectrum multimodal network. (August 2020)

• Record Type: Journal Article
• Authors: Chen, Xiaoyu; Sun, Bei; Zhang, Chenrui; Lou, Xiaotian; Zhao, Zhuang; Han, Jing
• Is Part Of: Measurement. Volume 160(2020)
• Publication Date: 2020-08
• Subjects: Wire component -- Shielding gas flow -- Multimodal -- Image and spectrum; Weights and measures -- Periodicals; Measurement -- Periodicals; Measurement; Weights and measures; Periodicals; 530.8
• DOI: 10.1016/j.measurement.2020.107797
• Languages: English
• ISSNs: 0263-2241

Abstract

Highlights: A multimodal monitoring network is proposed to online analyze composition of welding wire and flow of shielding gas. A loss function is improved to balance training samples with different scales of ground truths. The proposed method is proved to have adequate accuracy on welding monitoring tasks. Abstract: The existing monitoring of arc welding quality mainly uses single sensor data, which has poor monitoring reliability and limited scope. Multi-sensor fusion provides multiple dimensional data for quality monitoring, showing impressive advantages over single sensor. As the abnormality of welding wire composition and shielding gas flow affect the quality of welding seam, a multimodal convolution neural network based on fusion of image and spectrum feature (ISFNet) is designed to better monitor the welding process. ISFNet contains image and spectrum branches to extract more robust feature for online welding analyzing. In addition, an improved loss function Normalized Mean Square Error reduces the interference of unbalanced sample magnitude of labels, so that all the samples can be uniformly trained. The experiments show that ISFNet has higher accuracy and robustness than other methods, the minimum RMSE values of wire composition and shielding gas flow prediction results can reach 0.0499 and 0.0586.