MVIF-Net: An online monitoring method for weld surface quality in weaving welding based on long-duration video

Weaving welding is an important process in medium- and thick-plate welding, during which the molten pool morphology changes periodically with torch weaving. Static-image or short-duration video methods have difficulty distinguishing normal molten pool variations from weld surface defects, resulting in misjudgments of weld surface quality. Therefore, this paper proposes a long-duration video-based method for online weld surface quality monitoring. Inspired by the dual-pathway spatiotemporal modeling of SlowFast networks for action recognition, and considering the locality of defect locations in weaving welding and the interference caused by molten pool variations, a multiscale video-image fusion network, MVIF-Net, is constructed. The network adopts a slow branch to capture spatial detail information of the molten pool region and a fast branch to extract dynamic information of the welding process. Through a cross-branch information exchange module, dynamic features are progressively introduced into the spatial representation process, establishing the association between static appearance and dynamic behavior. The terminal video-image fusion module enhances the long-duration evolution representation of the slow branch and the local dynamic detail modeling capability of the fast branch, respectively, and aggregates the dual-branch features through feature concatenation, thereby achieving complementary fusion of spatial structural information and temporal dynamic information. Experimental results show that MVIF-Net achieves a recognition accuracy of 99.2% with an average inference latency of only 2.65 ms. Compared with existing methods, the proposed method combines high accuracy, low latency, and a lightweight design, while reducing high-frequency image processing pressure in online monitoring.