Radiographic testing of welds plays a critical role in ensuring the quality of welded manufacturing, because X-ray imaging technology can clearly reveal the internal structure of the weld area. However, existing mainstream detection methods rely on manual inspection or supervised detection, both of which are susceptible to limitations imposed by subjective factors and model generalization capabilities, respectively. Therefore, this paper proposes a two-stage unsupervised detection framework based on reconstruction to achieve fast and accurate detection of welding quality. First, an algorithm for generating simulated defects based on real welding defect characteristics is designed. A dataset encompassing multiple defect types is constructed, and image quality is further optimized through data augmentation algorithms. Second, a high-quality diffusion model (H-DiffuM) based on residual learning is proposed, which achieves accurate reconstruction of weld defect images through a residual-guided noise scheduling mechanism. Finally, by combining the gated mechanism with frequency domain features of X-ray images, a multi-scale frequency domain attention fusion module (MFDAFM) is designed and embedded into the discriminative network (Seg-net), thereby enhancing detection accuracy. The final experimental results demonstrated that the proposed method achieved 97.80% in pixel-level AUROC and 93.34% in AP, which surpassed the current state-of-the-art unsupervised detection approaches. Meanwhile, the inspection method described in this paper offers the advantages of rapid detection speed and high precision, demonstrating its potential for application in the rapid assessment of welding quality.
扫码关注我们
求助内容:
应助结果提醒方式:
