A Mixed Samples-Driven Methodology Based on Denoising Diffusion Probabilistic Model for Identifying Damage in Carbon Fiber Composite Structures

Abstract

X-ray imaging is a common non-destructive detection method for Carbon Fiber Composite Structures (CFCS) that is useful in identifying damage in CFCS-cored wires. In recent years, deep learning models that incorporate classification and objection detection have become frequently utilized by the non-destructive testing industry. These models typically rely on the assumption that there are sufficient annotated failure samples from history that have been measured and can be used for training. Unfortunately, in real-world measurements, it is often challenging to obtain these types of samples. To address the issue of small sample size in such scenarios of real-world field testing, this paper propose a mixed samples-driven methodology based on the Denoising Diffusion Probabilistic Model (DDPM) for identifying damage in CFCS. First, new samples are synthesized through DDPM module to improve the robustness of a small sample size. Then, the synthesized sample, along with a small number of authentic samples measured from real-world testing, are then integrated and fed into a DenseNet-based module. Lastly, the mixed samples-driven architecture is then constructed and employed to diagnose the damage of CFCS. The effectiveness of this approach is demonstrated through experiments in real-world field testing.