Cross-Modality Transfer Learning in Diffusion Models for Data Augmentation in Metal Additive Manufacturing

Abstract Anomaly detection and in-situ monitoring are essential in additive manufacturing (AM) quality control. Artificial intelligence (AI) approaches can be used to leverage AM process data for quality improvements. However, limited data availability is one of the challenges in the applications of AI techniques in AM. Generative models (GMs) can be leveraged for data augmentation, enabling model performance improvement. Nevertheless, these models still require a sufficiently large dataset to generate high-quality synthetic samples. In AM, data collection can be very costly and time consuming. Therefore, transfer learning (TL) can be used to leverage publicly available datasets as source datasets, to transfer the knowledge obtained from the source domain to the target domain with a small-scale dataset. However, the high diversity of sensors for data collection in both source and target domains makes the transfer of knowledge across different sensor modalities more challenging. Therefore, a cross-modality transfer learning approach is proposed by leveraging a publicly available dataset of AM melt-pool optical images to improve the generative capabilities of a Denoising Diffusion Implicit Model (DDIM) by implementing model-based TL for data augmentation of AM thermal history. The real-world case study shows that the proposed approach enhances the performance of thermal image generation, especially when there is limited target data available.