Ok so one of our customers sent us their Cap study, with out of spec dimensions, on a stamping part that we built the tool for. Their Cap Study is showing CPKs ranging from -1.059 to 81.961 is this even possible? This is a 1 piece cap study.
I am aware of a stamping company that calculates Cpk based on each of their 5 piece samples rather than on the run. They are using a screwdriver as a chisel. Just because they do it doesn't make it right - but they have convinced themselves of the practice. As others have said - especially with the number of parts one makes in the stamping process - it is absurd. If one had their individual values from their entire trial run and could correctly analyze the variation it would be of must more usefulness.
First of all, Cpk is the best value of Cpl = (Mean - LSL)/3*Std.dev and
Cpu = (USL-Mean)/3*Std.dev. So, one needs the standard deviation and the mean of the sample to calculate it. What is the standard deviation on one specimen? The mean would end up being the specimen data. Not sure what data manipulation they are doing to even come up with Cpk with one piece.
The closer the mean is to the center of the specification, the larger the Cpk. To get an 81 you need to have the data very close to the center of the specification and a very small standard deviation. It is difficult to give you an example with one part, as the standard deviation of one part, again, is nonsense.
As was mentioned before, to get a negative number the mean of the data needs to exceed the spec.
Capability studies are a thumbnail view of the capability of a process, and are of little value for much else. As described before, sequential data from a controlled process is far more important analytical tool. In fact, if this is precision stamping and the major variation form the process is tool wear, Cpk would not apply because the distribution is
non-normal. It should never be negative, but it will be very low as the tool should be made at one end of the spec to allow it to wear to the other end before replacement.
That is not normal distribution. Poor measurement and lack of understanding of distributions allows Cpk to exist in the industry, as wrong as it is. But with a little information we have here, we don't even know if the variation is a tool issue or maybe even a measurement issue. When measurement error exceed the process variaon, the data will
appear normal (measurement error is a normal distribution.) That problem has fooled many, many people.