Attention-Driven Explainable PatchCore for Real-Time Anomaly Detection on Resource-Constrained Edge Devices

In this research, a lightweight PatchCore-based explainable attention mechanism (LW-PC-EAM) is proposed as a framework for real-time industrial visual inspection, incorporating a lightweight and interpretable anomoly detection system. A compact CNN backbone, PatchCore’s efficient memory bank, and an explainable attention mechanism are all integrated into LW-PC-EAM, which was designed exclusively for edge deployment. This allows for precise defect detection at the patch level while simultaneously preserving computational resources. On factory assembly lines, the model generates visual heatmaps for abnormalities that have been spotted, providing operators with clear and actionable insights that improve both trust and the speed with which decisions are made within the system. The improved detection accuracy, low latency, and capacity to operate effectively on resource-constrained edge devices of LW-PC-EAM have been validated by extensive evaluations on real-world industrial datasets. The implementation, such as the model architecture, memory size, and coreset sampling mechanism, are addressed to guarantee reproducibility and clarity about the requirements of the system. The proposed method tackles a critical real-world challenge enabling manufacturers to deploy trustworthy, real-time anomaly detection on resource-limited devices while ensuring operators can understand and verify the system’s decisions in quality assurance workflows.