Implementação de monitoramento inteligente para detecção e classificação de defeitos superficiais em pré-formas obtidas por manufatura aditiva por deposição a arco.

Abstract

At a time when flexibility, combined with production speed and material economy, it is well known that such characteristics make Additive Manufacturing (AM), also known as 3D printing, gain prominence in the industry. Among metal AM processes, Wire Arc Additive Manufacturing (WAAM) uses an electric arc as a heat source to enable material fusion, which generates high deposition efficiency and the possibility of automated operation at a low cost. In contrast, repetitive heating by layer superposition and excessive heat generated in WAAM can lead to the generation of defects such as high levels of tensile residual stresses, distortions, poor surface finish, destabilization of deposition geometry along the layers, as well as surface roughness and porosity. Thus, neural network models for visual control are increasingly being used in production lines to meet the demand for high-quality manufactured products. Consequently, this work aimed to implement in-situ monitoring of surface defects such as porosity, humping, and spatters in thin walls manufactured by WAAM, through the application of Machine Learning using the YOLO (You Only Look Once) model. The classification and recognition process was validated by comparison with manually obtained results and applied through a graphical user interface designed with Python algorithms. Using the pre-trained YOLOv8n model, available on the Ultralytics website, it was possible to perform training with transfer learning. The training was carried out for 300 epochs in three different models, B1, B2, and B3, which differed in the quantity of images. Through artificial augmentation techniques, it was possible to expand the amount of data in the databases for better performance. In addition, the tuning technique was applied, which uses combinations of hyperparameters to find an ideal configuration for the model. For understanding the effectiveness of the network, the concepts of evaluation metrics and confusion matrix were applied, specifically mAP, Recall, and Precision. Through the analysis of the results, it was obtained that database B1 achieved the best overall performance in defect detection, with 74.5% for precision, 61.2% recall, 65.7% for mAP, and 71.5% for the model. Overall, the neural network designed in the algorithm proved to be a potential tool for defect classification and localization.