Vibration-signal-driven generative framework for automated nut thread quality inspection in tapping processes

Tapping is a widely used threading process in manufacturing, where variations in machining conditions often lead to deviations in nut internal thread quality. Traditional inspection methods rely on either direct imaging or contact gauges, both of which are time-consuming and difficult to integrate into real-time production. To address these challenges, this study proposes two vibration-signal-driven generative frameworks for nut internal thread image synthesis. The first is a Wasserstein Generative Adversarial Network (WGAN)-based approach with a class-aware discriminator that leverages vibration signals as conditional embeddings in the generator, while incorporating nut quality labels in the discriminator to enhance class fidelity. The second is a Conditional Latent Diffusion Model (CLDM)-based framework that employs a Transformer encoder to effectively embed tapping vibration signals into the diffusion process for improved diversity and distributional coverage. A comprehensive evaluation protocol was established, integrating perceptual metrics (MS-SSIM, LPIPS), distributional metrics (Precision, Recall, Density, Coverage), and statistical metrics (FID and a domain-specific Custom FID). Experimental results show that while the CLDM-based method achieves superior fidelity and diversity, the WGAN-based framework offers higher computational efficiency and flexibility in conditional embedding. Importantly, both approaches preserve discriminative features sufficient for downstream nut quality prediction, with tapping-phase vibration signals proving most informative. The proposed frameworks provide practical and non-destructive solutions for automated nut quality inspection. By converting readily available vibration signals into realistic thread images, they reduce inspection time, labor cost, and dependence on direct imaging, thereby enhancing efficiency and scalability in industrial manufacturing.