Arc sound signals are considered appropriate for detecting penetration states in cold metal transfer (CMT) welding because of their noninvasive nature and immunity to interference from splatter and arc light. Nevertheless, the stability of arc sound signals is suboptimal, the conventional feature extraction methods are inefficient, and the significance of arc sound attributes for determining penetration statuses is often overlooked. In this study, a compact convolutional neural network (CNN) model is proposed for the adaptive extraction of features from arc sound signals. The model uses the Mel spectrum diagram of an arc sound signal obtained through a short-time Fourier transform (STFT) and a Mel filter bank conversion step as its input. To improve the recognition capabilities of the model, a novel CNN-selective kernel (SK) model for weld penetration recognition is introduced, which integrates the dynamic selection kernel network (SKNet) into the CNN architecture. The experimental results indicate that the CNN-SK model outperforms the traditional models, achieving an accuracy of 98.83% on the validation dataset. This model holds promise for assessing weld penetration in CMT welding applications. The project is available at https://github.com/ZWL58/data/tree/master.
Copyright: © 2024 Zheng et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.