Linking Fabric Wear to Final Product Defects : A Cross-Stage Unsupervised Vision Framework

Sammanfattning

This study introduces a cross-stage computer vision framework for unsupervised defect detection and spatial correspondence analysis in rubber manufacturing which is applied for defect detection for thermally aged separator fabric using two embedding‐based anomaly detection models PatchCore and PaDiM models. To connect these upstream defects with downstream defects in the rubber, the study suggests a framework that relies on NTP synchronized timestamp, physical delay compensation, and spatial metrics such as Intersection over Union and Transfer Ratio to quantify defect propagation across production stages. Additionally, the experimental results demonstrate excellent anomaly detection performance, where the AUROC for the rubber images is 0.9789 and the AUROC for the fabric images is 0.9651 for the PatchCore model. Also, the cross-stage analysis successfully differentiated between upstream propagating defects and defects that occurred downstream. Finally, the proposed framework specifies the physical interface between independent instances to enable a proven proof-of-concept for automatic investigation in industrial areas with a validated workflow. The framework could also help to minimize the use of material by making decisions on the retirement of the fabric based on the condition of the material in addition to detecting defects, while also introducing further operator ergonomics and removing the need for manual trace of defects with an automated correspondence analysis.