Dynamic penetration prediction based on continuous video learning

Online penetration monitoring of complex grooves remains challenging due to steel plates’ groove instability and welding heat distortion. Penetration is an accumulation process of material deposition. Temporal signals, such as video, can provide a more comprehensive characterization of the melt pool state. A deep learning method based on continuous video is designed to monitor groove welding penetration in-process. The proposed Fast Video-feature Extraction Net (FVENet) consists of a video extraction module and a multi-feature screening module. The efficient network can quickly extract high-dimensional data features in complex arc environments and achieve accurate results for backside melt width prediction. The feature extraction process of the network is explored by visualizing the results of different network layers. Experimental results indicate that the mean squared error (MSE) of FVENet reaches 0.0634 mm, outperforming other mainstream deep learning frameworks. The inference time under video input reaches 100 FPS. The network structure designed in this paper has the potential to become a universal template for processing melt pool images.