iSixSigma

Measuring process capability

Six Sigma – iSixSigma Forums Old Forums General Measuring process capability

Viewing 36 posts - 1 through 36 (of 36 total)
  • Author
    Posts
  • #30587

    Pankaj A
    Participant

    Hello folks – In my plant, we have a machine shop producing components for making compressors to be used for Refrigerators. We want to measure the process capability of the machines used. One machine generates more than one CTQ on the component – say the Diameter, the perpendicularity and surface finish. Each CTQ is important but it is not known relatively which has a greater impact on the product (compressor)quality. My concern is – do we individually cp/cpk for each CTQ and if I want to find overall capability do I take the minimum of all the values for CTQs. If individual CTQ approach is taken it will result in close to 200 cp/cpk measurements. Is that the content of work to be done or can someone give a smarter solution. I would be grateful if u could help out

    0
    #79786

    abasu
    Participant

    In situations like yours determine DPMO of each CTQ.  Then add the DPMOs together to give total DPMO for the compressor.  Now calculate sigma level of the compressor.
     

    0
    #80040

    Raju
    Participant

    I guess you can try using a multivariate process capability index covering all the quality attributes your are interested in. The method mentioned in this article {Wang et al Using Principal component analysis in process performance for multivariate data. International Journal of Management Science 2000;185-194} worked well for my situation and the assumption of multivariate normality is not a major problem for this method.  If you need more help in implementing this procedure please send a mail. Hope this of any help to you.
    Raju

    0
    #80050

    john beaudoin
    Participant

    You have a good problem here.  Let me try to respond with my knowledge:
    1) When a part is produced, it is either within specifications or it is not.  Using this approach, you can get a Cpk for the machine as a whole.
    2) If the machine Cpk is not satisfactory to you, you are going to have to dig deeper in the reason, and thus you are going to have to ask why a part is not in spec.  This will eventually lead to you evaluating the individual measurements anyway.
    3) You can reduce measurements by looking at random samples and selecting a confidence interval.  You may also decide that certain measurements have so little variation that you only want to check the measurements where you have the most problems.  If you can improve the machine by replacing wearable parts, etc., based on the worst offending measurement areas, the whole of the machine may improve to an acceptable level of output.

    0
    #80053

    Marc Richardson
    Participant

    Try Keki Bhote’s book “World Class Quality, Using Design of Experiments to make it happen” It gives a clear and concise explanation of Dorian Shainin’s techniques, which are uniquely suited to determine what the key product characteristics are. In your case, what specific characteristics of which components contribute to “compressor quality.” How do you measure compressor quality? Have you identified the primary characteristics that lead to customer satisfaction in a compressor?
    Marc RichardsonSr. Q.A. Eng.

    0
    #80058

    Mikel
    Member

    Forget that Bhote / Shainin nonsense. Did they teach you DMAIC?
    The M&A are all about narrowing the focus from all of the things the designers designate as critical because they don’t  know to the things that are really critical.
    Marc, the only Shainin tool (really a Gillette tool from the 50’s) taught routinely as part of M&A is MuliVari and the focus is routinely narrowed quite nicely. I don’t have a clue what you mean about the Shainin stuff being unique. Multivari is just variable grouping and doing a gross visual analysis. B vs. C is Tukey end counts. Iso plots are just regression (and an extremely limited view of MSA).
    Shainin has more hype in it than the 6 sigma stuff (which is a really high bar to get over)

    0
    #80073

    Marc Richardson
    Participant

    Hi Stan,
    >the only Shainin tool (really a Gillette tool from the 50’s) taught routinely >as part of M&A is MuliVari and the focus is routinely narrowed quite >nicely. Multivari is just variable grouping and doing a gross visual >analysis. B vs. C is Tukey end counts. Iso plots are just regression (and >an extremely limited view of MSA).
    Thanks for your observations. The Multivari tool is similar to stratification; looking at the data set through different identifying characteristics like within part, part to part, time to time, machine, operator, material supplier, tools and so on to see if there is a trend. However, I think that the BOB-WOW and Component Search is particularly effective for finding out which component is causing the problem in an assembly. Is it rocket science? No, but that’s the beauty of it. They are simple, effective techniques that can be taught to anyone (non-BB’s) and that can increase your business’s ability to solve its problems.
    Marc RichardsonSr. Q.A. Eng.

    0
    #80078

    Pankaj A
    Participant

    Dear Raju
    I would like to know more and even go ahead with the Multivariate analysis tool you have mentioned.
    Can you help out with more inputs on it?
    Thanx for your reply. It has given me a different insight to the issues.

    0
    #80080

    Raju
    Participant

    PankajI have enclosed a presentation on this topic which may help you to understand better. If you are measuring all those CTQ’s or variables (i.e. Diameter, the perpendicularity and surface finish) of a component and you have this data for atleast sufficient number of components (around 30) then you can try this procedure mentioned in the presentation which will reflect the overall process capability of the machine in producing the component satisfying all the CTQ’s specification.Step1: Take the dataset (data on diameter, perpendicularity and surface finish for 30 or so components manufactured from the same machine)Step2: Run a Principal component analysis (PCA) – available in most of the statistical software’s like Minitab – you will get principal component loadings for each variable and eigen values for each principal component (similar to slide 8). You can choose 2 or 3 principal components depending on the results, normally that would explain more or less 90% of the variation (this will appear as % of explained total variability).Step 3: Convert your specification on each of the CTQ’s or variables into specification for each principal components (see slide 8 – simple linear combinations).Step4: After obtaining specifications for each principal component, the compute Cp for each principal component (as given in slide 9)Step5: Compute multivariate process capability as geometric mean of individual principal component Cp’s. For example if you have two principal components, PC1 & PC2 then you would have obtained two Cps one for PC1 & another PC2 as in step 4. Then to get a overall Cp compute geometric mean of Cp_pc1 and Cp_pc2. i.e. Overall Cp = SQRT [Cp_pc1 * Cp_pc2].{as in slide 6}.Hope this helps. Multivariate Process Capability Indices Presentation Download [Microsoft PowerPoint, 405 KB] Multivariate Process Capability Indices Presentation Download [Adobe Acrobat File, 43 KB]

    0
    #80093

    Marc Richardson
    Participant

    Raju,
    Your attachment didn’t make it. If you contact the folks at isixsigma, they will publish your presentation on this site so that everyone can benefit.
    Thanks,
    Marc Richardson
    Sr. Q.A. Eng

    0
    #80102

    Raju
    Participant

    Marc
    Thanks for the feedback. The document is now attached.
    Raju 

    0
    #80103

    Pankaj A
    Participant

    Dear Raju
    I have picked up the presentation u sent. It seems to be the answer to the issues which we are grappling with. Let me attempt the method you have suggested and I will surely write about the success level

    0
    #80110

    Marc Richardson
    Participant

    Raju,
    Thanks for the presentation. It will surely increase my understanding of this topic.
    Marc Richardson
    Sr. Q.A. Eng.

    0
    #80146

    John J. Flaig
    Participant

    Marc,
    About five years ago a paper was published in JQT or EQ that assessed the effectiveness of the component search technique as compared to other statistical search procedures. The results indicated that the component search techniques was not as good as some other approaches.
    John J. Flaig, Ph.D.
    Applied Technology (e-AT-USA.com)

    0
    #80176

    John J. Flaig
    Participant

    Raju,
    Let me see if I understand this procedure. I take a bunch of non-robust process capability metrics (i.e., Cp or Cpk). I weight them based on PCA analysis, combine the results, and I get a good multivariate capability metric. Does this make sense to you?
    John J. Flaig, Ph.D.
    Applied Technology (e-AT-USA.com)

    0
    #80213

    Raju
    Participant

    John
    No, I guess there is a difference. It is not obtaining Cp or Cpk on the univariate first and weighting them by PC’s to get a multivariate Cp. It is about obtaining Cp or Cpk for each principal component first and combining them to get one overall Cp. This procedure exploits PCA properties – each individual principal component will be normally distributed and are independent of one another. The procedure had been published in the scientific literature (see reference). This preocedure seems to work well for our case. Of course you need to understand your data set well and accordingly choose one which will answer your question rather than complicating the question further.
    Raju

    0
    #80216

    John J. Flaig
    Participant

    Raju,
    I can’t find the referenced article in Omega do you have a copy? Your Powerpoint document does not say how Spc is computed. I read the article in JQT but it did not cover the PCA approach that Dr. Wang proposed.
    I actually like the approach but I have have the following concern. You compute the Cp for the major PC’s and then you combine them. The problem is the use of Cp. Cp is the “potential process capability” not the “actual”. You can have a process with a Cp of 6 and it can be producing 100% defective product. When you combine metrics that have this sort of serious problem the resultant metric is very questionable in my mind. Now, if a “better” capability metric can be found (and it can be — see my 1999 paper in Quality Engineering or the Demo on my website), then Dr. Wang’s approach would definite merit.
    John J. Flaig, Ph.D.
    Applied Technology (e-AT-USA.com)

    0
    #80218

    Ashman
    Member

    You want to the process capability of a measurement machine, or
    a manufacturing machine? cpk for a measurement machine can not
    be defined. Also, you have a cp/cpk of 200? You have got to
    check your SLs —they probably are not sensible.

    0
    #80219

    Matthews
    Participant

    I have been looking into the capabilities of the machining process also. I have designed a part to be machined that has all features that can be machined (approx 30 features) These vary between deep pockets, diameters, radius, thin web profiles,etc. I have then manufacured the part thirty times on each machine.
    I have then calculated the process capabilities for each feature on a sliding tolerence scale(Changing the USL + LSL) This in turn has given me process capabilities for each machine on each feature. Then in turn I can look at a part that is going to be machined on paper put in what features are on that part and quantity and tolerance bands on each and before I even cut the part I can see the propabilities for defect on the part.
    The machines are now running a batch of five of these parts every month to update capabilities and tell maintenance of any problems this just goes into a XBAR and R Chart for each Machine.
    Hope This Helps,
    All The Best

    0
    #80220

    Mikel
    Member

    Marc,
    I agree that Component Search and Variabile Search are nice tools to teach. I find them particularly effective to teach to technicians in test and design.
    Thanks for reminding us. They are not taught by most BB training courses.

    0
    #80224

    Gabriel
    Participant

    Sure, you didn’t understand.
    He didn’t mention a measurement machine. He has several machines and each one makes several characteristics at the same time (I imagine multispindle turning, for example). If he went to calculate the Cp/Cpk values for each charcteristic in each machine, he would have more than 200 Cp/Cpk values (he said something like “… I would have more than 200 Cp/Cpk measurements”, which is not “a Cp/Cpk of 200 for a measurement machine”)

    0
    #80226

    Michael Boyd
    Participant

    Marc,I agree on your views toward Shainin-DOE. I worked over 4 years as Lead MFG/Quality Engineer at the world-leader in optical networking components/systems in the world, and Shainin worked best for me in a fast-paced MFG environment.In the real, fast-paced hi-tech, MFG environment where quarterly revenue MUST be met, an engineer/manager that can quickly respond to failures is the one that will succeed.Mike

    0
    #80227

    Michael Boyd
    Participant

    Stan,I agree with your statements. I just sent an e-mail to Marc Richardson w/ my explanation.Please share w/ me some of your Shainin experiments/studies! Send to me @ [email protected],
    Michael

    0
    #80228

    Michael Boyd
    Participant

    Correction…I agree with Shainin tools. These worked extremely well for me in a hi-tech MFG environment where quarterly revenue must be met.I like to take a little of what fits best with my vision and incorporate it all (where possible). 6 Sigma, Shainin, Lean, 5S, etc…

    0
    #80245

    John J. Flaig
    Participant

    Marc and Mike,
    I suggest you read the article I referenced in Quality Engineering. The results of the research indciated that the component seach technique worked pretty well but that it sometimes failed. There were other techniques that proved to be more robust. It’s been a long time since I read the article so I don’t recall if they compared time to completion.
    I’m glad component search worked well for you but its successful application in a finite number of cases does not prove that it is the fastest, the most accurate, or the most robust method. It might be good but is it the best?
    John J. Flaig, Ph.D.
    Applied Technology (e-AT-USA.com)
    PS- I’ll see if I can find the article, if your interested.

    0
    #80250

    Mike Carnell
    Participant

    John,
    Nobody cares if it is the best. Did it fix the problem? So why would you choose to try to diminish what they accomplished with some rhetoric about some article or if it was best.
    Shanin came into our facility in the mid 80’s used component search and fixed something that the engineers had been working on for over 10 years. The people who wrote the check could have cared less if his technique was best worst or whatever. They paid for results not some pontification on some study of possibly limited applicability.
    You need to relax.

    0
    #80267

    John J. Flaig
    Participant

    Mike,
    Your response is an indication of exactly why the “engineers had worked on the problem for 10 years without success”. You need to have a open mind like Mr. Shainin — this discussion is NOT about religion. I’m not trying to say that his technique does not work. In fact, I’m sure it does, what I’m saying is that you need to know when it works and when it does NOT work.  I get the impression that you think that component search always works and this is just NOT true.
    I noticed that you did not ask me to provide you with the referenced article and you attacked the research without even reading it.  Nonetheless, I will provide the reference in hopes the, if not you, then others might benefit.
    Title: Counterexamples for the Component Search ProcedureCopyright: 1993, Marcel Dekker, Inc. and ASQC Author: Amster, Sig; Tsui, Kwok-Leung; Organization: AT&T Bell Laboratories, New Jersey Subject: Binary response; Component search; Confounding; Experimental designs; Fractional factorial; Interactions; Series: Quality Engineering, Vol. 5, No. 4, APRIL 1993, pp. 545-552 Abstract: This article investigates the experimental design used in the “Component Search” (Bhote) technique, a procedure for identifying the significant factorial effects from a larger number of factors. Three counterexamples show that the “component search” technique may give no information for reducing the number of candidate factors when third- or higher-order interactions exist. It was found that the procedure failures were caused by the confoundings of the factorial effects in the experimental design. These confoundings were discussed in detail by Cotter. Similar to the traditional fractional factorial designs, the component search technique should be used with the understanding of the potential confoundings. But when third- or higher-order interactions are negligible, the component search technique gives less ambiguous results than the traditional fractional factorial designs.

    0
    #80269

    Mike Carnell
    Participant

    John,
    Your impression is wrong. I don’t always use component search and I didn’t say that.
    I really don’t care much about your impressions. What is really rubbing a lot of people wrong is your arrogant “I am the light and the way attitude.” A couple guys did a good thing but Dr. John and his paternalistic attitude can’t be satisfied. Let me find some little piece to pick at. It is the same self destructive behavior that has pervaded the quality discipline for years. The big bad guys eating the young. It is to bad that people like you with an extensive academic background can’t figure out how to make yourself useful. You have a piece to add but you are not the end all and be all to this industry.

    0
    #80277

    Jim Athon
    Participant

    John,
    In the article you referenced, the abstract states that
    “Three counterexamples show that the “component search” technique may give no information for reducing the number of candidate factors when third- or higher-order interactions exist”
    Although an interesting statement, the fact remains that 3rd order (or higher) interactions are rare, and I dare say extremely rare in mechanical systems, where Component Search is most often applied. 
    They propose DOE as a better solution to the hypothetical problem.  To adequately observe and quantify 3rd order interactions, for a 2k DOE one must use a resolution 7 at a minimum.  Point is, testing say 8 fators would require 128 treatment combos (at 2^8-1)  you likely would not see the signal for the noise you failed to control because it took so long to somplete the DOE.
    The abstract continues:
    “…But when third- or higher-order interactions are negligible, the component search technique gives less ambiguous results than the traditional fractional factorial designs. “
    Since 3rd order and higher are very uncommon, it is arguable that Component Search is PREFERRED over traditional DOE techniques because it “…gives less ambiguous results …”
    Remember, “In theory there is no difference between practice and theory, but in practice there is.”
    Incidentally, when offered “counterexamples” of how Ppk>Cpk in practice, said examples were immediatly discredited as ‘can’t happen’ or ‘massaged the data’ etc.  Did not the authors of this article do the same thing?
    Regards, Jim

    0
    #80281

    Mike Carnell
    Participant

    Jim,
    Great find. You had more patients than I did. The failure to acknowledge the success (or even effort) just puts me off.
    I am not sure what you see but I have never seen a third order interaction in a mechanical situation. The chemical processes seem to have them but I can’t imagine how a person would apply component search in that situation (how would I take a good one and a bad one apart and reassemble them?).
    Great input. The last paragraph was great.
    Regards,
    Mike 

    0
    #80286

    John J. Flaig
    Participant

    Jim,
    I agree with almost everything you are saying. I’d say that the engineer should think about the possability of higher order interactions. If they perform an analysis and do not get good results, then they have a possible explaination.
    You would only need a resolution IV design to tell you if there was siginficant variation that was not explained by the main effects and two-factor interations. If that was the case, then you would know that the component seach teachnique would not be the right tool to use.
    What do you think.
    John
    PS- Jim, Cpk and Cpk(hat) are two different things.

    0
    #80308

    Mike Carnell
    Participant

    John,
    You really don’t seem to get what it is like out in industry. Nobody is walking around worried about higher order interactions. They are so rare it isn’t worth worrying about unless you are in a chemical/mixture type process. Component serch won’t work in those processes.
    You do something like component search because it is faster, cheaper and easier to control than something like a 5 factor DOE. We aren’t writing a thesis. BB’s are out here to get results and move on to the next process we aren’t doing science projects.
     

    0
    #94594

    Rahayu Indonesia
    Participant

    Dear Raju,
    I would like to ask about the article ” Using Principal component  Analysis in process performance in multivariate data”  by Wang et al. :
    1. Do you know the basic idea that PCA can be used as an approach for obtaining the multivariate indices capability process?
    2. Does the result of using the PCA approach will be better than previous methods like in the article “Comparison of three multivariate process capability indices” by Wang et al.?
    3 If You have the full paper of ” Using Principal component  Analysis in process performance in multivariate data”  by Wang et al., would you like to send it me please? because I need for supporting material my final task (script).
    I’m looking forward to my address email.
    Thanks for your time and attention.
     
     
     
     
     
     
     
     
     
     
     
     
     
     
     
     
     

    0
    #95256

    mohamamd zadeh
    Participant

    I reading the article” Comparison of Three Multivariate Process Capability Indices” and finding that already 3 method is propused by different authors (shahriari,taam & chen).i am also reading the article about principal componenet analysis in process capability.
    can you help about other methods in multivariate process capability &the comparison of these methods.

    0
    #95307

    Birajashis Pattnaik
    Participant

    You can also refer the MPCI proposed by K.S Pern,W.L.Pern and P.C.Lin(2003)
    Mail me [email protected]

    0
    #96866

    Birajashis Pattnaik
    Participant

    Another MPCI can be based on yield index suggested by Boyels R.A,  refered as Spk.. Check a technical paper presented by K.S Chen, W.L.Pern, and P.C.Lin named ” Capability measures for Process with Multiple Characteristics” in QUALITY AND RELIABILITY ENGINEERING INTERNATIONAL.
    Birajashis Pattnaik

    0
Viewing 36 posts - 1 through 36 (of 36 total)

The forum ‘General’ is closed to new topics and replies.