Analyzing Attribute Data - Mechanical Pull Testing of Joint

[Conquer]

Fellow Elsmar members, I need some input.

At my company we were trying to validate a process for joining two plastics together. In the process, we will using mechanical pull testing of the joint to ensure it meets the 20 N minimum.

After the validation was completed, I reviewed the data and presented by the team and determined that the data was not normally distributed. The root cause of this was that the test method didn't always result in a failure of the joint, but rather a failure of the joint used to prepare the sample for testing.

So, now I am left with a bunch of data that I can only treat as attribute data and am trying to determine at what confidence I know that my pull strength will meet the 20N minimum. What statistical tools for attribute data can I use to provide me with confidence that I will meet the 20 N minimum.

Any help you could provide would be greatly appreciated.

[Bev D]

If you post your data we can provide better help. you do not necessarrily have to treat your data as categorical.

[Statistical Steven]

In Reply to Parent Post by Conquer

Fellow Elsmar members, I need some input.

At my company we were trying to validate a process for joining two plastics together. In the process, we will using mechanical pull testing of the joint to ensure it meets the 20 N minimum.

After the validation was completed, I reviewed the data and presented by the team and determined that the data was not normally distributed. The root cause of this was that the test method didn't always result in a failure of the joint, but rather a failure of the joint used to prepare the sample for testing.

So, now I am left with a bunch of data that I can only treat as attribute data and am trying to determine at what confidence I know that my pull strength will meet the 20N minimum. What statistical tools for attribute data can I use to provide me with confidence that I will meet the 20 N minimum.

Any help you could provide would be greatly appreciated.

You can use the binomial confidence interval to determine the limit on the percent outside the 20N limit.

[bobdoering]

In Reply to Parent Post by Conquer

At my company we were trying to validate a process for joining two plastics together. In the process, we will using mechanical pull testing of the joint to ensure it meets the 20 N minimum.

After the validation was completed, I reviewed the data and presented by the team and determined that the data was not normally distributed. The root cause of this was that the test method didn't always result in a failure of the joint, but rather a failure of the joint used to prepare the sample for testing.

So, now I am left with a bunch of data that I can only treat as attribute data and am trying to determine at what confidence I know that my pull strength will meet the 20N minimum. What statistical tools for attribute data can I use to provide me with confidence that I will meet the 20 N minimum.

I have this very same issue with weld pull tests, and it does render the data questionable. Typically, the idea is to understand the strength of the weld, not the material - so when the material breaks first the weld is good, but the data does not represent the weld strength. It must by its nature under-represent the weld strength. One would want to throw out that data point because the break did not represent the weld, but you do not know the actual weld strength. That is important, because you may be throwing out high data that would have beneficially affected your capability.

If the question is will the joint system meet 20 N, (so breakage at any point), then you can develop your statistics from the data you have using non-normal distribution statistical analysis for the appropriate distribution model.

The bottom line is you surely can not tell the capability of the joint strength itself accurately with this kind of data - although people will attempt to in order to appease myopic customers.

Ultimate Tensile Strength is usually the worst for repeatability. Often people will use yield to get a more consistent data point for tensile data, but that may still be confounded with the base material yield force in your case distorting the conclusion. I would not likely help your case.

I agree with Bev D - it would be nice to see the data. Also, do you ever get the tensile curves? They would also be nice to see.

[Conquer]

Thank you Bev, Steven and Bob for your reply. I have attached my data to this post to provide some insight.

How the data was run was in 3 seperte lots at nominal processing conditions. As you can see from the histograms for the lots, the data is skewed because some break at the actual material joint while others break at the joint created to be able to conduct the test

In reply to Steven, I checked into the binomial confidence interval and all looks good accept that I have a p value of 1. In fact, I will always have a p-value of 1 because if anything fails my product fails.

Also, in the last tab on the sheet I have tried to use an equation for calcuating the sample size needed based on a zero-acceptance sampling plan. I used the equation to work backwards and determine based on a total sample size of 90 peices and a confidence interval how many pieces I could expect to be non-conforming. Not sure if any of you have a comment on this concept.

[Statistical Steven]

In Reply to Parent Post by Conquer

Also, in the last tab on the sheet I have tried to use an equation for calcuating the sample size needed based on a zero-acceptance sampling plan. I used the equation to work backwards and determine based on a total sample size of 90 peices and a confidence interval how many pieces I could expect to be non-conforming. Not sure if any of you have a comment on this concept.

Here is the analysis of your data using the upper confidence interval on the binomial. It uses the SAME approach as your last sheet (though instead of solving for n, I solve for p in the binomial for a given sample size.

N Confidence Reliability PPM Upper Limit
150 90% 98% 19835 2.0%
95% 98% 24319 2.4%
99% 96% 35223 3.5%

30 90% 91% 94803 9.5%
95% 89% 114726 11.5%
99% 84% 160327 16.0%

[bobdoering]

As a quick test I put all of the data together into Distribution Analyzer and attached the results. Remember, this data does NOT represent the joint strength, just the joint system strength (including failures other than the joint itself)

[Jim Wynne]

In Reply to Parent Post by bobdoering

I have this very same issue with weld pull tests, and it does render the data questionable. Typically, the idea is to understand the strength of the weld, not the material - so when the material breaks first the weld is good, but the data does not represent the weld strength. It must by its nature under-represent the weld strength. One would want to throw out that data point because the break did not represent the weld, but you do not know the actual weld strength. That is important, because you may be throwing out high data that would have beneficially affected your capability.

If the question is will the joint system meet 20 N, (so breakage at any point), then you can develop your statistics from the data you have using non-normal distribution statistical analysis for the appropriate distribution model.

This is a good reason that cohesion failure (the material coming apart before the weld does) should be accounted for. In many instances of pull testing it's expected that the adhesive bond will be stronger than the cohesive properties of the materials joined material(s). If a bond is expected to withstand a pull force of x units, and there's cohesive failure before that limit is reached, the bond is considered adequate. This results in attributes data, which should be considered an acceptable way of validating the process.