Nuclear fuel rods operate under extreme service environments in reactor systems, hence precise inspection of cladding tube and end plug welds represents a critical nuclear safety safeguard. Traditional detection methods face significant challenges including the scarcity of high-quality training datasets due to low real defect occurrence rates and high X-ray imaging equipment costs, alongside existing models' inability to meet real-time detection requirements.

This study aims to develop an intelligent weld defect detection algorithm based on imbalanced convolution feature extraction that achieves both high detection accuracy and real-time performance for fuel rod weld defect identification.

A novel defect detection model, named as YOLOv8n-WIOU-Fasternet, was designed through comprehensive architectural modifications. Firstly, a distance-aware attention mechanism was integrated into the bounding box regression loss function, incorporating a two-level distance penalty mechanism to focus on ordinary-quality anchor boxes while mitigating adverse gradients from low-quality anchors. Secondly, distribution focal loss (DFL) was incorporated to refine edge-level position estimation, enabling more precise boundary localization through cross-entropy optimization of probability distributions around target labels. Meanwhile, by collecting digital radiography (DR) images of 500 fuel rods, a batch of samples containing abnormal defects such as porosity, gas expansion, incomplete penetration, tungsten inclusion, and blockage were prepared. The area near the end plug weld was selected as the region of interest (ROI), and open-source image annotation tool Labellmg was applied to manual annotation of the defect area to obtain 720 defect images. By systematically expanding the dataset, samples with "pseudo defects" were generated, and a total of 7 200 DR images of fuel rods containing different types of defects were constructed. Among them, the number of different types of defects such as pores, tungsten inclusions, and incomplete penetration was the same in the training set, and were divided into training and validation sets in an 8:2 ratio. An additional 72 fuel rod defect DR images that did not participate in the training and validation process were collected as an independent test set to evaluate and experimentally validate the performance of the defect detection model. Finally, traditional convolutional modules were replaced by a newly designed partial convolution (PConv) structure that selectively applies convolution to a subset of input channels while retaining others, followed by pointwise convolution for spatial information fusion and maintaining representational completeness.

The experimental validation results demonstrate that the proposed YOLOv8n-WIOU-Fasternet model achieves false negative and false positive rates both below 5%, representing significant performance improvements across multiple metrics and substantial reductions compared to baseline models. The maximum F1 score reaches 0.947 at a confidence threshold of 0.285, with corresponding false positive and false negative rates of 3.1% and 3.5%, respectively. The average precision (AP) performance significantly surpasses both traditional feature extraction methods and the original YOLOv8 model across various IoU thresholds.

The proposed model successfully achieves an optimal balance between detection accuracy and computational efficiency through its innovative architectural design. The integration of distance-aware attention mechanisms and partial convolution structures reduces computational overhead while maintaining superior detection performance. This comprehensive approach provides a robust and practical solution for automated fuel rod defect detection in real-world nuclear fuel manufacturing applications, meeting both precision requirements and real-time processing constraints essential for industrial deployment.