Sorting Out Attribute Data - What is this data telling me?

[TownDawg]

To all - I'm attaching an excel spreadsheet of "attribute" data.. would any of you have any thoughts as HOW to analyze what (if anything) the data is trying to tell me?

[DrM2u]

Let's see ... Seems like there was a sample of 330 parts analyzed. The part is 290.2436 (mirror?). There is the time stamp for every part when produced or inspected (not sure). There is a C weld and (4) M welds that are done by the same operator or different operators. Each weld was inspected against an unspecified criteria and was either approved (APP), rejected (REJ) or NDF if the part failed the C weld. Here are some observations:
- the inspection criteria is not specified but there must be something to help decide if parts are good or bad
- a part is NDF (not completed?) if it does not pass the C weld
- a part is complete M welded if it passes the C weld
- you can count how many C failures you had in the whole sample
- you can count how many of each M failures you had in the sample
- you can count how many failures you had per part in average
- you can look at potential relations between time and failure rates and/or types
- you can decide if the parts meet your acceptance criteria

[TownDawg]

Good points.. and lots of good direction.


"Seems like there was a sample of 330 parts analyzed."
Yes, as I recall.

"The part is 290.2436 (mirror?)"
The 290 is the series.. and the 2436 is the 24"36" size. There are a few other sizes at the end.

"There is the time stamp for every part when produced or inspected (not sure)."
Time stamp is when the part was produced/inspected -- because it is operator in-line inspection of weld.

"There is a C weld and (4) M welds that are done by the same operator or different operators."
C is "corner" and M is "miter". Often the same operator, but not always.

"Each weld was inspected against an unspecified criteria and was either approved (APP), rejected (REJ) or NDF."
Criteria is based upon visual criteria, and supported by boundary samples -- and you are correct, if the "C" weld failed, the remaining 4 "M" welds are not performed until the mirror is reworked.

"how many C failures you had in the whole sample, how many of each M failures you had in the sample, how many failures you had per part in average, look at potential relations between time and failure rates and/or types."
Yes, but is there a simple way to do this? I was "on the hunt" for some software (or even Excel) that would do some calculations for me.. I was feeling lazy I guess?

[DrM2u]

Originally Posted by TownDawg

Yes, but is there a simple way to do this? I was "on the hunt" for some software (or even Excel) that would do some calculations for me.. I was feeling lazy I guess?

Try the COUNT and COUNTIF functions in Excell. They work both for rows and columns. This way you can count how many instances of each APP, REJ and NDC you have in each row and/or column. From there, is simple math. You can also create a parallel set of columns for the data that have 1 & 0 for the APP and Failures by using the IF function. Then you can plot the values against time for additional studies. You can also calculate the time difference from sample to sample for estimated cycle times. Hope this helps a little.

[TownDawg]

Actually, it was a little laborious, but I used a matrix of "dcounts". As you know, sometimes the data is not self-evident. However, sometimes seeing the information graphically (sliced/diced) helps to visualize some possibilities.

[Caster]

I'm surprised it hasn't been said yet..... "plot the dots"

I always start with a simple run chart, time vs events.

It might show a pattern that repeats over time, could be operator, machine, whatever.

Get it out of columns and into graphs.

Spreadsheets are for accountants, graphs are for problem solvers.

Next possibly do a good old Pareto, by cost rather than count if possible.

Always try to monetize defects, how much are these rejections costing, is it even worth going after them, or are there bigger dragons to slay?

[DrM2u]

So, seems like you have about 8% initial failure rate (C-weld) that means the parts are scrapped. There is another 5% failure rate on the M-welds but I am not clear if these are scrap or not. Interestingly enough, the higher failure rate is on the first M weld (3%) and no failures on the last M weld. Also worth mentioning that the lowest failure rates are during the hours when most people are most efficient, in the morning. The highest failures seem to take place around lunch time then they dwindle down again in the afternoon. This brings up the question of how much impact on product quality is due to the operator, since we already know that the detection is operator dependent. Are the operators more aware later in the day and therefore detect more failures, or are they less careful and cause more failures?

Just like Caster suggested, I also suggest to further analyze the causes of failures. A number of charts, including Plot and Pareto, are very good tools for this. I like the graphs and they definitely help you see the failure rates.

Very nice example of analyzing data to identify opportunities for improvement!

[TownDawg]