Modeling of 2D and 3D Assemblies Taking Into Account Form Errors of Plane Surfaces
ABSTRACT The tolerancing process links the virtual and the real worlds. From the former, tolerances define a variational geometrical language (geometric parameters). From the latter, there are values limiting those parameters. The beginning of a tolerancing process is in this duality. As high precision assemblies cannot be analyzed with the assumption that form errors are negligible, we propose to apply this process to assemblies with form errors through a new way of allowing to parameterize forms and solve their assemblies. The assembly process is calculated through a method of allowing to solve the 3D assemblies of pairs of surfaces having form errors using a static equilibrium. We have built a geometrical model based on the modal shapes of the ideal surface. We compute for the completely deterministic contact points between this pair of shapes according to a given assembly process. The solution gives an accurate evaluation of the assembly performance. Then we compare the results with or without taking into account the form errors. When we analyze a batch of assemblies, the problem is to compute for the nonconformity rate of a pilot production according to the functional requirements. We input probable errors of surfaces (position, orientation, and form) in our calculus and we evaluate the quality of the results compared with the functional requirements. The pilot production then can or cannot be validated.
 [Show abstract] [Hide abstract]
ABSTRACT: In the variational modeling of assemblies it is important to define the location of a part both in absolute terms and with respect to the position/orientation of other assembled parts. The present paper proposes a programming optimization approach to solve this problem. The algorithm, by using the heuristic NelderMead technique  combined with a penalty function  simulates and solves sequential assembly strategies to find the optimal geometric configuration of a rigid part with variational features satisfying all the assembly constraints in the given sequence. The algorithm best aligns mating features avoiding, at the same time, featuretofeature interferences, and automatically calculating the amount of movement the part being assembled must obey to satisfy assembly constraints, at that state of the assembly process. Thus, different assembly sequences can be simulated also including variational features.Procedia CIRP. 01/2013; 10:169177. 
Article: Skin model shapes: A new paradigm shift for geometric variations modelling in mechanical engineering
[Show abstract] [Hide abstract]
ABSTRACT: Geometric deviations are inevitably observable on manufactured workpieces and have huge influences on the quality and function of mechanical products. Therefore, many activities in geometric variations management have to be performed to ensure the product function though presence of these deviations. Dimensional and Geometrical Product Specification and Verification (GPS) are standards for the description of workpieces. Their lately revision grounds on GeoSpelling, which is an univocal language for geometric product specification and verification and aims at providing a common understanding of geometric specifications in design, manufacturing, and inspection. The Skin Model concept is a basic concept within GeoSpelling and is an abstract model of the physical interface between a workpiece and its environment. In contrast to this understanding, established models for computeraided modelling and engineering simulations make severe assumptions about the workpiece surface. Therefore, this paper deals with operationalizing the Skin Model concept in discrete geometry for the use in geometric variations management. For this purpose, Skin Model Shapes, which are particular Skin Model representatives from a simulation perspective, are generated. In this regard, a Skin Model Shape is a specific outcome of the conceptual Skin Model and comprises deviations from manufacturing and assembly. The process for generating Skin Model Shapes is split into a prediction and an observation stage with respect to the available information and knowledge about expected geometric deviations. Moreover, applications for these Skin Model Shapes in the context of mechanical engineering are given.ComputerAided Design 05/2014; · 1.52 Impact Factor  SourceAvailable from: Borhen Louhichi[Show abstract] [Hide abstract]
ABSTRACT: In order to improve digital mockup, a tolerancing phase should be integrated in the geometric models. However, in CAD software, tolerances are represented by annotations, which are neglected as well as the tolerance impact. Thus, the system malfunction is generated. For these reasons, in this paper a tolerancing phase is integrated in the numerical model to form a realistic model, where worst case configurations of assemblies are determined from the tolerances assigned to the nominal model. The proposed model incorporates tolerances on CAD models in the case of planar face, cylindrical face and planar face with non quadratic loop. In addition, the model ability to respect the maximum material condition (MMC) and the requirement of datum priority order in the CAD models is shown. Finally, functional requirement of a linear guide mechanism is inspected by using the proposed model.01/2013; 14:191206.
Page 1
1
Modeling of 2D and 3D assemblies taking into account form
errors of plane surfaces
Serge Samper  PierreAntoine Adragna  Hugues Favreliere  Maurice Pillet
Université de Savoie – Polytech’Savoie – SYMME – B.P. 80439 – 74944 – Annecy le Vieux Cedex FRANCE
serge.samper@univsavoie.fr; pierreantoine.adragna@univsavoie.fr; hugues.favreliere@univsavoie.fr;
maurice.pillet@univsavoie.fr
ABSTRACT.
The tolerancing process links the virtual and the real worlds. From the former, tolerances define a variational geometrical language
(geometric parameters). From the latter, there are values limiting those parameters. The beginning of a tolerancing process is in this
duality. As high precisions assemblies cannot be analyzed with the assumption that form errors are negligible, we propose to apply
this process to assemblies with form errors through a new way allowing to parameterize forms and solve their assemblies. The
assembly process is calculated through a method allowing to solve the 3D assemblies of pairs of surfaces having form errors using a
static equilibrium. We have built a geometrical model based on the modal shapes of the ideal surface. We compute the completely
deterministic contact points between this pair of shapes according to a given assembly process. The solution gives an accurate
evaluation of the assembly performance. Then we compare the results with or without taking into account form errors. When we
analyze a batch of assemblies, the problem is to compute the NonConformity Rate (NCR) of a pilot production according to the
functional requirements. We input probable errors of surfaces (position, orientation, and form) in our calculus and we evaluate the
quality of the results compared to the functional requirements. The pilot production can then be validated or not.
KEYWORDS: 3D assembly, form errors, modal analysis, positioning force, statistical analysis, non conformity rate.
1 Introduction and purpose
Variational geometry is the first step of any tolerancing process even if most of the time this step is not conscious.
Assumptions are made in order to simplify the “real” features. There are many ways to give geometry to some
parameters. For example, we could say that a sphere could be given by one scalar if we do not care about its form
errors. We could add to this parameter several errors like sphericity, cylindricity, conicity, and define deeply the
geometry. In an extreme way, we could give the whole set of coordinates of the measured points as parameters. How
many parameters do we need to define a given geometry? We propose a method allowing to define several levels to
adjust the set of form parameters based on a geometric filtering.
A first answer is to start from the function from a topdown analysis (from the need to the solution). Then we could
say that a minimum set of parameters could be sufficient, and the difference between the measured objects and the ideal
has to be constrained within allowable values. We could want to know more about those deviations in order to better
understand the influence of form errors on the function (finer analysis). The second answer is given by a bottomup
analysis where the manufacturer would like to know the forms of his production according to the customer’s needs.
When the set of parameters is fixed (geometric model), the functional requirements (most of the time, precisions and
assembly conditions) fix their limits. Tolerancing leads to quantifying the allowable limits fixed to those parameters.
There are two ways to solve the compatibility of limits, the worst case and the statistical analyses. Simulations and
industrial applications showed that statistical cases authorize larger tolerances. Thus, in this paper, we will show how
we analyze a batch of assemblies by taking into account form errors. The aim is to validate a pilot production of
assemblies.
In order to take into account form errors, the matting of surfaces has to be solved. If there are no defined criteria, the
matting is undefined. We propose a method allowing to solve the matting of a pair of surfaces from the skin model by
solving the static equilibrium of the pair depending on the matting force and the assembly device.
The solved values can come from a batch of measured surfaces of a pilot production that gives a batch of forms. In
this paper, we use simulations that have the same statistic properties. Then we compute the mean and standard deviation
of a probable batch of assemblies, which is checked with the functional requirements. This process gives NCR of a
supplier assemblies according to the functional requirements of his customer taking into account full geometric
Page 2
2
deviations (dimension, position, orientation and form) of a pilot production in order to optimize the risks and costs of
the function.
2 Assembly analysis model of form errors
2.1
Literature review
Parts assembly simulation is a heterogeneous research domain. Some papers are related to the stack of tolerances
analysis but very few of them propose to take into account form errors.
Models for geometric description of features are the first step for a precision analysis. The complexity of the
description is associated to the corresponding solving. From the most simple model based on a onedimensional
description assembly to a several dimensional one (discrete decomposition of features), we can find increasing data
(and process) models. The onedimensional models ([1],[2]) are used to solve size variations. With those scalar models
the analyses can be done with the worst cases or the statistic solving.
In order to take into account orientations and positions of associated surfaces, literature on the subject gives several
models using a set of parameters. Requicha [3] uses a geometric model with possible offsets of surfaces. Srinivasan [4]
modified this model changing the offset by a sweeping sphere. Wirtz [5] introduced the vectorial parameter concept
which Gaunet [6] added the TTRS [7] (Technologically and Topologically Related Surfaces) parameters of Clement.
Sacks and Joskowicz [8] developed a method for mobile planar mechanisms. The Jacobian Torsor of Desrochers and
Laperriere [9] is a kinematic model associated to an interval method. The TMAP® (Tolerance Map) model of Davidson
and Shah [10] allows to define those parameters in a constrain space. The Small Displacement Torsor (SDT) model of
Bourdet and Clement [11] is a kinematic based model that unifies the threetranslation and the threerotation
parameters. Giordano and Duret [12] [13] have developed the domain model in order to limit the SDT components to
respect the zone concept of GPS standards [14].
The skin model [15] is a measurable surface, which must be parameterized (filtered) in order to take into account
errors. The filtering can be made by several methods. Most of them use periodic functions that can build simple to
complex shapes. Fourier series is the first used decomposition method used in order to define periods on circular ([16])
and linear shapes. Discrete harmonic transforms are well known in signal processing. 2D fourier is used by Capello et
al. ([17][18]) used to define form parameters of rectangular shapes. Huang et al. [19] developed a modebase method
for form error decomposition and characterization by using Discrete Cosine Transform (DCT). In the case of disk
shapes, the periodic method of Zernike Polynomials [20] can be very useful especially in optics.
A Tchebychef Fourier Series model is used by Henke ([21]) in order to describe the forms of a cylinder and identify
specific types of error shapes. Some models based on fractals [22] and wavelets [23] give solutions on local simple
geometry (roughness to waviness on rectangular shapes). In the case of cylindrical shapes, Fourier Tchebychef models
used by Gouskov [24] defines specific form errors such as eccentric or conical.
We proposed in [25] a new way to define form error parameters based on the eigenshapes of natural vibrations of
surfaces. The originality is that the set of form parameters can be computed for any kind of shape. The modal shapes
have vectorial space properties and are defined by using a discretization of the ideal surface. This method is resumed in
section 2.3 .
In more general cases, some authors propose to define a basis built on the shape parameters of the CAD system as
Cubeles ([26]) and Pottman ([27]) who propose to define a complex shape by a mechanism built on the set of control
points of NURBS. Those methods based on polynomials can be used for a complex shape but they do not define
systematically a basis with its properties and are difficult to use for a set of shapes. They are mainly used to define
simple form error shapes.
Most of the models are topdown type (“a priori” shapes), they can be used to simulate shapes without knowing the
production process. But some form error models uses bottomup methods (“a posteriori” shapes) in order to use
informations from measurements. A model of eigenshapes derived from a Principal Component Analysis (PCA) is
proposed by Summerhays and Henke in [21] and [28]. Camelio and Hu [29] use a PCA of a production data to study
sheet metal assemblies. In order to analyze 3D medical images, Nastar, in [30] uses the modal shapes to define a set of
instances (over time) of a human organ. In image processing [31] the PCA eigenshapes is often used and gives
interesting research directions.
In order to assess assemblies, some authors ([32],[33]) uses a form error zone which is added to a position and
orientation zone. This type of method does not allow to mix position form and orientation and can increase costs. In
Page 3
3
order to analyze assemblies with a better accuracy, some works give matting solutions for circular [16] or prismatic
joints [34]. The solution of matting between two surfaces is computed by minimization of a geometric criterion [16]
[35]. Huang and Kong [36] use Monte Carlo method in order to define variations of assembled surfaces described by
DCT. An other way is to study the static equilibrium of assemblies (with an assembly force) with form errors [37].
2.2
Position and orientation models: SDT domain model
The skin model is reduced to its associated surfaces [12]. A frame is attached to each surface. Their possible deviations
are size, position and orientation. In order to model those last two types of deviations, we can use a kinematic based
model (displacements parameters). Bourdet and Clement [11] defined the Small Displacement Torsor (SDT) that gives
to positions and orientations of frames (of associated surfaces) a unified model. The SDT model allows calculating
deviations without taking into account the observation frame. We can make simple operations on SDT (addition,
transportation…).
The model of SDT domains of Giordano [12] is based on the duality of contact conditions and SDT components
inequalities. It builds a mathematical representation of those inequalities by a convex hull in a space of 6 dimensions (or
less as in the case of planar or axisymmetric mechanisms with 3D domains). Each specification or functional
requirement is expressed as a domain. The functional requirement (FR) verifications consist in checking that the
resulting SDT of all the involved deviations are within the FR domain (inclusion test in a 6D space).
2.3
Form errors parameters: modal model
2.3.1 Calculation of the modal model
We propose to define geometry variation by using a geometrical basis built on the mode shapes. They are calculated by
using the dynamic equations of either continuum or discrete mechanics. We summarize those equations in the
following.
⎪⎧
ρ
Γ = div([σ]) at each point of the volume
→
= fn
Equations (1) are the local dynamic equations and the associated boundary conditions, where ρ is the density, →
acceleration vector, →
fv is the field forces vector (nil here) [σ] is the stress tensor, n
surface and fn
The solving of the set of equations (1) for a continuum domain leads to obtaining the displacement solutions. We can
use this method for some simple geometry (as line, square or disks) cases [38]. In most of cases, there are no analytic
solution. We have to use a numerical method in most cases. As we compute modal form bases for any type of shape, we
use the Finite Element Analysis (FEA) [39] in order to determine the modal shapes.
The discrete analysis is resumed in the following equations (28).
The form error modes are the eigenshapes of the perfect surface. If we use a FEA, the nodes of the meshing will be
the corresponding measured points. The calculus of the natural mode shapes is made by the solving of the dynamic
conservative equilibrium of the discrete geometry. The following equations (2) and the boundary conditions (which are
described in displacements here) will give the eigenshapes of the geometry.
⎩⎪
⎨
→→
[σ]{n}
→ dS at each point of the outer surface
(1)
Γ the
→ the normal vector to the dS local
→ the applied local pressure on dS.
M q ¨ + K q = 0 (2)
M is the mass matrix and K the stiffness matrix of the structure and q the nodal displacements vector. The solutions of
(2) are written as follows:
qi=Qi Cos(ωi t) (3)
Qi is the amplitudes shape vector and ωi the corresponding pulsation. The linear system (4) gives the n natural
pulsations ωi and modes shapes Qi.
⎝⎛ ⎠⎞
M1 K  1
ωi² Id Qi = 0 (4)
Page 4
4
If there are no boundary conditions we obtain as rigid body modes as allowed global displacements. Those modes
would be used, in our model, to define global position and orientation of the associated surface.
2.3.2 Projection operator
We suppose that we have measured (by a CMM for example) a given geometry in order to obtain a set of errors
(displacements of measured points). This set of Euclidian displacement vectors becomes the V vector of the measured
feature. If we do measure the position of nodes, we have to make an interpolation in order to set V. As we have defined
the modal basis Q, we can obtain the projections of V in this basis by calculating the projection equation (5) in a non
orthonormal basis:
n
λi Qi = Q λ V = ∑
i=1
(5)
Here, Q is the matrix of eigenvectors Qi. λ is the vector of modal coefficients λi. We choose to give Qi an
infinity norm (Qi∝ = 1) in order to give λi a metric sense. We obtain the vector λ by solving equations (6).
(Qt Q)1 Qt V = λ
The λi coefficient can be showed as a histogram chart called modal spectrum. This representation is understood for
metrologists and designers.
We compute the vector residue of the projection in the reduced basis by
m
λi Qi with m<n
(6)
R(m) = V  ∑
i=1
(7)
The corresponding scalar value is given by equation (8).
r(m) = R(m)
As the natural shapes are defined by modal masses and modal stiffness their complexity are globally growing with i
(growing order of the frequency). This property is very useful in order to filter shapes, by keeping lowest values of λi.
Most of measured shapes have got a decreasing modal spectrum.
(8)
2.3.3 Properties of modal bases
In Fig. 1 we show the first seventeen modes of a square plate with free boundary conditions. Solving can be made by a
Finite Element Analysis (FEA) or with the help of analytical solutions [38]. The six first are rigid bodies that will reveal
the position and orientation errors. We will convert them into SDT components of the associated surface. From the
seventh, we can see form error modes. At a higher level of mode number, we will observe undulation and further,
roughness. Form, undulation and roughness of a surface are distinguished by the level of periods of a flexure behavior.
In addition, we could solve, size variations (depending on the size parameters) by using membrane behavior. We can
see with this elementary example the properties of the modal bases. The mode shapes are sorted by the value of the
associated natural frequency. As we do not care about this value, we keep this sorting operator that naturally gives
mode shapes in increasing complexity order.
Page 5
5
Fig. 1 Mode shapes of a free BC’s square
In Fig. 1, a Finite Element Model (FEM) is shown, the square is discretized by 441 nodes (21*21). We could solve
2646 (441*6) mode shapes. Sorting them by increasing frequency give increasing complexity. Since the first six mode
shapes are rigid body type, we can write a linear operator [αij] in order to represent the corresponding modal
coefficients λRi into the equivalent SDT components (9). As there are 6 rigid body mode shapes for a free structure and
6 kinematic Degrees Of Freedom (DOF), [αij] is 6*6 constants. The [αij] matrix is nonsingular and its opposite gives
SDT by knowing the λRi coefficients. It is then easy to move the position and orientation of frames.
⎧
⎬
λR5
λR6
The geometric parameterization built on mode shapes has the following properties.
Geometric Basis: Qi and Qj are independent if i ≠ j
Uniqueness: To a set of coefficients, λi corresponds only one form given by V
Stability: Continuum parameters (no conditional test, bifurcations, …).
Growing complexity: Qj is a more complex form than Qi if j>>i.
Exhaustiveness: Any variation of a shape can be decomposed in the modal basis
Metric of shape variations: λi represents the metric coefficient of the mode shape Qi in the shape V
⎩
⎨
⎭
⎫
λR1
λR2
λR3
λR4
=
⎣
⎢
⎡
⎦
⎥
⎤
α11 α12 … α16
α21 α22 … α26
. . ... .
. . ... .
α61 α62 … α66
⎩
⎨
⎧
⎭
⎬
⎫
TxA
TyA
TzA
RxA
RyA
RzA
(9)
2.3.4 Analytic or numeric method?
We are making software that decomposes form errors of a measured shape. We can use analytic solutions of (1) in the
cases of simple geometries. This way is easy to program, but in most of cases, we have to solve an FEM of eigen
shapes of a given geometry with the help of (4). Those shapes are input in a matrix Q of vector parameters. This one
can be an input of the software. Then the measured (or simulated) shape V (thus numeric) decomposition is obtained by
the same solving for the two types of eigenshapes (if eigenshapes are analytic we discretize them). If possible, we
measure displacements of points that are taken from the FEM meshing (points are nodes). If not, we use interpolations.
2.3.5 Separation of geometric errors
Dimension, position and orientation errors are easy (as concepts) to separate to form, waviness and roughness. Form
errors are not easy to define, compared to waviness and roughness. Those geometric errors can be defined by the same
type of mathematic model but do not lead to the same mechanic behavior. Surfaces sizes can be very different
compromising a size separation criterion. We propose to define form errors as the first mode shapes having the main
Page 6
6
influence on assembly results (i.e. final position of surfaces in contact). If the contact force is small and the surfaces are
filtered to remove waviness and roughness, the two surfaces are assembled with a number of points equal to the
possible components of the contact torsor. In the case of planes, we can apply one force and two moments thus there
must be three contact points according to a small contact force and two form error surfaces.
The waviness is defined as intermediate mode shapes of the surface that are involved in a lot of small contact zones.
They should remain smooth and could be deformed (elastically or plastically) with small forces.
Roughness is defined as numerous very small contact zones (spots) that should (in most of cases) have a very small
influence on the assembly position.
In this paper we focus on position, orientation and form errors of surfaces.
2.4
Single assembly of two form errors
2.4.1 Contact determination
First we make a filtering of each shape by keeping the firsts (i<m) coefficients λi. We obtain the deviation surface by
subtracting the two matting shapes. We keep the m first coefficients considering the residue r(m). We can then build a
meshing adapted to m (the NyquistShannon evaluation is used in order to define a minimum number of FEM nodes).
By this way, the number of possible matting facets is lower and their sizes are bigger.
Then we define a difference surface (surface #1 becomes perfect and #2 is the difference surface). In order to select all
the possible matting facets, we compute the convex hull of this surface. The solving of the intersection of the matting
force whit this convex hull gives the matting facet thus the contact configuration.
2.4.2 Position parameters and FR verification
Assembly is obtained by knowing contact points and their displacements. The rigid coefficients λRi are deduced and
equations (9) gives the corresponding SDT. As explained in 2.2 , we can verify by an inclusion test if those SDT
components are inside the FR domain or not. The assembly of form errors can be compared with the assembly of
associated surfaces (model without form errors) in order to know the influence of form errors on the functional
requirement.
2.5
Multiple assemblies verification and NCR
Our proposal is to evaluate the assembly of two batches having form errors. The statistical characteristics of the batches
are supposed to be known (from measurement of the pilot production batches for instance). Then, the assembly
deviation or the NCR can be evaluated through simulations. The following part 2.5.2 presents a solution to simulate a
batch of form error based on its statistical characteristics. As this paper deals with no measure, the part 2.5.1 presents
our solution to simulate a virtual batch characteristics (mean and covariance matrix).
2.5.1 Generation of statistical characteristics of a virtual batch
We have developed routines in order to make virtual skin models. We built virtual assemblies that have the properties
of observed batches in order to explain our method even if we can study any kind of batch. The statistical description of
a batch is based on a mean value of position, orientation and form and a covariance matrix of these parameters. The
statistical characteristics can be obtained from modal descriptions of a measured or simulated batch. The modal
characteristics of the simulated pilot production (mean vector µλ and the covariance matrix Cλ of modal coefficients)
are obtained with the following method:

A mean shape is identified, called the mother shape of the batch. Based on observed modal analyses,
modal coefficients λi follows a decreasing distribution law with a µ0/i type evolution (in Fig. 2, µ0=0.2
mm). The number of modal coefficients is limited to m that could be associated to a measurement filtering.

Based on this mother shape, a smallsimulated batch is randomly drawn. The λi modal coefficient are
normally drawn with a σ0/i standard deviation (in Fig. 2, σ0=0.01 mm). The size n of the simulated batch
directly affects the covariance matrix. Because random modal coefficients are independently obtained, a
batch of a large size of shapes will have a quite noncovariant covariance matrix. Then, the smaller n is
(n≥2), the higher the covariance factors are.

The mean modal signature µλ and covariance matrix Cλ (Fig. 2) of the simulated batch are computed to
create the simulated batch of n shapes.
Page 7
7
Fig. 2 Generation of the modal statistical characteristics of a virtual batch
2.5.2 Generation of a virtual batch based on statistical characteristics
As the aim is to check that the pilot productions respect the functional requirement, simulations are used. The NCR
evaluation is based on the statistical characteristics of the pilot production. Our approach is to simulate two batches with
a large number of parts based on the two statistical descriptions of the two pilot productions. One batch example is
presented in Fig. 3.
The generation of one virtual batch is made as follows:

A set of N random modal signature Λr (of m modal coefficients) is drawn. Each modal coefficient follows
a standard distribution law (for instance, normal distribution with a mean µλi = 0 and a standard deviation
σλi = 1).

We compute a diagonal matrix Cλ diag (that contains the variance on the main direction) and the transfer
matrix P (from the diagonal variance matrix to the batch covariance matrix) from the covariance matrix Cλ
of the pilot production batch. A modal signature Λv of a virtual form error of the pilot production is
obtained from the random modal signature Λr with the following relation:
Λv = P . Cλ diag . Λr +µλ (10)
Mean modal
signature
Modal coefficient
standard deviation
The mean modal
signature : µ µλ λ
The modal
covariance
matrix : Cλ λ
Fig. 3 Generation of a virtual batch based on statistical characteristics
The assemblies can be analyzed as presented in section 2.4 . The NCR is computed by an inclusion test of the assembly
SDT components inside of the FR domain.
Page 8
8
3 Application on an assembly of two parts
The following example is a classic chain of parameters closed by a functional requirement (FRB) of the B pair of
surfaces. The pair of surfaces A (A1 form part 1 and A2 form part 2) presented in Fig. 4 has to keep a given distance. The
location specification t gives the value of the functional requirement (between surfaces B). We assume that the pair of
faces of the functional requirement Bi has no geometric errors. We will compute how deviations on the contact pair Ai
influence the relative positions of Bi. The positioning device is supposed to be perfect and the assembly is rigid.
Fig. 4 Functional Requirement of the assembly
If the supplier is asked to give interchangeable parts 1 and 2, tolerancing for each part should be made and the
corresponding NCR should be higher according to the same production than with the noninterchangeability one
because some compatible form errors assemblies would not be allowed. In this paper, we focus on the taking into
account of form errors on the assembly NCR without interchangeability condition.
In this example, the contact surface Ai is a square which side has a length of 40 mm (LAX=Lz). The functional
requirement surfaces Bi is a rectangle of 20*40 mm (LBX * LZ) side lengths and the intermediate surface is a 40*40 mm
square. The other dimensions of the parts do not influence our analysis.
The two parts are not supposed to be the same and have the same type of deviations. We propose to use the tolerances
of Fig. 4.b. Then the question for the designer is “what minimum values can we input for the tolerance t ?” The question
for the producer is “are my products compatible with those specifications and those functions?”. In order to help this
relation between the customer and the supplier, we will make in the following some simulations of virtual assemblies.
We will analyze a simple assembly (Fig. 5) in order to show the method. The two matting surfaces A1 and A2, may have
position, orientation and form errors. The matting position of a pair of parts is given by the position set composed of an
external position device and a matting force. If there is no mobility of the assembly, this set stops six DOF. In the
example shown here, the external device is composed of three rods that could give three contact points because the
plane joint is dominating (3 fixed DOF). The position of the resultant force F determines the three other DOF. We use
this principle in order to compute the relative position SDT of the parts according to the matting set and their form
errors. We simulate probable form errors of A1 and A2 with decreasing values of their λ1i and λ2i coefficient according to
their i mode number (classic evolution of shapes).
a) 2D view of the assembly b) 3D view of the assembly
Fig. 5 Assembly
Page 9
9
3.1
2D assembly
We assume here that geometric errors are nil in the x direction. In order to model position, orientation and form errors,
we build the associated structural model by a freefree beam. As the measuring is on the z direction, the FEM DOF are
Ty and Rz (translation about y and rotation around z). The six first modes are presented in fig. 6. The modes #1 and 2
are rigid in order to identify the position and rotation parameters. All the modes have an infinite norm; the maximum
displacement of each one is equal to ±1. For example, mode #2 is a rigid body rotation and the maximum value of a
node’s displacement is 1. The modes represent possible forms of each surface.
Fig. 6 Modeshapes of a freefree beam
If we measure a form (“A1” surface of Fig. 7) in order to extract its modal parameters, we obtain a vector of values
shown here as a histogram. The most significant mode is the first mode (0.01 mm amplitude of position), the second
one is the third mode (0.007 mm amplitude “banana” form), the third significant is the second (0.003 mm rotation
amplitude) and the fourth significant is the ninth (0.003 mm amplitude). If we filter by using the lowest modes, we
obtain the filtered shape of the focus in Fig. 7.b. This method uses the same scheme as Fourier and is well understood
by geometric quality users.
a) Modal projection
b) Measured and associated surfaces
Fig. 7 Modal parameters
3.1.1 Contact facet
We assume that the contact between a pair of surfaces is given by a static solution. Thus we have to analyze the static
equilibrium of the two parts linked by the matting surface. This one is ideally plane (in this example) and has three
kinematics DOF in 3D analysis. The DOF are all removed by external links (fig. 5). A joint force F creates the matting
pair.
Page 10
10
Our approach consists on identifying all the possible contact points of the two faces regarding their form deviations and
a given prepositioning mechanism. A positioning force identifies the contact points that define the stable positioning.
The first step of the method is the identification of the possible contact points. We define the difference surface A
linking surfaces A1 and A2. This surface corresponds to the difference of the corresponding form deviations. The set λA12
of modal parameters λ1A2i of the equivalent surface A is obtained by the difference (eq. 11, Fig. 8) between the modal
parameters λA2 (surface A2) from λA1 (surface A1).
λ1A2 = λA2 – λA1
(11)
Fig. 8 Modal coefficients of the difference surface
On Fig. 9.a surfaces A1 and A2 in their measured frames are shown. Fig. 9.b presents the difference surface (A2A1). By
this way, we could consider that A1 is perfect and A2 has equivalent geometric errors. The second step of our approach
is the identification of all (two here) the possible contact points between the two surfaces. The proposed solution is the
use of the convex hull that identifies all the possible contact facets (segments here), hence all the possible contact
points. Fig. 9.b shows the convex hull (polygon in 2D) of the difference surface (curve) and the potential contact points
(small circles).
2020 151510 10550055 10 10 1515 2020
0.05 0.05
0.040.04
0.030.03
0.020.02
0.010.01
00
0.010.01
0.02 0.02
0.030.03
0.04 0.04
0.05 0.05
(mm)(mm)
Deviation (mm)
Surface A1Surface A1
Surface A2 Surface A2
Deviation (mm)
a) Shapes A1 and A2 with errors
b) Difference surface, associated convex surface
and contact facet.
Fig. 9 Contact facet determination
When the convex surface is computed, the contact facet is selected by solving the intersection with the axis of the
matting force.
Page 11
11
a) Matting of associated surfaces (without form
errors)
b) Matting of filtered surfaces (with form errors)
Fig. 10 Comparison between assembly methods
In Fig. 10, we can observe assemblies with (Fig. 10.b) or without (Fig. 10.a) form errors. In Fig. 10.a we can see
interpenetrations and gaps when associated surfaces are assembled (meansquare criterion). If we don’t analyze form
error contacts, we can either assemble meansquare surfaces or shifted associated surfaces, where the shift is two times
the flatness value (one for each surface). Those two models enclose the values founded by our method.
3.1.2 Form to position parameters
By using equations (9), we translate the rigid λi coefficients in the corresponding SDT. In Fig. 11 the SDT of associated
surfaces A1 (ERA1) and A2 (ERA2) are relative deviations from the nominal surface. The SDT components of ER1A2 (Error
Rigid SDT) of A2 from A1 are shown in the 2D (TyA,RzA) space of SDT coordinates (Fig. 12.a). Here, we can observe the
influence of the form error on the position of this plane joint.
Fig. 11 Deviations SDT of associated surfaces
3.1.3 Functional requirement verification of a single assembly
In Fig. 12, we present the resulting SDT of the assembly of associated surfaces (without taking into account form
errors) of A1 and A2. ER1A2 = ERA1ERA2 is the rigid SDT of the assembly (without taking into account form errors). E1A2 is
the assembly SDT with form errors. We have added EF1A2, the difference between E1A2 and ER1A2 in order to evaluate the
taken into account form errors. As those torsors are calculated in the centre of surface A and the functional FRB
requirement is calculated in the centre of surface B, we move the E1A2 SDT in order to be compared to the FRB domain.
This domain is built by the four inequalities of each point of B2 from B1 (zone condition). If each (four) summit SB2i of
B2 perimeter must remain within the segment of t range. The translation (along y axis) of each summit (called TySB2i)
respects the following inequalities then FRB is verified.
t/2<TySB2i<t/2
(12)
Page 12
12
Those eight inequalities (12) are coupled so that four are sufficient. The corresponding FRB domain is shown Fig. 12.c.
The two assembly SDT (with or without taking into account form errors) are first given in the local frame attached in
the centre of the surface A (Fig. 12.a). The set of torsors EX1A2 are transported in the same frame as the FRB domain (Fig.
12.b). The functional verification is made by an inclusion test showed in Fig. 12.c. In the presented case, deviations are
compatible with the functional requirements.
a) Assembly of A1 and A2 observed
in the centre of surface A
b) Assembly of A1 and A2
observed in the centre of surface B
c) FR checking observed in the centre of
surface B
Fig. 12 Resulting SDT and Functional Requirements of one assembly
The two SDT components verify the FR but we can see in this example that results are quite opposite.
3.1.4 Functional requirements verification of a simulated pilot production
The supplier has made a pilot production in order to check the NCR. He produces two independent batches of one
hundred parts each (for part 1 and part 2). If we assume that any part 1 can be assembled with any part 2, we will have
to test all the hundred assemblies.
We created simulated a set of random shape having decreasing values of modal coefficients λi (an hyperbolic law was
used) and covariances such as a batch production should have. This batch is usefull to assess our method but we can
analyse other set of shapes. Cvetko et al. [40] gives an estimation of the NCR dispersion by the simulation dimension.
In our example of 100 assemblies, the standard deviation of the NCR is about 0.5%.
a) Assemblies of associated surfaces
b) Assemblies of form error surfaces
Fig. 13 Conformity of a set of 100 assemblies
Here, we fixed t=0.1mm (L=20mm). As shown in Fig. 13, the conformity of this batch is better without taking into
account form error in the assemblies. The risk is then to deliver some (9% here) nonconform assemblies without
knowing the NCR. As those assemblies (Fig. 13.a) are made by using the associated surfaces (position, orientation but
no form errors), the corresponding deviations are lower than those with form errors (Fig. 13.b). In addition, the results
give the mean STD values (vertical cross in the centre of the ellipse) and the corresponding sixsigma ellipse. The
supplier can identify the component involved in the NCR (here the mean of the batch is shifted on the translation
Page 13
13
direction at 0.025 mm). Here, the effective (with form errors) batch deviation could be inside the FRB if the mean were
nil.
3.2
3D assemblies
The same assembly is presented with the assumption that the form error shape is a 3D one. The DOF are Ty (translation
upon Y axis), Rx and Rz (rotations upon X and Z axis).
3.2.1 Surfaces form errors
We have built one hundred different form errors for A1 and A2. The modal basis is computed by using the natural mode
shapes of a free BC’s (Fig. 1) of a square surface. We create probable (decreasing coefficients) form errors
(Fig. 14.a,b,c) for A1 and A2. The three first λi are associated to translations and rotations of surfaces. We solve the
assembly of the associated surfaces (Fig. 14.d) in which form errors remains. Some interferences between A1 and A2 are
shown in Fig. 14.d.
a) 35 firts λi A1 coefficients of A1
b) A1 surface
20
10
0
10
20
20
0
20
0.004
0.003
0.002
0.001
0
A2 mating A1 and contact facet of the mean assembly
c) A2 surface
Fig. 14 Surfaces A1, A2 and assembly of associated surfaces.
d) Assembly of A1 and A2 associated surfaces
3.2.2 FCR verification of an assembly
We build a difference surface (Fig. 15.a) with the same method as the 2D one. A convex hull computation of this
surface gives the set of possible contact facets. The intersection of the F axis with the convex hull gives the contact
facet (thick triangle). Then we identify the three contact points belonging to this facet. Their z values give the three
“rigid” coefficients of A2 (λ1 A2, λ2 A2, λ3 A2). By moving A2 from A1 with those three coefficients, we obtain the matting
conditions.
Page 14
14
a) Top view of the difference surface and the convex
surface
b) Top view of the contact facet
Fig. 15 Contact facet on the assembly
Fig. 16 3D view of the contact
In Fig. 15, the contact facet is defined by three points. In order to see both surfaces and the contact facet, we can see, in
Fig. 17 the topographic top view of the two surfaces (A2 in filled lines and A1 in dashed lines) and the contact facet i.e. a
triangle defined by the three summits (P1,P2,P3). Surface A2 is repositioned with the three (λ1A2, λ2A2, λ3A2) coefficients
that are solution of the contact solving. In the two zooms, we can observe that in P2, the topographic lines have the same
color (or gray) value (contact is effective) and in P4, different values (gap).
Page 15
15
Fig. 17 Assembled surfaces and contact points, a 2D view
We calculate the SDT of relative position of A2 from A1 corresponding to those displacement values. Then we transport
this torsor in the centre of the specification B. The functional requirements are respected if the SDT components are
within the FRB domain (Fig 18). The assembly of form error surfaces is closer to the limits than the one with associated
surfaces. It is easy to add SDT to several joints and check the FRB.
The contact conditions of the FRB are given the by the eight inequalities of (12). Here, as there are three DOF
parameters, those eight inequalities (half spaces) define an octahedron.
Fig. 18 FRB domain and assembly verifications
3.2.3 Functional requirements verification of a pilot production
After the virtual assemblies of one hundred pair of surfaces, we can compute the Non Conformity Rate of this batch.
In Fig. 19 the two analyses of assemblies with or without taking into account form errors are shown (with mean SDT
components and six sigma ellipsoid). In the case of associated surfaces, the whole batch is acceptable (within the FRB
domain). The NCR of assemblies with form errors gives 14% out of FRB assemblies.
View other sources
Hide other sources
 Available from Maurice Pillet · May 22, 2014
 Available from arxiv.org