The Design limits include Upper Specification Limit USL and Lower Specification Limit LSL. The measurements include sample mean m', lot mean m, estimated standard deviation s', and standard deviation s. Calculate Cp which is a process index that relates the specification range to process variation. Cp = (USL-LSL)/6s' Cpk is a process index that relates the process mean to the nominal value of the two sided specification. First determine the difference between the process mean and the specification limits. XUSL = (USL-m') XLSL = (LSL-m') Select the minimum of the two values. XMIN = min[XUSL, - XLSL] Cpk is found by the following. Cpk = XMIN/3s' Relation between Cp and Cpk Cpk = Cp(1-K) K is found with Target value T, which is the center of the specification range, and Specification width W, which is the specification range. K = (m' - T)/(W/2) T = (USL – LSL)/2 W = USL-LSL Ppk is the same as Cpk, however replace m' and s' with m and s. Its a difference of sample verse lot. Ideal Test Limits You might think that big Cpk numbers mean good test limits, but that's wrong. Cpk= 1.33 (4 StdDev) is about ideal, a little larger is often needed in electronics to allow component variably that can't be controlled. You should consider reducing the test limits if Cpk is 1.67 (5 StdDev) or larger. Some Quality people would say I was nut's for saying this but the job of test is to find things that are bad, or put another way unlikely to be good. If the test is passing some units outside 4 StdDev they probability have a problem. Production Outliers What are production Outliers? This is a unit that falls in the gray area of being statistically unlikely but within the design limits. Outliers are detected with probability. Some time back I design an automatic test system that tested and passed over 12,000 power processing devices, it passed 63 units outside four standard deviations, i.e. Outliers. Assuming a normal distribution only 0.0063 percent should fail, less than one part Statistically. I am aware of reliability problems inherent in outliers. I have performed analysis on some of these units and saw things like increased resistance due to an open in one of the parallel windings, partial solder flow, wrong value parts. A well tested product looks at characteristics that maximize the use of statistics. One of the best examples I know is using bias levels in circuits that come into play only under cretin conditions like light load, thermal shutdown, current limit, input lockout and output lockout. These bias level test reveal circuit failures as outliers, without adding complex test methods. Probability Between +/- 1 standard deviation 68.27% +/- 2 standard deviation 95.45% +/- 3 standard deviation 99.73% +/- 4 standard deviation 99.9936% +/- 5 standard deviation 99.999943% Reference Statistical Quality Design and Control 1992 Prentice-Hall, Inc. Process Measurement By Hawthorne SMT February 1997 CHANGE LOG: History $Id$ Release New Initial release Copyright Notice Copyright (C) 2004-2008 Ronald Steven Sutherland To report errors or omissions http://epccs.com/contact.php Permission is granted to copy, distribute and/or modify this document under the terms of the GNU Free Documentation License, Version 1.2 or any later version published by the Free Software Foundation; with no Invariant Sections, no Front-Cover Texts, and no Back-Cover Texts. A copy of the license is included in the section entitled "GNU Free Documentation License". GFDL taken form http://www.gnu.org/copyleft/fdl.html