Physics-guided label-free ultrasonic testing of high-speed rail wheels with diffusion models

Ensuring the structural integrity of high-speed rail wheels is essential for operational safety, yet automated ultrasonic testing is constrained by scarce defect samples, costly pixel-level annotation, and complex field noise. This study proposes a physics-guided ultrasonic anomaly-detection framework with defect-label-free training based on diffusion models. A physics-guided multi-channel representation incorporates probe configurations, ultrasonic propagation paths, overlapping inspection regions, and wheel geometry. A frequency-domain adaptive loss preserves defect-sensitive spectral information, while an uncertainty-aware scoring strategy combines reconstruction residuals, diffusion-induced predictive uncertainty, and physically defined regions of interest. The diffusion model is trained exclusively on 6000 normal ultrasonic B-scan images without defective training samples or pixel-level defect annotations. For independent evaluation, defect-related echo regions in test images are annotated and cross-checked by three experienced ultrasonic specialists. Pixel-level evaluation achieves an AUROC of 0.972, an F1-score of 0.953, an IoU of 0.815, and an FPR of 0.024 at a TPR of 0.95. Preliminary on-site trial testing and offline validation using data from an operational high-speed rail wheel inspection system demonstrate the potential of the proposed method for reliable defect screening under noisy field conditions.