Outlier elimination: how reliable or are there pitfalls?

[Jan d.]

My machines generate a fair amount of outliers (I have real bad castings and sometimes I take points in areas where I have a lot of material missing). The only way to get the system to throw them out automatically is to play around with the multiple of the standard deviation. Sometimes a sinlgle point may be out by as much as 0.050" (all the others are within 0.0003"; plane constructed through vector points). The only way to get PC-DMIS to take that one out is to use a a 2.0 for the standard deviation multiple (3.0 wouldn't work). That seems very steep. Gets me worried about what else may be going on.

Does anybody have any experience with throwing out outliers automatically? How reliable is it? Any pitfalls? Any tips? Or just use-as-is???


Thanks.

[JamesMannes]

My machines generate a fair amount of outliers (I have real bad castings and sometimes I take points in areas where I have a lot of material missing). The only way to get the system to throw them out automatically is to play around with the multiple of the standard deviation. Sometimes a sinlgle point may be out by as much as 0.050" (all the others are within 0.0003"; plane constructed through vector points). The only way to get PC-DMIS to take that one out is to use a a 2.0 for the standard deviation multiple (3.0 wouldn't work). That seems very steep. Gets me worried about what else may be going on.

Does anybody have any experience with throwing out outliers automatically? How reliable is it? Any pitfalls? Any tips? Or just use-as-is???



Thanks.

Jan,
Two things,
1.) you say you have a lot of castings with material missing? If the surface you are measuring is controlled by profile then you want these OOT points included don't you?(they have an assignable cause that is true to the part and not a known error).
2. As far as removing outliers, do you mean through the generation of feature sets? Or how? Using the outlier removal tool in feature sets does not?did not work well through 3.7mr3. It would only work with a straight line. Not cool hey? Well, I take another look at this thread later or tomorrow. Chaio
P.S.
If the software allows different types of outlier removal now, could you provide a screen capture? Intersted to see it.

EDIT...
Jan sorry couldn't think of the name last night. I meant to say constructed filter set. Is that correct not feature sets.

[Jan d.]

1.) you say you have a lot of castings with material missing? If the surface you are measuring is controlled by profile then you want these OOT points included don't you?(they have an assignable cause that is true to the part and not a known error).
2. As far as removing outliers, do you mean through the generation of feature sets? Or how? Using the outlier removal tool in feature sets does not?did not work well through 3.7mr3. It would only work with a straight line. Not cool hey? Well, I take another look at this thread later or tomorrow. Chaio
P.S.
If the software allows different types of outlier removal now, could you provide a screen capture? Intersted to see it.

EDIT...
Jan sorry couldn't think of the name last night. I meant to say constructed filter set. Is that correct not feature sets.

1) Yes, you are right. So I DO include them in my report. But of the 30 vector points I measure on the plane, this 1 point skews the whole plane that I construct through all of them. And that plane is being used as a datum for the next planes. I have a chain of datums. Problem is, it is always another point that falls off. My plane consists of 6 small pads that have extremely minimal casting support features (they are skinny and long). On top of that, they are full of tapped holes and dowel pins. So I can never predict which point may become an outlier.

So what I was planning to do was output all vector points (just Z-value) to look for core shifts. Then contruct a plane through all and throw out the obvious outliers, which I apparently can do in V4.1 (never worked with an earlier version). However, I am worried about the method that PC-DMIS is using to identify the outliers. If I find it to work for one set of data, does the same setting work for another???

2) Go to the contruction of a plane and it offers outlier removal as a standard option in V4.1. It uses the standard deviation as the way to determine what is and what is not an outlier. I find that it is very difficult to guess, based on the standard deviation, the point at which PC-DMIS will eliminate the point. I have found that when it works for one point out, say 0.200", it may not work on another that is only 0.050" out. Then when I get it all to work work for the 0.050", I throw out almost all points! Very confusing.

I think my problem is that I measure these 6 pads. On every pad I have 5 points and this probably gives a "weird" distribution, which cuases my standard deviation to be very big every now and then. Although all are very close on 1 pad, if you take all 6 pads, you do see very significant steps between the pads.



Jan.

[JamesMannes]

Wow, I'm glad I responded to your post! Learned something new today about the 4.1 version (outlier removal in a constructed plane). What if you were to use 3 points per pad? That's how I do my datum pads that have a diameter describing the datum pad? Perhaps that will help with you SD?