This week Nissan hosted their Tech Innovation conference in Smyrna, where I got the opportunity to speak:
'From Impossible to Possible, Next-Gen AI Vision Tackles Tough Manufacturing Inspections'
And for those of you in Manufacturing, you know Vision's been a problem for a long time.
For tricky inspections - think stamping splits, incomplete welds, or soft-seated connectors..
Many of you have experienced a system working great in a conference room but failing in production.
One OEM controls engineer told me 33% of the vision systems they buy end up sitting unused after 6 months!
So how are we at Overview AI tackling that problem?
Well, these tricky inspections CAN be solved!
Cutting-edge AI models, given MILLIONS of training images, can learn to accurately detect the differences between good and bad, even when variations in defects get very subtle or complex.
But of course, that's great in an ideal world, so how you do you get that working in production?
You need to give the powerful AI Model the horsepower it needs to train it on your parts:
-> A top-line NVIDIA GPU, on the camera, using parallel processing to dissect those massive training sets and teach the model.
Other Vision providers like Cognex or Keyence choose CPUs for most of their Vision Systems... which would take months to train the exact same model.
That won't work in your production environment.
But... having an accurate model is only half the model.
What happens when lighting changes, an operator bumps it, or you need to add an inspection and reteach the model?
You need to react quickly.
Having a GPU on the device means we can retrain our models in minutes, sometimes even seconds. This means you can add new defects to your training set as they pop up, iteratively improving it's accuracy.
In our workshop, I highlighted how our system is automating tricky inspections in production for Stamping, Black-on-black, Welding, and Soft-seated Connectors.
Then audience volunteers set up a soft-seated connector inspection - end to end - having never seen our vision system before.
I asked them to try to break the inspection.
And... they did! On the big screen, in front of the whole crowd.
But then they saw just how easy it was to take false positives, add them to your training set, and retrain the AI model to greatly improve the accuracy!!
We had an hour... but these smart volunteers got it working in 40 minutes!
If you're interested in how we might automate your vision inspections too, give me a shout. Happy to take a look.
