Computer-Aided Design 70 (2016) 193201

Contents lists available at ScienceDirect

Computer-Aided Design
journal homepage: www.elsevier.com/locate/cad

Statistical geometric computation on tolerances for dimensioning

Songgang Xu , John Keyser
Department of Computer Science and Engineering, Texas A&M University, United States

highlights
A generalized RSS method is proposed for modeling geometric representations of tolerances in the statistical way.
A set of basic operations over the new tolerance model are proposed to enable tolerance compositing and cascading.
A set of examples demonstrate applications of the new model in tolerance estimation.
A tolerance allocation framework based on optimization is also proposed by utilizing analytical forms of the new model.

article info abstract

Keywords: Dimensions are used to specify the distances between different features in geometric models. These di-
Statistical tolerance modeling
mensions will often be expressed as a range of allowable dimensions. When considering a group of tol-
Tolerance analysis
eranced dimensions, these ranges can be analyzed as either a worst-case bound on allowable ranges, or
Tolerance allocation
as a statistical measure of expected distribution. This paper presents a new geometric model for repre-
senting statistically-based tolerance regions. Methods for tolerance estimation and allocation on a geo-
metric model are provided by generalizing root sum square (RSS) methods for compositing and cascading
over tolerance zones. This gives us a geometric interpretation of a statistical analysis. Tolerance regions
are determined by probabilities of variations of dimensions falling into the region. A dependency graph
over dimensions can be represented by a topological graph on which the tolerance cascading and toler-
ance allocation can be processed. To illustrate applications of this geometric method, we provide exam-
ples of tolerance estimation and tolerance allocation on our model. The estimation examples utilize the
compositing and cascading operations provided in the analysis method. The allocation examples present
an automatic tolerance allocation procedure on the tolerance model. As opposed to existing methods,
our allocation method allows us to specify not only a numerical objective of the optimization, but also a
statistically-based objective for the geometric shape of the tolerance.
2015 Elsevier Ltd. All rights reserved.

1. Introduction estimation [2,3]. The statistical variation model relaxes tolerances

The quality of parts manufacturing is determined by both
design and manufacturing tolerances, which affect the geometric
and functional features of finished parts [1]. Engineers are usually
required to select appropriate models to describe tolerances so
that, on one side, tolerances can be estimated with respect to
the design diagram, and on the other side, tolerances can be
allocated to each part of the design diagram to meet the target
tolerance requirement. Traditionally, there are two major types of
tolerance modeling: worst case estimation and statistical variation
by considering distributions of variations so that each tolerance
is associated with a distribution. Tolerance propagation can be
achieved by summing up related tolerances with statistical rules.
A statistical analysis model estimates a distribution describing the
target tolerance rather than computing extreme values described
in the worst case estimation. This paper focus on presenting a new
framework for modeling tolerances by statistical variations as well
as operations that perform tolerance compositing and cascading
on this new model. Since our work is on modeling tolerances, we
demonstrate that our model is not only appropriate for tolerance
estimation [4] but also for tolerance allocation [5].

This paper has been recommended for acceptance by Scott Schaefer and Charlie 1.1. Prior work
C.L. Wang.
Corresponding author. Tolerance analysis covers techniques that compute the vari-
E-mail addresses: sxu@tamu.edu (S. Xu), keyser@cse.tamu.edu (J. Keyser). ations of tolerances for worst case estimation or for statistical
http://dx.doi.org/10.1016/j.cad.2015.06.012
0010-4485/ 2015 Elsevier Ltd. All rights reserved.
194 S. Xu, J. Keyser / Computer-Aided Design 70 (2016) 193201
variation estimation. General reviews on this topic are available
in [6,4,7]. A key part of tolerance modeling is representing the tol-
erance zone in an appropriate way so that variations of tolerances
can be obtained and propagated. Worst case estimation models
the uncertainties of tolerances by simple geometric entities that
are guaranteed to bound variations of dimensions [8,9]. Geomet-
ric objects employed are usually higher dimensional polytopes or
dual-cones, which represent the allowed region as intervals of
coefficients in their algebraic parameterization [1012]. Part fea-
tures, such as form, orientation, and size of tolerances can be
obtained from such models relatively easily, though further com-
putations on tolerances can still be very complicated. Statisti-
cal models consider distributions of variations of tolerances so
that each tolerance is associated with a statistical distribution
model [1315,3,16]. Computations on the statistical model, such as
tolerance compositing, cascading, or allocation, are far more com-
plicated than those of the worst case estimation.
Several prior methods have proposed addressing tolerance
compositing and cascading. Tolerance compositing on statistical
models focuses on describing distributions of tolerances as well as
providing geometric interpretations (tolerance zones) on models.
Tolerance cascading on statistical models studies methods for
propagating tolerances along the dimensioning chain or the
tolerance dependency graph. Tolerance charting methods are the
traditional method for estimating tolerance propagation using
engineers experience. An alternative is computer-aided tolerance
charting [17], which aims to reduce the number of iterations of
physical trial-and-error runs. However, this method cannot handle
complex high dimensional tolerance propagation nor geometric
tolerances. Many methods for modeling more general tolerances
in both the worst case approach and the statistical variation
approach have been proposed [1820,13,2126]. One particular
approach has used Small Displacements Torsor (SDT) [27,28] to
model tolerances [18]. Another approach has used Technologically
and Topologically Related Surfaces (TTRS) to form any part as a tree
representing the succession of surface associations [24,19,21,23].
Statistically, the cascading of tolerances of parts could be simulated
by a Monte Carlo method [20,13].
Several prior methods have also proposed addressing tolerance
allocation. Tolerance allocation is an inverse operation to tolerance
estimation, in which tolerances are allocated to dimensions with
respect to target tolerances as well as design requirements. Chase
et al. [5] present an allocation method based on the root square
method (RSS). Choi et al. [29] also present an optimal tolerance
allocation method on a statistical model. Another allocation
method based on optimization is proposed by Forouraghi [30]
using a special multi-objective particle swarm optimizer. Loof
et al. [31] present an allocation method for linear dimensions by
formulating an analytical cost function and analytical constraints
to minimize the manufacturing cost, which is hard to extend to
handle geometric tolerances. Singh et al. [32] propose an allocation
method built on the T-Map model. However, pure geometric
operations are hard to implement robustly and to associate directly
with probability distribution functions for statistical tolerance
analysis.
Recently, some new methods for modeling tolerances and their
applications have raised research interest. A representative work
is presented by Sahani et al. [33] that describes a systematic
solution for the tolerance stack up problem involving geometric
characteristics on both the worst case model and the statistical
model. Gayton et al. [34] present a method for predicting the defect
probability for all allowable production batches on the statistical
tolerance model. Beaucaire et al. [14] aim to evaluate a predicted
quality for the designer by using the statistical tolerance analysis.
Tsai et al. [16] study non-normal distributions and presented a
method to analyze the resultant tolerance specification. Barkallah
et al. [13] apply statistical analysis methods to study how
manufacturing tolerances should be determined for milling
operations. Qureshi et al. [15] propose a mathematical formulation
of the tolerance analysis integrating quantifier so that a single
description of the geometrical requirement can be obtained.
1.2. Our work
Traditional methods usually use Monte Carlo simulation to
apply statistical rules to tolerance estimations. Though the method
implementation is easy, the simulation time cost is usually very
high, which limits the usage over complicated dimensioning
diagrams. In addition, as the inverse of tolerance estimation, it is
hard to design tolerance allocation methods based on simulations.
In this paper, we present a new geometric model for statistical
tolerances that avoids simulations. Our goal is to decompose the
complicated analytical computation of the conventional statistical
tolerance analysis into a series of simple geometric computations,
without losing the intrinsic statistical meaning. We first use a
multivariate statistical model to represent or approximate the
tolerance distribution for one single part in the system, which is
associated with several dimensions that describe part features. For
a given confidence value, we can determine a geometric region that
describes the tolerance zone. Those regions are usually bounded by
high dimension ellipsoids. We propose a set of operations on the
statistical model so that geometric zones from different parts can
be cascaded and tolerances can be propagated along the tolerance
dependency graph. We also propose an operation to extract the
lower degree of freedom (DOF) information from a high DOF
representation so as to facilitate the tolerance estimation, such as
the clearance estimation of an assembly graph. Since we propose a
new model for statistical tolerance computation, we demonstrate
its applications both on tolerance estimation and on tolerance
allocation. We argue that our new model contributes to the
computation of tolerances in both statistical and geometric ways.
To summarize contributions of this paper:
We generalized the RSS method from the space of one DOF
to the space of high DOFs and proposed complete geometric
representations of tolerances without sacrificing the underlying
statistical meaning.
On our tolerance model, we described a set of operations as a
part of the generalized RSS method to enable tolerance com-
positing and cascading. Those operations have both geometric
and statistical explanations.
We proposed applications for tolerance estimation with
our tolerance model. The estimation utilizes operations on
tolerance zones so that tolerances can be composited and
propagated.
We also proposed applications for tolerance allocation with
our tolerance model, that utilize the analytical representation
of our model and provide an automatic allocation paradigm
based on optimizing cost. As opposed to conventional methods,
our allocation method allows us to specify not only numerical
constraints, such as zone areas, but also geometric constraints,
such as zone principal directions.
This paper is organized as follows. Section 2 explains tolerance
compositing. We build tolerance zones for primitives within
this section. Section 3 explains the generalized RSS method and
a set of operations based on the statistical model. Section 4
proposes applications on our tolerance model, including tolerance
estimation and tolerance allocation.
2. Tolerance compositing
In this section, we use the multivariate normal distribution
(MND) to approximate distributions of tolerances. We also use the
2 distribution to yield geometric representations of tolerances.
 S. Xu, J. Keyser / Computer-Aided Design 70 (2016) 193201
(a) High dimension PDF. (b) Confidence region.
Fig. 1. Multivariate normal distribution. (For interpretation of the references to
color in this figure legend, the reader is referred to the web version of this article.)
2.1. Zone compositing
Assume an object is constrained by n DOFs which formulate a
vector Rn , and assume the correlation between pairs of DOFs
in is defined by a covariance matrix Rnn . We use the mul-
tivariate normal distribution (MND) to formulate the uncertainty
zone of vector . We describe this zone using a probability density
function (PDF) [35] that gives the probability of some variable x for
a given and . The PDF is given by
1 1
P (x; , ) = exp (x )T 1 (x ) .
n
(2 ) 2 | | 2
1
2 be approximated more accurately with a shifted version of the
We write x N (, ) for short. The MND is related to the 2
distribution by
P ((x )T 1 (x ) p2, ) = 1 ,
where p2, indicates the chi-square distribution with p degrees of
freedom, evaluated at . What this states is that x will fall into
a particular region with confidence (1 ). The shape of this
confidence region of the MND will usually be an ellipsoid of high
dimension. The center of the ellipsoid is decided by , and direc-
tions and lengths of the ellipsoid axes are decided by the covariance
matrix .
If given real manufacturing data, we can easily obtain the cor-
relation matrix . Given two sets of random data for two ran-
dom variables, we use the technique in [36] to compute the
correlation coefficient. Correlation coefficients between elements
of a given random vector, which form a correlation matrix, can
be obtained by computing the correlation between random vari-
ables pairwise. For 2D
cases, consider the covariance matrix
2
= 1 1 2
. Twoeigenvalues 1,2 = 21 (12 + 22
1 2 22
4 2 2 2 + 14 222 12 + 24 ), yield axes with length li =
2 1
i p2, , where p = 2 for 2D ellipsoids. Corresponding eigenvec-
tors define axis directions.
Fig. 1(a) shows the PDF of a binormal distribution, a specific case
of the MND. Fig. 1(b) shows the confidence region for confidence
values 0.95 (blue ellipse) and 0.995 (red ellipse). Black dots
are simulation results. The motivation for modeling uncertainty
probabilities of tolerances with the MND is based on several
considerations:
Distributions of tolerances can usually be represented or
approximated by the MND. Some previous work (that we
will build on) presented the RSS method for linear dimension
stacking over normal distribution models [6]. MND, as a
generalization of the normal distribution, can be used to model
tolerances in the space of high DOFs as well as the inter-
correlations between different DOFs.
195
(a) Symmetric zone. (b) Shifted symmetric zone.
Fig. 2. Symmetric zone modeled by normal distribution. (For interpretation of the
references to color in this figure legend, the reader is referred to the web version of
this article.)
(a) Weibull distribution. (b) Approximated distribution.
Fig. 3. Asymmetric zone modeled by normal distribution.
Some asymmetrical distributions can be approximated by
shifted MNDs. The basic MND has a symmetrical form over
tolerance distributions, however it can be easily adjusted (as
we show below) via a shift. Some asymmetric distributions can
MND.
Analytically, the MND has a compact mathematical form for
facilitating the computation over tolerance zones. The confi-
dence region can be naturally viewed as a geometric tolerance
zone if the confidence rate is given. Tolerance cascading can be
achieved easily by summing up two MNDs.
These observations inspired us to use MNDs to model tolerances
in a statistical way and to use confidence regions to model
corresponding tolerance zones. Below, we show how to composite
a tolerance zone for a given confidence rate, which serves as the
base for building tolerance cascading.
2.2. Symmetric and asymmetric tolerance zones
A symmetric tolerance zone can be naturally modeled with an
MND for which the expected value is the nominal value of the
dimension and variances can be obtained from statistical data.
Fig. 2(a) shows the symmetric zone for this condition. The red
dashed line indicates the nominal value of the dimension. The tol-
erance zone is bounded by two black solid lines. As a reminder, the
position of these black lines will be determined by the confidence
rate.
However, we can approximate certain asymmetric distribu-
tions by shifting the nominal value to be different from the
expected value of the distribution. Fig. 2(b) shows the correspond-
ing tolerance zone for such a case. The red dashed line still shows
the nominal value of the dimension, which does not overlap with
the symmetric axis of the distribution. Two black solid lines, not
symmetric to the red line, show the tolerance zone.
We want to use the (inherently symmetric) MND as the basic
model for describing the tolerance zones in this work. When
we need to approximate an asymmetric tolerance distribution
with the symmetric model, we can use a nonlinear regression
to find what shift will produce the closest (shifted) MND to our
asymmetric distribution. We can then use this shifted version in all
further computations. Fig. 3(a) shows a Weibull distribution, which
has been shown to be useful for describing asymmetric tolerance
uncertainties caused by worn-out tools [37]. As shown in Fig. 3(b),
we can approximate this distribution closely with a shifted MND.
196 S. Xu, J. Keyser / Computer-Aided Design 70 (2016) 193201
(a) Tolerance cascading. (b) Tolerance cascading.
Fig. 4. Tolerance propagation using cascading rule. (For interpretation of the
references to color in this figure legend, the reader is referred to the web version of
this article.)
3. Generalized RSS method
The original RSS method is built on the normal distribution for
computing tolerance cascading along a tolerance chain of dimen-
sions in a single degree of freedom (when we refer to linear
below, we mean a single degree of freedom). Cascading of two
tolerance variances 1 and 2 can be represented by
Though the method is simple and straightforward, there are some
limitations.
Since the RSS method works on linear dimensions, tolerance
chains are required to be composed by 1D dimensions. Accord-
ingly, variances and tolerance zones are also linear, which lim-
its the representation of tolerance zones to be only a tolerance
interval rather than a geometric region. Therefore, tolerance
zones provide no geometric meaning, such as principal direc-
tions, shapes, etc.
High DOFs are required to be decomposed before composing the
tolerance chain, which complicates the computation of the de-
pendency graph for tolerance cascading.
The independence between a pair of DOFs is assumed since di-
mensions are decomposed, which ignores the representation of
covariance within DOFs.
There is thus a strong motivation to generalize the RSS method
to overcome all those limitations. Below we decouple our gener-
alized RSS method into several basic operations. We explain each
operation with its assumptions and limitations. By applying those
operations in a proper sequence, we achieve the tolerance com-
positing and cascading on high DOFs in a systematic way. We also
explain how all limitations of the traditional RSS method are over-
come with our new generalized RSS method.
3.1. Cascading
The generalized RSS method is built on the linear combina-
tion of multivariate normal distributions. In detail, suppose we
have two independent sets of random variables (vectors) x
N (x , x ) and y N (y , y ), where x, y Rn . The linear com-
bination z = Mx + Ny + L has the probability distribution z
N (M x + N y + L, M x M T + N y N T ), where M and N Rn Rn ,
and L Rn . We give a short proof for this assertion.
Proof. Since x and y are independent, the joint probability
distribution for x and y is
x x
x 0
N , .
y y 0 y
The linear combination z = Mx + Ny + L turns to be
x
z = M N +L
y
x x MT
0
+ L, M
M N N
y 0 y NT
N (M x + N y + L, M x M T + N y N T ).
This combination has two assumptions. One is that both x and
y use the same bases in the space of DOFs. The other is that both x
and y have the same number of DOFs. With those assumptions we
can cascade tolerances.
Fig. 4 shows two examples of tolerance cascading. Tolerance
zones are composited with the confidence rate 0.95. We assume
all MNDs have the form
1 1 1 2
2
P (x; 1 , 2 , , , ) N
2 2
, . (1)
1 2
2
1 2 22
In Fig. 4(a), the black ellipse is the tolerance zone of distribu-
tion P (x; 1.0, 2.0, 0.2, 0.3, 3/5) and the green ellipse is the toler-
ance zone of distribution P (y; 5.0, 7.0, 0.3, 0.2, 1/5). The blue
ellipse (circle) is the cascaded tolerance zone of x and y. In
Fig. 4(b), the black ellipse is the tolerance zone of distribution
P (z ; 1.0, 2.0, 0.2, 0.3, 3/5) and the green ellipse is the tolerance
12 + 22 .
zone of distribution P (w; 5.0, 7.0, 0.3, 0.2, 3/5). The blue ellipse
(circle) is the cascaded tolerance zone of z and w .
However, the previous two assumptions are not always
satisfied. For most cases, we neither ensure different parts use
the same bases of DOFs nor ensure that they have the same
number of DOFs. When the first assumption is invalid, we use an
Alignment operation to adjust the frame of DOFs; when the second
assumption is invalid, we use Projection to combine part of all DOFs.
3.2. Alignment
The alignment operation is used when two tolerances are not
spanned by the same bases of DOFs. Basically, we will apply a
rotation to one set of bases, and calculate the distribution that
provides the best match for that rotated set of bases. Given the
tolerance chain in Fig. 5(a), the bases for the tolerance formed by
dimension a b is different from those for the tolerance formed
by dimension b c. Cascading these two tolerance zones requires
us to transform bases, which is the essential motivation behind the
alignment operation. Let the tolerance zone for a b follow the
distribution P (x) N (x , x ) and the zone for b c follow the
distribution P (y) N (y , y ). Assume the orientations of bases
for the two tolerance zones are consistent. We need a rotation R
to align the two bases. Without loss of generality, we can view
the bases of P (x) as the static reference. The probability of P (y) is
aligned to be P (y) = (y , RT y R), where R is the rotation matrix
to rotate the bases of P (y) to align with P (x).
We show analignment example in Fig. 5(b). P (y)
1.0 0.2 0
2.0
, 0 0.3
. The green ellipse shows the tolerance zone
with confidence rate
0.99.T We align the vertical axis of the bases
to the direction ( 3, 1) . The rotation matrix turns out to be
0.5 0.866
, which yields the new aligned distribution P (y)
0.866 0.5
1.0
, 00..0433 0.0433
275
2.0 0.225
. The black ellipse shows the new zone
with confidence rate 0.99. This aligned tolerance can be cascaded
with the zone of P (x) using previous cascading techniques.
Since tolerances of a part associated with several DOFs can be
modeled by MND, we can cascade two tolerance zones having
different bases of DOFs, with the help of the alignment operation.
 S. Xu, J. Keyser / Computer-Aided Design 70 (2016) 193201
(a) Tolerance chain. (b) Zone alignment.
Fig. 5. Tolerance alignment. (For interpretation of the references to color in this
figure legend, the reader is referred to the web version of this article.)
(a) Projection. (b) Cascading.
Fig. 6. Tolerance projection. (For interpretation of the references to color in this
figure legend, the reader is referred to the web version of this article.)
Note that avoiding the DOF decomposition is our motivation
for the alignment operation. The alignment operation keeps the
correlation of DOFs so that the tolerance estimation is more
accurate than the traditional method requiring decomposition.
3.3. Projection
When cascading two tolerance zones with different numbers
of DOFs, we provide the projection operation to combine them
together. Let P (x) N (x , x ) have dx DOFs and P (y)
N (y , y ) have dy DOFs, where dx > dy . We assume P (x) and
P (y) use the same set of bases, i.e., the bases of P (y) are a subset
of the bases of P (x). The process to cascade P (x) and P (y) can be
finished in three steps. First, reorder matrix y to be y1 so that
the order of bases used by P (y) is the same as that of P (x). Second,
augment matrix y1 to y2 to have the same number of DOFs as x
so that corresponding rows from two matrices indicate the same
DOFs. Empty spaces are filled with zeros. Last, merge x and y2
using cascading rules. Since this operation projects bases of P (y) to
corresponding bases of P (x), we call it Projection.
Fig. 6 is an example using projection to cascade tolerances of
one DOF and two DOFs. Assume the tolerance of two DOFs has
probability P (x; 1.0, 2.0, 0.2, 0.3, 3/5) and the tolerance of one
DOF has probability P (y) N [(5, 7)T , 0.5]. Note that we will need
to know which of the two bases of x that the basis of y corresponds
to. Fig. 6(a) shows the results of projections onto different DOFs.
The black ellipse is the zone of x. The blue ellipse is the result of
projecting y to the first DOF of x. The green ellipse is the result
of projecting y to the second DOF of x. Fig. 6(b) shows the final
cascading result for the two different interpretations. All those
zones have a 0.95 confidence rate.
197
(a) Sliced curves. (b) Conditional distribution.
(c) Density functions. (d) Confidence interval.
Fig. 7. Slice operation over tolerance zone.
3.4. Slicing
Previous operations allowed us to obtain tolerances repre-
sented by high dimensional ellipsoids. More generally, we would
often like to know the variance on lower DOF measurements, such
as the clearance of an assembly. To do this, we need a slicing oper-
ation to reduce the PDF from higher DOFs to lower DOFs. Fig. 7(a)
shows three cuts of the 2D MND PDF.
Note that the slicing operation is not equivalent to the
conditional distribution of the MND. The conditional distribution
is obtained by fixing some DOFs and considering the distribution of
the remaining DOFs. Though the nominal values of the conditional
distribution are changed according to fixed value points, the
variances are constant. Fig. 7(b) shows the conditional distribution
PDFs for the three cuts in Fig. 7(a). Given a confidence rate, all three
PDFs report confidence intervals having the same width, which is
not consistent with the ellipsoidal zone of the 2D MND. Thus, we
need a different approach than the conditional distribution, which
we call our slicing approach.
We illustrate the PDF compositing of slicing on 2D DOFs, which
can be easily extended to high
DOFs. Consider the covariance
12 1 2
matrix = . Assume we fix the second DOF,
1 2 22
corresponding to 2 . We shrink 1 to be 1 1.0 d/w , where d is
the offset of the second DOF from the nominal center and w is the
maximal offset of the second DOF constrained by the ellipse. For
example, if we are taking a slice halfway from the nominal center
to the boundary of the ellipsoidal region, then d/w = 12 . Using
the new 1 , we compute the conditional distribution and take it as
the distribution
produced by the slicing operation, which yields the
variance 1 d/w1 (1 2 ). Note that the variance diminishes
as we move farther from the nominal point of the MND. Thus,
we know that the width of the confidence interval must also be
shrinking as we move farther away (unlike the pure conditional
distribution). Fig. 7(c) shows the composited PDFs for three cuts.
Fig. 7(d) shows the corresponding three confidence intervals.
Several points are worth noticing. One is that we do not
have to slice in the directions of bases. With the help of the
alignment operation, we can adjust the covariance matrix of the
MND model so that we can slice in any direction, which facilitates
the estimation of target dimensions. Another point is that the slice
198 S. Xu, J. Keyser / Computer-Aided Design 70 (2016) 193201
operation do not have to go through the center, which helps us
estimate when using a shifted MND model. Third, since the sliced
result still follows an MND, it can serve as a part of further tolerance
cascading.
3.5. Shifting
The shifting operation solves the problem of shifted distribu-
tions. We change the nominal value of dimensions having shifted
distributions. In detail, if the nominal value of the distribution is
and the shifting offset is vector , we use + as the new nominal
value for the tolerance cascading. By applying the shifting opera-
tion, we can apply our previous cascading rules onto shifted distri-
butions with almost no extra effort.
3.6. Inflation
Users might want more conservative tolerance estimations
under some conditions. Although this is mainly a hack to allow
users to inflate tolerance zones, it can be used to provide a more
conservative bound that accounts for numerical error in cascading
operations. We allow users to inflate tolerances in one or more
DOFs. The inflation operation on a tolerance zone consists of
multiplying the variance of bases in the covariance matrix of the
distribution by a constant. For example, given a distribution as
formula (1), the inflation on base one is P (x; 1 , 2 , s1 12 , 22 , ),
where s1 > 1 is a constant. We can also inflate two bases by two
constants s1 > 1 and s2 > 1 as P (x; 1 , 2 , s1 12 , s2 22 , ). Using
the same confidence rate, the new tolerance zone will be larger
than the original one, therefore the tolerance estimation becomes
more conservative than before.
To conclude this section, we revisit the limitations of the
traditional RSS method mentioned previously. The limitation of
numbers of DOFs has been naturally overcome by the generalized
version, since the new method is built on high DOFs using the
MND model. DOFs associated with one part do not need to be
decomposed before applying cascading operations. Alignment and
projection can be viewed as auxiliary operations for preparing
DOFs before cascading. Slicing provides the tolerance estimation
of lower DOFs. Last, the new method allows us to represent
the correlation among DOFs by covariance matrices. All those
properties imply that the generalized RSS provides a flexible
framework for tolerance analysis on a statistical model.
4. Applications
We have presented the generalized RSS method for estimating
tolerance propagation along the tolerance chain. In this section, we
illustrate the procedure to apply the generalized RSS method to
both tolerance estimation and tolerance allocation.
4.1. Tolerance estimation
Before estimating tolerance zones of general parts, we need a
directed acyclic graph (DAG) of dimension dependencies indicating
the relative dependencies of individual parts. Each directed edge
indicates a dependency from the reference part to the target part,
where the reference one is viewed to be static to the target one.
As opposed to the traditional graph of the tolerance chain, we
allow nodes of the DAG to be geometric primitives, such as points,
segments, or general parts having complicated geometric shapes,
which facilitates the formulation of DAG quite a bit. Tolerances
of primitives can be achieved using operations in the generalized
RSS method. And tolerances over general parts can be formulated
with related dimensions. We illustrate this idea with an example
in Fig. 8.
(a) Scheme A. (b) Scheme B.
Fig. 8. Tolerance compositing for boxes.
Fig. 9. Segment tolerance model. (For interpretation of the references to color in
this figure legend, the reader is referred to the web version of this article.)
When compositing the tolerance for a general part (such as a
box), we must notice that different types of dimension schemas
yield different tolerance zones. Fig. 8 gives two dimension schemas
for the box. The left one has two dimensions x and y. Therefore, ev-
ery point on the whole box has the same covariance matrix, which
yields tolerance zones with the same size and shape. Centers of
those zones are coordinates of corresponding points. The right fig-
ure has four dimensions. The four corners of the box are defined
as (x1 , y1 ), (x1 , y2 ), (x2 , y1 ) and (x2 , y2 ). In other words, the four
corners have their own tolerance zones. For points on edges, we
use the tolerance model for the segment to describe the tolerance,
as shown in Fig. 9. Tolerances of points on the segment can be
obtained by interpolating tolerances at the two ends. Assume the
zones of q1 and q2 have been given. The zones of p3 and p1 are also
represented by the MND model, with the interpolated mean values
and variances of q1 and q2 . We assume p2 depends on p1 of which
the relative tolerance zone (i.e. assuming p1 is fixed) is colored in
green. We can use the cascading technique to propagate the toler-
ance from p1 to p2 . The cascaded region is colored in pink.
Fig. 9 also shows a good example for handling geometric
tolerances such as angularity and perpendicularity. The variation
zone of p2 can be spanned by two directions that are along
(denoted by 1r) and orthogonal to (denoted by 1 ) segment p1 p2
respectively. In geometry, variance of , the angle between two
segments, actually indicates the angular variation of segment p1 p2 .
Therefore, the configuration of angularity and perpendicular are
adaptive to our tolerance model. Other linear features could be
handled in a similar way.
Next, we apply our method to assemblies of some simple
geometric parts. Fig. 10(a) shows an example of tolerance
cascading. The zone of the box below is colored in green while
the zone of the box above is colored in blue. Since the box above
depends on the box below, we propagate the green tolerance to the
blue one and obtain the cascaded pink zone. Note that each box is
related to two orthogonal dimensions that do not correspond in
this case to the dimensions of x and y. Therefore, all points on each
box share the same tolerance zones. Variances of d1 and d2 can
be obtained by applying the slicing operation in two orthogonal
directions.
 S. Xu, J. Keyser / Computer-Aided Design 70 (2016) 193201 199

(a) Box cascading. (b) Cascaded zone.
Fig. 10. Cascaded zones for boxes. (For interpretation of the references to color in
this figure legend, the reader is referred to the web version of this article.)
Fig. 11 shows an example of axis alignment in tolerance
cascading. The box and the triangle have different frames of DOFs
in Fig. 11(a). The axis alignment and the tolerance cascading
results are shown in Fig. 11(b). The orange zone, which shows
the zone of the box, has a frame which is not aligned with the
global coordinate system. By applying the alignment operation, the
covariance matrix is changed so that the frame is rotated to be
aligned with the global system. The green zone, which shows the
zone of the triangle, has a frame aligned with the global system.
Therefore, we can cascade the green zone and the aligned orange
zone. The result zone is colored in pink. The variance of dimension d
can be obtained by applying the slicing operation on the pink zone.
For a part of a general shape, such as the part in Fig. 12, we
compute its tolerance zone with respect to related dimensions.
We will illustrate by briefly describing how we would use the
operations we have described to estimate the height of the circle
center P. We first estimate tolerance zones of points A, B, C and D.
The four points are related with dimensions x1 , y1 and z1 . We use
the compositing technique to build four 3D ellipsoids indicating
tolerance zones for each point. Next we estimate tolerance zones
of points E , F , G and H, which are related to dimensions x1 , y3 , and
z1 . We know four 3D ellipsoids describing their tolerance zones.
We propagate tolerances from A, B, C and D to E , F , G and H using
the cascading technique and get final variation zones of E , F , G and
H. We relate point P with Q , the middle point of edge GH. The
tolerance zone of point Q is the interpolation of the zones of G and

(a) Point cascading. (b) Point cascading.

Fig. 11. Cascaded tolerance zone for the box over a slope.

(a) Assembly. (b) Separated parts.

(c) Front view. (d) Right view.

Fig. 12. Tolerance estimation for real parts.

200 S. Xu, J. Keyser / Computer-Aided Design 70 (2016) 193201

H, which is also a 3D ellipsoid. We propagate this tolerance to P via

dimension y4 using the projection operation, since now we have
mismatched bases of DOFs of tolerances. We have a 3D ellipsoid to
describe the variation of point P. Finally the slicing operation yields
the variation of P in the vertical direction.
Note that we do not consider over-constrained dimensions
within our current work, which means we can always find unique
related dimensions to composite the tolerance zone of a given
point. We know that the tolerance of one vertex is decided by two
dimensions within a 2D graph and is decided by three dimensions
with 3D cases. Therefore, our method can be generalized to cases
of high DOFs easily.

4.2. Tolerance allocation

Assume that we have two stacked objects. Previous techniques Fig. 13. Tolerance allocation. (For interpretation of the references to color in this
provide methods to propagate the tolerance from one object to the figure legend, the reader is referred to the web version of this article.)
other. This process is tolerance estimation. Tolerance allocation is
the reverse process. Given the cascaded zone, we need a method
tB is l3 l4 = 3 4 p2, . The optimization problem becomes
to decompose the zone to two zones for two objects respectively.
There are infinite solution configurations for the allocation without max( 1 2 2
p, + 3 4 p2, ), subject to the constraint A +
any constraints, so people search for allocation schemes that B = . This nonlinear optimization can be further reduced to be
optimize certain merits. Chase et al. [5] presented basic guidelines
max (21 22 + 23 24 )
for tolerance allocation by analytical methods. The basic idea is
to find an applicable tolerance configuration within the whole st. A + B = , (3)
possible tolerance space. Preferences for some tolerances are
differentiated by weighted allocations. However, this scheme is which can be solved with a numerical nonlinear optimization
limited by the designers experience as well as by the number of solver.
DOFs of tolerance constraints, since only numerical ranges of the Fig. 13 showsa numerical
example for one level 2D allocation.
1.0 0.8
tolerance are allowed to be constrained. We have no chance to We have = 0.8 2.0
, A = 0.5 and B = 0.7. Applying
specify geometric information, such as zone shapes or principal
directions. In contrast, our method is automatic so that no designer
the
optimization w1 = w2 = 1, we got A =
with weights
0.11 0.079 0.89 0.88
experience is required. Meanwhile, we allow designers to specify 0.079 0.22
and B = 0.88 1.78 . The black ellipse is the
geometric information over tolerance constraints. area zone of , blue is for A and green is for B .
We associate the tolerance zone with a scalar measurement, We revisit Fig. 12 for illustrating the allocation technique over
such as the area or the manufacturing cost, so that a cost real parts. To simplify the analytical form in discussion, we only
function taking the scalar measurement as the parameter can be consider the tolerance within the plane formed by points C , G, D
formulated. Solving the problem by optimizing the cost function and H. We study how a prescribed tolerance of P can be allocated to
by numeric methods yields the optimal allocation scheme. We those four points. Tolerance dependencies within the plane include
assume that we want to maximize weighted areas of tolerance C G, D H, G and H Q , and Q P. Therefore, the variance
zones for each part on the tolerance chain, which formulates the of P can be obtained using estimation techniques as
objective function of the optimization. The reason to choose this
merit cost is that a bigger tolerance zone implies more flexible 1

P = (C + G + D + H ) + P .
requirements on designing and processing. Meanwhile, we allow 2
users to add preferences over different tolerances by adding
weights. The optimization over the tolerance allocation becomes Note that P is the final tolerance of P which is formed by
tolerances propagated from other parts. Similarly, we use absolute
max wi Area(ti ). values of singularity values to estimate the area of tolerances
zones. For symmetry in this example, we require that C and D have
st. t0 = T0 (2)
the same tolerance zones, and G and H have the same tolerance
where the ti are tolerance zones related to each node on the zones too. By specifying the same weights to all zones, we yield
tolerance chain. t0 is the given tolerance zone specified by the optimization system as below.
designers with the geometric value T0 .
We illustrate this problem on an example of 2D tolerance zones. max (2C1 2C2 + 2D1 2D2 + 2G1 2G2 + 2H1 2H2 + 2P1 2P2 )
The tolerance allocation problem becomes how to decompose the P = and C = D and G = H ,
st. (4)
given covariance matrix so that A + B = while Area(tA ) +
Area(tB ) is maximized. encapsulates all geometric information where , distinguished by point names, indicates singularity values
of the constrained zone. for the covariance
12 A 1 2 matrix of each point.

Consider the tolerance variance A = of the 2.0 0.6
A 1 2 22 Let = 0.6 3.0
and let C = 0.3, D = 0.3, G =
reference.
are 1,2 =
Absolute values of two singularity values
0.2, H = 0.2 and P = 0.15. By plugging-in those values into
|| 21 12 + 22 422 12 A2 + 14 222 12 + 24 ||. The area of formula (4) and optimizing the system,
we have thetolerance zone
1.008 0.375
for each point as below. C = D = , G = H =
0.375 1.550

the tolerance zone is l1 l2 , where li = i p2, . The area of 0.729 0.181 0.262 0.044

0.181 1.122
, and P = 0.044 0.372
.
tA is l1 l2 = 1 2 p2, . Similarly, we assume the area of
 S. Xu, J. Keyser / Computer-Aided Design 70 (2016) 193201
5. Discussion and conclusion
We have presented a new generalized geometric method for
statistical tolerance representation, propagation and allocation by
generalizing the conventional RSS model. Our method includes
a set of operations on the statistical model so that tolerance
cascading can be achieved without losing the intrinsic statistical
interpretation of the tolerance zone. Compared with conventional
tolerance methods, this method provides both geometric and
statistical representations of tolerance zones, including boundary,
orientation and area information. To justify the adaptivity and
flexibility of our model, we presented examples of both tolerance
estimation and tolerance allocation that utilize our tolerance
model. Results of those applications demonstrate that our method
provides a complete and systematic procedure for tolerance
related computation in a statistical framework. Because of its
novelty in modeling the tolerance, our method brings several
advantages to tolerance analysis.
Our tolerance model provides a geometric explanation of
high DOFs of tolerances. We use high dimension ellipsoids to
represent tolerance zones so that our method is not limited by
the number of DOFs within the tolerance space.
We provide operations to increase the number of DOFs
(projection) or reduce the number of DOFs (slicing) in geometric
space so that tolerances within different spaces of DOFs can be
cascaded in geometric and statistical ways.
Since our model has both geometric and statistical explana-
tions, it is not only a powerful tool for tolerance estimation
but also provides full analytical structures for automatic toler-
ance allocation. In this paper, we proposed a primitive alloca-
tion method based on optimizing the merit cost, which could
be a very interesting research topic in the future.
Our tolerance model will support most of the tolerance desig-
nations supported in the ASME Y14.5 standard. This includes
non-linear tolerances such as those related to angularity and per-
pendicularity, as we described earlier. However, some forms of tol-
erance, such as a single-variable measure of circularity, are not
directly supported. Our initial analysis shows that our approach
should be extensible to such measures, but it may require the addi-
tion of further techniques for transforming between different sets
of Degrees of Freedom.
There are some limitations of our method which provide
potential for future research. First, we use the MND to model
tolerances, which can only approximate many varieties of real
distributions. One area for our future work is to incorporate other
distributions of tolerances and make them work seamlessly with
the current tolerance framework. Second, tolerance zones are
described by ellipsoids so that direct geometric computation on
the geometric shape cannot be applied easily. We are forced to
use analytical forms to build operations on the resulting geometric
shapes. A fully direct geometric method based on statistical
distributions remains an elusive goal in tolerance studies.
References
[1] Standard A. Dimensioning and tolerancing, ASME Y14.5M. New York:
American Society of Mechanical Engineering; 1994.
[2] Fischer BR. Mechanical tolerance stackup and analysis. CRC Press; 2011.
[3] Ryan TP. Statistical methods for quality improvement. John Wiley & Sons;
2011.
[4] Hong Y, Chang T. A comprehensive review of tolerancing research. Int J Prod
Res 2002;40(11):242559.
201
[5] Chase KW. Tolerance allocation methods for designers.
[6] Scholz F. Tolerance stack analysis methods. In: Research and technology
boeing information & support services, boeing, seattle. 1995. p. 144.
[7] Shen Z, Ameta G, Shah JJ, Davidson JK, et al. A comparative study of tolerance
analysis methods. Trans ASME-S-Comput Inf Sci Eng 2005;(3):24756.
[8] Akella S, Mason MT. Orienting toleranced polygonal parts. Int J Robot Res 2000;
19(12):114770.
[9] Brost RC, Peters RR. Automatic design of 3-d fixtures and assembly pallets.
In: Robotics and automation, 1996. Proceedings, 1996 IEEE international
conference on, Vol. 1. IEEE; 1996. p. 495502.
[10] Roy U, Li B. Representation and interpretation of geometric tolerances for
polyhedral objects. II.: Size, orientation and position tolerances. Comput-
Aided Des 1999;31(4):27385.
[11] Mujezinovi A, Davidson J, Shah J. A new mathematical model for geometric
tolerances as applied to polygonal faces. J Mech Des 2004;126:504.
[12] Davidson J, Mujezinovi A, Shah J. A new mathematical model for geometric
tolerances as applied to round faces. J Mech Des 2002;124:609.
[13] Barkallah M, Louati J, Haddar M. Evaluation of manufacturing tolerance using
a statistical method and experimentation. Int J Simul Model 2012;11(1):516.
[14] Beaucaire P, Gayton N, Duc E, Dantan J-Y. Statistical tolerance analysis of over-
constrained mechanisms with gaps using system reliability methods. Comput-
Aided Des 2013;45(12):154755.
[15] Qureshi AJ, Dantan J-Y, Sabri V, Beaucaire P, Gayton N. A statistical tolerance
analysis approach for over-constrained mechanism based on optimization and
monte carlo simulation. Comput-Aided Des 2012;44(2):13242.
[16] Tsai J-C, Kuo C-H. A novel statistical tolerance analysis method for assembled
parts. Int J Prod Res 2012;50(12):3498513.
[17] Lehtihet E, Ranade S, Dewan P. Comparative evaluation of tolerance control
chart models. Int J Prod Res 2000;38(7):153956.
[18] Villeneuve F, Legoff O, Landon Y. Tolerancing for manufacturing: a three-
dimensional model. Int J Prod Res 2001;39(8):162548.
[19] Desrochers A. A cad/cam representation model applied to tolerance transfer
methods. J. Mech. Des. 2003;125:14.
[20] Huang SH, Liu Q, Musa R. Tolerance-based process plan evaluation using
monte carlo simulation. Int J Prod Res 2004;42(23):487191.
[21] Ngoi B, Lim L, Ang P, Ong A. Assembly tolerance stack analysis for geometric
characteristics in form controlthe catena method. Int J Adv Manuf Technol
1999;15(4):2928.
[22] Mansuy M, Giordano M, Hernandez P. A new calculation method for the worst
case tolerance analysis and synthesis in stack-type assemblies. Comput-Aided
Des 2011;43(9):111825.
[23] Ambu R. A method with a statistical approach for the evaluation of tolerance
chains. Adv Mater Res 2013;651:6016.
[24] Ameta G, Serge S, Giordano M. Comparison of spatial math models for
tolerance analysis: tolerance-maps, deviation domain, and TTRS. J Comput Inf
Sci Eng 2011;11(2):021004.
[25] Li H, Zhu H, Li P, He F. Tolerance analysis of mechanical assemblies based on
small displacement torsor and deviation propagation theories. Int J Adv Manuf
Technol 2014;72(14):8999.
[26] Zhang J, Qiao L. Three dimensional manufacturing tolerance design using
convex sets. Proc CIRP 2013;10:25966.
[27] Clment A, Bourdet P. A study of optimal-criteria identification based on
the small-displacement screw model. CIRP AnnManuf Technol 1988;37(1):
5036.
[28] Ghie W, Laperrire L, Desrochers A. Statistical tolerance analysis using the
unified JacobianTorsor model. Int J Prod Res 2010;48(15):460930.
[29] Choi H-GR, Park M-H, Salisbury E. Optimal tolerance allocation with loss
functions. J Manuf Sci Eng 2000;122(3):52935.
[30] Forouraghi B. Optimal tolerance allocation using a multiobjective particle
swarm optimizer. Int J Adv Manuf Technol 2009;44(78):71024.
[31] Lf J, Hermansson T, Sderberg R. An efficient solution to the discrete least-
cost tolerance allocation problem with general loss functions. In: Models for
computer aided tolerancing in design and manufacturing. Springer; 2007.
p. 11524.
[32] Singh G, Ameta G, Davidson JK, Shah JJ. Tolerance analysis and allocation for
design of a self-aligning coupling assembly using tolerance-maps. J Mech Des
2013;135(3):031005.
[33] Sahani A, Jain P, Sharma SC, Bajpai J. Design verification through tolerance
stack up analysis of mechanical assembly and least cost tolerance allocation.
Proc. Mater. Sci. 2014;6:28495.
[34] Gayton N, Beaucaire P, Duc E, Lemaire M. The apta method for the tolerance
analysis of productscomparison of capability-based tolerance and inertial
tolerance. KMUTNB: Int J Appl Sci Technol 2013;4(3):2537.
[35] Eaton ML, Eaton M. Multivariate statistics: a vector space approach. Vol. 198.
New York: Wiley; 1983.
[36] Watkins JC. Correlation and regression. In: An introduction to the science
of statistics: from theory to implementation. University of Arizona; 2015.
p. 3362.
[37] Pearn W, Wu C-W, Wang K. Capability measure for asymmetric tolerance non-
normal processes applied to speaker driver manufacturing. Int J Adv Manuf
Technol 2005;25(56):50615.