DOE Keys to Success

Quick reference

• 00:04 Hi, I'm Ray Sheen, and it's now time to wrap up this course.
• 00:08 Let me leave you with a few reminders of the keys to successful Design
• 00:13 of Experiments study.
• 00:15 We started started the course with the emphasis on the study objective,
• 00:18 let's revisit that now in light of what you've learned.
• 00:22 Understanding the objective will directly lead to a decision about the study design
• 00:27 and factor selection.
• 00:28 If you want to screen out minor factors, you want a study design that includes many
• 00:33 factors so that you will be able to identify which ones are significant.
• 00:37 But if you're trying to optimize performance,
• 00:39 the objective points to what the response variable must be, and
• 00:43 the nature of the optimization will often dictate what factors to include.
• 00:46 While if instead the goal is a root cause identification,
• 00:50 a different response factors be chosen, and different control variables.
• 00:54 Or at least different high and low levels on the variables may be selected. In fact,
• 00:59 the objective tells us what type of DOE is needed.
• 01:02 If the objective is performance optimization,
• 01:05 then a fractional factorial is used.
• 01:07 If the objective is to get a complete design space equation,
• 01:10 then a full factorial.
• 01:11 If the objective is robust manufacturing process control
• 01:14 by the operator will then it's Taguchi.
• 01:17 And if the objective is to sift through dozens of potential factors to focus on
• 01:21 the vital few, then use a Plackett-Burman DOE.
• 01:24 Unclear objectives lead to missed expectations on the part
• 01:27 of the stakeholders.
• 01:29 They're expecting one type of answer, and you bring the a different type of answer.
• 01:33 They may be expecting a fast and quick Plackett-Burman or Taguchi study, and
• 01:37 you're bringing them a cost and budget for a full factorial study.
• 01:41 A clear objective avoids this type of confusion, the next point is something we
• 01:45 stated early on, and that's the need for a quantitative response variable.
• 01:50 For some reason, this seems to be a difficult element for stakeholders.
• 01:53 When asked what's the most important attribute to measure, they tell us,
• 01:57 well just make it work right.
• 01:59 I hope you can see by now that we need a measurable
• 02:01 value in order to calculate the design space equation,
• 02:05 and all of our graphs and charts are based upon that equation.
• 02:09 A pass-fail objective can’t be analyzed unless they can give
• 02:13 us a specific pass-fail threshold in a quantitative factor.
• 02:17 Of course,
• 02:18 the response factor should directly correlate with the study objective.
• 02:21 So the that the design space equation essentially is an equation for
• 02:25 the study objective.
• 02:27 So a study for
• 02:28 reliability will be measuring the failure rate as the response variable.
• 02:32 Another challenge with obtaining accurate variable response factor data,
• 02:36 is the measuring system. Don't assume it's capable of giving you reliable data.
• 02:41 I had one client who had spent months conducting analysis about an optical
• 02:45 alignment problem with one of their products,
• 02:47 including conducting a DOE to identify the most significant factors.
• 02:52 With inconsistent results and no significant improvement,
• 02:55 I finally convinced them to do a measurement system analysis
• 02:58 on the optical alignment measurement system.
• 03:00 They assured me up until then that everything was just fine.
• 03:03 Well it turned out the system error was nearly ten times larger than the allowable
• 03:08 span from the minimum to the maximum of their tolerance on the response factor.
• 03:12 All that data that they had been collecting was worthless, and
• 03:16 they had to invest in a new optical alignment measuring system
• 03:19 before they could proceed.
• 03:21 If you think back to the design space equation that was generated, it started by
• 03:25 taking an average of the output when the control factors were either plus 1 or
• 03:29 minus 1.
• 03:29 That average is far more meaningful when the response factor
• 03:34 is a variable output and not a zero or one pass-fail answer.
• 03:39 We talked about study design elements early in this course,
• 03:41 it's worthwhile now to revisit some of them again.
• 03:44 When structuring your DOE consider the different design elements,
• 03:48 they'll help with the accuracy and confidence that you have in your study and
• 03:52 the design space equation.
• 03:54 Replication reduces the impact of noise, and by that I mean,
• 03:58 statistical noise not the Taguchi DOE noise.
• 04:01 If you know that there are many factors that you can't control,
• 04:04 you should add replicates to accommodate their likely impact.
• 04:08 Also more replicates will increase the confidence in your statistical analysis.
• 04:12 Remember in those design space equations,
• 04:15 the factors are calculated by using the average of the high and low factors.
• 04:19 Well, once you get up to the point where you have at least 15 or 20 high and
• 04:23 lows of each factor, your statistical confidence in that average is quite high.
• 04:28 And you can add data points by adding replicates, another concern could be that
• 04:32 the experimental test process changes some as the operators are conducting the test.
• 04:37 They may not be consciously changing things,
• 04:39 but repetition creates habits which can influence how the tests are performed.
• 04:43 The use of center points and randomization of the test runs will both tend to
• 04:48 minimize the effect of drift in the process.
• 04:51 If you have a very complex test process, or one that takes a long time,
• 04:55 consider the use of center points.
• 04:58 And blocking was a way to accommodate real world constraints on the test process.
• 05:03 If there's a logical reason why the testing must be done in different groups
• 05:08 like different locations or timing or
• 05:10 equipment, then use blocking, otherwise don't.
• 05:14 The last point to make about the study design goes back to
• 05:17 the resolution number in confounding or aliasing of interaction effects.
• 05:22 Be sure that you understand the level of confounding and
• 05:25 that it fits within the study objective, this will many times dictate the level of
• 05:30 fractionality that you're willing to accept.
• 05:33 My final key to success is one that has become a critical success factor in many
• 05:38 design and development processes over the past ten years: iterate to succeed.
• 05:43 The most common type of DOE is fractional factorial which normally runs
• 05:48 in three phases or iterates through screening, refining and optimizing.
• 05:53 One of the reasons it is so popular is that with each iteration the analysis
• 05:57 provides more clarity and accuracy to the design.
• 06:00 Each phase uses lessons learned from the previous phase to focus the
• 06:05 analysis of the next phase.
• 06:07 Another benefit of iterative approach is related to the fact that it is
• 06:11 a statistical analysis, statistics can be screwed up by outliers.
• 06:16 So the screening study just looks for
• 06:18 big effects which will probably be driven by many data points,
• 06:22 then only truly significant points are analyzed in the refining study.
• 06:27 But that usually has more runs and
• 06:29 the impact of an outlier is therefore minimized.
• 06:32 Bottom line is that you can have confidence in the statistical data and
• 06:37 not be fooled by a lucky run.
• 06:39 And one other thing, the Minitab results in the session window will also provide
• 06:44 you a level of statistical confidence in your analysis.
• 06:47 Finally, there's the confirmation phase of the study,
• 06:51 that last check to make sure that your calculations did not have an error.
• 06:55 It's a great comfort to have that result come in at or
• 06:58 near the level of your predictive performance in the design space equation.
• 07:03 And when you go back to the stakeholders and say, here's the equation,
• 07:07 this is what's predicted, and here's the actual results,
• 07:11 you have made a strong case for a new design or improvement.
• 07:15 So those are the keys for the DOE success, a clear objective,
• 07:19 a quantitive response variable, robust study design that takes
• 07:24 advantage of the design elements, and iterate to success.

Lesson notes are only available for subscribers.