Issue link: https://iconnect007.uberflip.com/i/1379105
70 SMT007 MAGAZINE I JUNE 2021 direct impact on the choice of hardware on which the AI model runs. e IT architecture therefore must be care- fully designed to address all the requirements above. Edge Computing Solution Edge computing is defined as "a distributed computing paradigm that brings computations and data storage closer to the location where it is needed, to improve response times and save bandwidth." [1] In large scale manufacturing, inspection data are generated at a large num- ber of inspection points (equipment or station) within a plant and across plants, edge comput- ing is a natural fit to distribute trained AI mod- els close to inspection points so that inspection results can be returned to the inspector very quickly. A solution architecture was designed for AI model and edge service deployment for qual- ity inspection in manufacturing, based on IBM Edge Computing Reference Architecture[2]. e architecture was carefully designed for efficient model deployment as well as to fulfill all the requirements discussed in the previous section. Figure 1 shows the system context of the solution architecture design, with consider- ation for external factors and constraints. It also ensures user experience is a focus item in the design. e solution supports three persona groups: • Model Engineer: Train and optimize AI models, manage the lifecycle of the AI models, and deploy AI models to edge devices • Edge Manager: Set up and manage edge devices (or edge servers) • Quality Inspector: Trigger and monitor the inspection process e personas interact with these three exter- nal systems: • Data warehouse: Database for all inspection images and results manufacturing plants in many locations. e system should have the capability to be scaled out easily, i.e., adding a computer vision (for AI model training) instance, either on prem- ises or on Cloud, as well as adding edge devices. Performance (speed, exception handling, etc.) has to be considered to support users around the globe. User authentication and authorization An important feature of the manufacturing quality management system is that only autho- rized and trained operators are allowed to per- form the quality inspection. erefore, user authentication and authorization for different user roles are required. Model management and device management Once deployed, lifecycle management of the AI models and devices (on which the AI models run) become critical to operations. Users require an easy and efficient way to manage AI model versions used in produc- tion. Device monitoring and recovery are also important to minimize disruption to manufac- turing schedules. Data security Inspection data (images and results) is con- stantly generated from the inspection pro- cess. is data has to be securely stored and archived as critical and confidential data. e data shall be easily consumed by other appli- cations (such as analytic applications, dash- boards) in the short term and easily retrieved upon request in the long term. Cost In any IT deployment, the associated cost has two components: fixed cost and variable cost. Fixed cost includes hardware and infra- structure set up or purchase required. Variable costs include cloud consumption, if any, and maintenance cost of the hardware and infra- structure. For example, the inferencing mem- ory usage of the selected model type has a