Um método semi-supervisionado baseado em GAN para detecção e classificação de defeitos em soldas

Abstract

Detecting defects in welding beads is critical for ensuring industrial components’ structural integrity and quality. However, the limited availability of labeled datasets for training robust machine learning models presents a challenge in this domain. This master thesis proposes a semisupervised method for weld defect detection and classification, leveraging Generative Adversarial Networks (GANs) to generate massive synthetic data. Our method focuses on a lightweight GAN architecture that reduces computational complexity while maintaining fidelity in synthesized images. Experiments on two weld bead datasets, lweld-long and lweld-short, contained over 200,000 artificially generated weld bead samples with diverse defect types, such as pores, deposits, discontinuities, and stains. Results showed improvements in model performance when GAN-generated samples were included. For example, in the lweld-long dataset with a 10% confidence threshold, the 𝐹-score improved from 66% to 69% and the mean Average Precision (mAP) from 61% to 63% with GAN samples. At a 50% confidence threshold, the 𝐹-score increased from 57% to 69% and the mAP from 41% to 56%. Similarly, on the lweld-short dataset, a slight gain was observed with GAN samples at a 10% confidence threshold, increasing the 𝐹-score from 49% to 50%, while the mAP remained at 41%. At a 50% confidence threshold, GAN samples led to more significant gains, with the 𝐹-score rising from 36% to 49% and the mAP from 20% to 33%. Across all defect categories, incorporating GAN-generated samples increased true positives and reduced false negatives, as demonstrated by improved confusion matrix outcomes across multiple experiments. These findings highlight the potential of synthetic data generation to enhance weld defect detection and quality control in industrial applications.