BOE Technology Group Deploys AI Defect Detection Across Display Panel Production with 96% Accuracy and 70% Labor Substitution

BOE Technology Group, one of the world's largest display panel manufacturers, deployed a new AI-driven defect detection system across its display panel production lines. The system is built on Intel Xeon scalable processors deployed at edge servers, with OpenVINO toolkit providing software optimization for the OpenCV image processing library. The cloud-edge collaborative AI defect detection system has been rolled out across BOE's factories. It achieves a 96% inspection accuracy rate on display panels and reaches a 70% labor substitution rate versus the previous manual workflow. BOE also developed a defect root-cause analysis system built on the same cloud-edge architecture and big-data platform. The system ingests map data, inspection data, production histories, and equipment parameters from sensors, industrial PCs, and PLCs at the edge; after ETL preprocessing, the data flows to a centralized big-data platform for clustering and root-cause analysis. In practice, edge servers select appropriate models for inference and execute image annotation, defect classification, defect recognition, trend analysis, and quality-warning tasks, with results displayed on engineering monitoring dashboards. The motivation was operational. Training a single skilled inspection engineer takes months, and even experienced engineers face accuracy degradation when reviewing 10,000 to 20,000 complex images a day. As panel processes advanced, defect types multiplied — a single large display panel may contain hundreds of defect types, some repairable and some that require scrapping — and any critical defect that goes undetected raises the overall defect rate. To maintain quality, at one factory BOE previously needed 36 inspection engineers a day, working in three rotating shifts, which drove large labor costs and unstable accuracy. BOE developed the automated defect classification (ADC) system in-house using deep-learning methods, partnering with Intel to use edge computing to lift inspection efficiency. Historical images and new images captured by automated optical inspection (AOI) equipment are preprocessed by edge servers, uploaded to the data center, and used to train image-detection and classification models (ResNet, Faster-RCNN); the trained models are then deployed back to the edge servers.