PDA

View Full Version : RF30 with Shumatech DRO: test cube data



rgsparber
06-24-2007, 11:12 PM
My intension here is to show what can be done with typically HSM machines and NOT as a platform for bragging. I hope the reader accepts these findings with that in mind.

I decided to start a new thread on this subject because it has taken me a few weeks to refine my technique and make a decent test cube.

My goal was to see how accurately I could machine a test cube using just my RF30 mill/drill and a Shumatech DRO equipped with Harbor Freight calipers. My finish cuts are done with an ultra fine feed on the mill/drill and reading the HEX display from the DRO in order to get around the fact that decimal displays are only shown to 0 or 5 tenths in the smallest digit. I use an HP32S calculator to convert to decimal inches rounded to the nearest tenth.

I recently had the great fortune of receiving a "spoiled" 0.800" Jo block good to 6 millionths of an inch. It has a ding in it but is still plenty good over the rest of the surface. I cut my test block to 0.800" and first verified my mic with the Jo block and then on the test cube. All measurements were conducted at a temperature of 80F and 3% relative humidity (I live in the Sonoran desert).

I measured the thickness of each of the 3 dimensions at the 4 corners of each dimension and in the middle. My mic has a slip clutch and I use it for each reading on the Jo block and on the test cube.

I have not yet tested the block for squareness.

Across all readings I had an average of 0.7999" with a standard deviation of 0.0002".

The full details are shown below:

for dimension
dim UL UR Center LR LL average SD
1 799.7 799.8 800 800.1 799.9 799.9 0.2
2 799.5 799.9 799.9 799.9 799.6 799.8 0.2
3 800.2 799.5 800 799.9 800.4 800.0 0.3

average 799.9
SD 0.2

Dimensions 1 and 2 were entirely cut with the end of the end mill. Dimension 3 had one face side milled while the other was cut with the end of the end mill. You can see that side milling is not as consistent as using the end of the end mill.

mklotz
06-25-2007, 01:23 PM
Rick,

What formula are you using to compute standard deviation? Just for giggles, I ran your first series of measurements,

799.7, 799.8, 800.0, 800.1, 799.9, 799.9

through one of my programs and I get 0.14 as the SD, not the 0.2 that you quote.

Also, how do you arrive at your final overall mean and SD? If I treat all 18 measurements as drawn from a single distribution (which I think is wrong), then that distribution has (mean, SD) = (799.888, 0.234) which agrees with what you show. However, treating the three sets of six measurements as independent distributions (which I think is fair) and then combining them into one composite Gaussian distribution, that distribution would have a SD which is the RSS (root sum square) of the individual SDs,

RSS (0.2, 0.2, 0.3) = 0.412

or about twice what you show.

Regardless of my comments, I think that, if you're going to quote statistical results, you make clear exactly what you are assuming and doing to arrive at those results.

pcarpenter
06-25-2007, 05:46 PM
I guess I still don't understand what this shows, in terms of the abilities of the equipment. Machines are intentionally evaluated for the tolerances they can hold across some unit of length. Maybe .0005" per foot is a reasonable expectation for a mill to hold, for example. I don't know that you can conclude anything about the abilities of your machine from something .79" long.

One could argue for example that if you had an end mill with a diameter of say 2 inches and you brought it straight down on the cube that you could expect it to be near perfect. That in no way evaluates the quality of the machine in terms of its ability to hold tolerance over its work envelope as the table moves....which is the real test of real-world work on a mill.

Get yourself a long bar and try that and see what you get. You would need to not flip it lengthwise, but rather flip it in place and mill from one end to the other. If there is angular error in the way surfaces, this should show that value X2.

Paul

rgsparber
06-25-2007, 06:44 PM
Rick,

What formula are you using to compute standard deviation? Just for giggles, I ran your first series of measurements,

799.7, 799.8, 800.0, 800.1, 799.9, 799.9

through one of my programs and I get 0.14 as the SD, not the 0.2 that you quote.

Also, how do you arrive at your final overall mean and SD? If I treat all 18 measurements as drawn from a single distribution (which I think is wrong), then that distribution has (mean, SD) = (799.888, 0.234) which agrees with what you show. However, treating the three sets of six measurements as independent distributions (which I think is fair) and then combining them into one composite Gaussian distribution, that distribution would have a SD which is the RSS (root sum square) of the individual SDs,

RSS (0.2, 0.2, 0.3) = 0.412

or about twice what you show.

Regardless of my comments, I think that, if you're going to quote statistical results, you make clear exactly what you are assuming and doing to arrive at those results.

Marv,

I'm really glad you saw this post since I know you have a strong math background. My knowlege of how to apply statistics to machining is minimal. What is the best way to characterize a test cube so it give useful results?

I used the SD equation from Excel and it matched the results from both my HP32S built in program and the one from an old text book of mine that gives the definition of SD.

Thanks in advance,

rgsparber
06-25-2007, 07:03 PM
I guess I still don't understand what this shows, in terms of the abilities of the equipment. Machines are intentionally evaluated for the tolerances they can hold across some unit of length. Maybe .0005" per foot is a reasonable expectation for a mill to hold, for example. I don't know that you can conclude anything about the abilities of your machine from something .79" long.

One could argue for example that if you had an end mill with a diameter of say 2 inches and you brought it straight down on the cube that you could expect it to be near perfect. That in no way evaluates the quality of the machine in terms of its ability to hold tolerance over its work envelope as the table moves....which is the real test of real-world work on a mill.

Get yourself a long bar and try that and see what you get. You would need to not flip it lengthwise, but rather flip it in place and mill from one end to the other. If there is angular error in the way surfaces, this should show that value X2.

Paul
Paul,

I think of machining a surface as defining a point on the surface and then seeing how the surface behaves around this point. Initially I focused on how well I could cut the point. That is a function of the repeatability of my machine and its DRO. The behavior of the surface is a function of how the test block or bar is held in the machine and how well the machine moves along its XY plane.

Many machinists on this BBS encouraged me to cut test cubes to a precise thickness. Obviously I could have set up a large diameter cutter to a given height and fed the block under it, turn the block 90 degrees and repeated the operation for all 3 dimensions. That proves nothing as you point out.

You will have to take my word for the fact that I started with a rough 1" cube and cut it down to a square 0.850" cube before going for the finished size of 0.800". This is my 3rd test cube. They have been steadily improving in accuracy as I learned more about my set up.

I believe that this test cube demonstrates that I can cut to the desired thickness 3 times in a row plus hold the surface variation to the stated tolerance. It is not meant to show how well the mill cuts over a large distance. That is a different, and interesting, test. Thanks for the suggestion.

When I do cut a bar to a given thickness over say 12", I would still have to be able to set the thickness accurately. So the lessons learned from cutting this test cube do apply.

The "general wisdom" I heard on RF30 mill/drills is that they are good to 0.002" on a 1" cube. I also know that the DRO with HF calipers has an error limited to 0.001" for movements less than 4". That was motivation enough to find out what it really is if I tried my best.

pcarpenter
06-25-2007, 09:10 PM
I think that my point was not to discuss the merits of using a cutter straight down, but rather to point out that a .8 inch cube is not a good test of the precision of the mill. To make the point about machining straight down and getting a cube, perhaps I should have said that you could do that with a drill press. That test too, would not adequately test the travel of the tables which is a primary source of inaccuracy in a milling machine (along with cheap spindle bearings etc.) Neither does moving the table only .8 inches in any direction unless you only intend to use it to make really small parts.

Its important to create test conditions that accurately reflect the way a tool will be used, or they don't produce useful information. Can you make .8" cubes within .0002? I would hope so. That extrapolates to someting like .003" per foot (if the error were consistent) which is not all that great for a mill. The question is whether it will machine within some standard expectations for milling machines. I think I got the .0005/foot value from Machine Tool Reconditioning as I recall. In reality, on a mill-drill, I would not expect that, but I would be happy to be within say twice that or .001/foot.

Paul

rgsparber
06-25-2007, 10:06 PM
I think that my point was not to discuss the merits of using a cutter straight down, but rather to point out that a .8 inch cube is not a good test of the precision of the mill.

There are many tests of precision for a mill. It all depends on what you plan to do with it. If I only made parts 0.8" long, then my little test cube would be fine. If I make parts 1 foot long, then testing 1 foot long bars would make more sense to do.

One thing I want to add is that a test cube tests both the machine and the machinist. If I am not 100% focused on the work, I do not get results this good.


To make the point about machining straight down and getting a cube, perhaps I should have said that you could do that with a drill press. That test too, would not adequately test the travel of the tables which is a primary source of inaccuracy in a milling machine (along with cheap spindle bearings etc.) Neither does moving the table only .8 inches in any direction unless you only intend to use it to make really small parts.

Running the drill press straight down or a mill/drill quill straight down does measure your ability to accurately set depth of cut. Without that ability, you can't cut a precise thickness. It is necessary, not sufficient.


Its important to create test conditions that accurately reflect the way a tool will be used, or they don't produce useful information. Can you make .8" cubes within .0002? I would hope so.

It might be very easy for you but it took me months to figure out how to do it. I had many factors that prevented me from getting repeatable results and each one had to be understood and neutralized.

Obviously you are a more accomplished machinist than I am. We are all at different places on this journey.



That extrapolates to someting like .003" per foot (if the error were consistent) which is not all that great for a mill.

I do not recall seeing a slope in one direction but rather random fluctuations across this surface. It is possible that I would get an average thickness error of 0 with some variation around this value along a 12" test bar.

One thing I found that works great to cancel a slope error is to use machined in place soft jaws. Many people from this BBS and Yahoo groups helped me write an article on this topic. It can be found on my web site (rick.sparber.org).



The question is whether it will machine within some standard expectations for milling machines. I think I got the .0005/foot value from Machine Tool Reconditioning as I recall. In reality, on a mill-drill, I would not expect that, but I would be happy to be within say twice that or .001/foot.

My 4" wide mill vise soft jaws are machined in place. They show an error of less than 1 tenth from end to end using a Starrett Last Word DTI that was recently reconditioned at the factory. The fact is, I run my DTI along the edge of these softjaws and the DTI needle does not move at all. This is an example of an error that I had to understand and neutralize.

pcarpenter
06-26-2007, 12:03 AM
I guess what I was trying to convey was that measuring variation in a .8" cube will help you know what you can expect whenever you machine parts .8" or smaller. Devising a test that works across say 12" or the entire envelope of a machine lets you know that you can expect that degree of accuracy on any part smaller....or perhaps better as the part gets smaller. You can safely apply your minimum expectations to a smaller work area if you tested it over a larger one, but you can't necessarily tell anything about the larger case from the smaller. It's hard for me to imagine only machining tiny parts.

If I understood the test correctly, it was to establish real accuracy limitations of the mill itself. In effect, testing the accuracy of the machine across its machining envelope will tell you in the broader sense what can be done. Does it have to be a whole foot? Not necessarily, but .8" as a percent of the total travel of the mill (looks like 17" for an RF-30) doesn't demonstrate a lot. If you cut 15" you can calculate the variation per foot on average. If you measure a tiny fraction of a foot, it requires a great leap of faith to extrapolate out to the larger case. If you already machined 4" vise jaws that showed no variation when measured with a tenths reading indicator, then you already demonstrated more that way than with the tiny cube.

If in cutting the cube you were to use say a 3/8" end mill then in effect, you only moved less than three times the width of the cutting face. From my previous discussion, if you did not move the work at all, you should be able to expect a near perfect flat surface across the end of the cutter because it rotates and any variation in the grind of the end mill should be cancelled by the other flutes. In your current scenario, I would maintain that you are as much testing the end mill as the milling machine. It gets a little better if you test it over a larger area and better still over the full travel of the mill.

ckelloug
06-26-2007, 12:19 AM
Rick,

The first problem that mklotz is talking about is that you have not computed the standard deviation where you say. The .2 that you give is in fact the variance which is the square of the standard deviation.

The problem is further complicated by the fact that there are two kinds of standard deviation: the sample standard deviation and the population standard deviation. The sample standard deviation is the sqrt( (1/(N-1))*sum (x_i-average)^2). The population standard deviation is sqrt( (1/N)*sum (x_i-average)^2). For very large samples, the values are almost identical but for small samples, it makes a big difference. the .14 that mklotz gives is the sample standard deviation which I was taught in statistics is almost always preferable as it converges to the actual value of the standard deviation.

Computers and calculators are dangerous in statistics as mklotz points out. When I was in school, they would not let us use any statistical tool other than pencil and paper unless we could derive the formula that it was using in our lab book and prove that it got the expected answer.

Marv,

What combining operation are you doing on the data to get a single distribution, what distribution is it and what do you want to show with this distribution? I know if you compute the distribution of the sum of three normals that the standard deviation adds that way but I seem to be having a brain fart right now about what you're trying to show. (Not in the critical sense but in the hmm it seems like I should remember something from school a long time ago and I don't).

Rick,

What, specifically are you trying to show with your data? Assume I've come down with rhetorical and statistics idiocy today and explain from square 1 if you care to. If you can explain what you want to demonstrate, perhaps interested parties can suggest more appropriate statistics.

Generally, I'm afraid marv is right that since the three sides of the cube are measurements of three separate properties, it's a bit statistically unsound to simply assume you have 18 samples from the same distribution. He's right that you have 6 samples each from three different distributions.

The only thing I can think of is that you're trying to show that you had 18 measurements that each should have been .8000 and that they are all very close to that value. Off the top of my head, I'm not sure whether there is a rigorously valid approach to this that uses all 18 points simultaneously without adding apples oranges and pears even though they look alike.

I hope you don't take this constructively aimed criticism as a flaming as it isn't what I intend. Between the frightfully smart people around here, the pedants and the asshats, it's dangerous to post a finding around here without Faraday cage encapsulated asbestos underwear. It is a quick way to see if you have made an error or dieties forbid, a typo.

Kudos for investigating some interesting points.

--Cameron

rgsparber
06-26-2007, 12:29 AM
I guess what I was trying to convey was that measuring variation in a .8" cube will help you know what you can expect whenever you machine parts .8" or smaller.

Makes sense to me.


but you can't necessarily tell anything about the larger case from the smaller.

I agree if you are talking about machine variation along an axis. However, it does say a lot about setting the depth of cut accurately.



It's hard for me to imagine only machining tiny parts.

Oh, I make larger parts, but lately my focus has been on understanding how the machine and DRO work for just the test cubes. I am making a small beam engine and must machine the top of the column which is less than a 1/2" cube.


If I understood the test correctly, it was to establish real accuracy limitations of the mill itself.

As you said, this test cube only characterizes the mill and DRO for parts that are no larger than the cube. My biggest surprise in this work was that the DRO was able to guide me to cuts within 2 tenths. This DRO uses Harbor Freight calipers that have an error limit of 1 thou and a repeatability limit of 5 tenths. At least for my caliper, it is more like 2 tenths and this assumes zero error for the mill. So maybe the mill is 1 tenth and the caliper is 1 tenth. Still amazes me that it is so good.



If you measure a tiny fraction of a foot, it requires a great leap of faith to extrapolate out to the larger case.

No, I would not stick my neck out that far. The caliper has an error limit that is a function of distance so that alone would prevent ataining this kind of accuracy beyond an inch or two.



If you already machined 4" vise jaws that showed no variation when measured with a tenths reading indicator, then you already demonstrated more that way than with the tiny cube.

Not really. These are machined in place jaw and were not cut to any specific size. It does show consistency of the X axis gib but then I run my ways rather tight.


If in cutting the cube you were to use say a 3/8" end mill then in effect, you only moved less than three times the width of the cutting face. From my previous discussion, if you did not move the work at all, you should be able to expect a near perfect flat surface across the end of the cutter because it rotates and any variation in the grind of the end mill should be cancelled by the other flutes.

I used a 5/8" end mill and made two passes over the surface. I can see the effect of having the cutter not exactly perpendicular to the table. It is just under a tenth. For an 8" tram circle this comes out to a 1.2 thou offset which is what I have.

rgsparber
06-26-2007, 12:56 AM
Cameron,

Good to hear from you. What I want to gain from this data is some prediction of expected accuracy. Mostly I am interested in depth of cut like the data I sent you off line. As Paul pointed out, cutting a test cube of 0.8" should be able to tell me how well I can cut a cube of no larger than 0.8". I would not dare extrapolate to larger sizes.

I can just use my worst case variation but would prefer to know something of how likely I will hit my target within tighter limits. I thought that calculating SD would give me that feel. For example, can I use the provided data to say that there is a 95% probability that I will be able to hit within X tenths of my target value?



The problem is further complicated by the fact that there are two kinds of standard deviation: the sample standard deviation and the population standard deviation. The sample standard deviation is the sqrt( (1/(N-1))*sum (x_i-average)^2). The population standard deviation is sqrt( (1/N)*sum (x_i-average)^2). For very large samples, the values are almost identical but for small samples, it makes a big difference. the .14 that mklotz gives is the sample standard deviation which I was taught in statistics is almost always preferable as it converges to the actual value of the standard deviation.

So I should have used the sample SD which appears to give a smaller number.

By my thinking, one number per dimension represents the error associated with the DRO. Say it is the center measurement. I actually measured just to the right of the center but it is almost the same value. The other 4 numbers per dimension tells me about the variation of the milling process. Should I then take the sample SD of the 3 center measurements to characterize the DRO? If so, then I guess I would take the 4 other measurement for a given dimension to tell me how well the mill holds true. Why wouldn't I be able to take the 4 measurements from each dimension and combine them into a single sample SD?



What, specifically are you trying to show with your data? Assume I've come down with rhetorical and statistics idiocy today and explain from square 1 if you care to. If you can explain what you want to demonstrate, perhaps interested parties can suggest more appropriate statistics.

Hopefully my above text will help.



I hope you don't take this constructively aimed criticism as a flaming as it isn't what I intend.

Not at all. I put the data out there so I could learn the right way to evaluate it. Believe me, I've been flamed big time and this sane and scientific discussion couldn't be further from such experiences.

I certainly do not like to be flamed but often find wisdom among the harsh words as long as they don't get personal.

Funny you should mention Faraday cages and asbestos underwear. I always imaging sitting on top of a barn with my long copper pole during a lightning storm as I submit a new post with my latest data and theories. Sure glad that underwear is washable :-0

oldtiffie
06-26-2007, 09:36 AM
Deleted/edited-out

rgsparber
06-26-2007, 11:00 AM
the discussion and theory re. measurements and statistics is so far above my head that it is on a much higher ethereal plane altogether than that upon which I can and do try to operate.

The only thing this fancy math is really doing is trying to find ways to describe a collection of numbers so they makes sense. Beyond the complex theory, the equations are usually rather small. I really like the idea (maybe dream) that I can make a bunch of test cuts and then say that future cuts will be within certain error limits. Knowing what my machine plus me can do is important for me to know.



"Side milling" is a different proposition, as assuming that you were "conventional" milling, the cutter, no matter how sharp and no matter how rigid the quill etc. set-up is, will "rub" on the work until such time as the section of the cutting edge presented to the cutter and the resistance to bending/deflection of the cutter itself, will eventually "raise an edge" that the cutter can deal with. This depth of cut and the size of the chip will progressively increase until that "tooth" on the cutter exits the work. Further, the cutting edge is a spiral/helix and (usually) being right-handed, will progressively "climb" up the job as the cut on that tooth progresses toward exit from the cut. Further, as the entire cutting edges are helical, there will be times (depending on the spiral and length of the "side" cut) that more than a single cutting edge will be engaged with the work-piece. This is the same as a lathe to some extent as regards "getting started" but a lathe cut will continue at the same depth until it exits the cut.

I started by removing 5 thou and ran the end mill back and forth untill no more particles of metal came off. So in one direction I was climb milling and in the other down milling. Is there another way to do it in order to minimize the error? After I side milled that one side I turned the block so my soft jaws supported the side milled end. On my last block I took off half of the material and turned the block over. In this way I end milled all faces. On the block I just reported on, I left the side milled face in order to get an idea of how bad it was. I am talking about a tenth or two here.


(This is the reason why I will only use the "short" series of end-mills. The "long/er" series is only used as a last resort).

I sure have seen this effect!



Further, I think you said your tram error was of the order of 0.0012/8" which translates to 0.00012/0.8", say "1 tenth per 0.8", which is a given input without any deflection of the cutter when side milling.

I think I see what you are driving at here. Given a tilt of the end mill due to tram error, I should expect a taper of a tenth over the face of the cube. I'll take a closer look and see if that is true. I plan to also look closer at the other 2 dimensions to see any signs of tram error. Looking at the surface finish of the cube, I don't see any dramatic step where one pass of the cutter runs over the last pass. Maybe the vibration of the mill during cutting washes out this tiny tram error.



Climb milling is the preferred method of cutting in CNC - hence the rigidity of the CNC set-ups and the need for as near to absolute zero back-lash

I learned the hard way how important it is to lock unused axes.


"Tramming" the spindle of the quill to the table is all well and good if and only if the table top face/surface is parallel to the mill X and Y slide-ways as well. In other words if the table is not parallel to these ways/axis and the spindle is trammed to the table the spindle is not trammed to the machine. The best and probably only way to verify this is the put a DTI on the table and traverse it throughout the X and Y slide-way limits. Any error found will have to be compensated for when using the table as a reference for tramming the Z axis spindle.

Wouldn't any "table not parallel to XY plane of ways" error be canceled by using soft jaws that are machined in place and set to the same position as when in use?

My table top is flat - I lapped it a year ago. I can see that this is not enough if the ways are not true.



I hope this helps.

It does and I thank you.

ckelloug
06-26-2007, 11:40 AM
Tiffie,

Don't worry about the statistics stuff above. Statistics is a very technical subject. I got my degree in general engineering with a concentration in probability and mathematical statistics: which is to say that I don't know a huge amount about statistics but that I have a bigger pile of books to look stuff up in than many folks and I have two more classes in it than my classmates most of whom I would describes as geniuses to begin with (not being sarcastic: went to elite private engineering school). Knowledge or no knowledge, I measured the wood doors I built for my shop attic yesterday and found that they're made to better than a 32nd of an inch of precision but the design length is an inch short of what it should have been. Can you say nail in a thicker threshold and more weather stripping? I can!

At any rate, I think your post is entirely on topic. There are some here like me that specialize professionally in analyzing errors to death. It tends to scare the errors away and sometimes the onlookers too. You on the other hand know a lot about correcting errors that got made and building things with good practices and no errors to begin with. That's stuff that I'd like to learn. I for one love to see your posts here as they are invariably well written and insightful.

I gotta go do some real work now so have a good day and I haven't forgotten about the statistics here, I just can't look right now.

DR
06-26-2007, 12:43 PM
My intension here is to show what can be done with typically HSM machines and NOT as a platform for bragging. I hope the reader accepts these findings with that in mind.

The full details are shown below:

for dimension
dim UL UR Center LR LL average SD
1 799.7 799.8 800 800.1 799.9 799.9 0.2
2 799.5 799.9 799.9 799.9 799.6 799.8 0.2
3 800.2 799.5 800 799.9 800.4 800.0 0.3

average 799.9
SD 0.2

Dimensions 1 and 2 were entirely cut with the end of the end mill. Dimension 3 had one face side milled while the other was cut with the end of the end mill. You can see that side milling is not as consistent as using the end of the end mill.


I guess I don't understand how the cuts were made. The end of the end mill? That implies you moved the block between cuts, right? Using the end of the cutter aren't you only testing the Z axis?

What I'd like to see is mounting the blank in the vise. Side mill one edge, zero the DRO, crank over in X ( or Y) a distance of .8", plus the diameter of the end mill, as shown on the DRO. In your case the crank over would be .8 + .625". I realize this will only allow you to mill a raised boss on the blank, but I believe this is the way most milling machine accuracy tests are done.

Note that side milling accuracy is drastically affected by cutter run-out, where end milling isn't to any practical extent. What amount of run out do your holders have? They must be extremely accurate to achieve the result of test three.

Also, it would be nice to tell us the material, rpm, cutter style, feedrate, DOC, etc, etc.

Test number three shows an accuracy of essentially .00035"+/-. That would exceed what I expect on my machine costing 50 times as much as your RF30. Actually, if I get a customer print with tolerances for side milling of closer than .001"+/- I pass on it.

BTW, this is not flaming. It's just curiosity (and doubt) based on many years of precision machining on very accurate CNC machines.

mklotz
06-26-2007, 01:28 PM
Caveat: I am neither a metrologist nor a statistician.

First some random remarks.

You do need to compute *sample* standard deviation, most especially with the small sample sizes you are using. The sample variance is obtained from the formula:

variance = [sum (x(i)-mean)^2]/(N-1)

where the x(i) [1 <= i <= N] are the N measurements of the sample. The sample standard deviation is the square root of the sample variance.

As ckelloug pointed out, don't trust any canned computations (Excel, calculator) unless you have proved to yourself what equations are used and you understand the significance of those equations. In fact, that's excellent advice for anything mathematical you do.

After more reflection, I don't think it's fair to combine the statistics of your three sets of six measurements into a single mean and standard deviation until you are able to prove that that is statistically meaningful. I don't think you'll be able to prove that. [Not a slam, I don't think anyone can prove that.}

It's very important that we all agree on what we're testing when we talk about the "accuarcy of a mill". If I were assigned that task, I'd cut 20 or more (larger sample size) pieces of stock to rough size, then side mill each of them *separately* to some pre-agreed size. After that, I'd measure the error of each and develop the *sample* statistics. Then I'd repeat the entire experiment with the single change of end milling the pieces.

[It's fair to assume that the errors in both these experiments will be normally distributed. That's true because there are so many contributors to the error and, while each may have its own (possibly non-Gaussian) distribution, the sum of many random errors almost always approaches a Gaussian distribution.]

The two experiments outlined duplicate the way the way the mill is actually used in practice and furthermore separate the effects of side and end milling, which, as ckelloug points out (and you've already noticed), bring different error mechanisms into play.

Finally, statistical analysis of experimental data is a real mathematical quagmire. Get yourself a good book or two on the subject and do some reading, paying special attention to the subjects of hypothesis testing and goodness-of-fit testing. Until you can back up your observations with some solid math you won't know what you have and you won't be able to convince anyone that your conclusions are valid.

Norman Atkinson
06-26-2007, 01:55 PM
Hang on-fellas!

Of course, I could set a statistical sampling of results of meaningful measurements. Or I could!

I, too, would be wanting far more detailed measurements- like you, Marv but!

I have had a RF30- a cheap jack machine and one which no professional would tolerate to earn his coins- or to set QC limits. It's a joke!

However, this is a hobby- for me and a mill of doubtful quality is an asset- cheap and nasty though it is.

To add a bit of possible accuracy and ease, I bought a cheap jack Shumatech
which I am building up. Enough to see the lights blink and a few brackets being fabricated.

Looking at it and the results - skewed as the are- do suggest what is possible with the thing.

If we want to go into to the realms of measurement and quality tooling, not with a RF30-eh?

Norm

mklotz
06-26-2007, 02:23 PM
Rambling comments and personal anecdotes aside, Norm, I'm not addressing the utility of the experiment or its practical application.

I'm making the point that, if one is going to attempt to statistically characterize machine errors, of whatever ilk, it's important to design a meaningful experiment and do the mathematical analysis properly.

Frankly, while most of us have some feel for the "accuracy" of our mills, I doubt that any of us have ever attempted to quantize it, separating the machinery-induced errors from the operator-induced errors and mathematically modeling the former.

Rick, I believe, is trying to do that and, since this is a hobby venue, it's not up to us to criticize what he does in his spare time.

In fact, if the final outcome of this discussion is a solid mathematical model that can be used to characterize machine tool errors, we will all have benefited.
A valid methodology for characterizing machine errors would be as useful for an RF30 as it is for a Deckel.

Norman Atkinson
06-26-2007, 02:54 PM
Marv,

If the subject gets to a point where measurement for endless parts, is heading- and I feel that it is- it should be conducted elsewhere. Home Shop Machinists have garden sheds rather than factories.

I have a crap RF30, a crap pair of myfords and a crap operator with even more worn out equipment. I find that in tooling that I am two standard deviations - on the positive side- from the rest of readers.

I hope that others may not be too offended my opulence

Norm

mklotz
06-26-2007, 03:20 PM
Norm, I don't think *you* want to get into a discussion about the relevancy of posts to this forum. Trust me on this.

If the subject annoys you, don't read it.

ckelloug
06-26-2007, 04:37 PM
According to recent studies, regularly exercising your brain helps prevent Alzheimers. I'm a young guy so I'm not worried about it yet.

Those reading my posts here and on CNCZone should realize I often work through seemingly simple problems with the most powerful theoretical tools I have available. It keeps my skills sharp for real problems and on occasion provides useful answers. It's the equivalent of continuing education as it prevents me from remaining at the technical level I was at when I graduated from engineering school a decade ago.

What I love about the HSM board is that I have never before found so many people who are so broadly knowledgeable in the same place outside of where I went to school.

A poster name Geof over on CNCZone has a sig that reads, "An open mind is a virtue as long as all of the common sense has not leaked out!"

At any rate, Marv, thanks for your comments on the statistics. I'll review them and see if I can lend Rick a hand on them now that your comments and Ricks latest ones have cleared up some of the mystery both in what was wrong with the experiment and what was actually to be demonstrated by it.

Regards all,

Cameron

Norman Atkinson
06-26-2007, 05:13 PM
Well Cameron,
I have got to 77. I can still go into a workshop and do my thing- albeit more slowly and with increasing need to be able to see measuring instruments and split the hairs in verniers etc. A DRO - cheap and cheerful- was thought sensible like getting zenon lights in my newest car.

I have discussed this with my family- sort of consultant medics and tooth fairies, musicians and linguists. My maths was done with a 'guessing stick' but the rest was honed on two practices. One was Trachtenburg Maths and the other was Pelmanism. Neither needed a laptop, nor a calculator or even a piece of paper and pencil. Dear old Prof Jackow Tractenberg fled Communism and then he spent 7 years in concentration camps- and paper was a list to the gas chamber.

It may be 'all our yesterdays' but did it work? Well, I retired at 55 and I suppose that I am still a multi-millionaire inspite of huge dollops to family.

Maybe there isn't a mathematical formula- Marv- but do I get a nod from my stockbroker!

I must be getting it right- somehow.

Cheers

Norm

PS, I hadn't realised that Pelmanism was now available in pdf form.
I had the original grey books etc
Worth remembering- write it down

ckelloug
06-26-2007, 07:08 PM
Norm,

It is not the work of fools with candles to haul a vernier off to the shop and make dust, the death of us. I'd be doing that right now if I didn't need to haul a router up a 20 foot ladder and finish building the shop first. (Two little doors and power and wall sheeting and I'm done). I find doing the statistical analysis on the caliper almost as entertaining as making chips so I'd contend that I just happen to like an odd facet of the machining hobby: analyzing things.

I am probably the last student on earth that used a guessing stick on college homework. (I've still got my Pickett 3ES 12 incher). I usually used my HP48GX calculator but when it got stolen, it was fun and entertaining to use the slide rule. I'm just bitter that a couple of prized 6 inchers were also in my bag and stolen with my calculator that bad day.

Back to the brain exercises,

Cameron

Norman Atkinson
06-26-2007, 07:28 PM
Message appreciated. Time for an old man to get his priorities right.
There is now real work to be done- please see my other post.

I have already spent half the night giving guidance etc to 'my' people.

madman
06-26-2007, 09:44 PM
Duh??????????????

rgsparber
06-26-2007, 10:10 PM
I guess I don't understand how the cuts were made. The end of the end mill? That implies you moved the block between cuts, right? Using the end of the cutter aren't you only testing the Z axis?
Except for one cut on the 3rd dimension, all cuts were with the end of the end mill. I moved from the front left to the back, then over 0.4" and from back right to the front. I am mostly testing the accuracy of setting the Z axis but if there were major problems with the Y axis, I would see that too. In fact, when the X axis gib got loose, I saw the shift right away.



What I'd like to see is mounting the blank in the vise. Side mill one edge, zero the DRO, crank over in X ( or Y) a distance of .8", plus the diameter of the end mill, as shown on the DRO. In your case the crank over would be .8 + .625". I realize this will only allow you to mill a raised boss on the blank, but I believe this is the way most milling machine accuracy tests are done.

I think there are dozens of milling machine accuracy tests. My journey started out trying to understand the limits of my DRO with RF30 and figured out early that side milling was the sloppiest. My plan is to first understand the errors associated with milling with the end of the end mill and later understand the limitations of my DRO and mill on side milling.


Note that side milling accuracy is drastically affected by cutter run-out, where end milling isn't to any practical extent. What amount of run out do your holders have? They must be extremely accurate to achieve the result of test three.
I don't agree. The tram of the head directly effects how square the end cut will be relative to the reference surfaces but run out can be canceled by just taking a reference cut first and measuring how much is left. Then just zero the DRO.


Also, it would be nice to tell us the material, rpm, cutter style, feedrate, DOC, etc, etc.
I am cutting 6061 with a new 5/8" HHS ENCO 2 flute endmill. The RPM is around 770. I am only taking about 3 thou on the finish cut and the feed rate is around 1 IPM.


Test number three shows an accuracy of essentially .00035"+/-. That would exceed what I expect on my machine costing 50 times as much as your RF30. Actually, if I get a customer print with tolerances for side milling of closer than .001"+/- I pass on it.

I would certainly go broke trying to make money doing this level of accuracy. It has taken me many months of "pealing the onion" to understand and compensate for my mill and DRO's errors.


BTW, this is not flaming. It's just curiosity (and doubt) based on many years of precision machining on very accurate CNC machines.

I certainly don't see your email as flaming. It is healthy to have a good measure of doubt. For what its worth, I took my last two test cubes to a professional machinist friend of mine last weekend and he pulled out the fanciest digital mic I have ever seen. Only after a lot of measuring did he finally accept my claims. My last cube, reported in this thread, is far better and took a lot less time to do. My first 2 cubes took about 15 hours each. This last one used only my DRO and mic and took me less than 2 hours.

Knowing exactly how well me set up can make a cut is important to me but the journey may have been more important. I learned a lot along the way. A few of my major lessons learned:

1. use machined in place soft jaws
2. run the mill while adjusting the DRO (prevents things from sticking)
3. all surfaces must be absolutely clean
4. always use the thimble clutch on the mic
5. lock all axes that are not going to move
6. on DRO: read the HEX value and do your own conversion to decimal inches (avoids a truncation bug)

oldtiffie
06-26-2007, 10:45 PM
Delted/edited-out

rgsparber
06-26-2007, 11:03 PM
You do need to compute *sample* standard deviation, most especially with the small sample sizes you are using. The sample variance is obtained from the formula:

variance = [sum (x(i)-mean)^2]/(N-1)

I thought I was supposed to use 1/N not 1/(N-1).




As ckelloug pointed out, don't trust any canned computations (Excel, calculator) unless you have proved to yourself what equations are used and you understand the significance of those equations. In fact, that's excellent advice for anything mathematical you do.

I agree. I first hand calculated based on my Statistics book equation and then ran the same numbers with Exel and the HP. My REAL problem appears to be a lack of understanding, not poor math tools.



It's very important that we all agree on what we're testing when we talk about the "accuracy of a mill".

Very true. If I was in the business of making 0.800" cubes, then monitoring the variation of this product would be necessary and sufficient. I do think that making a test cube does provide a basic idea of quality: am I cutting to 0.02" or 0.001".



If I were assigned that task, I'd cut 20 or more (larger sample size) pieces of stock to rough size, then side mill each of them *separately* to some pre-agreed size. After that, I'd measure the error of each and develop the *sample* statistics. Then I'd repeat the entire experiment with the single change of end milling the pieces.

Sounds like a good scientific approach but far beyond what I am willing to do.



Finally, statistical analysis of experimental data is a real mathematical quagmire. Get yourself a good book or two on the subject and do some reading, paying special attention to the subjects of hypothesis testing and goodness-of-fit testing. Until you can back up your observations with some solid math you won't know what you have and you won't be able to convince anyone that your conclusions are valid.

I can see that this is a messy business. From what you have said, even if I did have the correct mathematical tools, I would have to collect a lot of data to have any valid results. Sounds too much like work to me right now.

As for convincing others of what I am able to machine, well, I guess that is really not so important to me. I would like to inspire others to try pushing their machines to their limits and share what they have learned. Those "arm chair machinists" that kept telling me that I could not do any better than 0.002" with my RF30 were certainly wrong.

oldtiffie
06-27-2007, 02:36 AM
Deleted/edited-out

Norman Atkinson
06-27-2007, 04:43 AM
Oh, No! Now I am getting notes that people are attemping to make money with their RF30. The plot does thicken. My impression was that this was a Home Shop Machinist's Forum where 'one off's' were produced- with limited resources. So others are interested in that nasty emotive thing called money.
I thought that it was just me.

The limits of my thoughts were people who didn't know how to set up a lathe tool or what it was like to drink brake fluid. Now we are having those who are doing statistical analysis on the square root of not a lot.

There are more pressing needs- which I am trying to be a postive part .

In another posting, I have tried to explain where my real priorities lie.

Meantime, talk amongst yourselves.