In high-end manufacturing and aerospace equipment, reliable identification of welding defects is the key premise for intelligent fault diagnosis and structural health management because undetected or misclassified defects may lead to serious safety risks. However, the actual welding inspection scene is usually subject to some problems, such as limited training data, serious imbalance of categories, and low contrast of defect patterns, which significantly reduce the robustness of traditional vision-based inspection methods. In order to meet these challenges, this paper proposes a tail-sensing-enhanced welding defect detection framework based on YOLOv8, which is specifically aimed at the distribution of common long tail defects in industrial inspection tasks. On the premise of not changing the original network architecture, this method systematically enhances the learning ability of underrepresented defect categories through collaborative optimization in data, training, and loss levels. The specific measures include tailperceived data enhancement, convergence-oriented training strategy, and loss re-weighting of classification priority. The experimental results on multi-class welding defect datasets show that the proposed framework significantly reduces the false alarm rate and inter-class confusion of small samples and low contrast defects, and improves the reliability and robustness of the overall diagnosis. These results show that the proposed method provides an effective and practical solution for intelligent welding fault diagnosis and has great potential to be integrated into aerospace manufacturing quality inspection and health monitoring systems.