Experimental validation of saliency maps for understanding deep neural networks for weld penetration prediction

Abstract

Deep neural networks (DNNs) have achieved high-level accuracy on many imaging-based penetrations to predict tasks during the welding process. However, the deep neural network decision mechanisms are often considered impenetrable leading to a lack of relevance to the welding pool behavior. To solve this problem, we build a deep learning convolutional neural network based on the reflected laser pattern image to predict the GTAW weld penetration and use the saliency map to understand the neural network. Then, we systematically analyze the relevance of saliency maps against the welding pool behavior. Experimental results show that the developed convolutional neural network model can accurately classify unpenetrated, critical penetrated, and fully penetrated with an accuracy of 98.1 %. The interpretable model for the neural network shows that the winged areas on both sides of the pattern were taken into account by the DNNs in the unpenetration, the striped area in the center of the image was highlighted in the full penetration, and the neural network focused on both the center and the sides in the critical penetration. It proves that there is a strong correlation between the welding penetration state and the pool surface oscillation information extracted by the convolutional neural network and the weld pool behavior.