摘要
Abstract
To address the challenges of limited sample acquisition,significant cross-scene imaging differences,and diverse defect shapes and scales in nuclear power plant weld radiographic images,a scene transfer defect recognition method for nuclear power plant weld radiographic inspection is proposed.Firstly,a denoising diffusion probabilistic model is pre-trained on multi-scene weld data and iteratively fine-tuned with the target scene data to construct a target style generation set,thereby alleviating the adverse impact of domain deviation and limited samples on model training.Secondly,a multi-branch detection structure with a shared backbone and feature pyramid network neck is designed,which fuses anchor-based and anchor-free prediction mechanisms,to enhance the adaptability to key defects in nuclear power plant scenes while main-taining the general feature expression ability.Finally,in the inference stage,adaptive window width and window level en-hancement,wide-directional overlapping block inference,and result fusion strategies are introduced to meet the application requirements of long,high-resolution weld radiographic images.Compared with mainstream single-stage detection models,the proposed method achieves an improvement of 1.9 to 5.8 percentage in mean Average Precision at IoU threshold 0.5(mAP@0.5)in multi-scenario condition.The method attains the defect detection rate of 94.1%and the quantitative deviation rate of 3.7%in the nuclear power plant scenario,which are comparable to the evaluation level of professional inspectors.Moreover,the film evaluation efficiency is improved by 12 times,meeting the practical engineering requirements for as-sisted evaluation.关键词
核电焊接/射线检测/焊缝缺陷识别/去噪扩散概率模型/场景迁移Key words
nuclear power welding/radiographic testing/weld defect identification/denoising diffusion probabilistic model/scene transfer分类
信息技术与安全科学