“Sorting this material by hand will take two guys about six months and cost us around $60,000,” said my client.
Someone on the production line had installed the wrong component leading to one dimension being off enough to be unusable, and 3 million parts had gone through before they realized the problem. Even though only 1 in 22 parts were bad, the whole batch was tainted and headed for scrap unless the good parts could be sorted and verified. At 7 cents per part, that added up to around $200k.
“So we want you to modify your image analysis code to measure for this defect and see if there is enough separation to accurately reject the bad parts,” he continued.
Luckily the images we were already taking were the right angle and the analysis code similar in nature to what we needed. Five minutes of coding later, I was ready to try it out, thanks to modular code and a good architecture. Below are the results. Above the black line are good parts, below are the bad.
Five more minutes of code had the bad parts being kicked out of the system. They ran the system overnight and made it through 400,000 parts during that shift. Spot testing showed no bad parts passing and very few false rejects.
“This works so well, we want you to roll it out to the other production lines permanently so we can pick up this same common problem before parts get through next time,” he said the next day.
Another five minutes of code, and I started installing the update on the first new line. Except the change wasn’t working as well on this line. One of the 22 stations was showing a measurement way out of spec from the others. Then realization hit.
“Um, it looks like one of the stations on this line has the same problem,” I quickly informed the line manager.
They shut it down, and sure enough, same wrong component. Luckily the rest of the lines were problem free.
“You just saved us a ton of money,” my client gratefully commented. “And two guys a very boring six months,” I replied.