Find radiator defects using machine vision

No production can do without quality control. For a long time, the only possible option for manufacturers was to visually inspect products by specially trained personnel. However, this method requires enormous human resources, long training, increased attention and is very dependent on the human factor. Employee fatigue and carelessness lead to the release of marriage.

Advantech in partnership with the company Smasoft developed solutions to fully automate visual quality control of manufactured products. These solutions already work in real production today. The article tells about the successful experience in implementing a visual quality control system using machine vision on the production line of cooling radiators.

System description

The customer is engaged in the production of copper cooling radiators for heat removal from microprocessors. For radiators, an extremely important parameter is the quality of the work surface in contact with the chip. If this surface has defects, the cooling quality can significantly decrease, and the final device will fail. In addition, chips can cause corrosion and damage the radiator.

It is important for the manufacturer to monitor such quality parameters:

  • Uniformity, flatness – any longitudinal deformation will reduce the quality of the radiator’s fit to the surface.
  • Dents, chips, scratches – irregularities on the surface itself that impair heat transfer
  • Marking damage – for an automatic assembly line, the marking on the components must always be read

The main types of radiator defects that the machine vision system detects

For continuous automated quality control, a line has been developed that checks the surface of radiators using machine vision in several stages, which works in conjunction with a robotic arm with a vacuum pump that removes defective parts. To eliminate recognition system errors, images are taken from several cameras at different angles.

The line consists of a circular rotating platform, where one test is performed for each rotation of the platform. The first unit installs new radiators on the platform using a vacuum pump. Next, the product is checked for flatness using a high-precision laser range finder, which runs along the perimeter of the device. In the next step, the camera photographs the surface of the radiator at a right angle. For additional verification, in the next step, another camera photographs the surface at a different angle. The real-time process is shown in the video below.

Quality control line of radiators in action. Description of items counterclockwise.

At the same time, each type of marriage is moved to a separate tray, so that in the future it would be more convenient for specialists to investigate the causes of marriage and adjust production lines.

System components

The computing module for controlling the entire system as a whole works on the basis of a compact industrial computer Advantech MIC-770. We already talked about this series of computers in the article Fanless Performance Computers MIC-7000.

A computer MIC-770 collects readings from all system components

Advantech computers are used to process high-resolution images received from cameras. MIC-730AI powered by the NVIDIA Jetson Xavier platform, specifically designed to run neural networks and machine learning systems. Previously, for such tasks it was necessary to use entire clusters of graphic processors (GPUs) with large active cooling systems. Today, such clusters are replaced by a single computer with completely passive cooling.

Advantech Computer MIC-730AI based on the NVIDIA platform, Jetson Xavier implements image processing using a neural network

Ainavi – Advantech’s machine learning framework for defective parts, designed specifically for Nvidia Jetson hardware platforms.

Video of the system from a different perspective


Today, the introduction of a machine learning system is available to any manufacturer much cheaper than a few years ago. Ready-made hardware platforms fit in a single industrial computer. You no longer need to deploy clusters of video cards. Dozens of typical trained models are already able to use machine learning software frameworks, so the customer does not need to develop the system from scratch.

Similar Posts

Leave a Reply