Академический Документы
Профессиональный Документы
Культура Документы
August 30th
_____________ 03
, 20 _____
Rami A. Musa
I,______________________________________________,
hereby submit this as part of the requirements for the
degree of:
Master's of Science
________________________________________________
in:
Industrial Engineering
________________________________________________
It is entitled:
Simulation-Based Tolerance Stackup
________________________________________________
Analysis for Machining
________________________________________________
________________________________________________
________________________________________________
Approved by:
________________________
Dr. Samuel Huang (Chair)
________________________
Dr. Sam Anand
________________________
Dr. Richard Shell
________________________
________________________
Simulation-Based Tolerance Stackup Analysis in Machining
MASTER OF SCIENCE
In Industrial Engineering
August, 2003
by
Rami A. Musa
Dimensional and geometric tolerance can result from either process variation and/or process
is interrelated with tolerance control, tolerance allocation and setup planning. During
machining operations of a part, tolerance stackup is inevitable most of the time. Therefore,
tolerance stackup must be studied accurately and efficiently. In spite of this, traditional
methods for analyzing stackup (statistical and worst-case methods) have some drawbacks that
reduce their accuracies. These drawbacks are discussed in details. This study presents a novel
and inspection process using Monte Carlo simulation along with major manufacturing errors.
It overcomes the argued drawbacks in the traditional methods. Further, it is proved that both
the statistical and worst-case methods are conservative compared to the proposed one.
chances to accept process plans that are usually precluded using the traditional ones. Three
illustrative examples are presented to compare the results of the simulation with the
traditional methods.
I
II
Acknowledgments
First of all, I wish to offer my sincerest gratitude to my advisor; Dr. Samuel Huang who was
an outstanding advisor in all measures during my work with him. His professionalism,
True thanks to Dr. Sam Anand and Dr. Richard Shell for serving as committee members in
my thesis defense, words of encouragement and appraising my effort. Also, I would love to
thank and recognize my friends: Mohammad Hamdan, Mohammad Younis and Zain Dewaik,
who introduced and encouraged me all the way to go for my graduate study. Also, I would
like to extend my thanks to my colleague and friend Anshum Jain who contributed
my family for their encouragement and support. I always felt I am the luckiest person in the
world to have such a family; my late father, my loving mother and my brothers: Naji and
Husam.
This work has been gratefully sponsored by the National Science Foundation and thankfully
III
Contents
1. INTRODUCTION 6
2. LITERATURE REVIEW 14
1
3.2.1 DATUM EVALUATION 36
5. ILLUSTRATIVE EXAMPLES 58
6.1 SUMMARY 68
BIBLIOGRAPHY 71
2
List of Figures
Figure 1-1: One way clutch mechanism [Chase, Gao and Magleby (1994)].............................8
Figure 1-2: One way clutch mechanism vector loop [Chase, Gao and Magleby (1994)] .........8
Figure 1-4: 2-D Closed vector loop for one way clutch mechanism [Chase, Gao and Magleby
(1991)]..............................................................................................................................10
Figure 1-5: 3-D Closed vector loop for crank slider mechanism [Chase, Gao and Magleby
(1991)]..............................................................................................................................10
Figure 2-1: With Cp=1, only 2700 part per million (PPM) defects are expected ....................17
http://www.ymp.gov/documents/ser_b/figures/chap4_2/f04-174.htm)...........................30
3
Figure 3-10: Non-rejected datum [Wilhelm (1998)] ...............................................................40
Figure 3-11: Sample points located in a square feature using different approaches ...............45
Figure 3-12: Sample point locations using random and low-discrepancy methods [Davis and
Martin (1998)]..................................................................................................................45
Figure 3-13: Benchmarked results at 1 billion iterations [Cvetko, Chase and Magleby (1998)]
..........................................................................................................................................47
Figure 5-1: Example 1 (design requirements and the machining sequence) ...........................58
Figure 5-4: Example 2 output (The simulation output for the distributions of the distances
between surfaces).............................................................................................................61
http://www.wzl.rwth-aachen.de/WM/SIMON/deliverables/DA0/DA0_02D.htm) .........62
Figure 5-11: Rejection areas comparison when allocating concluding links tolerance using
4
List of Tables
Table 3-3: Recommended sample size for different geometries [Henzold (1995)].................44
Table 4-4: Variance comparison between simulation and experiment for smooth part ..........57
Table 4-5: Variance comparison between simulation and experiment for a rough part..........57
Table 5-5: Part per million (PPM) rejections comparison when allocating tolerance using
5
1. Introduction
Tolerance is a common arguing point between design and manufacturing. Design engineer
tends to tighten the tolerance to meet functional requirements whereas production engineer
tends to loosen (relax) it to satisfy resource availability. Nevertheless, the most important
factor to be considered is the cost. Cost increases hysterically by tightening the tolerance.
machining, raw part, fixture unit and measurement machine, it has to be compromised by
different departments in the companies. Obviously, tolerance problem is kind of promoter for
One serious problem in process planning is that some good plans (plans that lead to design
requirement satisfaction) could be rejected and some bad plans could be accepted due to
defined as the accumulation (or stackup) of errors when machining a part using different
operational datum than the ones specified in the blueprints. The two traditional methods used
nowadays to analyze tolerance stackup in machining are: worst-case and statistical methods.
These methods are believed to have major drawbacks that reduce the accuracy of tolerance
6
3. Tolerance stack between features is preformed in one dimension; which does not
represent the actual three-dimensional features of interest. 3-D simulation must be the
driving force behind the entire dimensional management process [Craig (1996)].
Additionally, we found that both of the traditional methods evaluate tolerance stackup
conservatively. In this work we developed a more accurate method for evaluating tolerance
stackup in machining that can lead to more cost-effective (less conservative) and/or less
machining and inspection processes along with major manufacturing errors using Monte
Carlo simulation. It will be shown in chapter 5 (illustrative example 3) that using our method
for stackup evaluation will result in much less rejects expectations per million parts compared
Some research works have been done in assembly tolerance stackup using Monte Carlo
simulation in the literature. Although this seems quite close to our work here in tolerance
stackup for machining, there are exclusive differences between the two problems, their
formulations and applications. Component (part) and assembly designs are the two major
tasks in any design department. Component design provides a single component drawing that
components are: shaft, gear, pulley, etc. However, it is unlikely to have a component
functioning alone as there is a need to assemble it with other components. Assembly design
7
studies the suitability of two or more components to meet machine functions. When
assembling parts together, there should be some manufacturing variation in the part that will
cause assembly tolerance stackup. Figures 1-1 and 1-2 show an example of an assembled one
way-clutch mechanism. The mechanism consists of: four rollers, a hub, four springs and an
outer ring. The objective of the tolerance analysis here is to study the effect of manufacturing
This problem has been studied extensively by Chase, Magleby and Gao in Brigham Young
University. They developed computer software (CATS) that applies their methods in
assembly tolerance analysis. Another system has been developed by Variation System
Analysis (VSA).
The first step in evaluating assembly tolerance stackup in the literature is to find an explicit
function of the dimension (tolerance) to be controlled in terms of the other components using
8
trigonometric functions. The following are the explicit functions of the assembly dimensions
for the mechanism shown in figure 1-2 [Gao, Chase and Magleby (1997)]:
a+c
Φ1 = cos −1 ( ) (1.1)
e−c
b = (e − c ) 2 + ( a + c ) 2 (1.2)
Some assembly tolerance stackup methods in the literature that assumes the availability of
3. Quadrature,
5. Reliability index,
6. Taguchi method.
Normally, it is very hard or even impossible to get explicit assembly equations for a typical
dimensions in an assembly that can be used to find a set of implicit assembly equations.
Figures 1-3, 1-5 and 1-6 show examples of closed vector loops for 1-D, 2-D and 3-D
mechanisms.
9
Figure 1-3: 1-D assembly Figure 1-4: 2-D Closed vector Figure 1-5: 3-D Closed vector
mechanism [Law (1995)] loop for one way clutch loop for crank slider mechanism
mechanism [Chase, Gao and [Chase, Gao and Magleby (1991)]
Magleby (1991)]
The following are the governing assembly equations for the closed loop one-way-clutch
hx = 0 = b + c sin(Φ1 ) − e sin(Φ1 )
h y = 0 = a + c + c cos(Φ1 ) − e cos(Φ1 ) (1.3)
hθ = 0 = 90 − 90 + 90 − Φ1 − 180 + Φ 2 + 90 = −Φ1 + Φ 2
From the third equation in the previous set of equations (1.3), it can be seen that Φ1 = Φ 2 = Φ .
This reduces the equations into two as follows [Gao, Chase and Magleby (1997)]:
hx = 0 = b + c sin(Φ ) − e sin(Φ )
(1.4)
h y = 0 = a + c + c cos(Φ ) − e cos(Φ )
It is apparent that it is very difficult to convert these equations into explicit form. The main
two methods available in the literature to solve this problem for implicit assembly functions
are: Direct Linearization Method (DLM) and Monte Carlo Simulation. First order Taylor
10
series linearizes the assembly constraints in DLM to have a set of linear simultaneous
equations and then linear algebra is used to solve them. Afterwards, assembly tolerance
stackup are estimated using statistical or worst case methods. Monte Carlo simulation method
includes the following steps: (1) generate random variates for each variable in the assembly
constraints (2) Select appropriate nonlinear solvers to solve the constraints (3) fit the output
numbers with a distribution and get its parameters (first four moments: mean, variance,
skewness and kurtosis.) Chase, Gao and Magleby use Crystal Ball software to solve the
problem. Crystal Ball is spreadsheet Monte Carlo Simulation software that can solve implicit
Gao, Chase and Magleby (1995) made a comparison between the two methods. It turned out
that the concern regarding the DLM is the accuracy and the concern regarding the Monte
Carlo simulation is the huge number of iterations needed to solve the problem.
Noteworthy, the following are the differences between using Monte Carlo simulation for
machining tolerance stackup analysis [Musa and Huang (2003)] and assembly tolerance
manufacturing variations. The two analyses are close in the sense that we are trying to
(2) Objective. The objective of assembly tolerance analysis is to assign tolerances for all
11
goodness of a process plan and/or assign proper contributing link tolerances
(3) Sequence. Assembly tolerance analysis comes after machining tolerance stackup
analysis.
Improving quality and reducing cycle time and cost are the main objectives for competitive
manufacturing these days. In other words, achieving minimum tolerance possible using the
available resources, reducing trial and error procedures and taking economical issues into
consideration can lead to the ideal process which all industries aim at (figure 1-6). These
Tolerance control involves controlling the tolerance stackup via proper choices of processes,
The objective of this study is to present a novel, less conservative and more accurate
evaluation method of tolerance stackup compared to the existed analytical ones (worst case
and statistical methods) in the literature. This method is based on simulating machining and
inspection process using Monte Carlo simulation along with major manufacturing errors.
12
IDEAL PROCESS
ITERATIVE
TIGHTEST
DESIGN/ ECONOMICALLY
TOLERANCE
MANUFACTURE FEASIBLE
POSSIBLE
AVOIDANCE
The thesis is divided into six chapters. It starts in chapter 1 with the introduction that explores
background of the problem, motivation and objectives of the work. Then, chapter 2 reviews
some basic concepts and terms that are commonly used in later chapters and discusses the
problem of tolerance stackup by defining it, presents the traditional analytical methods
available in the literature and discusses tolerance chart method. Monte Carlo simulation
manufacturing error categories and models, sample plan and stopping criteria for the
for evaluating: machining, fixture unit and raw part errors are outlined in chapter 4. In
chapter 5, three illustrative examples are demonstrated and solved using the proposed method
and comparisons are made between the traditional methods and the proposed one. Finally,
13
2. Literature Review
The following are some commonly-used terms and concepts in this thesis:
Feature: Any surface in the machined part (e.g. hole, slot, boss, tab).
Tolerance: The permissible amount of variability in geometry. Limit of size and plus-minus
tolerances are two methods used to specify tolerances. Limit of size means that an upper and
lower limit are given for a specific dimension. As for plus-minus tolerance, a nominal (target
Fixture unit: A unit that is used to constrain the workpiece from movement during machining.
Size tolerancing (coordinate dimensioning and tolerancing) used to be the only approach for
dimensioning and tolerancing. In this approach, the dimension and its tolerance are
represented by the distance and its variation between two features or points. Although this
approach was found to be successful for many design cases, there were major shortcomings.
These shortcomings showed up because of the increased demand and need for high quality
2. It does not represent the tolerance zones properly in some cases. An example is that
14
3. The functional requirements (e.g. assembly) in manufacturing the part are not valid.
In order to remedy these shortcomings when using coordinate tolerancing, long written
Dimensioning and Tolerancing (GD&T; ASME Y14.5M-1994) can be used and can solve all
1. Obtaining clear instructions for inspection and manufacturing (by using the datum
concept).
Geometric tolerances include fourteen types of tolerances that are usually categorized into
five categories; namely: form, orientation, profile, location and runout. Form tolerances
include: parallelism, angularity and perpendicularity. Profile includes: profile of a line and
profile of a surface. Runout tolerances include: circular runout and total runout. Finally,
location tolerances include: position, symmetry and concentricity. They can be further
classified into datum-dependent and datum-independent tolerances. Table 2-1 depicts all the
15
Table 2-1: Geometric Tolerances (ASME Y14.5M-1994)
Process Capability: The process is considered capable if the process variability is equal or
less than the design specification (tolerance). Usually, it is represented by the Cp index which
T USL − LSL
Cp = = (2.1)
6σ 6σ
The USL and LSL are the upper and lower specification limits. Referring to figure 2-1,
considering the design tolerance equals to 6σ (Cp=1) implies that we are satisfied with about
2700 PPM rejects. Nevertheless, Cp index assumes that the process does not drift from the
mean (refer to figure 2-2). Six sigma quality strategy (developed by Motorola in 1980s)
16
assumes ±1.5σ as a typical mean drift. Another index called Cpk considers this drift. It is
defined as:
USL − µ µ − LSL
C pk = min{ , } (2.2)
3σ 3σ
Figure 2-1: With Cp=1, only 2700 part per million (PPM) defects are expected
17
2.2 Tolerance Stackup; Definition and Applications
In general, tolerance results from both process tolerance and the tolerance stackup
[Whybrew and Britton (1997)]. The latter is the accumulation (buildup) of error (tolerance) in
a dimension between features resulting from taking operational datums that are different from
the ones indicated in the design specifications. In other words, if the datum indicated in the
design drawings is the one used for locating and clamping, then a stackup-free dimension will
result and there will be no tolerance stackup in this specific dimension. Consequently,
tolerance analysis and tolerance control will not be necessary since the tolerance will depend
solely on the process capability [Huang (1995)]. However, in practice, due to economic
reasons and resource constraints, design datums are not always used as locating and clamping
datums. Therefore, some of the blueprint dimensions will be machined indirectly. Hence, in
The way of machining a part determines the stackup in a dimension. There are three main
surface is used as a datum to machine the next surface. This will result in the greatest
accumulation of tolerance.
2. Base-line: This is how parts are machined in a single setup using NC machines. In this
approach, the operational datum is fixed (zeroed) by the coordinate system in the NC
machine for each machining cut. Using this approach decreases the tolerance stackup.
3. Mixed of chain machining and base-line: This happens when parts are machined in
multiple setups.
18
Tolerance allocation is a crucial step in setup planning. Figure 2-3 shows a typical approach
for designing a setup plan. It can be seen that tolerance stackup analysis and tolerance
allocation play important roles in setup planning. Thus, tolerance stackup behavior needs to
be studied carefully and analyzed accurately in order to generate cost-effective setup plans.
During the setup planning, in order to maintain the required tolerances provided in the
blueprints, proper choices of the contributing ones (increasing and decreasing tolerances)
must be made. Achieving this with simulation is possible if we think of the problem in an
opposite way. Rather than providing tolerances for the contributing tolerances to get the
concluding one, the required (concluding) tolerance is provided in order to get tolerances of
contributing ones. Simulation can be run a number of times for a range of the modeled
manufacturing error values to find what the tolerance for each case. A more general
setup planning). Setup planning can be defined as the act of preparing instructions to machine
a part. Decisions usually taken by the setup planner are: proper datums, machined surfaces,
operations and sequence of operations. The input of the problem is: design requirements and
available resources (tools, machines and fixtures). Essentially, this is an optimization problem
that aims at decreasing: cost and tolerance stackup. Tolerance stackup is part of the cost of
material removal operation. Simulation can be used here to check the goodness of a given
setup plan by examining if the proposed plan leads to acceptable tolerances or not.
19
Setup Plan
- Setup formation
- Datum selection
- Setup sequences
Process
Tolerance
Tolerance
Stackup Analysis
Analysis
No
Tolerance
Allocation
Feasible
Unconstrained Yes Feasible?
Plans
The general relation of a distance in the x, y and z space can be expressed as following [Lin
d = f ( xi , y j , z k ) (2.3)
Where:
20
Dimension chain (sometimes called tolerance chain) is a closed loop of interrelated
dimensions. It consists of increasing, decreasing links and a single concluding link. In figures
2-4 and 2-5, link i is the increasing link, d is a decreasing link and c is the concluding link.
Apparently, the concluding link c is the one whose tolerance is of interest and which is
produced indirectly. Increasing and decreasing links (both called contributing links) are the
ones that by increasing them, concluding link increases and decreases; respectively.
d c
i
Operational datum
Machined surface
The equation for evaluating the concluding link dimension is [Lin and Zhang (2001)]:
l m
c = ∑i j − ∑ dk (2.4)
j =1 k =1
Where:
21
m: number of decreasing links.
c =i−d (2.5)
l
∂c m
∂c
∆c = ∑ | | ∆i j + ∑ | | ∆d k (2.7)
j =1 ∂i j k =1 ∂d k
Referring to figure 2-5 and equations (2.6 and 2.7), the deviation of the concluding link is:
l
∂c m
∂c
∆c = ∑ ( ∂i
j =1
∆i j ) 2 + ∑ (
k =1 ∂d k
∆d k ) 2 (2.9)
j
Here, the tolerance is considered as the difference between two or more independent random
variables (links) which is calculated by adding variances up. Referring to figure 2-5 and
equations (2.5 and 2.9), the deviation of the concluding link c is given by:
22
2.4 Tolerance Chart
Tolerance chart (developed by Wade in 1983) is a formal, graphical record for a proposed
process plan that is used to verify proposed process plans by identifying tolerance chains and
then checking tolerance stackup in the dimension chains existed in the plan, stock removal
allowances and stock removal tolerance stackup [Whybrew and Britton (1997)]. Firstly,
tolerance chains are identified by the chart and then tolerance stackup are evaluated for the
concluding links. Tolerance chart includes a plenty of chains that are needed to be verified
(1) Part drawing (at the top of the chart) that includes surfaces of interest,
(6) Machining direction descriptions (Middle of the chart) that are represented by arrows
that start with the operational datum which are represented by donut symbols ( ) and
ends with machined surfaces that are represented by arrow symbol heads ( ),
(8) Resultant dimensions (Bottom right) that are needed to be compared with blueprints.
Figure 2-6 shows an example of tolerance chart taken from [Xue and Ji (2002)]. As it can be
noticed in the figure, there are three chains in this plan that must be checked out. After
calculating the tolerance stackup in the critical dimensions, they are compared (in the bottom
23
of the chart) with the design requirements (blueprints). Process plan here satisfies the design
24
3. Simulation-Based Tolerance Stackup Analysis
mimic the real world behavior. We need to tackle the problem of machining tolerance
stackup by simulating the inspection process, after simulating the machining process in terms
of material removal and manufacturing errors. Since manufacturing errors have random
characteristics that can take any probability distribution function (pdf), Monte Carlo
The idea of this simulation is to represent the features of interest by sample points (Figure 3-1
as an example). Then enough number of parts are then virtually machined according to the
intended material removal and the manufacturing errors and inspected according to the
standard CMM (coordinate measuring machine) inspection procedures by tracking the spatial
changes of the features. For more details about the simulation methodology and its
applications, readers should refer to reference [Liu and Huang (2001)]. Simulation is a proper
choice for this problem since other different types of errors can be incorporated in the model.
Furthermore, simulation is not restricted to normal error distributions only; rather, it can take
any probability distribution function (Normal, Uniform, Weibull, Triangular, etc) depending
25
Figure 3-1: Part representation by sample points
Figure 3-2 is a flowchart that illustrates the general simulation system architecture we are
using in this study. The components of the flowchart are further explained as follows:
NO
NO
NO Verification?
Validation?
YES
Feasibility?
YES
End
26
(1) Setup plan
The flowchart starts with a proposed setup plan for machining the part. Setup plan can be
defined as the instructions for machining a part in order to meet the design requirements by
choosing proper: setup formation, datum and operations sequence. The aim of this planning is
to develop the way of machining a part with the minimum cost and the least tolerance stackup
possible.
In order to represent our parts, we use the same concept used in the coordinate metrology by
representing features by sample points in the space. Since manufacturing processes are far
from perfect, there is no way to yield 100% accurate parts. Therefore, we need to make
representative sample points for the features by choosing proper sample size and sample
Since our simulation is based on simulating manufacturing errors, we need to identify the
contributing error sources that shape up the features in the space. In our model, as it will be
shown later, we adopted the following error sources: (a) cutting tool deviation that includes:
workpiece-tool interaction and cutting tool repeatability and (b) setup error that includes:
fixture unit error and raw part inaccuracies. These errors were categorized (section 3.1.3) and
some evaluation procedures were developed in chapter 5 in case they are not available.
27
Simulation starts from here by considering a virtual part, shaping its form and orientation in
the space by the sample points and keeping track of the changes of feature representation due
Validity of the Monte Carlo results depends highly of the number of iterations executed.
Unfortunately, if the number of iterations (number of virtual parts here) is not big enough,
overly misleading results will show up. Therefore, there should be some metrics or criteria
that are used to determine the number of iterations (sample size of the virtual part batch) to
achieve certain accuracy. This will be discussed more in details in section 3.4.
After collecting enough data (or sample points/dimensions), tolerances can be evaluated
using the standard methods. This usually includes: datum evaluation, dimensional and
geometric tolerance evaluation. This will be discussed more in details in section 3.2.
(7) Verification
It is the task of ensuring that the simulation is modeled properly. It is also known as
debugging the model. If a bug was found in the code, a review must be done from the start of
(8) Validation
It is the task of ensuring that the simulation model is close enough to the real world behavior.
This is mainly done by conducting real experiment that make physical machining and
inspection for the same part requirements and setup plan and then checking the closeness of
28
the simulation results with the physical ones. As usual, the closeness can be checked by
making statistical inference tests (t and F tests). If a problem was caught at this stage, a
feedback will be given to the manufacturing error model to make another experiment to
check out the manufacturing errors or to lookup at any existed thing in the database.
(9) Feasibility
Simulation checks if the proposed plan is doable using the available resources and taking into
Monte Carlo methods are numerical methods used to solve probabilistic and deterministic
problems by taking samples from contributing populations and plugging them in the
governing function of the system. Another definition is [Kalos and Whitlock (1986)]: a
Monte Carlo Simulation can be further explained as follows: given input random variables
(X1, X2… XN) with their probability distribution functions (pdf’s) and the governing function
that relates them with the output random variable Y=f(X1, X2… XN), approximate behavior of
the output random variable can be found. After enough number of simulation iterations,
distribution of the output random variable can be found (Refer to figure 3-3). Apparently,
increasing number of iterations increases the accuracy of the output. Sample size and point
locations and number of iterations are important parts to be determined when working with
29
Figure 3-3: Monte Carlo Simulation (source: http://www.ymp.gov/documents/ser_b/figures/chap4_2/f04-
174.htm)
Researchers classified manufacturing errors according to different factors (refer to table 3-1).
1. Time. This classification accounts for the error variation with time. Quasi-static errors do
not change considerably (or change slowly) with time such as errors due to dead weights.
Dynamic errors change with time such as cutting tool wear error [Ramesh, Mannan and Poo
(2000)].
and random errors. Deterministic errors do not have considerable random nature; rather, they
30
cutting tool wear. On the other hand, random errors are the ones that change according to
3. Sources of errors. Geometric error sources represent the inaccuracy of surfaces moving
manufacturing inaccuracy [Ramesh, Mannan and Poo (2000)]. Thermal error accounts for
thermal deformation in the tool because of heat provided by cutting process, machine, people,
thermal memory (from previous environments) and cooling for the coolant. The third major
contributor to inaccuracy of machined part is the cutting-force induced errors that come from
4. Errors influence on geometric positions. This classification takes into account the effect
of the error on the finished part. Locating error accounts for the variation between the ideal
datum and the one after locating and clamping. And machining error accounts for the
variation between the ideal position of the machine tool and the actual one [Lin and Zhang
(2001); Huang (1995); Lin, Wang and Zhang (1997)]. This is the classification we adopted in
Factor Categories
- Quasi-static
Time
- Dynamic
- Deterministic
Randomness
- Random
- Geometric
Sources of errors - Thermal
- Cutting force-induced
- Machining (Machine motion
Errors influence on geometric
error)
positions
- Fixture (Setup error)
31
Researchers studied error models either in separate or combined with other types of errors.
Thermal error model and its compensation were studied extensively by [Ramesh, Mannan
and Poo (2000); Okafor, Ertekin and Yalcin (2000); Shuhe, Zhang and Zhang (1997); Yang,
Yuan and Ni (1996); Yang and Lee (1998); Chen, Yuan and Ni (1997); Chen (1996); Wang
et al. (1998); Yang, Yuan and Ni (1996); Elbestawi, Srivastava and Veldhuis (1995);
Krulewich (1998); Chen and Chiou (1995); Ahn and Cao (1999)]. Force-induced error was
studied by [Ramesh, Mannan and Poo (2000), Chiu and Chen (1997)]. Setup error and its
compensation were studied by [Gao, Chase and Magleby (1998); Satyanarayana and Melkote
(2002)]. Different approaches were used to model errors such as: finite element (FE),
artificial neural network (ANN), analytical methods. Table 3-2 shows various manufacturing
error models and their sources and some studies related to each category.
32
Since the final dimension is either related to one or two features, it is good enough to
consider errors existed in the features of interest. Generally, it is safe to say that
manufacturing inaccuracy can be owed only to two factors; that are: (1) cutting tool deviation
from its theoretical (ideal) path and/or (2) Setup error due to locating, clamping and raw part
inaccuracy. In this thesis, sometimes cutting tool deviation is called machining error (3.1.3.2)
while setup error is sometimes called workpiece error. Setup error can be further divided into
locating/clamping error (3.1.3.3) and raw part error (3.1.3.4). Also, machining error can be
further divided into: cutting tool repeatability and tool-workpiece interaction (refer to figure
3-4).
Dimensional tolerance and most of the geometric tolerances are datum-related. Some of the
geometric tolerances are not datum-related as shown in table 2-1. In the case of datum-
unrelated tolerances (such as flatness, straightness, etc.), cutting tool deviation is enough to
angularity, etc.), cutting tool deviation and setup error must be both considered.
Error Models
Cutting Tool
Setup Error
Deviation
(Workpiece Error)
(Machining Error)
Tool-Workpiece
Repeatability Fixture Unit Error Raw Part Error
Interaction
33
Figures 3-5 and 3-6 depict setup and machining errors effects on the cutting process.
Referring to figure 3-5, workpiece coordinate system (WCS) deviation from the machine
coordinate system (MCS) causes removing material we do not intend to cut and avoiding
material removal we intend to cut. The inclination in the machined surface shown in figure 3-
6 is caused by the setup error whereas machined surface irregularities represent machining
error effect.
This error accounts for cutting tool path deviation from its idea path. It is assumed in this
study that the deviation is limited to z-coordinate deviation as the cutting tool must travel in
parallel paths. Although, this assumption is valid for prismatic and rotational parts machining,
it is not valid for free-form (sculptured) part machining. This error can be further divided into:
cutting tool repeatability and tool-workpiece interaction error. In this work, we only
considered the tool-workpiece interaction error since the repeatability error is usually
34
3.1.3.3 Locating/Clamping Deviation (Fixture Unit Error)
This error accounts for surface in the workpiece deviation from its ideal location due to
and z and rotation around x, y and z. It is one of the contributors to the setup error.
Raw part error accounts for part datum inaccuracy contribution to setup error. Part
inaccuracy is represented in our study by the flatness values of the primary, secondary and
tertiary datums. Raw part error and locating/clamping error together establish the setup error.
It can be represented by six parameters; namely: translation in x, y and z and rotation around
x, y and z.
Although, there is a great amount in the literature about machining error modeling and its
compensation, very few researchers attacked the problem of synthesizing the error sources
for multi-operation machining in order to predict the quality of the finished part. This is
because of the complexity of the problem. Yao et al. (2002) developed a desktop virtual-
reality approach to represent the machining and measurement processes by including some
machining error sources in the model. Huang, Zhou and Shi (2002) studied the same problem
35
analytically to determine root-causes of machined part inaccuracy. We argue here about the
need of Monte Carlo simulation use to solve this problem. [Liu and Huang (2001)] presents
Standardizing and developing accurate methods for evaluation tolerances and datums are
very important. Since different interpretations for the same data can result in different results,
if the method is not accurate enough, some good parts can be rejected and some bad parts can
be accepted. It was mentioned previously that tolerance can be categorized into datum-
as profile, runout, parallelism, etc) must include datum evaluation. The next two sections
present standard methods to evaluate tolerances when discrete data points for the machined
It is an important task to evaluate the datum in order to find the tolerances related to this
datum. Generally, there are two approaches for evaluating the datum. These approaches are
36
Least squares (LS) approach: In this method, all the sampled points on a surface (datum)
found by the CMM are fitted and then the fitted plane is translated parallelly to the outmost
point from the material. This method is defined by ISO/WD 5459-3 as the following:
“Location of the datum is defined for planar datums as the plane which is parallel to the least
squares plane and contains the extreme point of the extracted datum feature as measured
from the least squares associated line of the hill in the direction of the outward normal from
the material.”
A procedural interpretation of this definition is as the following [Wilhelm et al. (1996, 1998)]
2. For vertices on the convex hull which are on the surface of the datum feature, not
within the material of the workpiece, fit a least squares plane. These points are the
3. Translate the least squares in the direction of its surface normal away from the
37
Candidate datum set approach: This method is given in ASME Y14.5 standard for GD&T.
According to this approach, two conditions must be met: Firstly, places considered as
candidate datums must be an external set of support. That is, the plane must contact at least
one point on the datum feature while remaining outside of the material part. Secondly, small
facets or sets of support near the center of a part are generally accepted as datums while small
facets near the edges are not [Wilhelm et al. (1996, 1998)].
1. Consider a candidate plane P which is an external set of support for the datum
feature. Let C be the set of contact points between the datum feature and P.
4. The datum will not be rejected if the contact points C are not located in one
segment.
5. Repeat the same procedures for all the lines in the datum.
38
Figure 3-8: Candidate datum set approach
The ASME Y14.5 standard does not give the details for applying this procedure. Wilhelm et
al. (1996, 1998) proposed methods for evaluating the planar datums and feature of size (FOS).
1. Construct the 3D convex hull for the sampled points. The convex hull consists of facets.
Each facet on the hull that is about the material side of the sampled points is an external set of
support.
39
Figure 3-9: 2-D projection of the convex hull [Wilhelm (1998)]
3. Evaluate the facet to be checked by the three region method. Referring to ANSI definition,
the facet corresponds to a plane are the vertices of the facet. All the lines L from the vertices
are constructed. L’ is found as was described before. Figure 3-10 shows an example of a facet
that is not rejected since the contact points between the facet and the feature datum are not
This approach is mainly a search approach that needs enough number of iterations to find the
appropriate datum. This means that every candidate plane (facet) shall be checked with
infinity number of lines from 0o to 360o. Nevertheless, It is worthy to mention that the lines to
be tested shall not be parallel since parallel lines will lead to the same result; which will
40
3.2.2 Evaluation Algorithms for GD&T
In order to evaluate a tolerance that depends on a datum, both the datum and the surface must
be evaluated so and the associated tolerance is found accordingly. The allowable variation of
the tolerances in GD&T is based on the envelope principle. The entire surface shall lie
CMM data must be further interpreted to evaluate the geometric deviations mathematically.
Usually, this is done by using the least sum of distances fitting, Least Squares fitting (LS) or
the Minimum Zone fitting (MZ). All of them are optimization problems with different
objective functions.
1/ p
1 p
min L p = ∑ ri 0< p<∞ (3.1)
n i
Least sum of distances is also known as the median-polish fit. The objective of this function
is equation 1 with p = 1. This fitting is less sensitive to the outliers than the Least Squares
fitting.
Least Squares method is the most widely used approach in CMM data analysis. The objective
41
The fitted feature for the data points (x, y, z) is called the substitute feature and the geometric
deviation is evaluated as the difference between the maximum and the minimum distances
Another widely used method is the minimum zone approach method. The objective function
of this approach (also called two-sided minimax fitting) is given by equation (1) with p → ∞.
The resulting fit is strongly affected by the data outliers. The objective function turns to be as
minimization form of the minimum zone. This approach is used to measure the size of the
feature rather than measuring the form deviation. It has two forms, depending if the feature is
internal or external.
Subject to ei ≥ 0
42
Subject to ei ≤ 0
Finding a proper sampling plan for the machined feature to be inspected is a crucial step in
coordinate metrology since the chosen points are considered as the only representative points
In order to have an accurate strategy for sampling points to be measured in a feature, the
minimum sampling size and the best sampling point locations must be found out. In general,
the sampling of a machined feature depends on the machining capability, the part dimensions,
the surface topography, the required tolerance to be found and the accuracy level.
Unfortunately, machined features can never approach the perfect. An awkward solution for
inspecting feature will be by measuring as many points as possible to figure out the shape and
Finding the proper sample size (number of points to be inspected) is a major research topic in
the literature. Increasing the number of sample points leads to a more accurate evaluation but
increasing the sample size increases the inspection cost. There are some recommended sizes
for different feature geometries (refer to table 3-3). In table 3-3, the mathematical column
refers to the number of points needed to define the given geometry mathematically and the
recommended values are the ones recommended for measuring the features of the given
geometries. As was mentioned earlier, the more points taken, the more accurate the results
43
are. Another method that can be used to evaluate the sample size is by checking the change of
the results by changing the sample size. Some researchers are proposing using the Shannon-
Nyquist theorem that is used to sample the signals. This theorem states that in order to have a
fair approximation to a wave (in our case is the feature topography), the sampling interval
Table 3-3: Recommended sample size for different geometries [Henzold (1995)]
Feature
geometry Mathematical Recommended
Straight line 2 5
Plane 3 9
Circle 3 7
Sphere 4 9
Cylinder 5 15
Cone 6 15
Concerning the sample point locations, the widely used approaches are: random, uniform
(equidistant), stratified sampling (randomized block or randomized grid), refer to figure 3-11.
In random sampling, the location of each point in the space has the same chance of being
chosen as then others. Uniform sampling distributes the points in the space with fixed
variations in the machined feature. In stratified sampling, the feature is divided into blocks
and a number of sample points are chosen randomly inside each of block. Stratified
distribution has a better coverage of the feature than the random distribution approach. Some
these sequences are: Hammersley and Halton-Zaremba sequences) [Woo et al. (1995)]. These
44
sequences were proven mathematically that they cover the area of interest with the least gap
a) Stratified b) Hammersley
Figure 3-11: Sample points located in a square feature using different approaches
Figure 3-12: Sample point locations using random and low-discrepancy methods [Davis and Martin
(1998)]
It is known that Monte Carlo Simulation is a good solution tool for problems that include
stochastic variables. However, it has the bad reputation as a computationally intensive tool
since it needs a large number of iterations to converge to an acceptable level of accuracy. For
45
our problem here, the number of iterations represents the number of virtually machined part
to be inspected. There are two methods to find a proper amount of iterations (number of
virtually machined and inspected parts); sometimes called terminating criteria. Making
approximate statistical calculations to find the sample size is the first method. The second one
Statistically, the minimum number of iterations can be calculated as follows. Suppose that
when running the simulation for no iterations, the half width (ho) of the confidence interval is
given by the following equation when sample standard deviation (so) is known:
so
ho = t n −1,1−α / 2
no (3.5)
When we want to achieve half confidence interval (h), then the number of termination
2
so
n = t 2 n −1,1−α / 2 2
h (3.6)
However, there is an apparent difficulty that the right hand side of the equation depends on a
prior knowledge of n. In order to overcome this problem, we can replace the t random
variable with standard normal critical values as shown in the following equation (this is valid
s2
n ≅ z 21−α / 2
h2 (3.7)
46
An easier but different approximation is given by the following equation [Kelton, Sadowski
2
h
n ≅ no o2
h (3.8)
Cvetko, Chase and Magleby (1998) developed new metrics to evaluate their simulation. One
method presented in their paper is by benchmarking the results of the simulation for a big
sample size (e.g. 1 billion). The objective of benchmarking the results at such a big number is
to evaluate the performance of the simulation at different sample sizes. When there are no
change in the fist four moments (mean, variance, skewness and kurtosis), the sample size of
the number of iterations is chosen (refer to figure 3-13). From the figure, it can be seen that at
Figure 3-13: Benchmarked results at 1 billion iterations [Cvetko, Chase and Magleby (1998)]
47
4. Manufacturing Error Evaluation
The objective of this chapter is to introduce the requirements and procedures of experiments
and algorithms for evaluating manufacturing error; which are the inputs of the simulation.
This will be required in case the manufacturing errors are not available. Afterwards, the error
distributions and their parameters (normal, uniform …) can be plugged in the simulation to
get the results. The discussed manufacturing errors here are machining and setup errors. As it
was previously mentioned, machining error can be further classified as: cutting tool
repeatability and cutting tool-workpiece interaction error. And setup error can be further
If the machining error (cutting tool deviation) is not available, an experiment must be
conducted to evaluate it. The following are the requirements and the procedure of a proposed
experiment we developed to evaluate machining error for a specific CNC machine. This
experiment can be used to evaluate both the: cutting tool repeatability and cutting tool-
Requirements:
2. A fixture unit.
48
Procedure:
2. Mill the surface with 2 mm depth material-removal (however, depth of cut can be
3. Measure enough number (say 30) of point heights (z-coordinate) using dial indicator
without removing the part from the fixture and without removing the fixture unit from
the machine table (refer to figure 4-1). Fill the measurements of the dial indicator and
the nominal heights of the part after machining in table 4-1. We need not to remove
the part from the fixture unit to isolate the effect of the fixture unit error here and take
on-machine measurements (OMM). Notice that the dial indicator (figure 4-1) is
attached to the machine body by a magnet and does not move. Machine table is
Trial/dial
Nominal
indicator 1 2 … 30
Height
measurement
1
2
…
30
5. Fill the table 4-2 with the differences between the measured heights and the nominal
49
Cutting Tool
Dial Indicator
Prismatic
Workpiece
Fixture Unit
Machine Table
Trial/dial
indicator
1 2 … 30
measurement-
nominal height
1
2
…
30
6. Fit the data in table 4-2using a proper distribution by finding the first four moments (mean,
50
Notes:
(1) Distributions of the measurements for the same machined surface (horizontal data points
in table 4-2) represent error due to cutting tool-workpiece interaction. On the other hand,
distributions of the height deviation measurement after each machining (vertical data points
measurements here rather than using a magnetic dial indicator. The OMM is a measuring
system that can be used as coordinate measuring machines by replacing the cutting tool with
a probe. This will increase the accuracy here as we will no longer depend on the magnet to
If the locating/clamping error is not available, an experiment must be conducted to evaluate it.
The following are the requirements and the procedure of a proposed experiment we
developed to evaluate this type of error for a specific fixture unit shown in figure 4-2.
51
Fixture
Unit
Prismatic
Workpiec
Requirements:
1. A prismatic standard smooth part. The part must be accurate enough (say at most 2
2. A fixture unit.
Procedure:
2. Mark the upper corner of the edge formed by three surfaces as workpiece origin.
52
Figure 4-3: Part Surfaces
3. Clamp the part in the fixture. Locate the fixture unit itself on the CMM table in order
to take measurements.
10 random points) for each of the following surfaces: xy, yz and zx.
5. Unclamp the part, remove it from the fixture and clamp it back in the same position.
6. Repeat step 4.
7. Deviation between coordinate system established in step 4 and 6 will give the
8. Repeat from 5 enough number of times (say 30 times) using the same part and enter
9. Fit the data in table 4-3 using a proper distribution by finding the first four moments
53
Table 4-3: Fixture unit error parameters
tx ty tz α β γ
30
Flatness values of primary, secondary and tertiary datums are the input of the raw part
inaccuracy to the setup error in our model. The algorithm used to evaluate flatness values
1. Generate three points to represent each datum using the given flatness values. The
distribution to generate the values from is normal with mean = 0 and standard
deviation = Flatness/6. The points must be 1/3 or more the surface dimension apart
2. Transform the generated sample points into machine coordinate system using the
found values for fixture unit deviations (clamping/locating error for a smooth part).
4. Find the point of intersection between the three fitted planes. Consider the intersection
5. Find the normal vectors of each datum (n1, n2, n3) using the following equations:
54
6. Repeat from step 1 for enough number of times (here we recommend at least 100,000
times). It is worthy to mention that the condition in step 1 will reduce the number of
In order to evaluate the performance of the proposed algorithm, an experiment was conducted
in Delphi Automotive System. The output of the program must match the experimental
results in order to consider the proposed algorithm representative. The input of the program
will be the flatness values of the primary, secondary and tertiary datums. The following are
Requirements:
2. A fixture unit.
3. Smooth part that is accurate enough (say average flatness less than 2 microns). It is
assumed that there is no effect of the smooth part flatness on the setup error.
4. Rough part.
Figures 4-4 and 4-5 show CMM, fixture unit and the part.
Procedure:
1. Evaluate the fixture unit error in terms of the 3 translational and 3 rotational
deviations using a smooth part (stone part) using exactly the same procedure
55
Probe
Fixture
Unit
Workpiec
Figure 4-4: Coordinate measuring Figure 4-5: Fixture unit with CMM probe
machine (CMM)
2. Evaluate primary, secondary and tertiary datum flatness values for the rough part
3. Find the rotational and translational deviations for the rough part using the same
The output of the program (simulation) was found to be consistent and close to the
experimental results. However, although the results found look fairly close and stable,
sometimes results show some very different behavior. In other words, sometimes results are
truly misleading. This can be justified by insufficient iterations of the simulation. A sample of
the program output and the experimental results are shown in tables 4-4 and 4-5. It is clear
that the simulation output and the experimental results are statistically the same since the p-
value is very high for all the cases except for θx in table 4-5.
Table 4-4 data are for extremely smooth part (all flatness values are 0). And table 4-5 is for a
56
H 0 : σ 2 simulation = σ 2 exp eriment
(4.2)
H 1 : σ 2 simulation ≠ σ 2 exp eriment
Table 4-4: Variance comparison between simulation and experiment for smooth part
Table 4-5: Variance comparison between simulation and experiment for a rough part
57
5. Illustrative Examples
The following are three illustrative examples that are used to evaluate tolerance stackup using
our simulation. The part in the first example is virtually machined within single setup since
there is no datum change during machining. In the second example, three setups for the same
part in example 1 are considered. Two setups are considered for the third example.
Figure 5-1 shows the design drawing of an example part with its dimensions and tolerances.
The sequence of operations and the working dimensions of machining are shown in the left of
the figure. The simulation is conducted to find out the stackup behavior of the dimension
between surfaces f2 and f4 after machining them in a single setup using the same datum, f0.
The errors included in the simulation are (e.g. provided from manufacturing capability
database):
58
• Machining error = N (µ=0, σ2=0.038 m2)
Figure 5-2 depicts the output of the simulation conducted for 500 virtual parts (500 iterations).
Notice that the variation of the concluding link was found to be less than the other links,
which contradicts to what were expected using traditional methods. According to traditional
methods of evaluating the tolerance stackup, tolerance of the concluding link (distance
between features f2 and f4) should be the summation of the two deviations of the other links in
the chain in the worst-case scenario and the square root of the sum of squares of the two
deviations in statistical analysis. Actually, getting such a lower variation in the concluding
link is justified in our point view since these two features are machined in the same setup.
There is no tolerance stackup in this case, as has been demonstrated in [Huang (1995)].
Machining error is the only error that causes the variation here. There is no contribution from
the setup error. Table 5-1 shows a comparison of the tolerance stackup evaluation using the
three methods.
59
Figure 5-2: Example 1 results
δ02 0.057648
δ04 0.052874
Worst Case Statistical Simulation
δ24 0.1105 0.0782 0.0126
The example is an extension of the first one (refer to figure 5-3). It involves four machining
operations with two changes of the machining datums, a total of three setups. Changing
datums will result in a tolerance chain. The errors included in the simulation of example 2 are
Figure 5-4 depicts the output of the simulation for 500 virtually machined parts (iterations).
Again, the results here do not agree or even close to either the worst-case or the statistical
60
Figure 5-4: Example 2 output
Figure 5-3: Example 2 (design requirements
(The simulation output for the
between surfaces)
The third example (shown in figure 5-5) is a housing part used by Bosch. A simplified
61
Figure 5-5: Example 3; ABS (Antiblock System) housing, Bosch (Source: http://www.wzl.rwth-
aachen.de/WM/SIMON/deliverables/DA0/DA0_02D.htm)
A simulation was conducted according to the setup plan described in figure 5-7. The plan
includes two setups. The part includes six surfaces of interest that are numbered from 1 to 6.
Milling is the process used to machine the surfaces. Hole drilling is not included in the setup
plan and simulation since it does not have an effect on the tolerance chain of concern. Figure
5-8 shows a tolerance chart of the part in order to predict tolerance stackup in the tolerance
chains. Here, we are interested in the dimension shown in line 8 in the tolerance chart as a
62
As was mentioned earlier, determining the number of iterations for the simulation is a crucial
task in Monte Carlo simulation. Here, we benchmarked the results of the simulation at
100,000 to calculate the errors in the first two moments when having less sample size.
Actually, 10,000 iterations are considered large enough by most of Monte Carlo parishioners
[Cvetko, Chase and Magleby (1998)]. The first two moments (mean and variance) at 100,000
Mean Variance
L1 84.9983246659704 0.000113812363943454
L4 59.9943255754447 0.000114461403303197
L7 49.9926265647646 0.000118204861229129
L8 74.9965883885753 0.000118607406206314
Flatness of the raw part: 0.05 mm (Flatness is considered to be representative for the raw part
error.)
63
4,5,6
4 4
4,5,6
5 5
6
100
z z z z
90
160
RAW PART 1 RAW PART, MATERIAL REMOVAL: 15 2 RAW PART, MATERIAL REMOVAL: 15 3 RAW PART, MATERIAL REMOVAL: 10
1,2,3
1 1 1
2 2 2
3 3
z z z z
y 6 y 6 y 6 y 6
x 5 4 x 5 4 x 5 4 x 5 4
4 RAW PART, MATERIAL REMOVAL: 10 5 RAW PART, MATERIAL REMOVAL: 15 6 RAW PART, MATERIAL REMOVAL: 10 FINAL PART
A comparison between the results of simulation, worst-case and statistical methods in finding
concluding link tolerance stackup is shown in table 5-4. Again, Monte Carlo simulation
results were found to be less than both the traditional methods’. The ratio of simulation
tolerance stackup to the worst case tolerance stackup was found to be 0.34 and the ratio of
simulation tolerance stackup to the statistical method tolerance stackup was found to be 0.59.
Standard
Tolerance=6σ
Deviation
L1 0.010668288 0.064009727
L4 0.010698664 0.064191982
L7 0.010872206 0.065233235
L8 0.010890703 0.065344216
64
Figure 5-8: Tolerance Chart of ABS part
Figures 5-9 shows the dimension histograms of the concluding link (L8) and the contributing
links (L1, L4 and L7). Number of iterations required to achieve certain accuracy can be
predicted from figure 5-10. This figure shows means and standard deviations values predicted
using the simulation in terms of number of iterations (x axis). Clearly, 4000 iterations seem to
have very close results to the 100,000 iterations. Therefore, 4000 iterations can be considered
as proper choice for the sample size virtually machined parts (iterations.)
65
Figure 5-9: Dimensions histogram using simulation
Suppose that we need to maintain 0.066 mm for dimension shown in line 8 in the tolerance
chart (figure 5-8). Then, we will need to allocate proper tolerances for the concluding links
(dimensions shown in lines: 1, 4 and 7 in the same figure). According to our simulation,
assigning 0.060 mm for each contributing link will be good enough to meet what we need.
However, if we need to make the allocation using the worst case and statistical methods,
66
If we choose processes that are capable to achieve our simulation requirements, then we will
be satisfied with 2,700 parts per million (PPM) rejects when having the process capability
index equals to 1. However, if we use the same process considering the traditional methods,
much more rejects per million will be expected (refer to table 5-5 and figure 5-11). This
shows the importance of having less conservative method for tolerance allocation.
Figure 5-11: Rejection areas comparison when allocating concluding links tolerance using worst case,
statistical and simulation methods
Table 5-5: Part per million (PPM) rejections comparison when allocating tolerance using worst case,
67
6. Concluding Remarks and Recommendations
6.1 Summary
1. A new method for evaluating tolerance stackup using Monte Carlo simulation was
developed in this work. Cutting tool deviation and setup error were simulated along with
machining and inspecting processes for virtual parts. This method gives less conservative
results compared to the traditional ones (worst case and statistical methods) and overcomes
2. Generally, it is safe to say that manufacturing inaccuracy can be owed only to two factors;
that are: (1) cutting tool deviation from its theoretical (ideal) path and/or (2) deviation of the
3. The only practical way for checking the possibility of getting the design requirement out of
errors involved, which are the root causes for dimensional and geometric inaccuracy of
machined components.
5. Worst-case and statistical methods give more conservative results compared to the
proposed one. Overestimating tolerance stackup could result in rejecting good process plans
that should be accepted. Therefore, accurate evaluation of the stackup can lead to cost-
68
6. Simulation is a proper choice for this problem because of its complexity. Furthermore,
simulation is not restricted to normal error distributions only; rather, it can take any
the actual error distribution. It is precious to mention that even though statistical tolerances
are assumed to be normally distributed; a lot of evidences in the real world defy this
7. Monte Carlo simulation is believed to be a powerful tool to solve problems that include
stochastic variables. However, the main critique to this method is the need for a quite large
number of iterations to converge to accurate enough results. 10,000 iterations are considered
as large enough by most Monte Carlo practitioners [Cvetko, Chase and Magleby (1998)]. In
our work here, we benchmark the results at large iterations size (like 100,000 or 1 million)
and consider these results as absolutely accurate ones. Afterwards, we calculate the errors by
increasing the sample size. We can choose a sample size that has close results to the
benchmarked ones.
1. Simulation validation and verification increase user’s confidence of the results’ accuracy.
Verification can be defined as the assessment of how close simulation results are to the
conceptual model. In other words, it is the task of ensuring that the simulation was built
69
Validation is used to ensure that the simulation model matches accurately enough with the
real world behavior. In our problem here, some experiments are recommended to be
conducted for validation. Some experiments must be done to determine simulation input in
case they are not already known. These inputs are the considered manufacturing errors
included in the simulation. Afterwards, a real machining must be done for a large enough
number of parts according to the same setup plan adopted in the simulation to evaluate the
tolerances of the dimensions. The output of the simulation experiment will match the results
of the experiment if the simulation model is valid. Typically, statistical inference tests can be
good to test the closeness between simulation results and the real world behavior.
70
Bibliography
[1] Ahn, K. G. and Cho, D. W., “In-Process Modeling and Estimation of Thermally
Manufacturing Technology, Vol. 15, Issue: 4, April 27, 1999, pp. 299 - 304.
Principles.
[3] Banks, J., Discrete-Event System Simulation, Prentice Hall, Third Edition, 2001.
[4] Carr, D., A Comprehensive Method for Specifying Tolerance Requirements for
[6] Chase, K.; Gao, J. and Magleby, S., “General 2-D Tolerance Analysis of Mechanical
[7] Chen, J. S. and Chiou, G., “Quick testing and modeling of thermally-induced errors of
CNC machine tools,” International Journal of Machine Tools and Manufacture, Vol.
[8] Chen, J. S.; Yuan, J.; Ni, J., “Thermal error modeling for real-time error
compensation,” Precision Engineering Vol. 20, Issue: 2, April, 1997, pp. 151.
[9] Chen, J.-S. , “Fast calibration and modeling of thermally-induced machine tool errors
in real machining,” International Journal of Machine Tools and Manufacture, Vol. 37,
71
[10] Chen, J.-S., “A study of thermally induced machine tool errors in real cutting
conditions,” International Journal of Machine Tools and Manufacture, Vol. 36, Issue:
[11] Chen, J.-S., “Computer-aided accuracy enhancement for multi-axis CNC machine
tool,” International Journal of Machine Tools and Manufacture, Vol. 35, Issue: 4,
[12] Chen, X.B.; Geddam, A. and Yuan, Z.J., “Accuracy Improvement of Three-Axis CNC
[13] Chiu, W.M. and Chan, K.W., “Design and testing of piezoelectric actuator-controlled
boring bar for active compensation of cutting force induced errors,” International
Journal of Production Economics, Vol. 51, Issue: 1-2, August 15, 1997. pp. 135-148.
[15] Cvetko, R.; Chase, K. W. and Magleby, S. P., “New Metrics for Evaluating Monte
Mechanical Engineering Conference and Exposition, Anaheim, CA, Nov. 15-20, 1998.
[16] Davis, T. and Martin, R., “Low-Discrepancy Sequences for Volume Properties in
[17] Drysdale, R.; Jerard, R. and Schaudt, B., “Discrete Simulation of NC programming,”
[18] Elbestawti, M. A.; Srivastava, A. K. and Veldhuis, S. C., “Modeling geometric and
Tools and Manufacture, Vol. 35, Issue: 9, September, 1995, pp. 1321-1337.
[19] Feng, S. and Hopp, T., “A Review of Current Geometric Tolerancing Theories and
72
Inspection Data Algorithms,” NIST, 1991.
[20] Gao, J. ; Chase, K. and Magleby, S., “Generalized 3-D tolerance analysis of
mechanical assemblies with small kinematic adjustments,” IIE Transactions, Vol. 30,
[21] Gao, J., Chase, K. W. and Magleby, S. P., “Comparison of Assembly Tolerance
Analysis by the Direct Linearization and Modified Monte Carlo Simulation Methods,”
[22] Hockenberger, M. J. and De Meter, E. C., “The Application of Meta Functions to the
Journal of Manufacturing Science and Engineering, Vol. 118, 1996, pp. 325-331.
[23] Hong, M. and Ehmann, K., “Generation of engineered surfaces by the surface-shaping
system,” International Journal of Machine Tools and Manufacture, Vol. 35, Issue: 9,
[24] Huang, Q.; Zhou, S. and Shi. J., “Diagnosis of Multi-Operational Machining Processes
[25] Huang, S. H. and Zhang, H.-C., “Use of Tolerance Chart for NC Machining,” Journal
University, 1995.
[27] Kalos and Whitlock, Monte Carlo Methods, Vol. 1, Wiley Sons, 1986.
[28] Kelton, W. D.; Sadowski, R. P. and Sadowski, D. A., Simulation with Arena,
73
[29] Kim, S. H.; Ko, T. J. and Ahn, J. H., “Elimination of Settling Error Due to Clamping
[30] Kim, W. and Ramanam, S., “On the selection of flatness measurement points in
[31] Krishnakumar, K. and Melkote, S., “Machining fixture layout optimization using the
genetic algorithm,” International Journal of Machine Tools and Manufacture, Vol. 40,
[33] Krulewich, D. A., “Temperature integration model and measurement point selection
for thermally induced machine tool errors,” Mechatronics, Vol. 8, Issue: 4, June, 1998,
pp. 395-412.
[34] Landry, M.; Malouin, J.L. and Oral, M., “Model Validation in Operations Research.”
[36] Lee, G.; Mou, J. and Shen, J., “Sampling Strategy Design for Dimensional
International Journal of Machine Tools Manufacture, Vol. 37, No. 7, 1997, pp. 917-
934.
[37] Li, S. ; Zhang, Y. and Zhang, G., “A study of pre-compensation for thermal errors of
NC machine tools,” International Journal of Machine Tools and Manufacture, Vol. 37,
74
[38] Lin E. and Zhang, H.-C., “Theoretical Tolerance Stackup Analysis Based on
[39] Lin, S.; Wang, H.-P. and Zhang, C., “Statistical Tolerance Analysis Based on Beta
Distribution,” Journal of Manufacturing Systems, Vol. 16, No. 2, 1997, pp. 150-158.
[40] Liu, Q. and Huang, S. H., “Prediction of Component Dimensional and Geometric
American Manufacturing Research Institution, SME, Vol. 29, 2001, pp. 525-532.
[41] Liu, S. C. and Hu, S. J., “An Offset Finite Element Model and its Applications in
[43] Mize, C. and Ziegert, J., “Neural network thermal error compensation of a machining
[44] Musa, R. A. and Huang, S. H., “3D Tolerance Stackup Analysis,” Proceedings of the
[45] Okafor, A.C. and Ertekin, Y. M., “Derivation of machine tool error models and error
compensation procedure for three axes vertical machining center using rigid body
kinematics,” International Journal of Machine Tools and Manufacture, Vol. 40, Issue:
[46] Ramesh, R.; Mannan, M. A. and Poo, A. N., “AI-Based Classification Methodologies
for Modeling Machine Tool Thermal Error,” Transactions of the North American
75
[47] Ramesh, R; Mannan, M.A and Poo, A. N., “Error Compensation in Machine Tools –
International Journal of machine tools and manufacture, Vol. 40, 2000, pp. 1235-1256.
[48] Ramesh, R; Mannan, M.A and Poo, A. N., “Error Compensation in Machine Tools –
A review, Part II: Thermal Errors,” International Journal of machine tools and
[49] Rong, Y.; Hu, W.; Kang, Y.; Zhang, Y. and Yen, D., “Locating Error Analysis and
[51] Schmitz, T. and Ziegert, J., “Examination of Surface Location Error Due to Phasing of
[52] Thimm, G.; Britton, G. and Cheong, F., “Controlling Tolerance Stacks for Efficient
[53] Wade, O. R., Tolerance Control, Chapter 2 in Tool and Manufacturing Engineers
[54] Wang, Y.; Zhang, G.; Moon, K. and Sutherland, J. W., “Compensation for the thermal
76
[55] Wang, Z., Modeling and Sampling of Work piece Profiles for Form Error Evaluation,
[56] Whybrew, K. and Britton, G. A., “Tolerance Analysis in Manufacturing and Tolerance
[57] Wilhelm, R.G. and Chui, B., “Computational Metrology for the Set of Candidate
[58] Wilhelm, R.G.; Bapat, S.; Reddy, P.V.R. and Yau, H-T., “Computational Metrology
[60] Woo, T.C.; Liang, R.; Hsiang, C.C. and Lee, N. K., “Efficient Sampling for Surface
Measurement,” Journal of Manufacturing Systems, Vol. 14, No. 5, 1995, pp. 345-354.
[61] Xue, J. and Ji, P., “Identifying tolerance chains with a surface-chain model in
[62] Yang, M. and Lee, J., “Measurement and prediction of thermal errors of a CNC
[63] Yang, S.; Yuan, J. and Ni, J., “The improvement of thermal error modeling and
Machine Tools and Manufacture, Vol. 36, Issue: 4, April, 1996, pp. 527-537.
77
[64] Yang, S.; Yuan, J. and Ni, J., “Accuracy Enhancement of a Horizontal Machining
[65] Yao, Y.; Li, J.; Lee, W.B.; Cheung, C.F. and Yuan, Z., “VMMC: a Test-Bed for
[66] Zhang, H., Advanced Tolerancing Techniques, Chapter 20, John Wiley & Sons, 1997.
78