Aiming at the problems of scarce datasets and the low identification accuracy faced in the field of weld-crack detection, this paper proposes an artificial-weld-crack preparation method based on the doping of dissimilar metal particles to augment the number of samples of weld-crack defects. Meanwhile, data augmentation methods such as random cropping, scaling and Mosaic are combined to further enhance the richness of the samples, so as to provide strong data support for the proposed weld-crack-defect detection model. Given the limitations of storage and computational resources in industrial application scenarios, this paper designs the lightweight detection network YOLOv6-NW. It is achieved by optimizing the YOLOv6-N model for width and depth compression to efficiently identify and locate weld-crack defects. The experimental results demonstrate that YOLOv6-NW significantly outperforms YOLOv5 in terms of both model detection performance and model size. Compared to YOLOv6-N, the number of model parameters in YOLOv6-NW is only 16% of that in YOLOv6-N, yet its model detection accuracy and recall rate remain on a comparable level with YOLOv6-N. Additionally, the precision of crack detection by YOLOv6-NW under low-resolution or low-illumination conditions remains above 0.9.

This paper primarily addresses the challenges posed by the difficulties in directly measuring the fusion width at the bottom of the weld and in real-time monitoring of the penetration state during the groove welding process. It focuses on the research of online penetration state monitoring technology, which utilizes multi-modal signals such as sound and image during the welding process. The multimodal network proposed in this paper, SIMNet, first employs the short-time Fourier transform (STFT) to convert the original sound signal into the time–frequency domain for preliminary feature extraction. Secondly, a visual feature extractor based on an attention mechanism is used to extract image features. Meanwhile, a cosine similarity loss function is introduced to align the features of the two modalities in the semantic space before fusion. Finally, the interaction and fusion of features are achieved through a cross-attention mechanism. The experimental results demonstrate that SIMNet achieves the best performance with a mean squared error (MSE) of 0.1141 mm, compared to other mainstream algorithms. Furthermore, the inference speed with multimodal input reaches 60 frames per second (FPS), enabling quantitative and real-time multimodal fusion intelligent penetration state monitoring.