Tricks for sample sizing given very low failure rate

[tgeorge]

Got a weird question. In trying to minimize cost, we are trying to pre certify a counting machine. Customer requirements are that machine must count all of the data with 99.8% accuracy over its entire life. Fairly strict!

Since we are talking only successes and failures this nets prohibitively large testing sizes (to get the test resolution to even say something with 99.8% confidence I think I need test sizes in the millions(darn that non continuous data)) using any of the standard discrete distributions.

Does anyone have any tricks for using smaller sample sizes? Distributions that we have to use are strictly binomial or discrete in nature.

Or maybe some interesting math?

 

[tomwalz]

I buy machines like that so have been testing them for about ten years. It seems to us that if they work they keep working with equal accuracy until a part wears out.

What seems to be a much greater problem for us is part size / weight variation.

To get 99.8% you need to have parts with maybe less than 1% variation if you are weigh counting.

The only way we get 99.9+ is with individual counting using a feeder passing an electric eye. The problem there is with feeder design to prevent stacking or shadowing.

Before you spend a whole lot of time I’d make sure you knew just what the actual tolerances were going to be on the parts being counted.

Also for us the life of the machine is how long it counts accurately. When it stops counting accurately it has reached the end of its usable life and needs to be rebuilt. This isn't necessarily that big a deal. One of our very small machines has a flipper gate to switch from one container to another. That flipper gate wears out about every other year. It is about $30 and 20 minutes and then it is “rebuilt”.

Not sure how clear all this is. Give me a call if you like and we can chat about it.

Tom Walz


Thomas J. Walz
Carbide Processors, Inc.

http://www.carbideprocessors.com