See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/261253057

Inverse Modelling

Chapter · January 2013

CITATIONS READS
0 1,620

4 authors, including:

Peter Maass Oltmann Riemer

Universität Bremen Universität Bremen
213 PUBLICATIONS 5,044 CITATIONS 207 PUBLICATIONS 1,704 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Iwona Piotrowska-Kurczewski on 02 April 2014.

The user has requested enhancement of the downloaded file.

Micro Metal Forming

2nd July 2012

Chapter 9

Simulation Technologies

9.3 Inverse Modelling

Iwona Piotrowska-Kurczewski, Christina Brandt, Peter Maaß, Oltman Riemer

The development and manufacturing of high quality micro machined structure parts goes together
with modelling, simulation and optimisation. In order to obtain optimal manufacturing results or to
determine the causes in case of unsatisfactory manufacturing results, efficient models and simula-
tions, depending on structures and process parameters, are needed. For analysing mathematical
methods for parametrisation and optimisation of simulation models two things are needed. The for-
ward model that depends on a certain system, structures, material parameters and process para-
meters, e.g. feed rates, depth of cut, stiffness and dumping parameters for machine structures,
has to be developed. Secondly, an appropriate inverse model is required. In real processes only a
rough approximation about some of these parameters might be available or these are not known at
all. In mathematics the relation between the parameters g and measurements f is described by an
operator equation
Q(g) = f (9.1)
and is usually called parameter-to-state mapping. The model allows the simulation of measure-
ments f for different parameters g. The process parameters g can be constant or functions de-
pending on time and/or space. The approaches used to fit the model to the data are so-colled
inverse modelling methods. These techniques allow to find feasible values of parameters g. In case
when parameters are chosen to be constant one considers the calibration of the model and the
equation (9.1) is usually well-posed. If the parameters g = (t, x) are functions then the parameter
identification methods need to be taken into account.
One well known inverse modelling approach is the least square method that determines the para-
meter by minimizing the deviation between the simulation and the measurements by

min Q(g) − f 2 . (9.2)

g

However, only a noised version f δ of the true data and the noise level δ with

f − f δ 2 ≤ δ

are known. In general the problem (9.2) is ill-posed what means that the operator Q is not surjective
on the space of all noisy data or that the problem does not have a unique solution or the inverse

1
Q−1 is not continuous. In this case, the so-called regularisation method has to be applied for the
determination of the parameter g. Probably the most well-known regularisation technique for linear
as well as nonlinear inverse problems is the method of Tikhonov regularisation [5]. Here the solution
is obtained by minimizing the Tikhonov functional

min Q(g) − f δ 22 + αΦ(g), (9.3)

g

where Φ is a non-negative penalty function suitably chosen to incorporate a priori information about
the true solution. Often the penalty functional
ˆ 1
u
u
Φ(g) = gu = |g(ξ)| dξ (9.4)

for 0 < u < ∞ is used. Here  · u for 0 < u ≤ ∞ means the norm of the Lebesgue spaces L u (R)
usually denoted by L p (R). In the case u = 2, the techniques of Hilbert spaces are used, which has
been widely investigated in the last years, see [5].
The presented parameter identification techniques are illustrated on an example for manufacturing
processes of micro parts. The quality of the product obtained by micro forming depends on the
tribological properties of the micro moulds. Hence, mathematical models are needed for describing
and characterizing the tribological surface properties. The micro milling operations are taken into
account. The model development is based on two steps, the numerical simulation for computing
the actual tool tip position and the inverse method to determine the optimal input parameters for a
required surface quality.

9.3.1 Forward model

In the following the model development of the tool tip position for micro milling operations will be
discussed. Only a sketch of the model is given, for further information the interested reader is
referred to [8]. The aim is the development of a model characterising the structures of cutting
processes with defined cutting edge. The model includes the dynamical changes of the tool caused
by the forces occurring during the process. These changes have a direct impact on the removal
material and hence on the machined surface. The interaction between the tool and the workpiece
is described by a system of ordinary differential equations (ODEs) yielding the actual parameters.
Changes of parameters lead to unexpected differences in positions of the tool and thereby in the
generated surface. On the other hand, in the micro range size effects have to be included, see [9].
The ball end milling cutter which is rotating and moving along the given path through the workpiece
[1] is analysed. The tool is moving along the x axis with a feed velocity v f and cuts the workpiece
with a depth of cut ap . Starting point is the description of the tool tip position at each process time
t. One of the major factors which cause the dynamical changes in work-tool position are forces
acting during the process. It is therefore essential to predict these forces precisely. For developing a
dynamical model and calculating the actual model parameters, the process parameters are defined
as time depending functions, which can be integrated later into the tool position model.
A coordinate system where the feed motion v f of the tool is in x-direction and the z coordinate
indicates the depth of cut ap as illustrated in Figure 9.1 is introduced. The tool tip position at the
time t is described by a vector (x(t), y(t), z(t)) T and is influenced by different effects. We take into
account the run-out η and the deflections δ r (t), δt (t), δa (t) of the tool in radial, tangential and axial
direction, see Figure 9.2. For each direction i = {r, t, a}, they are computed as the fraction of the
force Fi (t) over the stiffness kei by
 y z
Workpiece
n
R Fr
Tool vf
ae
Tool x
ap Workpiece
Ft
x

(a) (X, Y )-plane. (b) (X, Z)-plane.

Figure 9.1: Coordinate system in milling operations.

(0, 0, 0)
0
Tool
lh -1
F –2 ρx
lh
-3 ρy ρ2
aap (t) ap -4
-0.5 1
daap (t) δy δx 0 0
0.5 -1
δ
(a) Variation of the depth of cut. (b) 3-dimensional view (c) Model of a 2-fluted ball-end tool,
of the tool deflection cutting edge and the rotation axis with
with δx = −2 ,δy = 1. parallel offset.

Figure 9.2: Dynamical positional deviations of the work-tool position.

 Fi (t)
δi (t) = .
kei
By applying the coordinate transformation
⎛ ⎞
cos(ϕ(t)) − sin(ϕ(t)) 0
Ω(t) = ⎝ sin(ϕ(t)) cos(ϕ(t)) 0 ⎠ , (9.5)
0 0 1

the deflections δx (t), δy (t), δz (t) on the (X, Y, Z) system are obtained. Moreover, as an additional
effect the total tool deflection is considered. It is calculated by means of the resulting force F (t) =
Ft (t) + Fr (t) such that the variation of the depth of cut is given by

δ2 (t) δ2 (t) + δr2 (t)

avar
p = = t , (9.6)
2lh 2lh
where lh denotes the length of the deflected tool. For more details see the approach developed in
[7]. A similar method has been developed for precision turning process, see [3].
The rotational angle is given by ϕ(t) = 2πn rot t, where nrot is the rotational speed of the tool. In the
first modelling step the acting forces, which can be separated into three components, namely the
radial force Fr (t), the tangential force Ft (t) and the axial force Fa (t), are considered. The vector of
these force components is denoted by

F∗ (t) = (Fr (t), Ft (t), Fa (t))T .

The force model is based on the common approach that the force is proportional to the chip thick-
ness hchip (t) and the axial depth of cut a p (t). It can be extended by a second term which describes
the frictional force, see [6]. The presented tool position model is flexible and can easily be gener-
alized, for instance by taking into account other models for micro forces. Furthermore, let k be the
number of cutting edges. Thus, the total force at an arbitrary cutter rotation angle is evaluated by
adding the forces acting on each cutting edge, i.e.
k
 k

∗
F (t) = F∗i (t) = Bk aip (t)hichip (t) + Bm aip (t) (9.7)
i=1 i=1

where Bk = (Brk , Btk , Bak )T and Bm = (Brm , Btm , Bam )T are the vectors of cutting constants
which are dependent on the material properties and have to be derived from measurements. By
applying the coordinate transformation (9.5), the resulting forces in three orthogonal directions
F(t) = (Fx (t), Fy (t), Fz (t))T of the (X, Y, Z) system are obtained. Furthermore, the chip thickness
is calculated as
vfact (t)
hchip (t) = sin(ϕ(t)).
nk
The second part of the forward model is a model for the actual process parameters and the actual
tool path. Here, the approach introduced in [7] is followed. The tool path at the time t in the
coordinate system (X, Y, Z) is then given by
´t
x(t) = 0 vf (t) dt −δx (t) +ηx (t),
y(t) = −δy (t) +ηy (t), (9.8)
δr2 (t)+δt2 (t)
z(t) = ap (t) − 2lh −δz (t).
 36 400

Input feed velocity vf0 (t)

Input depth of cut a0p (t)

μm mm/s

200
34

100
33
0
0 0.05 0.1 s 0.2 0 0.05 0.1 s 0.2
Time t Time t
(a) Input depth of cut a0p (t). (b) Input feed velocity vf0 (t).

Figure 9.3: Given input parameters.

Here ηx and ηy describe the static run out vectors η = (ηx , ηy , 0) (see Figure 9.2) and for the run out
angle γ they are computed by

η(t) = (η2 sin(ϕ(t) + γ), η2 cos(ϕ(t) + γ), 0)T .

For given input depth of cut a p (t) and input feed velocity vf (t), z(t) can be interpreted as the real
depth of cut aact act
p (t) = z(t) and ẋ(t) =: vf (t) as the real feed velocity vf (t). Combining equations
(9.8) describing the tool path at the time t and equations (9.6) modelling the deflections of the micro
tool with the force model (9.7) and their derivatives, the system of ordinary differential equations

ϕ̇(t) = 2πnrot t
ẋ(t) = vf (t) −δ˙x (t) +ρ˙x (t),
ẏ(t) = −δ̇y (t) +ρ̇y (t),
˙ ˙
ż(t) = a˙p (t) − δt (t)δt (t)+
lh
δr (t)δr (t)
−δ˙z (t),
vfact (t) ˙ (t)
vact
ḣchip (t) = cos(ϕ(t))ϕ̇(t)
nrot k + nfrot k sin(ϕ(t)),
Ḟi (t) = ˙ (t)hchip (t)
Bik aact
p +Bik aact
p (t)ḣchip (t) +Bim ȧact
p (t),
Ḟi (t)
δ̇i (t) = Ki ,
δ̇x (t) = − sin(ϕ(t))ϕ̇(t)Ft (t)
− cos(ϕ(t))ϕ̇(t)Fr (t) − sin(ϕ(t))Ḟr (t)
+ cos(ϕ(t))Ḟt (t)
(9.9)
is derived. The solution of the ODE system (9.9) yields the actual process parameters as well as
the actual forces at each time step. With these results, we are able to compute the resulting position
of the tool tip. This gives a forward model

(x(t), y(t), z(x)) = Q(ap (t), vf (t)), (9.10)

where Q is the operator solving the system of ODEs for input parameters a p (t) and vf (t).
Numerical example for the forward model
MATLAB routines were used to solve the system of ODEs numerically. The solution is the vector of
the actual process parameters and the actual forces at each time of the process. The model was
validated in order to derive the parameters which are consistent with the measured cutting force.
 36
Input position z 0
300
Real position z
μm

Position x(t)
Position z(t)

μm
34

100
33 Input position x0
Real position x
32 0
0 0.05 0.1 s 0.2 0 0.05
0.1 s 0.2
Time t Time t
´
(a) Given depth of cut a0p = z 0 and real position (b) Given position x = vf0 (ξ)dξ and real posi-
0

z. tion x.

Figure 9.4: Resulting positions (x(t), z(x)) = Q a0p (t), vf0 (t) computed by the forward model for
input parameters a0p and vf0 from Figure 9.3.

For the comparison of the outcomes of the model simulations with measurements, we refer to [8].
The examples for the input parameters depth of cut a 0p (t) and feed velocity vf0 (t) are plotted in Figure
9.3. The input feed rate as well as the input depth of cut is reduced by the elastic defection of the
tool due to the cutting forces appearing during the process. In Figure 9.4 the real positions z(t)
and x(t) computed by the forward model for these input parameters are showed as a dashed lines.
They are compared with the input depth of cut and then ideal position x 0 (t) which corresponds to
the input feed velocity. The deviation in both curves is showed.

9.3.2 Parameter identification problem in micro cutting operations

An inverse parameter identification problem for cutting processes is introduced and solved by reg-
ularisation methods with sparsity constraints. The forward model (9.10) which allows to compute
the tool tip position f = (x(t), z(t)) for different choices of parameter g = g(a p , vf ) is taken into
account. Moreover, only noisy data f δ = xδ (t), z δ (t) is given, see Figure 9.5. To obtain the best
setting of the parameter g = g(a p , vf ), the minimizing problem (9.3) has to be solved. The problem
under consideration is ill-posed so that appropriate regularisation methods have to be applied. Be-
cause of restrictions on the cutting machine, we are looking for a solution to the problem (9.2) which
has a sparse representation, i.e. the solution can be expressed in an orthogonal basis or frame
ζ := {ζi }i∈Z by 
g= g, ζi  ζi ,
i∈Z

where only a finite number of coefficients g i = g, ζi  are non-zero. To achieve this, the sparsity
functional (9.4) is taken with u = 1 in the penalty term Φ, i.e
ˆ
Φ(g) = |g(τ )|dτ.

These optimisation techniques have recently become available. Especially, the method presented
in [4] is a breaking point in the theory of inverse problem and its application.
 Radius of the tool r = 0, 5 [mm]
Length of the ramp ε = 0.2 [mm]
Rotation speed n rot = 20 [rpm]
Feed velocity vf = 200 [mm/min]

Table 9.1: Experimental milling conditions

tool vf 35

Depth of cut ap
0.0
μm
Δap
0.2
0.4 25
Workpiece
0.6
0.8 20
1.0
50 1.2
0 15
0 1.4 0 0.1 0.2 0.3 s 0.5
200
400 Time t
(a) 3D representation. (b) Profile of the measured depth of cut.

Figure 9.5: Measured surface for a milling experiment with variation of depth of cut.

Based on the method developed in [2], analogous techniques for micro milling processes are
presented. The notations given there are kept on, and, since the methods are very similar, only
the general idea without any focus on the details is presented.
Assume that the positions x(t) and z(t) are given. From (9.9) it follows that δ i (t) for (i = r, t, a) and
δx (t) can be written as the function of v f (t) = ẋ(t), x(t) and z(t). Together with (9.8) we have
ˆ t
vf0 (ξ)dξ = x(t) + δx (x, ẋ, z) (9.11)
0

and
δr2 (x, ẋ, z) + δt2 (x, ẋ, z)
a0p (t) = z(t) + + δz (x, ẋ, z). (9.12)
2lh
We label these functions as f 1 (t) and f 2 (t) respectively . For the integral operator
ˆ
A1 : v → v(ξ)dξ

and identity operator

A2 : v → v
Equations (9.11) and (9.12) can be reformulated in terms of the general setting (9.1), i.e.

A1 (vf ) = f1

and
A2 (ap ) = f2 .
The regularisation methods with sparsity constraint, discussed above, for computing the best input
parameters are applied. The example covers the problem of computing the optimal input parameters
 desired a0p 35 z0
35 ap,α for zα
Depth of cut ap (t)

α = 10−2
μm

Position z(t)
μm
Δap = 19.52 Δap,α = 19.74
25 25

20 20

15 15
0 0.1 0.2 0.3 s 0.5 0 0.1 0.2 0.3 s 0.5
Time t Time t
(a) Input a0p and corrected optimal apα . (b) The corresponding initial positions z 0 and
zα (b)

Figure 9.6: Input and corrected depth of cut and their corresponding initial position and the corrected
position.

in order to achieve a desired height of the ramp. The goal is the computation of the optimal depth of
cut ap for a given height of the ramp. Therefore, the Haar wavelet basis is used for the representation
of the solution.
Assume that a1p and a2p are given and that a 1p > a2p . The milling tool starts to cut with a given depth
of cut a1p and at the given time t 1 change the depth of cut to the level a 2p . In Figure 9.5 we see the
measurements of such an experiment and its resulting depth of cut. The desired height difference
Δap = a1p − a2p was 20μm. Measured values are a 1p = 36.241μm and a2p = 15.393μm so that the real
difference is Δap = 20.848μm. The parameters of the experiment are collected in Table 9.1.
The task is the computation of optimal inputs a 1pα and a2pα such that the resulting surface has the
ramp Δpα = 20μm. The desired a∗p is plotted as a solid line in Figure 9.6(a) and is used to calculate
the function g2 . Considering the regularisation, we have to minimise the problem

apα = min A2 (ap ) − f2 2 + αap 1 . (9.13)

ap

The solution apα is plotted as a dashed line in Figure 9.6(a). We plug the a p and the optimal a pα into
the forward model and obtain the positions z ∗ and zα respectively. In Figure 9.6(b), the position z ∗
is plotted in time as a solid line and compared with position z α plotted as a dashed line. We observe
the improvement of the resulting height of the ramp.
 Bibliography

[1] Yusuf Altintas. Manufacturing Automation: Metal Cutting Mechanics, Machine Tool Vibrations,
and CNC Design. Cambridge Univ. Press, Cambridge [u.a.], 2000.

[2] Christina Brandt, Peter Maaß, Iwona Piotrowska-Kurczewski, Stefan Schiffler, Oltman Riemer,
and Ekkard Brinksmeier. Mathematical methods for optimizing high precision cutting operations.
Int. J. Nanomanufacturing, 2012.

[3] Christina Brandt, Jenny Niebsch, Ronny Ramlau, and Peter Maass. Modeling the influence of
unbalances for ultra-precision cutting processes. Zeitschrift für Angewandte Mathematik und
Mechanik, 91(10):795–808, 2011.

[4] Ingrid Daubechies, Michel Defrise, and Christine De Mol. An iterative thresholding algorithm
for linear inverse problems with a sparsity constraint. Communications in Pure and Applied
Mathematics, 57(11):1413–1457, 2004.

[5] Heinz W. Engl, Martin Hanke, and Andreas Neubauer. Regularization of Inverse Problems,
volume 375 of Mathematics and its Applications. Kluwer Academic Publishers Group, Dordrecht,
2000.

[6] Dietmar Hömberg, Carsten Mense, and Oliver Rott. A comparison of analytical cutting force
models. WIAS Preprint, 1151, 2006.

[7] Iwona Piotrowska, Christina Brandt, Hamid Reza Karimi, and Peter Maass. Mathematical model
of micro turning process. The International Journal of Advanced Manufacturing Technology,
Volume 45(1):33–40, 2009.

[8] Iwona Piotrowska-Kurczewski and Jost Vehmeyer. Simulation model for micro-milling operations
and surface generation. Advanced Materials Research, 223:849–858, 2011.

[9] Frank Vollertsen. Size effects in micro forming. Key Engineering Materials, 473:3–12, 2010.

View publication stats