Real-Time Recognition of Molten Pools Based on Improved DeepLabV3+ in Keyhole Tungsten Inert Gas Welding Applications

• DOI: 10.3390/electronics13020283
• Publication Date: 2024-01-08T10:21:38Z
• Authors: Dayong Yang; Penghui Dai; Shuwan Cui; Huzhe Song; Feiyang Liu; Xuan Zhou
• Keywords: Keyhole; Robustness; Weld pool

During the Keyhole Tungsten Inert Gas (K-TIG) welding process, a significant amount of information related to weld quality can be obtained from pool and keyhole topside molten image, which provides vital basis for control quality. However, image has unstable characteristic strong arc light, leads difficulty in contour extraction. The existing segmentation algorithms cannot satisfy requirements accuracy, timing, robustness. Aiming at these problems, real-time recognition method, based on improved DeepLabV3+, identifying more accurately effectively was proposed this paper. First, MobileNetV2 selected as feature extraction network with improve detection efficiency. Then, atrous rates convolution layers were optimized reduce receptive field balance sensitivity model pools different scales. Finally, convolutional block attention module (CBAM) introduced accuracy model. experimental results verified that had fast speed higher an average intersection ratio 89.89% inference 103 frames per second. Furthermore, trained deployed system achieved performance up 28 second, thus meeting K-TIG monitoring system.