Reducing false rejects

False rejects are the bane of the machine vision system developer’s life. If the level gets too high (and “too high” might be related to the cost of scrapped product or the impact on Overall Equipment Effectiveness,) production management will usually flick the “Off” switch, preferring to rely on human inspection instead.

So what’s to be done?

Well in Vision Systems Design (June 2^nd, 2011) Andy Wilson describes how Cognex are attempting to address this with ever-smarter software tools. However, I’m not sure this is the right way to go.

Over recent months I’ve been brushing up on my statistics, and I’m coming to the conclusion that we vision engineers can do a much better job of educating our customers, and ourselves, about how measurement uncertainty impacts inspection performance.

In particular, there were a couple of papers published on the Quality Digest website recently that led to this insight. These were written by quality statistics guru Donald J. Wheeler and, in my humble opinion, provide a very practical way of addressing our false reject problem.

The first paper you should read is “The Intraclass Correlation Coefficient.” This deals with determining the value of a measurement system for performing a particular task. (For our purposes, let’s assume that a vision system is a measurement system.)

Then, when you’ve plowed through and understood Dr. Wheeler’s mathematics, move on to the second paper, “100% Inspection and Measurement Error.” This relates the uncertainty in our measurement system to the level of false rejects and false accepts that can be expected. What I really liked about this is that it actually provides a numerical/mathematical basis for decision-making, rather than an approach based on beating up the manufacturing engineer and relying on unreliable human inspection.

As Dr. Wheeler points out most eloquently, the best way to reduce false rejects is to make less bad stuff, but failing that, let’s at least bring some science to bear. And if that doesn’t work, well perhaps then it’s time for smarter software.