MiKAD: Memory-Infused Knowledge Networks for Manufacturing Anomaly Detection

Many semiconductor industries have adopted smart manufacturing systems for defect detection, while others still rely on manual inspection methods that can compromise product quality and assurance. To address the limitations <br/>of manual inspection, we propose Memory-infused Knowledge Networks for Manufacturing Anomaly Detection (MiKAD), an unsupervised learning technique that combines multiscale knowledge distillation and a dynamic memory bank to detect anomalies of varying sizes and shapes in real industrial image datasets. The knowledge distillation framework consists of a teacher-student architecture, where the teacher is a pretrained network and the student is a trainable network that leverages EfficientNet-B7 as the multiscale backbone. A dynamic memory bank is integrated to support the student network during training by enhancing its ability to learn normal features by updating with new features and suppressing outdated ones. Discrepancy loss at multiple scales between the teacher and student ensures accurate detection and localisation of the anomalies across different sizes and shapes using real industrial datasets. Experiments on two real-world datasets, namely a Seagate Write Pole (WP) and BTAD, demonstrate that MiKAD achieves strong performance in both anomaly detection and localisation, with image level ROC_AUC scores of 97.77% and 96.85% respectively.