Determining Sample Size and Method

Hello,

I am new to this forum. I work for a dietary supplement manufacturing company. We primarily make powders that we fill in jars. We have a QA department and they do quality checks on the filling and bottling process hourly. QA will go into the room and randomly check jars for weight and seal compliance. There is no "set" number of how many jars are checked at this point. If no unsealed jars are found that hour , production continues. If unsealed jars ARE found, production stops and we adjust. Plain and simple. We are trying to advance our QA Program and want to implement a sampling plan that is more credible and valid. I realize that we need to calculate what a valid sample size would be. To go into further detail, we have been having issues finding jars that have not been properly sealed after the induction process. Currently if QA goes into do an hourly check and a bunch of jars are found to be not properly sealed, we stop the line and adjust whatever it is that we need to . Sometimes it is because the induction sealer got moved , sometimes it because the capper randomly decided to act up. That product is then reworked later and we continue as described...hourly checks...if we find jars to be unsealed we stop the line and adjust. What methods or sampling methods are out there that would help us determine how many jars we need to be checking an hour? Where do we start with not much data on the defect rates? I've looked up AQL but it doesn't seem to apply as we do hourly checks. I have also found calculators that tell you valid sample size and I have split that number by how many hours we are doing production. Nothing is really "making sense" for what type of sampling plan I need. Any suggestions would be greatly appreciated.

How many parts do you make in an hour? What is your historical reject rate?

Depending on what we are making , I would say about 2000 units an hour average. We are just now starting to collect data on defects. Is that what you mean by reject rate? We are simply doing hourly checks and stopping the line if we are finding they are not sealed all the way , fixing the problem and running again. I would say in a typical run we could end up reworking about 200 jars for seals. We check boxes of sealed product and if 5 or more are found to be open we move the whole box aside to rework later. Boxes aren't very large. Really there is no ryhme or reason to how we are sampling. So far I collected data on one product and the defect rate in seals was 1.33%. This is what we found by random hourly sample checking as defective for seals. Not sure how to proceed with this information.

The frequency of sampling should occur at least as often as the process is expected to change. Examine all the factors that are expected to cause change (sealer, capper, etc.). To be effective, sampling must occur at least as often as the most frequently changing factor in the process. If it is not clear how frequently the process changes, then collect data often, examine the results, and set the frequency accordingly. It is better to start off sampling more often than may seem necessary, and reduce the frequency based on what you learn, than to sample too little.

Sarahb333 said:

Hello,

If no unsealed jars are found that hour , production continues. If unsealed jars ARE found, production stops and we adjust. Plain and simple.

Click to expand...

This could actually be a very expensive and fruitless policy. Yes, plain and simple. but could be just simply wrong. Odd that I have another thread going on SPC limits, and invoking Dr. Deming's red bead experiment, and also we need to invoke the (lesser known) funnel experiment. If indeed there is a stable process creating unsealed jars, stopping the line after an individual result and searching for a unique answer COULD BE ineffectual, and worse, could be making the process worse (as predicted by the funnel experiment).

On the overall question of sample size, I'd suggest that the periodic results from whatever sample size you are currently using be plotted on a control chart, and see if the percent defective remains stable (and predictable) or if there are signals of special causes. The old MIL STD 105 did have some reasonable sample sizes for continuous processes, though you do have to declare a "batch size". A primary driver for sampling on a continuous process is to answer the question - how many defective product are you willing to make before you detect an issue?

I use (and have used for decades in various industries) an AQL sampling standard based on ISO 2859 for my final inspection. This isn't going to fix all your problems however it will allow you to quickly put something in place that will be more effective than random sample guesswork, and give you some sampling data to analyse.

Now I know I am probably going to get lambasted by more informed members of the forum by suggesting the use of a sampling standard without doing any initial analysis on the process, however my personal belief is that putting some structured sampling in place now and refining it when you understand more about the process would be better than waiting until that analysis is completed, to then take action.

I found this - Explaining Acceptance Quality Limit (AQL) for product inspection:
https://www.eurofins.com/assurance/consumer-products/resources/articles/explaining-acceptance-quality-limit-aql/

Scanton, I agree with you completely. Whether you choose ISO 2859 or MilStd 105 or some other sampling plan, at least it is a beginning. Choose SOME method of consistently sampling product, collect data, and then take action(s) that are appropriate - the PDCA cycle works for this as well as anything else.

There are so many pulibhsed theories and standards for sampling. ISO 2859 or MiStd 105 are so old and maybe fall behind the times.
There are many new sampling theories come into public awareness, for your production, ISO 28593: 2017 will be much suitable, you can calculate the sample size simply for any time period.

Sounds interesting - do you have any details on the newer standard?

ISO 28593:2017



$80 (CDN) for a 9 page booklet.
I think I'll keep my money and use some of the old sampling plans.