Metalsync: ferramenta de análise da conformidade de padrão de estampa em chapas metálicas por meio de redes neurais convolucionais

Abstract

The growing demand for quality and efficiency in the metallurgical industry has been driving the adoption of intelligent technologies for automated inspection. In this context, this study presents a computer vision–based solution for detecting nonconformities in stamped metal sheets using convolutional neural networks (CNNs). The research was conducted with real images of metallic parts obtained from the company Só Aço Industrial Ltda., comprising preprocessing steps, organization of data into training, validation, and testing sets, as well as the application of data augmentation techniques to enhance the robustness of the model. The CNN classifier was optimized to identify visual conformity patterns in stamped surfaces and evaluated using metrics such as accuracy, precision, recall, and F1-score. The results demonstrated remarkable performance, achieving an overall accuracy of 95.24%, precision of 98.33%, recall of 90.31%, and F1-score of 94.14%. Notably, for nonconforming parts, the model achieved 100% recall, highlighting its effectiveness in reliably detecting defects. The proposed approach proved technically feasible for real-time industrial applications by reducing the subjectivity of manual inspection and enabling integration into operational workflows through a mobile application with an intuitive interface. This study represents an advancement in the automated inspection of stamped parts, contributing to the modernization of the industry with an emphasis on quality, traceability, and productivity.