Generation of WAAM Defect Images Using a Hybrid CVAE-CGAN: A Data Augmentation Strategy for Small and Imbalanced Datasets

In industrial Wire Arc Additive Manufacturing (WAAM)., the scarcity and imbalance of labeled defect images limit the effectiveness of deep learning-based quality inspection systems. This paper presents a hybrid CVAE-CGAN framework designed to generate high-resolution., class-conditional molten pool images for data augmentation in small-sample settings. By combining a conditional variational autoencoder with adversarial training and integrating VGG19-based perceptual loss and sub-pixel convolution, the proposed model produces visually realistic and diverse synthetic defect images. Extensive experiments on a nine-class WAAM defect dataset demonstrate the model's ability to enhance classification performance, especially for underrepresented categories, offering a scalable solution to mitigate data limitations in intelligent manufacturing.