IMDRNet: An Interpretable Multiclass Defect Recognition Network Based on Arc Sound Signals for Robot Arc Welding

Abstract

Welding defects recognition based on arc sound signals remains a significant challenge in robot arc welding. Existing deep learning methods mainly focus on classifying penetration states rather than recognizing multiclass defects, such as slag inclusion and porosity. In addition, these methods lack physical interpretability of used neural network models, resulting in big risks if using in real welding applications. To address these issues, this article proposes an interpretable neural network named interpretable multiclass defect recognition network (IMDRNet) to recognize multiple welding defect types from arc sound signals. The original signals are first adaptively decomposed into several intrinsic mode functions (IMFs) with limited bandwidths and distinct center frequencies using the variational mode decomposition algorithm, which helps suppress noise interference and prevents the loss of useful information. Then, a multichannel continuous wavelet convolutional layer based on IMFs is designed as the first layer of the IMDRNet, which increases the depth of IMDRNet by using multiple parallel channels, thereby enhancing its capability to extract multiclass defect features. The proposed IMDRNet is tested and compared on two datasets with 9749 samples, and results indicate that the IMDRNet exhibits better effectiveness and interpretability in recognizing multiclass welding defects based on arc sound signals, achieving the highest accuracy of 97.09% on the SS355 dataset with six defect types and 91.70% on the AL5083 dataset with eight defect types.