Leveraging Denoising Diffusion Probabilistic Models for Generating Synthetic Datasets Pertaining to Condition Evaluation of Physical Assets

This study investigates the problem of data scarcity in the structural health monitoring (SHM), nondestructive evaluation (NDE), and condition evaluation of physical assets by presenting a comparative analysis of denoising diffusion probabilistic models (DDPMs), each trained independently on a structurally distinct dataset. The models are evaluated to assess their image generation performance and their ability to reproduce domain-specific visual characteristics observed in real data. Five distinct datasets were used to train separate DDPMs: 1) surface crack images typical of concrete structures and asphalt pavements (448×448 pixels), 2) their masked versions (448×448 pixels), 3) phased array ultrasonic linear scan images along weld lines in a stainless steel structure (256×256 pixels), 4) two-dimensional (2D) microstructural images of powder- bed fusion laser (PBF-L) additive-manufactured (AM) metals from light optical microscopy (LOM) imaging (patches 256×256 pixels), and 5) ground-penetrating radar (GPR) images (patches 256×256 pixels). The training process for each model was conducted using an NVIDIA H200 graphics processing unit (GPU), requiring approximately six days per dataset, with image generation taking approximately 45 seconds per image. The Fréchet Inception Distance (FID) was used to evaluate the quality of the synthetic images, yielding scores of 13.7 to 34.6 for the datasets, indicating high fidelity and diversity across all cases. By leveraging DDPMs, this research not only overcomes the limitations posed by limited datasets but also provides valuable contributions to advancing AI applications in diverse domains, including structural health monitoring, material characterization, and subsurface imaging.