'Representative' Sample of the Lot for a Continuous Process

[v9991]

in a blow-fill-seal process, involves simultaneous filling of 5-units (a single cassette packed as unit) ;
an inprocess controls and monitoring involves weight and volume checks wrt target.

for the final qc-testing, samples are selected for analysis; and

the question is, if we have to pickup 10 samples for a given test, should we pick up 2-casettes; or shall we pick up randomly 10-individual units randomly from the batch.

i would like to propose that, 10 samples for each of the simultaneous filling nozzles shall be randomly selected; (considered that the 10 samples are statistically justified and adequate)

 

[Tony_C]

A great question - There is probably an opportunity to reduce variation and eliminate the final inspection by collecting information in a strategic way.

I suggest selecting 2 cassettes of five in the following way.
- Select the first cassette of five at a shift change or another significant event in the process. Keep track of the five stations within the cassette.
- Select the second cassette of five at a different shift or other significant event in the process.
- Plot the two subgroups of five on a piece of paper using different symbols for each of the five stations. ex. station 1 = square, station 2=triangle, station 3 =X, etc.
- Try and determine if any preliminary conclusions can be performed about variation caused by the differing stations within the cassette, and variation caused between shifts.
- If all the data appear homogeneous, you could make a control chart with initial limits.
- Each time you are required to draw 10 pieces for inspection, repeat the above process, and add the points to the control chart revising the limits. Work with production to reduce variation based upon the data.
- Compare the data to the specification limits, and eliminate or adjust the inspection frequency.

 

[Steve Prevette]

I'm going to offer a pure random sample alternative. If you really want "representative" if you try to do a targeted sample (such as the suggestion of doing them at shift change) you do run a significant risk of not getting something representative of all items made.

A few ways to do this:

Generate a table of random times during the shift to take whatever five unit batch was most recently filled and then roll a die for which bottle to choose (yes, roll again if you get a six).

Another option is to have a random number generator button that is activated as a batch completes. If it chooses that lot to be sampled, then it can also run a random number between 1 and 5 and say choose that bottle. As an example, a facility I work at requires random inspections of persons entering the facility. As you are about to enter, you have to hit a button and it either shows a red (get inspected) or green (go on) light.

 

[v9991]

the approach of 'random number ' does bring a a structure to the sampling process, BUT
how does the random number generator ensure the
a. representation of sample across the batch
b. sample size relevant to that of process performance and batch size

or does the random number generation takes care of it?

 

[Steve Prevette]

the approach of 'random number ' does bring a a structure to the sampling process, BUT
how does the random number generator ensure the
a. representation of sample across the batch
b. sample size relevant to that of process performance and batch size

or does the random number generation takes care of it?

a. By getting a "pure" random sample, you do get a representative sample in the long run. Do compare that though to the polling in the last US election where maybe the people were chosen to take the poll was random, they could choose or not choose to be sampled.

b. You do need to decide at what rate you will do your random sampling, there are plenty of sample size tables out there. A lot depends upon how long you can wait until finding out there is a problem. And a lot depends on if the sample process is non-destructive or destructive. Basically become a cost-benefit or risk analysis that can be solved with computer simulation, or just choosing a sampling plan, such as from MIL-STD-105 or its current versions.