Real-time K-TIG welding penetration prediction on embedded system using a segmentation-LSTM model

Abstract

Keyhole tungsten inert gas (K-TIG) welding is capable of realizing single-sided welding and double-sided forming and has been widely used in medium and thick plate welding. In order to improve the accuracy of automatic weld identification and weld penetration prediction of robot in the process of large workpiece welding, a two-stage model is proposed in this paper, which can monitor the K-TIG welding penetration state in real time on the embedded system, called segmentation-LSTM model. The proposed system extracts 9 weld pool geometric features with segmentation network, and then extracts the weld gap using a traditional algorithm. Then these 10-dimensional features are input into the LSTM model to predict the penetration state, including under penetration, partial penetration, good penetration and over penetration. The recognition accuracy of the proposed system can reach 95.2%. In this system, to solve the difficulty of labeling data and lack of segmentation accuracy, an improved LabelMe capable of live-wire annotation tool and a novel loss function were proposed, respectively. The latter was also called focal dice loss, which enabled the network to achieve a performance of 0.933 mIoU on the testing set. Finally, an improved slimming strategy compresses the network, making the segmentation network achieve real-time on the embedded system (RK3399pro).