Automated Quality Control of Vacuum Insulated Glazing by Convolutional Neural Network Image Classification

Perfect accuracy achieved on test data: The automated system achieved 100% accuracy (AUC of 1.0 on the ROC curve) in classifying damage to vacuum insulated glazing, demonstrating that AI can reliably identify surface defects that threaten window lifespan.

Explainability methods show where the AI is looking: The researchers used Grad-CAM and Score-CAM techniques to visualize which parts of images the AI focuses on when making decisions, revealing that the system correctly identifies crack locations and geometry—an important transparency measure for quality control applications.

Efficient detection with limited training data: The proposed method can detect systematic defects "even without large amounts of training data," suggesting this approach could be practical and cost-effective for manufacturers who may not have extensive labeled datasets available.

Automated cropping improves efficiency: By using Faster-RCNN to automatically locate pillar positions and crop relevant image sections, the system reduces manual preprocessing work, making the quality control process faster and less dependent on human judgment.

Outperforms existing AI models: The custom classifier outperformed ResNet50V2 (a standard deep learning model) at identifying crack locations and geometry, indicating that specialized AI systems can be more effective than general-purpose models for specific manufacturing quality control tasks.