A thermographic data augmentation and signal separation method for defect detection

Non-destructive testing is a popular technique for defect assessment of composite materials, where machine learning models become more important in its data analysis. Nevertheless, deep learning, which has achieved state-of-the-art results in many tasks, has received less attention in this field. Herein, a generative independent component (IC) thermography method is proposed. In detail, a generative adversarial network is implemented for image augmentation, which generates fake thermal images that mimic the patterns of real measurements. In doing this, the sample size is enlarged and the defect information contained in the images is enriched. Then, both the real and fake thermal images are decomposed by IC analysis, which separates the defect signals represented by non-Gaussian sources and the non-uniform backgrounds caused by uneven heating. Consequently, the defect detection results are improved. The performance of the proposed method on a polymer composite specimen demonstrates its effectiveness. (paper)