Page 303 - Six Sigma Advanced Tools for Black Belts and Master Black Belts

OTE/SPH
OTE/SPH
August 31, 2006
JWBK119-18
Taguchi Methods
288 3:6 Char Count= 0
Table 18.4 Design of experiments vs. statistical process control.
Statistical process control Statistical experimental design
1. Used for ‘on-line’ quality control 1. Used for ’off-line’ quality engineering
2. Deals mainly with existing 2. Can be applied at process design and
processes development stage
3. Meant for routine application 3. Has a problem-solving dimension
4. Aims to maintain status quo . 4. Seeks improvements and best operating states
5. No new operational targets 5. Motivated by specific needs and seeks new
results
6. Nonintervention of physical system 6. Purposeful probing of physical system
7. Based on passive observation of 7. Depends on active manipulation of system
system output input--output linkages
8. Monitors known key parameters 8. Identifies key parameters
9. No forward planning element 9. Attempts to foresee and prevent problems
10. Waits for problems to happen 10. Identifies sources of problems and seeks their
elimination
11. No obvious sense of urgency 11. Efficiency is important
12. Carried out continuously 12. Carried out project by project

more controversial Taguchi procedures need not be introduced to nonspecialists in
first courses and, by the same token, less down-to-earth statistical arguments can be
reserved for those who have the inclination to probe deeper into the subject.
In place of an emphasis on completeness of coverage, efficiency of information
extraction and utilization should be highlighted as the prime concern. As is well
known, Taguchi methods users have to base their experimental designs on standard
tables that offer very little latitude in their adoption; thus one has to be careful not to
let design of experiments be reduced to an exercise in which a problem is made to fit
a standard orthogonal array. It is not unusual to see Taguchi case studies involving
multilevel designs right from the outset. In practice, however, fewer experimental runs
with two-level fractional factorials should be attempted first as (i) not all parameters
will turn out to be significant (again the 80--20 principle), so no purpose is served
by according to every one of them a multilevel treatment, thereby increasing the
size of the experiment; (ii) most changes in response are approximately linear with
respect to parameter values that vary within a realistic experimental range; (iii) even if
nonlinearities exist, they do not bias the estimates of main and interaction effects; (iv) it
is in fact possible to institute simple tests for the existence of significant nonlinearities
and, if they do exist, (v) it is a straightforward matter to augment a two-level design
by additional experimental runs to constitute a higher-level design. 12,29,30 Hence it is
unproductive to plunge into grandiose designs, complete with inner and outer arrays,
and bet on the final confirmation experiment to justify the entire effort, when fruitful
results can be obtained through a judiciously executed series of small, sequential
experiments.
Another important feature of an effective training program would be the elucida-
tion, best done by case studies, of the roles of different tools at different points in
the learning curve associated with the study of a particular physical subject. Taguchi