Statistical

Tolerance
Analysis
Basic Introduction

Fred Schenkelberg
Statistical Tolerance Analysis

Copyright © 2016 Fred Schenkelberg

Publication date July 2016

Attribution-NonCommercial-NoDerivatives

4.0 International License.

http://creativecommons.org/licenses/by-nc-nd/4.0

Feel free to email, share, and pass this ebook around the web

but please don’t alter any of its contents when you do.

Thanks!

FMS Reliability Publishing

17810 Comanche Trail
Los Gatos, California 95033
fmsreliability.com/publishing/
fms@fmsreliability.com

Join Accendo Reliability

www.accendoreliability.com

It’s free and you will find an expanding inventory of

reliability engineering professional development

articles, podcasts, webinars, and courses

Design for Reliability Series 2 Accendo Reliability

Statistical Tolerance Analysis

Contents
Objective 4

Why Do Tolerance Analysis? 4

Dilemma of Communication about Tolerances 7

Worst-Case Tolerance Analysis 8

Root Sum Squared Method 10

Monte Carlo Method 16

Where to Find the Data to Determine

Capability of Your Expected Tolerances 23

How to Construct and Interpret Histograms 25

Interpretation of Histograms 35

Probability Density Function 48

PDF Construction as an
Extension of a Histogram 50

Normal Distribution 53

Lognormal Distribution 59

Uniform Distribution 66

Triangle Distribution 70

Tolerance and Reliability 74

Design for Reliability Series 3 Accendo Reliability

Statistical Tolerance Analysis

Objective
Tolerances serve to communicate from the design team to the suppliers
and manufacturing teams the expected range of values for an element of a
product or system.

Why Do Tolerance Analysis?

The short answer is that everything varies.

The longer answer involves the agreement between what is possible and
what is desired.

If we could design a product and it could be replicated exactly, including

every element of the product, we would not need tolerances. Any part would
work with any assembly. We would simply specify the dimensions required.

Instead, variation happens.

Width, length, weight, roughness, hardness, and any measure you deem
worth specifying will vary from one part to the next. Manufacturing
processes impart some amount of variation among each item produced. In
many cases the variation is acceptable for the intended function. In some
cases the variation is unacceptably large and leads to failures. When the
design does not account for the variation, holes will not align, components
will not fit, or performance will be poor.

Design for Reliability Series 4 Accendo Reliability

Statistical Tolerance Analysis

When the designer understands the manufacturing process and naturally

occurring variation, the design tolerances balance what is possible with
what is necessary.

Functional System

The final product performance relies on each component functioning as

expected. Gears mesh and wheels turn. The inputs to the system provide the
desired output. Doors close securely, being neither too tight nor too loose.
Tolerances provide the range of values for each element of the design that
enables the desired results to occur with each product produced. Instead
of making the parts exactly to the drawing dimensions, the manufacturer
creates the item such that it is within the tolerances, so that the assembly
will perform as the designer intended.

The customer’s perception of quality relies on the components fitting and

working well together. Alignment, fit, and finish have to be just right, and not
noticeably off.

You may have noticed the difference between poor and well-crafted dresser
drawers: Poor drawers with poor alignment may bind when opening or
closing, whereas well-crafted drawers work as expected without binding
and do so smoothly.

Design for Reliability Series 5 Accendo Reliability

Statistical Tolerance Analysis

Manufacturable

Tolerances stack up to the extent that, if they are carelessly considered,

parts will not fit at all or fit poorly, leading to failures and scrap. When parts
do not fit, holes do not align, or connectors do not mate, the product is, at a
minimum, not assembled correctly, requiring repairs, or is simply scrapped.
In either case, the unnecessary costs mount.

The creation of individual parts, plus the variation of assembly processes,

leads to the need for generous tolerances. Yet the design may require tight
tolerances for the product to function as intended. The key lies in finding the
balance between performance and the ability to assembly the system. This
is an economic balance. It may cost more to maintain very tight tolerances.
The cost of advanced manufacturing processes or the cost of inspection and
scrap has to be balanced by the economic benefit of the customer-perceived
quality and willingness to purchase the product.

Cost of Individual Parts Appropriate to Requirements

There are many ways to form parts: from casting to stamping and from
die cutting to hand cutting. Each process has inherent variability. When
the process is stable and the equipment well maintained, the parts will
reflect the inherent variability of the manufacturing process. Each type of
manufacturing process has limits on precision and, generally, the more
precise methods are more expensive.

Design for Reliability Series 6 Accendo Reliability

Statistical Tolerance Analysis

Using only very tight tolerances may require using expensive manufacturing
processes for the parts. If not all part precision contributes to final
performance then some if not most of the part tolerances may be less
stringent. This may allow less expensive part manufacturing methods and
still create a product with the desired performance.

Dilemma of Communication about Tolerances

Design teams understand the need to include tolerances on part drawings
and component purchase specifications. The designing engineers need to
know the suppliers’ capabilities to achieve specific tolerances. Suppliers
desire to create parts that meet the design requirements. Everyone is trying
to do the best they can to together create a quality product. Nonetheless,
again variation occurs.

Ideally, there would be no variation and the design nominal values would be
sufficient. Variation adds complexity to the design.

Ideally, the design would accommodate the entire range of variation that
arises with manufacturing of the parts. There would be no scrap or wasted
material. Variation happens, and designs that can accommodate the most
variation are good (for the supplier).

Ideally, the part variation is stable and predictable and within the needs
of the design. This doesn’t happen by chance. Communication between
the design team and the supplier has to be clear and complete concerning

Design for Reliability Series 7 Accendo Reliability

Statistical Tolerance Analysis

design requirements and part variation. The two sides have to find a happy
balance that leads to an acceptable economic and quality solution.

The three principal methods of tolerance analysis are

1. the worst-case method,

2. the root sum squared method, and
3. the Monte Carlo method.

In this ebook we will discuss these three methods, with emphasis on the
Monte Carlo method. Each method has a role to play in most designs and
each has advantages. Of the three approaches, the Monte Carlo method
when done well provides an accurate analysis of the information you have
available. It provides a realistic estimate of the resulting assembly variation.

Being familiar with each method increases your capability to use the
appropriate tool as you engineer an appropriate design.

Worst-Case Tolerance Analysis

Worst-case tolerance analysis is the starting point when creating a
tolerance specification. It is a conservative approach as it only considers
the maximum or minimum values of part variation—whichever leads to the
worst situation.

In the worst-case method you simply add the dimensions using the extreme
values for those dimensions. Thus, if a part is specified at 25 ± 0.1 mm,

Design for Reliability Series 8 Accendo Reliability

Statistical Tolerance Analysis

then use either 25.1 or 24.9 mm, whichever leads to the most unfavorable
situation.

The actual range of variation should be the measured values from a stable
process. It may be based on vendor claims for process variation, industry
standards, or engineering judgment.

Simple Example

Let’s consider a stack of five plates and we want to estimate the combined
thickness. If each plate is 25 ± 0.1 mm, then the combined thickness will
be 5 times 25 mm, for 125 mm for the nominal thickness. The math for the
minimum and maximum is about as simple.

The tolerance is ± 0.1 mm; thus, combining five plates at maximum and
minimum tolerances provides a tolerance for five plates of ± 0.5 mm.

Thus, the stack of five plates will have a thickness of 125 ± 0.5 mm or a range
in thickness from 124.5 to 125.5 mm.

Best Practices and Assumptions

Worst-case tolerance analysis is quick and easy. You need just the
tolerances of the components involved. There is no need for distributions
or assumptions about distributions. We should have evidence that the
part tolerances are real though. If the plate is specified as 25 ± 0.1 mm,

Design for Reliability Series 9 Accendo Reliability

Statistical Tolerance Analysis

then verify that the measured values actually fall within the range of the
tolerance.

If the design function and the manufacturing process work using the worst-
case tolerance analysis, then that is a safe way to set tolerances.

There are, however, cases where the tolerance stack is too large for the
design or assembly process. In that case, consider conducting the analysis
using the root sum squared or Monte Carlo methods.

Root Sum Squared Method

The root sum squared (RSS) method is a statistical tolerance analysis
method. In many cases the actual individual part dimensions fall near the
center of the tolerance range with very few parts with actual dimensions
near the tolerance limits.

This of course assumes the part dimensions are tightly grouped and within
the tolerance range.

In the RSS method, one assumes that the normal distribution describes
the variation of dimensions. The bell-shaped curve is symmetrical and fully
described with two parameters, the mean μ and the standard deviation σ.

The variances, not the standard deviations, are additive and provide an
estimate of the combined part variation. The result of adding the means and
taking the root sum square of the standard deviations provides an estimate

Design for Reliability Series 10 Accendo Reliability

Statistical Tolerance Analysis

of the normal distribution of the tolerance stack. The formula to combine

the standard deviations of the stack is

v sys = | n
i=1
v 2
i

where σi is the standard deviation of the ith part and n is the number of parts
in the stack.

2.15% 13.1% 34.1% 34.1% 13.1% 2.15%

μ + 3σ μ + 2σ μ + σ μ μ + σ μ + 2σ μ + 3σ
68.2%
95.4%
99.7%

Design for Reliability Series 11 Accendo Reliability

Statistical Tolerance Analysis

The normal distribution has the property that approximately 68.2% of the
values fall within one standard deviation (1σ) of the mean. Likewise, 95.4%
fall within two standard deviations (2σ), and 99.7% fall within three standard
deviations (3σ). The plot above shows the probability for various regions
relative to the standard deviations away from the mean.

Simple Example

Using the same example as with the worst-case method, we have five plates,
each of which will have a different dimension. For any given set of five we
do not know the five individual dimensions, yet we can estimate what those
dimensions will be using statistics.

On average, the plates are 25 mm thick. Assuming that each part will be
slightly different than the average value and that the normal distribution
describes the variation, we then need to estimate the standard deviation of
the part thickness.

For this example let’s measure 30 plates and calculate the standard
deviation. If we find that the standard deviation is 0.33 mm we know that
most parts will have dimensions within the tolerance of 0.99 mm, if the
parts follow a normal distribution (and how to check this assumption will
be discussed later). This is our estimate of how the part thickness actually
varies.

Design for Reliability Series 12 Accendo Reliability

Statistical Tolerance Analysis

When you stack five blocks, the average thickness will 5 times the mean
thickness, or 125 mm.

We expect approximately 99.7% of the stacks of five blocks to have the

combined thickness to be within the range of plus or minus 3σ of the
combined plates. To combine the standard deviations we use the formula
to add the variances and convert back to standard deviation with a square
root.

In this case we add the five variances, 0.332, and take the square root of that
sum:

v sys = | 5
i=1
0 . 33 2
i = 0.7379

Since approximately 99.7% of the values are within ± 3σ, the range of
combined thickness values for the stack of five plates should be within 125 ±
(3 × 0.7379 or 2.2137) mm, or most fall between 122.79 and 127.21 mm.

To estimate the number of assemblies outside the desired tolerance we

can use the system normal distribution values; in this case the mean μ is
125, and the standard deviation σ is 0.7379. Within Excel you can use the
NORMDIST function. In general, construct the cell as follows:

=(NORMDIST(Mean+Tolerance, Mean, σsys)-0.5)*2

Design for Reliability Series 13 Accendo Reliability

Statistical Tolerance Analysis

where the mean is that of the combined means of the parts involved in the
stack. In this example the system mean is 125 mm.

The tolerance is the desired value. In this examples let’s assume that we
would like the total stack to be within 2 mm of the mean, or a tolerance of 2.

σsys is the standard deviation of the combined parts found by using the RSS
standard deviations of the parts involved.

We subtract 0.5 to find the one-sided probability of the result, being below
the maximum value (mean plus tolerance), and we multiple the resulting
probability by 2 to find the chance that the final assembly is either above or
below the desired tolerance.

In this example, for a tolerance of 2 mm, we would expect 99.33% of

assemblies to have a thickness within 125 ± 2 mm. This implies that we
should expect one assembly out of about 300 to result in a thickness either
thinner than 123 mm or thicker than 127 mm. By varying the tolerance in the
calculation we can estimate the scrap or defect rate and compare the cost of
scrap and failure to the cost of tighter individual part tolerances.

Design for Reliability Series 14 Accendo Reliability

Statistical Tolerance Analysis

Best Practices and Assumptions

The normal distribution assumption relies on the process variation having

many small perturbations that generally add to create the final dimension. It
is best to actually measure approximately 30 samples to estimate the mean
and standard deviation.

When gathering measurements is not feasible, then assuming that the parts
will have dimensions centered within the tolerance range and have ±3σ
across the tolerance range is a conservative starting assumption. Of course,
this implies that the part creation process is capable of creating 99.7% of
the parts within the tolerance specifications.

If measuring fewer than 30 parts to estimate the standard deviation, be sure

to use the sample standard deviation formula.

| ^xi - x h
N
i=1
v= N-1

where N is the number of samples, xi is the ith measurement, and x̄ is the

sample mean of the samples.

Design for Reliability Series 15 Accendo Reliability

Statistical Tolerance Analysis

Monte Carlo Method

In the Monte Carlo method one uses the idea that not all parts have the
same dimensions, yet a normal distribution describing the variation of the
part dimensions is not assumed. Although the normal distribution does
commonly apply, if the process includes sorting or regular adjustments or if
the distribution is either clipped or skewed then the normal distribution may
not be the best way to summarize the data.

The Monte Carlo method uses the part variation information to build a
system of randomly selected parts and determine the system dimension. By
repeating the simulated assembly a sufficient number of times, the method
provides a set of assembly dimensions that we can then compare to the
system tolerances to estimate the number of systems within a specific range
or tolerance.

In short, we randomly select part dimensions and apply a transfer function

(to simulate how the parts combine to create the final dimension) to create
the resulting assembly distribution of dimensions.

Design for Reliability Series 16 Accendo Reliability

Statistical Tolerance Analysis

x1

x2

f ^ x 1, x 2, g, x nh y

xn

The resulting distribution will be a normal distribution when all of the input
part dimensions are normally distributed. Also, if there are a large number
of parts in the stack, the result is also likely to be a normal distribution. For
a few parts with other than normal distribution of dimensions, the result will
likely not be normally distributed.

Design for Reliability Series 17 Accendo Reliability

Statistical Tolerance Analysis

Number of Simulation Runs

For a simulation to produce consistent results, it should use the number

of runs that allow the result to converge with some defined margin.
Commercial Monte Carlo software packages monitor the results after each
run to determine the convergence and stop when the value remains within a
pre-specified range.

One way to estimate the number of runs requires a two-step approach.

First, run the simulation for 1000 runs.

Calculate the standard deviation of the resulting values.

Then use the estimated standard deviation to estimate the number of runs m
needed to achieve a desired sampling confidence with 1% accuracy:

Za 2 # v 2
m =c m
Er ^nh

where Zα/2 is the standard normal statistic for a two-sided C confidence,

where α = (1 − C), with C being the statistical sampling confidence,
commonly set to 95%, and Er(μ) is the standard error of the mean and is
related to the amount of variation of the estimated mean.

Design for Reliability Series 18 Accendo Reliability

Statistical Tolerance Analysis

Dividing the numerator and denominator by μ converts the terms to

percentages:

JK Z a 2 # v NO 2
KK O
K n OO
KK Er ^nh OOO
m=K
K n OP
L

Thus, Er(μ)/μ is the percentage accuracy (convergence threshold), and σ/μ is

the relative standard deviation.

Simple Example

Step 1: Define the problem and the overall study objective.

What is the appropriate tolerance for individual plates used in a stack of five
to achieve a combined thickness of 125 ± 3 mm?

Given the variation of the plate thicknesses, how many assemblies will have
thicknesses outside the desired tolerance range?

Step 2: Define the system and create a transfer function that defines how
the dimensions combine: y = f(x1, x2, …, xn).

Assuming no bow or warp in the plates, we find that the stack thickness of
five plates is simply the added thicknesses of each plate: y = x1 + x2 + x3 + x4 +
x5.

Design for Reliability Series 19 Accendo Reliability

Statistical Tolerance Analysis

Step 3: Collect part distributions or create estimated distributions.

In this simple example, the five plates are the same and from the same
population having a mean value of 25 mm and standard deviation of 0.33
mm.

Step 4: Estimate the number of runs for the simulation.

The five plates have normally distributed thicknesses of 25 ± 0.99 mm (a

result that should converge to the same result as if using the RSS method).
Instead of setting up and running the Monte Carlo simulation to estimate
the standard deviation we can use the calculated standard deviation for the
combined five plates of σ = 0.7379.

Let’s use a confidence of 95%; thus, α = 1 − C = 1 − 0.95 = 0.05. Let’s conduct

sufficient simulation runs to have an accuracy of 1%.

We have, Zα/2 = 1.645 (in Excel = NORMSINV(0.95)), and σ/μ = 0.7379/125

= 0.0059, and Er(μ) is given as 0.01 or 1%, which means Er(μ)/μ is 0.01/125 =
0.00008.

The number of required runs, m, is then 1213852 = 14,735.

Creating 15,000 runs with Excel may be difficult, if even possible.

Thus you may require the use of a Monte Carlo package such as Crystal Ball
(an Excel add-on) or 3DCS Variation Analyst within SolidWorks.

Design for Reliability Series 20 Accendo Reliability

Statistical Tolerance Analysis

Step 5: Generate a random set of inputs.

For each of the five parts in the example, draw at random a thickness
value from the normal distribution with a mean of 25 mm and a standard
deviation of 0.33 mm. If you are using Excel the thickness for one plate
can be found with =NORMINV(RAND(),<mean value>,<standard deviation
value>). Do this for each part in the stack.

One run set of inputs may result in 25.540, 25.008, 24.565, 24.878, and
24.248 mm.

Step 6: Evaluate the inputs.

Use the input values in the transfer function to determine the run’s result.
In this example, we add the five thickness values to find the stack thickness.
The sum of the five thicknesses for one run in Step 5 results in 124.238 mm.

Step 7: Repeat Steps 5 and 6 for the required number of runs (Step 4) and
record each run’s result.

Step 8: Analyze the results using a histogram, confidence interval, best-fit

distribution, or other statistical method.

When all the input distributions are normal distributions, the resulting
distribution will also be normal. For the five plates the result is a normal
distribution with a mean of 125 mm and standard deviation of 0.7379 mm.

Design for Reliability Series 21 Accendo Reliability

Statistical Tolerance Analysis

As with the RSS method we can then calculate the percentage of assemblies
that have dimensions within 125 ± 3 mm. We find that 99.33% fit within the
defined tolerance when each plate is 25 ± 0.99 mm.

We can explore changing the part tolerance, if supported by manufacturing

processes, to determine the impact on final tolerance.

Best Practices and Assumptions

If you assume or actually have part dimensions that can be described by the
normal distribution, then the Monte Carlo method results will equal the RSS
method results.

If uncertain of the part dimension tolerance, use a uniform distribution. If

you suspect the distribution to be centered within the tolerance range for the
part, then use the triangle distribution.

Preferably, you should use an estimate of the actual distribution of part

variation based on measurements.

When that is not possible, then use engineering judgment and

manufacturing process knowledge to estimate the part variation
distribution.

Vendors may have measurements of similar parts that may provide to be a

suitable estimate for the distribution values.

Design for Reliability Series 22 Accendo Reliability

Statistical Tolerance Analysis

Where to Find the Data to Determine Capability

of Your Expected Tolerances
The RSS or Monte Carlo methods require an estimate of distribution
values. Ideally, the values are based on hundreds of measurements from
the actual assembly process. During the design stage, the final assembly
process doesn’t exist, so we need to find suitable estimates for the tolerance
analysis.

Engineering Judgment and Industry Standards

Some manufacturing, forming, cutting, and shaping processes have

established tolerances. Industry standards may provide a guideline to
determine the suitable distribution values. Also, your experience with
previous designs and their assembly provides a rich source of information
that provides a rough estimate.

Some organizations make a practice of measuring and recording

dimensions, thus forming a database of assembly or part variation dataset.
This provides the source data for tolerance analysis in the future using
similar manufacturing processes.

A common practice is to assume 6σ across your tolerance range (and later

check with real data).

Design for Reliability Series 23 Accendo Reliability

Statistical Tolerance Analysis

Given just a tolerance and no suitable estimate, one method is to assume(!)

that the part dimension will have a normal distribution with a mean
centered within the tolerance range and ± 3σ. This implies that the parts will
be within the tolerance range 99.7% of the time, if we assume no shift in the
part mean value.

For a conservative estimate to use in the Monte Carlo method you should
assume a uniform distribution.

Vendor Data

Ask your vendors whether they have data or recommended tolerances for
their specific parts. Ask whether they have estimates for the distribution of
the resulting parts.

Measure Parts

When possible measure parts and create an estimate for the appropriate
distribution. Ideally, you should measure hundreds of parts, yet 30 parts will
provide a fairly accurate distribution estimate.

Plot the data using a histogram to get a sense of the appropriate distribution
to employ.

Design for Reliability Series 24 Accendo Reliability

Statistical Tolerance Analysis

How to Construct and Interpret Histograms

Histograms provide a graphical way to display a set of quantitative
measurements. They provide a graphical summary of the univariate dataset
distribution. When confronted with a table of values, a great first step is to
construct a histogram. The graph provides visual information that allows
you to estimate the following:

• the center of the data,

• the spread of the data,
• whether the data are skewed and in which direction,
• whether there are outliers, and
• whether there are multiple modes in the data.

Histograms also provide a means to estimate probability of occurrence.

They provide the basis for the probability density function, which forms the
basis for many statistical tools.

There are three basic types of histograms: counts, normalized counts, and
relative frequency. Let’s explore the construction of each type using a set of
30 length measurements

Design for Reliability Series 25 Accendo Reliability

Statistical Tolerance Analysis

Lengths
4.5 3.4 4.3 4.6 3.8
4.5 4.3 4.5 4.0 4.5
3.9 4.2 4.6 4.4 4.0
4.6 4.3 4.8 4.4 4.0
4.1 4.7 3.9 4.5 3.7
4.2 4.6 5.2 4.2 5.0

Count Histogram

This basic form of histogram counts the number of values that fall within
bins. Bins are non-overlapping equal ranges along the measurement axis.
The bins may be assigned arbitrarily or by using a systematic rule (see Bin
Size Guidelines below).

It is common practice to establish between 5 and 20 bins that span the

range of the data. When first exploring a dataset try different bin sizes to
determine how many bins reveal useful information about the dataset.

Using the measurements above, you can determine the range. The highest
(longest) value is 3.4 and the lowest (shortest) value is 5.2. The difference
is the range. Divide the range by the desired number of bins, then round the
resulting bin size value to a number that is convenient to work with. This
result is the bin width.

Set the first bin to include the lowest value. It is good practice to select a
starting point that creates bin boundaries that do not coincide with any of

Design for Reliability Series 26 Accendo Reliability

Statistical Tolerance Analysis

the data points, thus avoiding any issues concerning to which bin a value
belongs.

Another common practice is to define the bin as including the first low end
of the bin range and not including the high end of the bin range (a low-open-
high-close chart).

The length data have a minimum of 3.4 and a maximum of 5.2 for a range of
5.2 − 3.4 = 1.8

Dividing 1.8 into 6 bins gives bin sizes of 0.3 (which is a rough estimate of the
standard deviation, which is actually 3.87).

Starting the first bin at 3.35 and adding enough bins to span all the data
creates seven bins ranging from 3.35 to 5.45 in steps of 0.3.

For each bin, count the number of measurements that fall within that bin
range.

Bin Interval Count

3.35–3.65 1
3.65–3.95 4
3.95–4.25 7
4.25–4.55 10
4.55–4.85 6
4.85–5.15 1
5.15–5.45 1

Design for Reliability Series 27 Accendo Reliability

Statistical Tolerance Analysis

Then plot the counts per bin across the measurement range of values. Note
that the count is the frequency of values occurring in the bin interval.

Histogram of Lengths
10
8
6
Count

4
2
0

3.5 4.0 4.5 5.0 5.5

Length

Design for Reliability Series 28 Accendo Reliability

Statistical Tolerance Analysis

Notice that the plot is roughly centered on 4.4, with a slight left skew (i.e.,
there are more values to the left of the mean).

The plot is close to being symmetrical, and altering the number of bins may
enhance or reduce the amount of skewness.

With five bins the plot is perfectly symmetrical.

Histogram of Lengths
14
12
10
8
Count

6
4
2
0

3.0 3.5 4.0 4.5 5.0 5.5

Length

Design for Reliability Series 29 Accendo Reliability

Statistical Tolerance Analysis

With nine bins the left skewness is pronounced.

Histogram of Lengths
8
6
Count

4
2
0

3.5 4.0 4.5 5.0

Length

Try a few bin counts to explore the data; there is no single correct way to
plot the data in a histogram. In general, try to set a bin count and size that

Design for Reliability Series 30 Accendo Reliability

Statistical Tolerance Analysis

minimize the number of bins with one or no values, yet reveal the overall
shape of the data. The first plot with seven bins seems reasonable.

Normalized Frequency Histogram

A histogram can show the normalized count or frequency where the count
in each bin is divided by the total number of observations.

Bin Interval Count % of Total

3.35–3.65 1 3.33
3.65–3.95 4 13.33
3.95–4.25 7 23.33
4.25–4.55 10 33.33
4.55–4.85 6 20.00
4.85–5.15 1 3.33
5.15–5.45 1 3.33
You can determine the % of Total by dividing the count in each bin by 30 and
multiply by 100%.

Design for Reliability Series 31 Accendo Reliability

Statistical Tolerance Analysis

Normalized Frequency of Lengths

30

20
% of Totla

10

3.5 4.0 4.5 5.0 5.5

Length

This construction provides an intuitive plot that permits comparison

between different total count datasets.

Design for Reliability Series 32 Accendo Reliability

Statistical Tolerance Analysis

Relative Frequency Histogram

A histogram can show the relative frequency where the count in each bin is
divided by the total number of observations times the width of the bin. The
resulting histogram has an area equal to one.

Bin Interval Count Relative

Frequency
3.35–3.65 1 0.111
3.65–3.95 4 0.444
3.95–4.25 7 0.778
4.25–4.55 10 1.111
4.55–4.85 6 0.667
4.85–5.15 1 0.111
5.15–5.45 1 0.111
You can determine the Percent of Total by dividing the count in each bin by
30 times 0.3. Thus for the first bin this is 1/(30 × 0.3).

Design for Reliability Series 33 Accendo Reliability

Statistical Tolerance Analysis

Relative Frequency of Lengths

1.0
0.8
Relative Frequency Density

0.6
0.4
0.2
0.0

3.5 4.0 4.5 5.0 5.5

Length

This construction provides a histogram that is similar to a probability

density function.

Design for Reliability Series 34 Accendo Reliability

Statistical Tolerance Analysis

Interpretation of Histograms
It does take some experience and experimentation to fully interpret
histograms, yet the basic structure of a distribution is often evident. Here
are eight different shapes that histograms may take (with brief explanations
of each).

In the classical bell-shaped, symmetric histogram (Histogram 1), most of

the frequency counts are bunched in the middle and counts drop off out in
the tails.

Histogram 1: Normal Distribution

From a physical point of view, the normal distribution is that distribution

which occurs most often in nature.

Design for Reliability Series 35 Accendo Reliability

Statistical Tolerance Analysis

For a symmetric distribution, the “body” of a distribution refers to the

“center” of the distribution, where most of the probability resides. The “tail”
of a distribution refers to the extreme regions of the distribution on either
side. The “tail length” of a distribution is a term that indicates how fast these
extremes approach zero.

For a short-tailed distribution, the tails approach zero very quickly. Such
distributions commonly have a truncated (“sawed-off”) look. The classical
short-tailed distribution is the uniform (rectangular) distribution in which
the probability is constant over a given range and then drops to zero
everywhere else.

For a moderate-tailed distribution, the tails decline to zero in a moderate

fashion. The classical moderate-tailed distribution is the normal (Gaussian)
distribution.

For a long-tailed distribution, the tails decline to zero very slowly and hence
one is apt to see probability a long way from the body of the distribution.

The classical long-tailed distribution is the Cauchy distribution.

In terms of tail length, the histogram (Histogram 2) would be characteristic

of a “short-tailed” distribution.

Design for Reliability Series 36 Accendo Reliability

Statistical Tolerance Analysis

Histogram 2: Uniform Distribution

The optimal (unbiased and most precise) estimator for location for
the center of a distribution strongly depends on the tail length of the
distribution. The common choice of taking N observations and using
the calculated sample mean as the best estimate for the center of the
distribution is a good choice for the normal distribution (moderate tailed),
a poor choice for the uniform distribution (short tailed), and a horrible
choice for the Cauchy distribution (long tailed). Although for the normal
distribution the sample mean is as precise an estimator as we can get, for
the uniform and Cauchy distributions, the sample mean is not the best
estimator.

For the uniform distribution, the midrange, defined as

midrange = (smallest value + largest value)/2

Design for Reliability Series 37 Accendo Reliability

Statistical Tolerance Analysis

is the best estimator of location. For a Cauchy distribution, the median is the
best estimator of location.

In terms of tail length, the histogram shown below would be characteristic

of a “long-tailed” distribution.

Histogram 3: Cauchy Distribution

The mode of a distribution is that value which most frequently occurs or has
the largest probability of occurrence. The sample mode occurs at the peak
of the histogram.

For many phenomena, it is quite common for the distribution of the

response values to cluster around a single mode (i.e., a unimodal
distribution) and then distribute themselves with lesser frequency out

Design for Reliability Series 38 Accendo Reliability

Statistical Tolerance Analysis

into the tails. The normal distribution is the classic example of a unimodal
distribution.

The histogram shown below illustrates data from a bimodal (two-peak)

distribution.

Histogram 4: Bimodal Distribution

The histogram serves as a tool for diagnosing problems such as bimodality.

Questioning the underlying reason for distributional nonunimodality
frequently leads to greater insight and improved deterministic modeling of
the phenomenon under study. For example, for the data presented above,
the bimodal histogram is caused by sinusoidality in the data.

Consider now the histogram shown below, also illustrating data from a
bimodal (two-peak) distribution.

Design for Reliability Series 39 Accendo Reliability

Statistical Tolerance Analysis

Histogram 5: Bimodal Distribution

In contrast to the previous example, in this case the bimodality is not due
to an underlying deterministic model but due to a mixture of probability
models. In this case, each of the modes appears to have a rough bell-shaped
component.

One could easily imagine the above histogram being generated by a process
consisting of two normal distributions with the same standard deviation
but with two different locations (one centered at approximately 9.17 and the
other centered at approximately 9.26). If this is the case, then the research
challenge is to determine physically why there are two similar but separate
subprocesses.

Design for Reliability Series 40 Accendo Reliability

Statistical Tolerance Analysis

A symmetric distribution is one in which the two “halves” of the histogram

appear as mirror images of one another. A skewed (nonsymmetric)
distribution is a distribution in which there is no such mirror-imaging, as
shown in the example below.

Histogram 6: Skewed Right Distribution

For skewed distributions, it is quite common to have one tail of the

distribution considerably longer or drawn out relative to the other tail. A
“skewed right” distribution is one in which the tail is on the right side. A
“skewed left” distribution is one in which the tail is on the left side. The above
histogram is for a distribution that is skewed right.

Skewed distributions bring a certain philosophical complexity to the very

process of estimating a “typical value” for the distribution. Suppose that an
analyst has a collection of 100 values randomly drawn from a distribution

Design for Reliability Series 41 Accendo Reliability

Statistical Tolerance Analysis

and wishes to summarize these 100 observations by a “typical value.” What

does typical value mean? If the distribution is symmetric, the typical value
is unambiguous: It is a well-defined center of the distribution. For example,
for a bell-shaped symmetric distribution, the center point is identical to that
value at the peak of the distribution.

For a skewed distribution, however, there is no “center” in the usual sense

of the word. Nonetheless, several “typical value” metrics are often used
for skewed distributions. The first metric is the mode of the distribution.
Unfortunately, for severely skewed distributions, the mode may be at or near
the left or right tail of the data and so it is not an ideal representation of the
center of the distribution.

As a second choice, one could conceptually argue that the mean (the point
on the horizontal axis where the distribution would balance) would serve
well as the typical value. As a third choice, others may argue that the
median (that value on the horizontal axis which has exactly 50% of the data
to the left and 50% to the right) would serve as a good typical value.

For symmetric distributions, the conceptual problem disappears because,

at the population level the mode, the mean, and the median are identical. For
skewed distributions, however, these three metrics are markedly different.
In practice, for skewed distributions the most commonly reported typical
value is the mean; the next most common is the median; the least common
is the mode. Because each of these three metrics reflects a different aspect

Design for Reliability Series 42 Accendo Reliability

Statistical Tolerance Analysis

of “centerness,” the analyst should report at least two (mean and median)
and preferably all three (mean, median, and mode) in summarizing and
characterizing a dataset.

Histogram 7: Skewed Left Distribution

The issues for skewed left data are similar to those for skewed right data.

A symmetric distribution is one in which the two “halves” of the histogram

appear as mirror images of one another. The example below is symmetric
with the exception of outlying data on far right (blue).

Design for Reliability Series 43 Accendo Reliability

Statistical Tolerance Analysis

Histogram 8: Symmetrical Distribution with Outlier

An outlier is a data point that comes from a distribution different (in

location, scale, or distributional form) from the bulk of the data. In the real
world, outliers have a range of causes, from as simple as

1. operator blunders,
2. equipment failures,
3. day-to-day effects,
4. batch-to-batch differences,
5. anomalous input conditions, and
6. warm-up effects

to more subtle causes such as

Design for Reliability Series 44 Accendo Reliability

Statistical Tolerance Analysis

7. a change in settings of factors that (knowingly or unknowingly) affect the

response or
8. that nature is trying to tell us something.

Proportion within Specifications Example

One way to use a histogram is to estimate the percentage of items within a

specification or tolerance range. Let’s use the 30 length measurements to
estimate the number of items within the tolerance of 3.5 to 5.0.

Here are the raw numbers:

Lengths
4.5 3.4 4.3 4.6 3.8
4.5 4.3 4.5 4.0 4.5
3.9 4.2 4.6 4.4 4.0
4.6 4.3 4.8 4.4 4.0
4.1 4.7 3.9 4.5 3.7
4.2 4.6 5.2 4.2 5.0

Here is the first histogram we plotted:

Design for Reliability Series 45 Accendo Reliability

Statistical Tolerance Analysis

Histogram of Lengths
10
8
6
Count

4
2
0

3.5 4.0 4.5 5.0 5.5

Length

The value of 3.5 falls about at the midpoint for the first bin and 5.0 is also at
about the midpoint of a cell. If we assume that the values within a bin are
equally distributed, we may count 0.5 for each split bin. Adding the full bin

Design for Reliability Series 46 Accendo Reliability

Statistical Tolerance Analysis

with a count of 1 for the full bin above 5.0, we have approximately 2 of the 30
readings outside the tolerance, or about 6.6%

Checking the actual readings we find 2 of the 30 readings fall outside the
tolerance, 3.4 and 5.2.

Bin Size Guidelines

The number of bins into which the data are to be separated is an arbitrary
choice. In general, between 5 and 20 bins is a good start. Experiment with
different bin counts as you explore the nature of the dataset.

There are a few published guidelines that provide a calculated starting bin
count.

Sturges’s rule is to set the number of bins, n, to

n = 1 + log 2 ^ N h

where N is the number of observations. Converting to log base 10, this

becomes

n = 1 + 3.3 log 10 ^ N h

One then rounds the result to the nearest integer.

Design for Reliability Series 47 Accendo Reliability

Statistical Tolerance Analysis

Rice’s rule is to set the number of bins to twice the cube root of the number
of observations:

n = 23 N

The best practice is to experiment to find meaningful view of the data

Probability Density Function

Just as the histogram reveals the distribution of a dataset, the probability
density function (PDF) provides a convenient summary for the data.
The PDF is a mathematical representation of a probability distribution.
For continuous data the PDF would be a smooth fit to the histogram bin
midpoints.

Consider the relative frequency histogram and imagine diminishing the

bin size to an infinitesimal width. With an infinite amount of data the
resulting histogram bin peaks would describe a smooth curve. Probability
distributions such as normal and lognormal have PDF formulas with just a
few parameters needed to describe their respective distributions.

The PDF formula provides a convenient way to create random numbers for
use in Monte Carlo studies.

Design for Reliability Series 48 Accendo Reliability

Statistical Tolerance Analysis

PDF Properties

For discrete data the probability function, p(x), has the following properties.

1. The probability P of x taking a specific value is p(x) or

P 6 X = x@ = p ^ x h

2. p(x) is non-negative for all real x.

3. The sum of p(x) over all possible values of x is 1, or

|p j =1
j

pj is the probability at xj, and j represents all possible values of x.

In short, p(x) will have a value between zero and one, or

0 # p^ xh # 1

For continuous data the probability function, f(x), has the following
properties.

1. The probability that x is between two points a and b is 1 is

p 6a # x # b@ = # f ^ x h dx
b

Design for Reliability Series 49 Accendo Reliability

Statistical Tolerance Analysis

2. f(x) is non-negative for all real x.

3. The integral of f(x) over all values is one, or

# f ^ x h dx = 1
3

-3

Likewise, f(x) will have a value between zero and one, or

0 # f^ xh # 1

PDF Construction as an
Extension of a Histogram
The histogram below shows the same 30 points used in the histogram
discussion. The red line is the PDF curve for the normal distribution with a
mean of 4.2 and a standard deviation of 0.3.

Design for Reliability Series 50 Accendo Reliability

Statistical Tolerance Analysis

Relative Frequency of Lengths

1.4
1.2
1.0
Relative Frequency Density

0.8
0.6
0.4
0.2
0.0

3.5 4.0 4.5 5.0 5.5

Length

The PDF doesn’t perfectly fit the histogram, even though the 30 data points
were generated from the normal distribution with a mean of 4.2 and a
standard deviation of 0.3.

Design for Reliability Series 51 Accendo Reliability

Statistical Tolerance Analysis

Changing the number of bins to five allows the PDF curve to appear as a
better fit.

Relative Frequency of Lengths

1.4
1.2
1.0
Relative Frequency Density

0.8
0.6
0.4
0.2
0.0

3.0 3.5 4.0 4.5 5.0 5.5

Length

Design for Reliability Series 52 Accendo Reliability

Statistical Tolerance Analysis

If the fit is close enough, the PDF function is easy to use to determine
probabilities of items outside a tolerance range.

The cumulative density function (CDF) is the integral of the PDF, thus also
ranging from zero to one, yet it provides the area under the PDF, which is
the cumulative probability from negative infinity to the point of interest. We
will use the CDF as a means to create random numbers from a selection of
distributions within Excel.

Probability distributions also have two other functions, the hazard function
and the reliability function.

PDFs or CDFs define the probability distributions. In the next section we

will examine four distributions that may prove useful for RSS and Monte
Carlo tolerance analysis studies: the normal distribution, the lognormal
distribution, the uniform distribution, and the triangle distribution.

Normal Distribution

When to Use

The normal distribution describes the variability of a process that has

many small perturbations. The perturbations combine to create the final
result. For example, when cutting a part, the sharpness, thickness, and
angle of the blade plus the technique used, the material properties, and any
other variable that impacts the final dimension will each contribute a small

Design for Reliability Series 53 Accendo Reliability

Statistical Tolerance Analysis

amount of variation. Adding the variations results in the final dimension

variation from one item to the next.

The normal distribution works when the data are symmetrical and exhibit a
bell-shaped curve. It works when the process is not constrained or adjusted
(centered) regularly. However, it also does not work well to describe items
that have been sorted into groups of different sizes or values.

PDF

For a normal distribution, the probability density function is

-^ x - n h
2

f^ xh =
e 2v 2

v 2r

where μ is the location parameter, also known as the mean, and σ is the scale
parameter, also known as the standard deviation

The following plot shows the normal distribution plotted with three different
mean values, 0, 5, and 12, each with a standard deviation of 1.5.

Design for Reliability Series 54 Accendo Reliability

Statistical Tolerance Analysis

The following plot shows the normal distribution plotted with three different
standard deviation values, 1, 2, and 3, and a mean of 12.

Design for Reliability Series 55 Accendo Reliability

Statistical Tolerance Analysis

When the data are normalized by subtracting the mean and dividing by the
standard deviation, the resulting distribution will have μ = 0 and σ = 1.

This normalized distribution is commonly called the standard normal

distribution.

The PDF for a standard normal distribution is

-x 2

f^ xh =
e 2

2r

A plot of a standard normal distribution is shown below.

Design for Reliability Series 56 Accendo Reliability

Statistical Tolerance Analysis

CDF

The cumulative density function is the integral of the PDF:

F^ xh =
1 -^ x - nh2
#
x
e 2v 2 dx
v 2r -3

Note that the integral of the normal distributor PDF does not have closed-
form solutions and must be calculated numerically.

For the standard normal distribution this reduces to

F^ xh =
1 -^ x h
#
x 2

e 2 dx
2r -3

Mean and Standard Deviation and Sample Standard Deviation

In most situations we do not know the population mean μ and standard

deviation σ.

We can use the data to estimate the mean and standard deviation, thus
estimating the normal distribution.

The mean x̄ is the average and is determined as expected with

| n
i=1
xi
n-x= n

Design for Reliability Series 57 Accendo Reliability

Statistical Tolerance Analysis

The standard deviation s is again as expected calculated with

| ^xi - x h
n 2
i=1
v-s= n-1

Note that most calculators and software packages calculate the standard
deviation by using the above formula for the sample standard deviation.
Some default to the population standard deviation, implying that you are
calculating the parameter will all of the data in the population, which is
rarely the case. The population standard deviation is

| ^xi - x h
n 2
i=1
v= n

You would only use the population standard deviation formula when working
with a small (n ≤ 30) population. With more than 30 values, the difference
between the two formulas is minor. You should always check you calculator
or formula to be sure you are using the sample standard deviation when
working with a sample of data from a population.

Creating Random Numbers from a Normal

To create randomly selected value from a normal distribution in Excel use

the formula

Design for Reliability Series 58 Accendo Reliability

Statistical Tolerance Analysis

=NORMINV(RAND(), μ, σ)

Substitute x̄ and s when the population parameters are not available.

Lognormal Distribution

When to Use

The lognormal distribution describes a range of distribution shapes. It is a

versatile distribution and suitable for skewed or exponential-decay-shaped
distributions.

The lognormal distribution is appropriate when the variation is the

multiplicative product of many small perturbations.

All of the contributions to the variation are positive. Values that fit a
lognormal distribution are all positive. The distribution is defined from zero
to positive infinity.

Repair time, additive coating processes, and delay times are examples of
measures that will often fit a lognormal distribution well.

If the log of the data is normally distributed then the lognormal is

appropriate.

Design for Reliability Series 59 Accendo Reliability

Statistical Tolerance Analysis

PDF

For a lognormal distribution, the probability density function is

^ ln^ x - i h m h2
-d n
f^ xh =
e 2v 2

; x 2 i ; m, v 2 0
^ x - i h v 2r

where σ is the standard deviation of the natural logarithm of the data values
(also referred to as the shape parameter), θ is the location parameter or the
offset or shift to the right for the distribution, and m is the median of the data
values (also known as the scale parameter).

The following plot shows the lognormal distribution plotted with three
different median values, 2, 5, and 12, each with a standard deviation of 1.

Design for Reliability Series 60 Accendo Reliability

Statistical Tolerance Analysis

The following plot shows the lognormal distribution plotted with three
different standard deviation values, 0.5, 1, and 2, with a mean of 12.

When θ = 0 and m = 1 the distribution is called the standard lognormal

distribution.

When θ = 0, there is no offset and the distribution is fully described with two
parameters.

The standard lognormal distribution equation is

ln^ x h2
-c m

f^ xh =
e 2
; x 2 0; v 2 0
x v 2r

A plot of a standard lognormal distribution is shown below.

Design for Reliability Series 61 Accendo Reliability

Statistical Tolerance Analysis

CDF

As noted before, the cumulative density function is the integral of the PDF.
For a lognormal distribution, it is given by

^ ln^ x - i h m h2
-d n
F^ xh = #
e x 2v 2

dx; x 2 i; m, v 2 0
- 3 ^ x - i h v 2r

Note the integral of the lognormal distributor PDF does not have a closed-
form solution and must be calculated numerical.

For the standard normal distribution this reduces to

ln^ x h
-d n
2

F^ xh = #
x
e 2v 2

dx; x 2 i; v 2 0
-3 x v 2r

Design for Reliability Series 62 Accendo Reliability

Statistical Tolerance Analysis

Estimating Parameters

The lognormal distribution has two parameters, μ and σ. These are the
natural log of data mean and the standard deviation. With this
parametrization μ=log(m). The μ and σ do describe the distribution, including
the reliability function, defined as

c ln ^ t h - n m
R^ t h = 1 - U v

where Φ is the standard normal cumulative distribution function, and t is

time.

One of the convenient features of the lognormal distribution is that

estimating the parameters is similar to estimating the mean and standard
deviation of the data using the same functions on our calculator or
spreadsheet.

There is one difference though: You must first calculate the natural
logarithm of each data value.

Let’s say we have the time to failure times for four heater elements. We
know that the time to failure distribution is lognormal from previous work.

We want to estimate the lognormal parameters and estimate the reliability

of this type of heater element at 365 days.

Design for Reliability Series 63 Accendo Reliability

Statistical Tolerance Analysis

Time to Fail ln(Time to Fail)

385 5.9532
427 6.0568
490 6.1944
705 6.5582

Calculate μ

In the table we have the time to failure data and the calculation of the
natural log of each data reading. To calculate μ we calculate the mean or
average value of the four ln(Time to Failure) readings:

5.9532 + 6.0568 + 6.1944 + 6.5582

n= 4 = 6.1907

Calculate σ

The calculation of σ requires a little more math. The formula for the
calculation of standard deviation includes the sum of values squared and
the sum of squares of the values:

n | i = 1 t 2i - _ | i = 1 t i i
n n 2

n ^n - 1h
s=

We need the sum of ln(Time to Failure) for the second summation term and
the sum of squares for the first summation term. Expanding the table to
make the calculations we find the two summation results.

Design for Reliability Series 64 Accendo Reliability

Statistical Tolerance Analysis

Time to Fail ln(Time to Fail) ln(Time to Fail)

Squared
385 5.9532 35.4411
427 6.0568 36.6846
490 6.1944 38.3706
705 6.5582 43.0100
Sum 24.7626 153.5063
As we have four readings, n = 4 in the example. Inserting the sums and n and
doing the math to find the value of σ, we find the second parameter for the
lognormal distribution:

4 ^153.5064h - 24.7626 2
4 ^4 - 1h
s= = 0.2642

Creating Random Numbers from a Lognormal Distribution

To create randomly selected values from a lognormal distribution in Excel

use the formula

=LOGINV(RAND(), μ, σ)

Substitute x̄ and s when the population parameters are not available.

Design for Reliability Series 65 Accendo Reliability

Statistical Tolerance Analysis

Uniform Distribution

When to Use

The uniform distribution has an equal probability of any value within a

finite range. The distribution may adequately describe part dimensions
that are sorted into specification ranges or are overly constrained in the
manufacturing process.

The uniform distribution is also appropriate as an assumed distribution

when you have only the specification or tolerance range and no other
information about the distribution of values. It provides a conservative
assumption for Monte Carlo modeling, yet it is not as conservative as the
worst-case approach.

Instead of using the tolerance or specification values as the values for the
uniform distribution, you could use the observed range of values or worst-
case range of values. The underlying assumption is that each part has an
equal chance to have any value within the defined range. In other words, the
larger the defined range the less chance a part will have a dimension at the
edge. Another assumption is that there is no chance of parts with values
falling outside the defined range.

Design for Reliability Series 66 Accendo Reliability

Statistical Tolerance Analysis

PDF

For a uniform distribution, the probability density function is

1
f ^ x h = ) b - a for a # x # b
0 otherwise

where a is the location parameter and b − a is the scale parameter or the

range over which the distribution is defined.

The following plot shows the uniform distribution plotted with three different
location parameters, 0, 2, and 4, each with a scale parameter of 1.

The following plot shows the uniform distribution plotted with three different
scale parameters, 1, 2, and 4, with a location parameter of 0.

Design for Reliability Series 67 Accendo Reliability

Statistical Tolerance Analysis

When a = 0 and b − a = 1, the distribution is called the standard uniform

distribution.

The standard uniform distribution equation is

1 for 0 # x # 1
f^ xh = '
0 otherwise

A plot of a standard uniform distribution is shown below.

Design for Reliability Series 68 Accendo Reliability

Statistical Tolerance Analysis

CDF

For a uniform distribution, the cumulative density function is again the

integral of the PDF:

x-a
F ^ x h = ) b - a for a # x # b
0 otherwise

For the standard uniform distribution this reduces to

x for 0 # x # 1
F^ xh = '
0 otherwise

The uniform distribution parameters a and b are defined at the endpoints.

These endpoints may present the tolerance range or specification, or they

Design for Reliability Series 69 Accendo Reliability

Statistical Tolerance Analysis

may be minimum and maximum observed values. They may also represent
the range of the measurements or be estimated directly from a histogram or
PDF plot.

Creating Random Numbers from a Uniform Distribution

To create randomly selected values from a uniform distribution in Excel use

the formula

=(b-a)*RAND()

Triangle Distribution

When to Use

A variation of the uniform distribution that acknowledges that the

distribution has different probabilities across the range of possible
values is the triangle distribution. This distribution can be symmetrical or
nonsymmetrical, thus estimating a centered or off-center mode.

You should use the triangle distribution when you know or suspect that the
distribution has a region of higher probability of occurrence than at the
edges of the region. It can be useful when you suspect a normal or lognormal
distribution yet do not have sufficient data to estimate its parameters. The
triangle distribution provides a conservative estimate for the normal or
lognormal distribution.

Design for Reliability Series 70 Accendo Reliability

Statistical Tolerance Analysis

PDF

For a triangle distribution, the probability density function is

Z] 2 ^ x - ah
]]
]] ^b - ah^c - ah for a # x # c
[ 2 ^b - xh
]
f ^ x h = ]]] for c # x # b
]] ^b - ah^b - ch
] 0 otherwise
\

where a is the location parameter, b − a is the scale parameter or the range

over which the distribution is defined, and c is the shape parameter or mode
(most likely value).

The following plot shows the triangle distribution plotted with three different
location parameters, 0, 2, and 4, each with a scale of 1 and shape parameter
at mid point of range.

Design for Reliability Series 71 Accendo Reliability

Statistical Tolerance Analysis

The following plot shows the triangle distribution plotted with three different
scale parameters, 1, 2, and 4, with a location parameter of 0 and shape
parameter at mid point of range.

When the mode c is midway between the a and b values, the triangle
distribution is symmetric, as is the orange triangular distribution plot above.

CDF

The triangle distribution CDF is

Z]
]] ^ x - a h2
]] ^b - ah^c - ah for a # x # c
]
^b - x h2
F ^ x h = ]] 1 -
[]
for c # x # b
]] ^b - a h^b - c h
] 0 otherwise
\
Design for Reliability Series 72 Accendo Reliability
Statistical Tolerance Analysis

For a symmetrical triangle distribution the CDF is

]] 2 ` x - a j for a # x # a + b
Z] 2

]] b-a 2
F ^ x h = ]]] 1 - 2 a b - a k for 2 # x # b
[ b-x 2
a+b
]]
0 otherwise
\
The triangle distribution parameters are defined at the endpoints a and
b. These may be the tolerance range or specification, or they may be the
minimum and maximum observed values. They may also represent the
range of the measurements or be estimated directly from a histogram or
PDF plot.

The mode parameter c is the most likely value. If you suspect that the
distribution is centered and symmetrical, then set c to a + b/2. If you suspect
that the distribution is not symmetrical, set c to the most likely value.

Creating Random Numbers from a

Symmetrical Triangle Distribution

To create randomly selected values from a symmetrical triangle distribution

in Excel use the formula

=a+(b-a)*(RAND()+RAND())/2

To create randomly selected values from a nonsymmetrical triangle

distribution in Excel, first set the following cells with values for a, b, and c.

Design for Reliability Series 73 Accendo Reliability

Statistical Tolerance Analysis

In another cell convert c to standard form, c’, with

=(c-a)/(b-a)

Substitute the cell location for the variables a, b, and c.

Create a random number r in a cell with =RAND().

Then the random number selected from a nonsymmetrical triangle

distribution is

= a + ( b - a ) IF( r <= c’, SQRT( r*c’ ),(1 - SQRT(( 1-c’)*( 1 - r )))

Substitute the cell locations for the variables a, b, c’, and r.

Tolerance and Reliability

Ideally, every part has exactly the same dimensions and each product
created fits together in exactly the same way, thus performing exactly
the same (and as expected). Unfortunately, that doesn’t occur because
everything varies.

Products fail because of variation that leads to excessive wear of parts

that are too large, or because the electronics run hotter than expected, or
because connectors fit too loosely, for example. There are many sources
for the variation and, if the design accommodates the expected range of
variation, the product will not suffer adverse effects of the variation that
occurs.

Design for Reliability Series 74 Accendo Reliability

Statistical Tolerance Analysis

One key aspect that reliability engineers should consider is the difficulty in
manufacturing parts within the specified tolerances. If it is likely to have
parts at or beyond these tolerances, premature failure can occur or poor
product performance can result.

By identifying the parts with tight tolerances or low process capability we

can work with the design team or supplier to improve the ability of the parts
to meet the needs of the design. At a minimum, we can at least take steps
to monitor and control the process to create parts that are within tolerance
specifications.

Design for Reliability Series 75 Accendo Reliability

Statistical Tolerance Analysis

Design for Reliability Series 76 Accendo Reliability

Statistical Tolerance Analysis

Design for Reliability Series 77 Accendo Reliability