In this case at AM/NS Calvert, the marking robot did not provide feedback. Poor or no marking creates rework in the coil yard, in addition to a potential safety hazard, so a solution had to be found that was reliable and wouldn’t require human intervention.

The solution was to implement a computer vision system that would detect missing or poor IDs. Microsoft Custom Vision was used to create and train a neural network, and it's run on site, using Docker. Already installed 4k cameras observe the critical area and if the system detects weak or no coil marking on the coil, there are two 2 types of alarms.

As of July 2020, the system is online and working. An image of every coil is saved pre and post detection, and a tool to review the results and images is provided. Post marking results and pictures are also integrated into the quality system. The model has been quite accurate, at 93.3% for outer wrap with glare accounting for 3% of errors. On the side wall of the coil, it's even better, with 98.4% accuracy. Coil form is the main issue for incorrect readings.

Smart people and digital products helped solve this challenge for AM/NS Calvert, increasing efficiency and saving money and time.