Enhancing Casting Manufacturing Quality Control with Vision Transformers

Ensuring quality control is a crucial task in modern manufacturing processes. Today, industries must guarantee the quality of their products to stay competitive. Automated visual inspection is one of the most effective ways to promote quality control, enabling the efficient inspection of 100% of the production volume. This paper deals with the quality control problem of casting discs for submersible pump impellers, ensuring quality at the casting stage. We propose an approach to detect defects in casting discs throughout the manufacturing process. Our methodology uses Vision Transformers (ViT), which employ self-attention mechanisms to capture spatial relationships between image patches, allowing the model to learn hierarchical visual representations. Both real and simulated experiments were conducted to validate the proposed approach. The results indicate an accuracy of 99.22% using the proposed ViT inspection. Moreover, the approach maintained high accuracy even in noisy images, with accuracies of 95.45% and 98.28% for Gaussian and Salt and Pepper noise, respectively, in the worst-case scenario. The experiments demonstrate the robustness and reliability of the approach, optimizing the manufacturing process.