P‐220: Conditioned Diffusion for Manufacturing Data: Improving Generation Controllability

In the manufacturing industry, enhancing production quality is a crucial aspect of business success. Rapid detection and elimination of various defects during the production process are essential for maximizing profits. The significance of data in this regard has been increasingly recognized, with AI technology emerging as an effective solution to numerous industrial challenges. However, data‐intensive companies with extensive know‐how face a considerable challenge in obtaining balanced datasets for training AI models due to the imbalance between defective and normal data. To address this issue, Generative Adversarial Networks (GANs) had been proposed as potential solutions. However, instability during GAN learning and mode collapse leading to insufficient diversity in generated results have limited their applicability i industrial settings. The emergence of diffusion models as a promising next‐generation image generation technology has drawn significant attention due to their stability and ability to generate high‐quality images that closely resemble real data. In this paper, we propose an innovative approach to generating balanced datasets for manufacturing industries using the advantages of diffusion models. By transforming existing model to enable take additional embedding effectively, the diffusion model could learn numerical information that comes from the microscopic images of the defective circuits. Our methodology involves training the model from zero to fully adapt the structured condition of the data, enabling its application in manufacturing contexts where free‐form prompts are not feasible. Specifically, we extracted coordinate information from defective image data using image processing, and used as additional features to learn. We demonstrate that this approach effectively generates datasets to address data‐imbalance problem with successful reflection of input information and the resulting images. By employing diffusion models to generate balanced manufacturing data, we aim to improve AI model performance in detecting and eliminating defects, ultimately leading to increased production efficiency and improved product quality.