EWMAST control chart for autocorrelation

randwick

Registered
Hi. I apologize if I have used this thread in a wrong way but since I'm not a frequent Elsmar visitor I've forgotten how to post a message. My question is if anybody has ever done an EWMAST control chart for dealing with autocorrelation. I have unsuccessfully googled hours and hours in order to find an example with real data. Since I recently read "ISO 7870-9 Control charts — Part 9: Control charts for stationary processes", I've seen that the standard has given much importance to EWMAST but the example is from an AR(1) time series simulated data set which I can't reproduce since it is based in random numbers.
Thanks in advance.
 

Miner

Forum Moderator
Leader
Admin
This chart is new to me, but your tie-in with an autocorrelated process caught my attention. I have over 40 years of experience in SPC. One (actually more than one) thing that I have found is that:
  • People make control charts way more complicated than Shewhart ever intended (i.e. the whole probability thing, etc.)
  • People make a big mistake trying to apply all of the extended rules. Most processes work great with rule 1 only. A few might benefit from the run rule OR from the trend rule, but very few processes need to be as tightly controlled as the other rules would drive.
  • Very, very few processes need the specialty charts (e.g., CUSUM, EWMA, etc.)
Early in my career, I was faced with applying SPC to an extrusion process. Extrusion processes are negatively autocorrelated, so sub-grouped control charts (e.g., Xbar/R, etc.) will not work because the control limits are too tight. Next, I tried using an I-MR chart, but the question was how to establish the frequency of checks. Fortunately, my company had a corporate statistician. Once I explained the problem, he suggested taking samples at 5-minute intervals for an hour then running the autocorrelation function (ACF) to determine when the data was no longer autocorrelated. We found that 20 minutes was that interval and set the frequency of sampling at 20 minutes. The I-MR chart worked great at that frequency.

Read Autocorrelated Data What does autocorrelation tell you about your process? by Donald J. Wheeler. He studied under Deming and is today's foremost expert in SPC. He covers both the positive and negative autocorrelation scenario and how the I-MR chart will work in both cases.

After reading that, ask yourself "Do I really need a specialty control chart that no one can understand?"
 

bobdoering

Stop X-bar/R Madness!!
Trusted Information Resource
I agree. What is the process, and what makes you think what you see is autocorrelation? There may be a special cause masking the true process variation, or the process may be a function rather than random and independent (needed for Shewhart charts).
 

randwick

Registered
I have to prepare a course on SPC for a customer who is using MES and so has big data. The sampling frequency is very short (1 measure of a diameter done with an automatic laser inspection equipment in every minute) and so I suspect the presence of autocorrelation (as literature reports). Since the MES s/w doesn't treat SPC for autocorrelated processes, I would like to illustrate the model-free approach (to which EWMAST belongs) for treating SPC for autocorrelated processes with a simple spreadsheet since EWMAST formulae are relatively simple.
Thanks for your previous message.
 

bobdoering

Stop X-bar/R Madness!!
Trusted Information Resource
If it is precision machining, it is not so much autocorelation as it is a function - which takes a different SPC approach than Shewhart charting, anyway. If the laser is only taking one diameter measurement, then it is already missing the roundness error - which in precision measurements masks the true process variation. If it is a different process - such as molding, then the process model could be a completely different distribution model.
Just because the data is "big" (plentiful) does not mean it is correct - so decisions made from it can still be erroneous. Just a cautionary tale.
 

Bev D

Heretical Statistician
Leader
Super Moderator
I have to prepare a course on SPC for a customer who is using MES and so has big data. The sampling frequency is very short (1 measure of a diameter done with an automatic laser inspection equipment in every minute) and so I suspect the presence of autocorrelation (as literature reports). Since the MES s/w doesn't treat SPC for autocorrelated processes, I would like to illustrate the model-free approach (to which EWMAST belongs) for treating SPC for autocorrelated processes with a simple spreadsheet since EWMAST formulae are relatively simple.
Thanks for your previous message.
I am concerned that you might not have picked up on what Miner was saying.
Additionally an Xbar or I, MR chart is model or distribution free...
and there is NO substitute for the graphical visualization of data...
 
Top Bottom