TSDD-UB: A Texture Simplification-Based Denoising Diffusion Model for Unsupervised Defect Detection Under Ultrasonic B-Scan Signal

The ultrasonic B-scan signal is widely used in the detection of pipeline defects. However, most existing methods face challenges in achieving high accuracy due to the amplitude attenuation phenomenon and the complicated texture structure of the ultrasonic B-scan signal. To address the above issues, a texture simplification-based denoising diffusion model is proposed for unsupervised defect detection under an ultrasonic B-scan signal. This model closely follows the principles of ultrasonic propagation and incorporates meticulous designs in both data processing and model architecture, aiming to achieve high-precision defect detection while training exclusively on normal samples. First, an adaptive time gain compensation (ATGC) method is proposed to mitigate the amplitude attenuation phenomenon of the B-scan signal. Second, a novel texture simplification network is proposed to simplify texture structures while preserving the echo characteristics of the B-scan signal. Third, a double noise scale reconstruction (DNSR) method is designed, which utilizes the denoising diffusion model to reconstruct signals with two different levels of perturbation, aiming to overcome the issue of low reconstruction quality. Experimental results in real pipelines demonstrate that texture simplification-based denoising diffusion model for defect detection under ultrasonic B-Scan signal (TSDD-UB) achieves an image-level area under the receiver operator characteristic curve (AUROC) of 100% and a pixel-level AUROC of 98.6%, outperforming state-of-the-art methods and indicating significant potential for practical applications.