Capability Analysis with various subgroups

It helps us determine whether your sampling and/or subgrouping approach is rational (appropriate to the process). Given your process, I would expect minimal variation within an order (setup), but much more variation between orders (setups). Your process would be setup dominant, with some potential impact from material lots.

Bev D said:

These are very low sample sizes - but that is a different discussion.

You can use all of the measurements (failed and passing) from the production data. Do NOT use any QC data as it is from a censored population and will not give you any idea of the capability. You will need around 25-30 batches. Do not censor or cherry pick which batches you use: use the last 25-30 batches produced. Calculate the capability using the Sd of all of the measurements ie the long term capability. There is no statistical manipulation at this point to calculating short term capability - the sample sizes are just too small to rely on any calculation without seeing the data in time series.

This calculation will not have any statistical or practical precision - it will only give you a rough idea of the capability; it’s a staring point.

If you would like you can post your data here (the actual measurements not the statistical summary) and we can provide better advice..

Click to expand...

Do you think that ANOM or ANOVA would be beneficial? By factor year or order?

New to statistics said:

Do you think that ANOM or ANOVA would be beneficial? By factor year or order?

Click to expand...

Valuable for what? is your goal to ‘do some statistics’ or to learn something useful about the process?

What are the specification limits for the data you posted? Why are the QC values so much higher than the production values? Are they for different characteristics? Or does this characteristic somehow get larger between the production measurement and the QC measurement?

You should first plot this data in time series (such that we see within sample values and the different samples. You need to put the specification lines on the graph as well. Then you and we can move to the next step.

Bev D said:

Valuable for what? is your goal to ‘do some statistics’ or to learn something useful about the process?

What are the specification limits for the data you posted? Why are the QC values so much higher than the production values? Are they for different characteristics? Or does this characteristic somehow get larger between the production measurement and the QC measurement?

You should first plot this data in time series (such that we see within sample values and the different samples. You need to put the specification lines on the graph as well. Then you and we can move to the next step.

Click to expand...

Between production and qc we have an extra process where product's heated, and change the hardness of the material.
USL 1.100
LSL 0

For example for comparison through the orders or through the years, how the quality of the products improved, stay the same or go worse.

That's why i referred to Anom amd anova analysis

SPC is a much better analysis for determining if the processes got better or worse.

Only the QC data matters as there is a heat treat process between production and QC.
Is there a historical model that predicts pre-heat treat specs to the post heat treat specs?

SPC you mean control charts?
Specs before and after the heating remains the same.

Miner said:

It helps us determine whether your sampling and/or subgrouping approach is rational (appropriate to the process). Given your process, I would expect minimal variation within an order (setup), but much more variation between orders (setups). Your process would be setup dominant, with some potential impact from material lots.

Click to expand...

Before ask you about sampling and subgrouping it will help you describe a little more.
The process is for seamless tube. For example you get the first billet with diameter of 50mm and with a series of machines goes 40 mm -> 30->20-> 12 mm (final machine) and then goes to annealing and QC received a sample 1/1000 kg.
Usually the specifications is at the last machine and at previous machines you have targets in order to watch the process.
From production we have only measurements such as diameter, wall thickness, length, weight.
In QC we have these but also Rm, Rp0.2 Rp0.5 expansion etc, measurements that production don't have the necessary equipment.

For the common measurements QC take at least one sample to confirm those.

Also for an order of 1000 kg production will measure 6 common characteristics (diameter, etc), QC will measure just 1.

For one machine in each shift will have at least 10 or more orders. Each order have different specifications.

Multiple the above by 8 (machines).

In conclusion the majority of common measurements it will received from the production.

I would like to ask consider the above and the tables that i attached, what statistical checks would you perform ( Cp, CPk Anom, Anova)?
What would be the suggestions about the sampling?

Entails what are the minimum requirements?

Thank you in advance

more detail is very help full.
you said: I would like to ask consider the above and the tables that I attached, what statistical checks would you perform ( Cp, CPk Anom, Anova)?
I will say again: statistics isn't about finding some data and doing math on it. Cp/Cpk and ANOM/ANOVA are statistical analyses that answer very different questions. So what is that you really want to know about this process? or are you just practicing the math?

Cp/Cpk will give you a ratio of the process data to the specification range. it provides some insight to the capability but ignores the most important aspect of the data which is in the time series relationship. (real statisticians and many quality professionals abhor this ratio because too many users conflate it the defect rate based on the Normal distribution and the falsehood that a precise prediction of data that actually lie beyond 3 sigma)

ANOM and ANOVA are not for time series data. They are are to determine if the average of 2-4 events created at the same time are different. To compare different machines, lines, fertilizers, etc.

Control charts are specifically designed for time series data.

read (broken link removed). then go to SPCPress.com and start reading. everything.

I have created a simple run chart of your QC hardness data. If you have given us the right data and spec limit (0-1.1) you are not capable as almost every reading is out of spec. (this is called a bloody obvious statistical test; no math needed) This process has not gotten better over time. it hasn't gotten any worse either. THIS is always the first analysis you do. Plot your data in time series. Then LOOK at it. Then THINK about it.

then go to SPCPress.com and start reading. everything.

Hello Bev D

Bev D said:

read (broken link removed). then go to SPCPress.com and start reading. everything.

Click to expand...

I try to find the books but i cant...do you know where i can download those? or similar that you believe its worthy to read?

Books

Papers