Micro-Interface Guided Ultrasonic Defect Detection in Magnetic Pulse Welded Joints using Dual-Channel Continuous Wavelet Transform Convolutional Transformer Network

Magnetic pulse welding is extensively employed for connecting power battery busbars, where the detection of defects in weld seam samples plays a critical role in quality assurance. Conventional approaches based on threshold segmentation, however, often suffer from high false positive rates. This study addresses the challenge of defect detection in aluminum-copper welded joints. We first perform microstructural analysis to elucidate fundamental differences in ultrasonic attenuation mechanisms between defective and defect-free samples. Based on these insights, we propose a novel deep learning architecture that integrates continuous wavelet transform with attention mechanisms—referred to as the continuous wavelet transform convolutional (CWTConv)-Transformer model. The proposed method exhibits two key innovations: first, it employs continuous wavelet transform to construct a localized time-frequency representation, overcoming limitations of traditional Fourier transforms in handling non-stationary signals; second, it adopts a parallel convolutional structure with differentiated receptive fields, forming a dual-branch feature extraction network capable of capturing multi-scale information. Finally, Transformer encoder layers are incorporated to establish global dependencies across frequency bands via self-attention. The experimental results indicate that the accuracy of weld defect detection using this model is 96.7%, with an F1 score of 96.26%. These metrics meet the detection requirements in actual industrial production settings.