Autoencoder-based defect detection in PVC profile manufacturing

This study develops an automatic defect detection system for polyvinyl chloride (PVC) profile manufacturing, addressing inefficiencies in manual inspection. It compares the proposed autoencoder model with other well-known unsupervised deep-learning methods, including GANomaly, f-AnoGAN, and the student-teacher network, for defect detection during extrusion. Utilising a defective PVC profile dataset, the study generates anomaly heat maps through reconstruction errors and assesses model performance using the area under the receiver operating characteristic (ROC) curve. The proposed autoencoder model is found to be optimal for this dataset, offering a balance between efficiency and accuracy. These findings have significant implications for enhancing quality control and reducing defects in PVC manufacturing, with potential applicability in other industrial settings.