Deep Learning based Optical Inspection with Centralized Analysis for High Volume Smart Manufacturing

Abstract

Increased capabilities in data storage and exploration provide significant insights for quality assurance in a high volume manufacturing environment. However, these opportunities are associated with great challenges in analytical model development, application deployment, system throughput and reliability. While no commercial software system fully meets the needs of recording head factories in Seagate, a novel strategy named optical inspection with centralized analysis has been developed to detect defects of trailing edges of the recording heads of hard disk drives, and fail the parts when necessary. Leveraging the state-of-the-art artificial intelligence technologies, a deep learning based semantic segmentation engine is built using convolutional neural networks for optical inspection. It has shown an improved accuracy to that of visual inspection performed by human. Meanwhile, a high performance computation engine has been built as a Kubernetes cluster with multiple GPU and CPU units. It is able to achieve the target throughput of three million high-resolution images in each day (i.e., 12 TB image data and 35 images per second). With the high fidelity offered by Kubernetes cluster, the developed applications (inference engine, preprocessor, postprocessor, etc.) serve as containerized microservices independently. Such an architecture ensures the vertical and horizontal scalabilities according to the computation of each individual deployment, while all deployments communicate through an Advanced Message Queuing Protocol cluster without human interference. This analytic framework enables Industry 4.0 recording head manufacturing by integrating advanced AI technologies with a robust edge computation architecture.