Building a robust machine vision system

Question: how many good parts do you look at when you’re doing your feasibility study?

Whatever your answer, it probably isn’t enough. Let me explain why.

Variation. Everything varies, in size and in appearance. It can vary with the time of day, the batch of material or the condition of the cutting tools, and your vision system has to deal with it. If you want a totally robust system you have to understand how much the good parts can vary over the long run.

If developers look at good parts at all, (and there are some who don’t,) it’s usually just three or four examples, usually consecutive parts taken from production. This is just a snapshot. All it tells you is what a good part looked like at that moment in time. Next time the machine is set up there could be some subtle differences. It might not look very different to you, but if the grayscale values have shifted you could end up with an increased level of false rejects, or worse still, shipping bad product.

So what do you do? At a minimum, I suggest obtaining samples of good production over several set ups. If you can cover a number of different batches of raw material in that, so much the better. Think about what might drift over a longer period of time. Coolant concentration is one factor that can affect appearance, humidity is another.

Bottom line? Figure out how to get a grasp of the natural, and acceptable, variation in your process, then acquire a set of good samples that reflect this variation. You’ll be glad you did.