Hi All, this is my first post so please bear with me.

I am looking at reducing the number of samples taken at each sample point.

The bones of the process are as follows:

Filling process

Fill weight check

Single fill head (fixed volume pump type)

4 Samples taken every 30 minutes

Pass/fail decision is made for each result, not for the average of the 4 results.

What I am hoping to do is to reduce the number of samples taken at each timepoint to 1. I think that if I can prove that there is no difference between each set of data i.e. use the sample number as a sub group, I can justify my decision.

Can anyone advise me as to what they think is the "best" statistical technique to use in this situation.

Many thanks

Pat

SPC sub grouping and frequency is a combination of Rational Sub grouping and economical process control. The goal is to establish the most economical balance of subgroup size and frequency that provides the desired state of control.

You could perform a long term capability study using your current method, then repeat the LT capability using the subgroup size of 1.

If your process shows signs of auto correlation (extremely tight control limits), you could perform an auto correlation analysis. If it exists, a sub group size of 1 would be correct. Auto correlation means that the results of any measurement is influenced by the measurement(s) immediately prior to it.

I would also recommend a short term capability study to get an understanding of the distribution of you current variation. Do you have any variables that affect the fill, such as batch to batch variables, temperature, viscosity, etc.? You may be better off finding what process variables need to be measured and controlled to assure that the fill is consistent, and therefore in need of less verification.:cool:

Guys

Thanks for your rapid responses.

What I have floating around in my head (along with lots of other useless information and a lot of empty space :)) is a rough plan that goes someting like this.

I will be looking at 5 different products/fill sizes. I am planning to get the records from the last 20 batches for each product and use a one-way unstacked ANOVA to compare the data sets for each set of samples, i.e. break the population into 4 subgroups based on the order of measurement at each time interval


I guess I'll have 25 - 40 time points for each batch so I'll have plenty of data to crunch.

If I can prove that there is no difference between the 4 data sets and also between each data set and the total sample population, can I justify that the additional 3 tests at each timepoint are not providing me with any additional control information and are therefore reduntant.

BTW the equipment used is pretty old and although not statistically in control, we do use "adjust limits" inside the acceptance limits to help minimise OOS results.

Thanks

Pat

You mention that a pass/fail decision is made for each sample. Are the actual measurements also recorded? You can generally get better information with less samples if you are dealing with the measurement (variables data) rather than the attribute (pass/fail) in SPC.

Hi Steve,

Thanks for your question.

Yes, we do record the actual values and make the decision based on that. If we do have a failure we have to reject back to the last successful IPC, reset the machine and inspect the next four samples off the line.

As I have said I expect to have approx 100 - 150 data points per batch for 20 batches and for 5 products, so I think I'll propably have a case of RSI and/or will be blind by the time I have all the data entered in Minitab.
:blowup:

Cheers

Is resetting the machine done on the basis of what the SPC is telling you, or on failure to meet specs? If being reset on the basis of specs, you could be having a version of Dr. Deming's Funnel Experiment occurring, and actually increasing the variability of the output.


Hi Steve,

Thanks for your question.

Yes, we do record the actual values and make the decision based on that. If we do have a failure we have to reject back to the last successful IPC, reset the machine and inspect the next four samples off the line.

As I have said I expect to have approx 100 - 150 data points per batch for 20 batches and for 5 products, so I think I'll propably have a case of RSI and/or will be blind by the time I have all the data entered in Minitab.

My recommendation would be to keep it as simple as possible.

Ensure that your data crosses all compoennts of variation: within filling event set-ups, set up to set-up, material lot to lot (including the material being filled and the material that is filled into. Then plot the data first. Then decide how to analyze it....

If the variation within the 4 pieces is practically less than the other components, then you can safely go to 1 per time period.

HOWEVER, since you already have data you could plot it. you don't have to plot every single data point. maybe just 3 sequential sets of 4, 3 sets - beginning, middle and end of a filling event. and select 3 different filling events. That's not a lot of data to enter.

Do you know the average standard deviation of the 4 pieces? and the total standard deviation? that will give us all a really quick look at the feasability of this.

By the way you might google "multi-vari" it is a fabulous visualization of the components of variation without statistics and distribution selection getting in the way. ANOVA and other approaches to determining the contribution of components of variation are just the mathematical quantification of the multi-vari plot.

It also would be handy to know how much of the variation is from the measurement system - in case it is masking the true difference that is occurring between the subgroup samples. It may not matter versus the specification, but it may matter when trying to compare one set of data with another. Hopefully, the measurement system has already been evaluated for its effectiveness. :cool:

By the way you might google "multi-vari" it is a fabulous visualization of the components of variation without statistics and distribution selection getting in the way. ANOVA and other approaches to determining the contribution of components of variation are just the mathematical quantification of the multi-vari plot.

The OP mentioned that he used Minitab. Minitab has the Multi-Vari option under Graphs.

Hi Guys

Thanks for all the responses so far, I have to say that I'm impressed with the feedback :)

I forgot to give some details of the measurement system - Quite simple actually. The line op takes 4 unfilled containers, pre-weighs them to get the tare weight, places them onto the line for filling and then removes and reweighs them to get the gross weight. The net weight is then calculated by simple subtraction. The balance used has a calibration check performed daily.

The project is part of a Continuous Improvement initiative that is attempting to reduce the number of IPC checks required to be performed by the line op. The issue is not really looking at the actual filling process per se, I want to be able to justify reducing the number of checks that are performed.

Looking at some previous process data I guess an average Cpk of 2 would be typical, with some batches as high as 4.5.

Hope this helps clarify things a little.

Thanks

Pat