Research on Weld Pool egmentation method for MAG Welding based on lightweight convolution Neural Network

https://elibrary.ru/HYHWHT

Abstract

During the welding process, the quality of the weld can be monitored by observing the morphology of the molten pool. However, interference such as arc light and spatter can affect the extraction of the molten pool morphology. This paper proposes a method for molten pool segmentation based on a lightweight convolutional neural network. First, a molten pool image acquisition platform was set up, and 292 images were collected. Secondly, a lightweight fully convolutional neural network (FCN) and a Deep Feature Aggregation Network (DFANet) were constructed. The experimental results show that among the FCN models, FCN-8 has the best segmentation performance. After using the Deep Feature Aggregation Neural Network, its segmentation performance was further improved. The accuracy of this network is 99.88%, the accuracy of the molten pool category is 92.17%, the intersection over union (IoU) for the molten pool is 85.89%, and the average IoU is 92.89%. Compared to FCN-8, it has improved by 3.37% in the average IoU metric. The use of DFANet for molten pool morphology extraction has practical promotion value.