MyAdvantech Registration

MyAdvantech is a personalized portal for Advantech customers. By becoming an Advantech member, you can receive latest product news, webinar invitations and special eStore offers.

Sign up today to get 24/7 quick access to your account information.

Robotic Visual Defect Inspection Powered by AI Technologies



For manufacturers, quality control systems have long relied on visual inspection. Traditional machine vision systems may fail to distinguish between defect types with high variation between similar parts. Tapping the power of AI and deep learning technology helps solve this problem by delivering high-precision detection results. Unlike machine vision systems using computer vision technology operated by rule-based algorithms, machine vision systems powered by deep learning software detect defects based on training data. Data-driven AI enables automated defect inspection with greater flexibility and accuracy, while lessening maintenance costs.


In this case, the customer was a robotic visual equipment builder. They planned to combine their robotic arms with AI capability to detect defects such as air bubbles and cracks among enamel-coated products. Traditional computer vision technology faced limitations: it could not distinguish between different defect types on similar product parts that only had nuanced differences. It was not flexible enough to update the existing defect inspection system to recognize new defect types. To leverage AI technology for visual defect inspection in real-time, the system required substantial computing power at the edge and a large storage capacity to store massive numbers of images captured from multiple production lines, as well as sufficient bandwidth to handle data transmission.


The AIR-300 AI system perfectly met the customer’s requirements. After the robotic arm accurately identified the placement of the thermal cup and picked it up for 360-degree photography, the captured images were sent to AIR-300 for real-time inferencing where defected products could be instantly identified. Complicated real-time AI inference and high-performance computing were done on AIR-300 locally with its Intel Xeon/ Core i3/i5/i7 CPU and 1x PCIex16 for high performance graphics card support. Regarding I/O and data storage volume, AIR-300 was equipped with 4x GbE ports 4x RS-232/422/485 and supported up to 20TB data capacity with a 4x 2.5” SATA III hard drive that provided full bandwidth and storage capacity to fulfill application needs. AIR-300 already had a built-in 850W power supply, so customers didn’t need to add an external power supply. AIR-300 can also be used as a local training server when the defect inspection system needs to be updated to inspect new products. The built-in vision system sends captured images back to AIR-300 for further AI model re-training. For example, the types of defects found in a cup product are likely different from that of a paper bag product. To change AI defect inspection systems from inspecting cups to inspecting paper bags, customers only need to prepare training datasets of defect types that paper bags have, retrain new AI models on AIR-300, and then deployed the trained model on AIR-300. With an AI re-training capability, updating the defect inspection system no longer requires the assistance of costly professional engineers.