基于改进YOLOv5s的控制棒导向卡磨损检测方法

Wear Detection Method of Control Rod Guide Card Based on Improved YOLOv5s

摘要:

针对核反应堆控制棒导向卡磨损检测中存在的小目标特征不明显、边界模糊及易误检等问题，对导向卡磨损检测方法进行了研究；基于YOLOv5s模型，采用轻量级骨干网络VoVNetv2s替代原CSPDarkNet结构，并在一次性聚合模块中嵌入卷积空间注意力机制，同时构建解耦检测头实现分类与定位任务分离，并引入EIoU损失函数对边界框回归进行优化；经实验测试，在mAP@0.5、mAP@0.5:0.95及召回率指标上分别达到98.9%、87.7%和96.2%，推理速度为168 FPS，模型参数量为6.31 M；结果表明该方法能够实现导向卡表面磨损的高精度实时检测，满足核反应堆设备检测对可靠性与实时性的工程应用要求；

Abstract:

Aiming at the problems of small target characteristics, blurred boundaries, and high false detection rates in the wear detection of nuclear reactor control rod guide cards, a detection method for guide card wear is investigated. Based on the YOLOv5s model, a lightweight backbone network, VoVNetv2s, is adopted to replace the original CSPDarkNet architecture. A convolutional spatial attention mechanism is embedded into the one-stage aggregation module, while a decoupled detection head is introduced to separate classification and localization tasks. In addition, the Efficient IoU (EIoU) loss function is employed to improve bounding box regression accuracy. Experimental results show that the proposed method achieves 98.9% mAP@0.5, 87.7% mAP@0.5:0.95, and 96.2% recall, with an inference speed of 168 FPS and a model size of 6.31 M parameters. The results demonstrate that the proposed approach enables high-precision real-time detection of guide card surface wear, meeting the requirements of reliability and real-time performance for nuclear reactor equipment inspection applications.