Design and Implementation of an Edge Computing-based AI Vision Inspection System

Abstract

This study proposes an edge computing-based artificial intelligence (AI) vision inspection system for manufacturing sites and validates it through a soldering inspection case study. While AI-based vision inspection offers higher accuracy compared with conventional rule-based methods, it suffers from increased computation time, and may fail to meet the required takt times. Furthermore, relying on external cloud servers introduces latency issues and data security risks. Although on-premises servers can mitigate latency and security concerns, a single server failure can trigger severe disruptions by shutting down the entire production line. Therefore, adopting an edge computing approach is desirable to simultaneously address latency, data security, and operational reliability. In this study, instead of utilizing a dedicated inference server, both the client and server are deployed onto a single local personal computer (PC) to implement a fast, cost-effective edge computing-based AI vision inspection system. The proposed system is then applied to and verified in a soldering inspection process. Additionally, to compensate for the degraded small-object detection performance inherent in AI models, a hybrid inspection framework integrating a rule-based algorithm is applied. Experimental results demonstrate an average computation time of 150 ms that successfully satisfied the 200 ms target tact time, and the hybrid inspection accuracy for solder ball defects substantially improved from 75% to 96%.