There seems to be disagreement on when to use fuzzy and when to use 
probability.  An example in manufacturing and production of machined parts 
may be helpful.  In manufacturing and production we use statistical process 
control (SPC) to maintain processes so that parts are produced within 
specified tolerances and will perform within specifications.  However, we 
know from experience that performance varies from within that tolerance 
range.  Those parts that perform the best and have the greatest reliability 
are those parts whose production parameters most closely match the nominal 
dimensions in the design tolerances.  As a consequence using SPC we attempt 
to reduce the process variance and keep the process means centered on their 
corresponding nominal tolerance values.  Although often confused, the 
tolerance values and the process values are not the same thing.  The 
process values are characterized by distribution parameters determined 
empirically through collecting event data on the process results.  The 
engineers who designed the product set the tolerances.  When we generate a 
plan for manufacturing a product we match  the available processes to the 
tolerance.  So a tolerance is not a probabilistic specification.  The 
nominal dimension is the location where full compliance with the design 
requirements exists.  The upper and lower tolerances are the upper and 
lower limits beyond which the performance is unacceptable (zero 
existence).  Within these tolerance boundaries, acceptable design 
compliance exists.  However, this acceptable design compliance is not 
uniform.  The closer to the nominal dimension the closer to full design 
compliance.  Thus, a gradient exists between the upper and lower tolerances 
and the nominal dimension.  Conceptually, this matches the description of a 
fuzzy set and these gradients can be modeled with fuzzy sets.  In addition, 
the actual performance of any manufactured part has the same 
characteristics.  Those parts whose parameters are closest to the nominal 
dimension will perform the best, those closest to the upper or lower 
tolerance will perform the worst.  Think of the "Monday/Friday car" that no 
one wants.  They are all within tolerance.

In moving a part to production we match the tolerances to a process by 
aligning its nominal dimension to a process mean and the upper and lower 
tolerance to their respective locations on the distribution.  The 
probability mass outside these limits determines the scrap rate.  The 
probability mass within these limits determines the "good" parts.  In 
accepting or rejecting a part we make no distinction between parts within 
these limits and statistically we treat them the same.  However, we 
recognize they are not the same by constantly trying to reduce the process 
variance while keeping the process mean centered on the nominal 
value.  This is to produce more parts with parameters close to the nominal 
dimension because they will more fully match the design intent and have 
better performance.  Thus, the process is probabilistic and we control it 
using SPC but the behavior of the design and the intent of the designer 
more closely matches the description of a fuzzy set.  With this in mind, we 
use fuzzy sets to model tolerances and assess designs, and probability to 
model processes and assess the outcome of production.  Hope this adds some 
insight.

Robert Young


=========================================================
Dr. Robert E. Young
Professor of Industrial Engineering
Department of Industrial Engineering, Campus box 7906
North Carolina State University
Raleigh, NC 27695-7906  USA
personal site: http://www.ie.ncsu.edu/young
research site: http://www.ie.ncsu.edu/wisdem
Phone: [+1] 919/515-7201
FAX: [+1] 919/515-5281
email: [EMAIL PROTECTED]
=========================================================

Reply via email to