SPC A-D TEST FAIL

Hi Masters,
Recently i had conducted SPC study for some pressing component having the critical dimension of 17 mm with -0.2mm tolerance. I had used recently calibrated 0.01mm L.C digital vernier caliper. And collected 10 Subgroup and total component used is 50. The result was CpK=1.42 & CP=1.69. As per my company norms the above mentioned result is in Green Zone but the study has been failed in A-D test and the Pvalue is 0.0069. I don't know what action i have to take to get the result OK & what are the root causes behind the failures? Please help me.

 

It sounds like the Anderson Darling test is assessing the normality of the data and indicating that it is not normally distributed so that using Cpk and Cp as indicators of your process capability will not be accurate.

Posting the data will allow others to help but you can start by graphing the data in a histograms and looking at how it is distributed (does it look like a bell curve or does it have multiple peaks, outliers, etc) and/or by Q-Q plots. These might show why/how it is not normal, not just that it is not. Doing a run chart to show drift on the process will also be useful although youay already see it (or not) on your SPC

Understanding of the process is also likely useful. For instance, if your pressing operation has multiple dies then you may be getting different populations from each. It may also vary from raw material batch to batch, etc.

David

 

Yes, I would post the data for review. I am suspicious as to what you indicated that your gage has way too little resolution for SPC. If you did a gage R&R, you need an ndc of about 10. I do not think that is possible with your spec and that gage. You indicated your tolerance was -0.2mm. Did you mean +/- 0.2 or +0.0/-0.2? For +/- 0.2 is a 400 micron (0.4 mm) tolerance. Just to measure the dimension, you should have 10:1 tolerance, or a gage that measures 40 micron increments (.04 mm). For SPC, you should be within 75% of your tolerance, so the gage should be even closer to 30 micron increments. You need this resolution to accurately flesh out the distribution. Chunky data from insufficient resolution will not calculate the distribution correctly. I think you find that with the correct resolution, the correct measurement technique - and if the expected distribution is normal - your data will be normal. For example, if you were doing precision machining, for example, the expected distribution is non-normal due to tool wear, and if you passed normality tests it would actually be a bad thing - usually excessive gage or measurement error masking the true distribution. Do you know what your true expected distribution should be? Rather than blindly doing a normality test, I prefer curve fitting, such as Distribution Analyzer from variation.com, to see if it is non-normal, and if so what distribution is a better model of the process variation.

As David indicated, never mix data from different dies. That will absolutely generate multimodal distributions.