IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 20, NO.

4, JULY 2012

901

Robust Active Chatter Control in the High-Speed

Milling Process
Niels J. M. van Dijk, Nathan van de Wouw, Ed J. J. Doppenberg, Han A. J. Oosterling, and
Henk Nijmeijer, Fellow, IEEE

Uncertainty model input/output.

AbstractChatter is an instability phenomenon in machining

processes which limits productivity and results in inferior workpiece quality, noise and rapid tool wear. The increasing demand
for productivity in the manufacturing community motivates the
development of an active control strategy to shape the chatter stability boundary of manufacturing processes. In this work a control
methodology for the high-speed milling process is developed that
alters the chatter stability boundary such that the area of chatterfree operating points is increased and a higher productivity can be
attained. The methodology developed in this paper is based on a
robust control approach using -synthesis. Hereto, the most important process parameters (depth of cut and spindle speed) are
treated as uncertainties to guarantee the robust stability (i.e., no
chatter) in an a priori specified range of these process parameters. Effectiveness of the proposed methodology is demonstrated
by means of illustrative examples.

Generalized plant input/output vector.

Actuator/cutter displacementsm.
System/controller state vector.
Periodic solution.
Perturbations about periodic solution.
Controller input signalm.
Scalar uncertainty.
Uncertainty set.
Structured singular value.
Delays.

Index TermsActive control, delay systems, high-speed milling,

machining chatter, magnetic bearings, robust controller synthesis.

I. INTRODUCTION
NOMENCLATURE

Axial depth of cutmm.

Constant for selecting controller input.
Tooth passing frequencyHz.
Actuator/tooltip forceN.
Spindle-actuator transfer function matrix.
Averaged cutting force matrixN/m .
Controller outputA.
Controller transfer function matrixA/m.
Spindle speedrpm.
Generalized plant.
Manuscript received February 02, 2011; revised April 17, 2011; accepted
April 18, 2011. Manuscript received in final form May 14, 2011. Date of publication June 20, 2011; date of current version May 22, 2012. Recommended by
Associate Editor R. Landers. This work was supported by the Dutch Ministry of
Economic affairs within the framework of Innovation Oriented Research Programmes (IOP) Precision Technology.
N. J. M. van Dijk was with the Department of Mechanical Engineering, Eindhoven University of Technology, 5600 MB Eindhoven, The Netherlands. He is
now with the Philips Innovation Services, 5656 AE Eindhoven, The Netherlands (e-mail: niels.van.dijk@philips.com).
N. van de Wouw and H. Nijmeijer are with the Department of Mechanical
Engineering, Eindhoven University of Technology, 5600 MB Eindhoven, The
Netherlands (e-mail: n.v.d.wouw@tue.nl; h.nijmeijer@tue.nl).
E. J. J. Doppenberg and J. A. J. Oosterling are with TNO Science and Industry,
2600 AD Delft, The Netherlands (e-mail: ed.doppenberg@tno.nl; han.oosterling@tno.nl).
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TCST.2011.2157160

HATTER is an instability phenomenon in machining processes which must be avoided at all times. The occurrence
of (regenerative) chatter results in an inferior workpiece quality
due to heavy vibrations of the cutter. Moreover, a high level of
noise is produced and the tool wears out rapidly. The occurrence
of chatter can be visualized in so-called stability lobes diagrams
(SLD). In a SLD the chatter stability boundary between a stable
cut (i.e., without chatter) and an unstable cut (i.e., with chatter)
is visualized in terms of spindle speed and depth of cut.
In the present day manufacturing industry, an increasing
demand for high-precision products at a high productivity
level is seen. This motivates the desire for the design of dedicated control strategies, which are able to actively alter the
chatter stability boundary and therewith enable high material
removal rates. Hereto, this paper presents an active control
strategy which alters the stability lobes diagram in a selective
spindle speed range and, therewith, ensures a priori chatter-free
milling operations for a predefined domain of process parameters (spindle speed and depth-of-cut) such that (chatter-free)
operating points of higher material removal rate become feasible. Herein, an important challenge is to transform the model
of the high-speed milling process (which in general is described
by a set of nonlinear time-variant delay differential equations)
into a generalized plant formulation making it suitable for
robust control design [1].
Basically three methods exist in literature to control chatter.
The first method to avoid chatter is to adjust process parameters (i.e., spindle speed, feed per tooth, or chip load) such that
a stable working point is chosen [2][4]. Although chatter can

1063-6536/$26.00 2011 IEEE

902

IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 20, NO. 4, JULY 2012

be eliminated by adaptation of process parameters, the methodology does not enlarge the domain of stable operation points
towards those of higher productivity.
A second method is to disturb the regenerative effect by continuous spindle speed modulation, see, e.g., [5], [6]. Although
the stability boundary is altered by spindle speed modulation
[7], the method cannot be used in the case of high-speed milling.
Namely, in order to disturb the regenerative effect, the spindle
speed variation should be extremely fast, while the speed of
variation is limited by the inertia and actuation power of the
spindle.
The third method is to passively or actively alter the machine
dynamics to alter the chatter boundary. There are passive chatter
suppression techniques that use dampers [8] or vibration absorbers [9]. Passive dampers are relatively cheap and easy to
implement and never destabilize the system. However, the practically achievable amount of damping is rather limited. Moreover, vibration absorbers require accurate tuning of their natural
frequencies and, consequently, lack robustness to changing machining conditions.
Active chatter control in milling has mainly been focused on
active damping of machine dynamics. In [10], active damping
of a milling spindle with piezoelectric actuators is demonstrated
for low spindle speeds. Kern et al. [11], [12] applied active
damping on a milling spindle equipped with an active magnetic
bearing (AMB). Minimization of the tooltip compliance using
-synthesis, which will result in an increase of stable machining
points, is presented in [13]. An approach taking a different perspective is presented in [14]. Herein, it is assumed that chatter
originates from workpiece flexibilities. Active damping is applied by using piezoelectric actuators and sensors, which are
mounted to the thin-walled workpiece. In general one can say
that damping the machine or workpiece dynamics, either passively or actively, results in a uniform increase of the stability
boundary for all spindle speeds.
All aforementioned active chatter control approaches aim
to attenuate chatter vibrations by applying damping to the
spindle or the tool in an active way. In general, one can say
that damping the machine or workpiece dynamics, either passively or actively, results in a uniform increase of the stability
boundary for all spindle speeds. To enable more dedicated
shaping of the stability boundary (e.g., lifting the SLD locally around a specific spindle speed), the regenerative effect
should be taken into account during chatter controller design.
In [15], an optimal state feedback-observer controller with
integral control in the case of turning was designed taking the
regenerative effect into account. The infinite-dimensional
regenerative delay term is written as a rational function via
Pad approximation. Recently, Chen and Knospe [16] developed three different chatter control strategies for the case of
turning: speed-independent control, speed-specified control,
and speed-interval control. Moreover, it has been shown that
significant improvement in tailoring the stability lobes can be
obtained using dedicated controllers, obtained via -synthesis,
as compared to proportional-integral-differential (PID)-like
controllers. Although the experimental setup discussed in [16]
exhibits some aspects encountered in high-speed milling, a
comprehensive active chatter control strategy tailored to the

full complexity of the HSM process is missing to this date.

Moreover, except for the work in [11] and [12], all research
on active chatter control is limited to low spindle speeds (i.e.,
below 5000 rpm).
In this paper, an active chatter controller methodology for the
high-speed milling process is presented, which can guarantee
chatter-free cutting operations in an a priori defined range of
process parameters such as spindle speed and depth of cut by
employing an active magnetic bearing as an actuator. Current
chatter control strategies for the milling process cannot provide
such a strong guarantee of a priori stability for a predefined
range of working points. In general, the existing techniques
require a posteriori calculation of the set of stable working
points. The methodology developed in this paper is based on a
robust control approach using -synthesis. Hereto, the most important process parameters (depth of cut and spindle speed) are
treated as uncertainties. The proposed methodology will allow
the machinist to define a desired working range (in spindle
speed and depth of cut) and lift the SLD locally in a dedicated
fashion. In practice the maximum actuator force is limited.
Hence, we propose a methodology for the robust stabilization
of high-speed milling operations while minimizing the control
effort. Effectiveness of the proposed control methodology is
shown by means of an illustrative example.
The paper is organized as follows. Section II presents a comprehensive model of the milling process. Moreover, stability
properties of the model will be discussed. Section III presents
the problem statement of the active chatter control problem.
Then, in Section IV, a comprehensive analysis is performed to
select an appropriate feedback signal for the active chatter controller input, such that the actuator forces needed for stabilization are significantly reduced. The model of the milling process
as presented in Section II cannot be directly used in the robust controller design procedure. Therefore, in Section V, some
model simplifications will be discussed in order to construct a
model suitable for controller design. Section VI present the robust control design procedure, based on a -synthesis approach.
Results of the proposed strategy, when applied to an illustrative
example, are presented in Section VII. Finally, conclusions are
drawn in Section VIII.
II. MILLING PROCESS
This section presents a comprehensive model of the milling
process which can be used to predict the occurrence of regenerative chatter. Moreover, stability properties of the model will
be discussed. The model of the milling process as discussed
in this section is originally presented in [17][20]. In Fig. 1,
a schematic representation of the milling process is given. A
block diagram of the milling process, with controller, is given
in Fig. 2. Each of the blocks in this figure will be explained in
more detail as follows.
As can be seen from the block diagram in Fig. 2, the milling
process is a closed-loop position-driven process. The setpoint
of the milling process is the predefined motion of the tool with
respect to the workpiece, given in terms of the static chip thickness
, where
is the feed per tooth
and
the rotation angle of the th tooth of the tool with
respect to the (normal) axis (see Fig. 1). However, the total

VAN DIJK et al.: ROBUST ACTIVE CHATTER CONTROL IN THE HIGH-SPEED MILLING PROCESS

903

tangential and radial direction for a single tooth are described

by the following exponential cutting force model:

(1)
where
and
are cutting parameters
which depend on the workpiece material and
is the axial
depth of cut. The function
describes whether a tooth
is in or out of cut
else
Fig. 1. Schematic representation of the milling process.

(2)

and
are the entry and exit angle of the cut, rewhere
spectively. Via trigonometric functions, the cutting force can
easily be converted to (feed)- and (normal)-direction (see
Fig. 1). Hence, cutting forces in - and -direction,
and
, respectively, can be obtained by summing over all teeth,
as shown in (3) at the bottom of the page, where
and

B. Spindle Dynamics and Actuator Dynamics

Fig. 2. Block diagram of the milling process.

chip thickness also depends on the interaction between the cutter

and the workpiece. Since, in general, the machine tool is not
infinitely stiff, the interaction between the cutter and the workpiece leads to cutter vibrations resulting in a dynamic displacement
of the tool which is superimposed on the predefined tool motion. This results in a wavy surface on the workpiece. The next tooth encounters the wavy surface left behind
by the previous tooth and generates its own waviness. This is
called the regenerative effect and results in the block Delay in
Fig. 2. The difference between the current and previous wavy
surface is denoted as the dynamic chip thickness
with
the delay. Hence, the total chip thickness encountered by tooth
,
, is the sum of the static and dynamic chip thickness:
. In the next sections, the components of the milling model will be described in more detail.
A. Cutting Force Model
The cutting force model (indicated by the Cutting block in
Fig. 2) relates the cutting forces acting at the tool tip of the
machine spindle to the total chip thickness. The cutting forces in

The cutting force interacts with the spindle rotor and tool
dynamics (block Spindle in Fig. 2). For the purpose of active
chatter control, an actuator is implemented in the spindle rotor.
The controller output
is dictated to the actuator which, in
turn, generates a force
on the spindle.
In general the spindle rotor and tool dynamics (jointly
called the spindle dynamics) can be modelled by a linear
multi-input-multi-output (MIMO) model. The model has four
inputs and four outputs. The inputs consist of the cutting forces
acting at the tool-tip in -/ -direction and the actuator forces
in -/ -direction induced at some point in the spindle, which
generally differs from the location at which the cutting forces
are acting (the tooltip). This leads to an inherent flexibility
between the actuator/sensor system and the cutting forces.
The outputs of the spindle rotor dynamics model are the
displacements
of the tooltip
and displacements
measured at some position on the
spindle, which are used for feedback. In this paper, the machine spindle-toolholder-tool dynamics is modelled by two
decoupled subsystems (representing the dynamics in two
orthogonal directions perpendicular to the spindle axis)
consisting of two mass-spring-damper systems to mimic the
inherent compliance between actuator and tooltip, see Fig. 3,

(3)

904

IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 20, NO. 4, JULY 2012

dynamic systems and in [11] feasibility of using such actuator

in the scope of high-speed milling has been shown. This motivates to pay special attention to this kind of actuator (model).
The nonlinear model of an AMB driven in differential mode is
given as follows [22]:

Fig. 3. Schematic overview of spindle dynamics model,

.
: forces displacements at tooltip.
forces displacements at actuator,

TABLE I
MILLING MODEL PARAMETERS

with masses

(5)

, eigenfrequencies
and dimension-

less damping ratios

. This is done in order to capture the inherent dynamics
between the actuator/sensor system (denoted by subscript )
and the cutting tool (denoted by subscript ). The parameters
of the machine spindle model and cutting force coefficients are
listed in Table I. Herein, the cutting model parameters (
and
) are taken from [21] and spindle parameters are chosen
such that these represent realistic machine spindle dynamics
for high-speed milling machines. The state-space equations
describing the rotor dynamic model are given as follows:

(4)
where
is the state vector (the order
of this model
primarily depends on the order of the spindle-tool dynamics
model, which in this case equals
),
,
, and
. The state-space matrices of the spindle rotor dynamic model are given in the appendix. Next to the spindle rotor dynamics, an actuator model
is included. The active chatter control design procedure will be
developed for a model incorporating a nonlinear actuator model
for an active magnetic bearing (AMB). As described in the introduction an AMB is a common type of actuator applied to rotor

are the specific AMB coefficients,

is the
where
so-called premagnetizing current (to compensate for gravity,
etc.),
the corresponding nominal gap displacement and
is the controller output (i.e., the input currents to the
actuator) and
the bearing displacements. In general, the
displacements in the actuator journal
are significantly
smaller than the gap width . In addition, the controller output
will be limited by the controller design methodology.
Then, for control design the nonlinear model of the AMB may
be linearized about
, which has already been
successfully performed for many applications as is described
in [22]. This results in the following linear model of the AMB:
(6)
where
(7)
(8)
C. Total Milling Model
In the previous sections, the submodels for the cutting force
and spindle rotor, toolholder, tool, and actuator dynamics representing the different blocks in the milling model as given in
Fig. 2 are introduced. Substitution of the cutting force model
and actuator model, given in (3) and (6), respectively, into the
model of the spindle rotor, toolholder, and tool dynamics, given
in (4), yields the total milling model, shown in (9) at the bottom
of the page. It can be seen that the model describing the milling
process is set of nonlinear, time-dependent delay differential
equations (DDE). In the next section, the stability properties of
the model will be analyzed.
D. Stability of the Milling Process
In this section, we will briefly address the stability analysis
exploited to determine chatter boundaries in the stability lobes

(9)

VAN DIJK et al.: ROBUST ACTIVE CHATTER CONTROL IN THE HIGH-SPEED MILLING PROCESS

diagram. In the milling process the static chip thickness is periodic with period time
. Here is the spindle speed in
revolutions per minute (rpm). In general, the uncontrolled (i.e.,
) milling model (9) has a periodic solution
with
period time [23]. To validate this fact let us adopt the following
decomposition of
:
(10)
where
is a -periodic motion that can be considered as
the ideal motion when no chatter occurs and
the perturbation term. When no chatter occurs,
and the tool motion is described by the following ordinary differential equation
(ODE):

905

which results in an uncertainty in the delay .1 Moreover,

chatter is defined as the loss of stability of this periodic solution
and stability of the milling process is based on the stability of
the model describing the perturbations of the milling process
around the periodic solution. The aim of this paper is to design
a finite-dimensional linear controller , which guarantees the
following:
robust stability of the milling process (12) for the given
uncertainties in depth of cut , time delay ;
performance by minimizing the total amount of actuator
energy needed to stabilize the uncertain milling process.
Hereby, it is assumed that the controller , with controller
input
and output current
, has the following
state-space description:
(14)

(11)
which follows from (9) by exploiting the fact that
. The ODE in (11) is a linear system with a periodic
excitation with period time . Hence, when
has
no eigenvalues at
, for
and all
, the
solution
exists, is unique and is -periodic [24]. Therefore, the periodic solution is (at least locally) asymptotically
stable when no chatter occurs and when chatter occurs it is unstable. Therefore, the chatter stability boundary can be found
by studying the (local) stability of the periodic solution
.
To this end, the uncontrolled milling model is linearized about
the periodic solution
which yields the following linearized
dynamics in terms of the perturbations
:

(12)
where

(13)
As can be seen from (12) and (13), the linearized model is a delayed, periodically time-varying system. Stability of these kind
of systems can be assessed using, e.g., the semi-discretization
method of [25]. The main point of semi-discretization is that
only the delay term is discretized, instead of the actual time-domain terms. Unless stated differently, all stability lobes diagrams, presented throughout this paper, are determined using
the semi-discretization method.
III. PROBLEM STATEMENT
Recall from the previous section that the SLD is determined
using the model which describes the perturbation vibrations
about the (chatter-free) periodic solution of the milling process.
Therefore, the controller design, as presented in this paper, will
be based on the model which is linearized about the periodic
solution with uncertainties in depth of cut
and spindle speed

,
,
,
,
Herein,
with
the order of the controller. The
and
choice of the controller input signal
will be discussed
in Section IV. The linearized uncertain model of the milling
process, given by (12), (13), cannot be directly used in the
standard robust controller design procedure. Therefore, after
discussing the selection of the controller input signal
, two
model simplifications will be presented in Section V such that
the infinite-dimensional time-varying model (12), (13), is transformed into a finite-dimensional linear time-invariant (LTI)
model. In this way, the model can be used in a robust control
design procedure, which will be presented in Section VI.
IV. CONTROLLER INPUT SIGNAL
An important part of any control system is the choice of the
feedback signal used for control.
From the discussion in Section II-D it becomes clear that the
nominal (chatter-free) solution of the milling model is periodic
with period time . Moreover, chatter is defined as the loss of
stability of this periodic solution and stability of the milling
process is based on the stability of the model describing the perturbations of the milling process around the periodic solution.
Then, two possibilities arise in selecting the feedback signal
which serves as an input to the controller
in (14), namely
as follows:
1) full output feedback, i.e., the total (measured) displacements
are used for feedback:
in (14);
2) perturbation feedback, i.e., the perturbation (chatter) vibrations
are used as feedback signal,
where
denotes the periodic solution of the nominal
model given by (9):
in (14).
In Section VI, the design of a linear dynamic output feedback
control law characterized by the transfer function
and with
a state-space description as defined in (14), is pursued. Next, the
controller input signal
will be denoted as
(15)
1Note that
and are not uncertain in practice, but since we aim to stabilize
, we treat the parameters
the milling process in a range of working points
as uncertainties in the control design.

906

IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 20, NO. 4, JULY 2012

where
is the periodic solution at the measured
output. Moreover, is a constant indicating whether full output
or perturbation feedback
is applied.
The implication of the choice for either one of the two controller input signals will be demonstrated next. It can easily be
shown that the stability properties of the closed loop, in case of
the linear actuator model are the same for both choices of the
feedback signal [either
, or
in (15)].
First, it will be shown that the chatter-free -periodic solution
of the closed-loop system (9), (14); will be
different for both choices of the controller input signal. To show
this, consider the following decompositions of the state vectors
and
:

(i.e.,
) the periodic solution
will be
equal to that of the uncontrolled system (and
, i.e.,
the nominal control action
is zero).
On the other hand, for full output feedback (i.e.,
)
the periodic solution
will be different from that of the uncontrolled system (and
will in general be non-zero, i.e.,
the steady-state control action
does not vanish).
Second, it can be shown that for both choices of the feedback signal the linearization of (9) with (14) about
is given by

(16)
(18)
Then, let us combine (9) with (14) and substitute
and
. Using the fact that
is -periodic (i.e.,
), this results in the following closed-loop
dynamics:

(17)
It can be seen that the closed-loop dynamics in (17) governing
the periodic solution is an LTI system with a -periodic disturbance since both
and
are periodic with . Then
it can be concluded that when full output feedback is applied
(i.e.,
), with the assumption that
has no eigenvalues at
, for
and all
, the solution
exists, is unique and is -periodic [24]. Note that this solution
differs from the periodic solution of the open-loop dynamics
given in (11). The conditions imposed on the eigenvalues of
will in general be satisfied, since the controller has to render
the closed-loop system stable also for
(i.e., the eigenvalues of
will lie in the open left half plane). However,
when
and consequently perturbation feedback is applied, the eigenvalues of
are given by the eigenvalues of
and
. Then, there exists a unique, -periodic,
solution
when
has no eigenvalues
at
, for
and
has no eigenvalues with real part equal to zero. The conditions imposed on
will typically be satisfied since, in general, the
AMB actuator is designed such that the decrease in stiffness,
due to the negative stiffness effect of an electromagnetic actuator, see [22], is significantly smaller than the stiffness of the
spindle rotor. Consequently, under such conditions, in case of
perturbation feedback,
is the only solution of (17) satisfying
. Therewith, the periodic solution
of (17) becomes equal to the solution of the open-loop periodic solution of (11). Hence, in case of perturbation feedback

with
as defined in (13). Clearly, for both choices of the
control input signal, the resulting SLD will be the same (since
the perturbation dynamics (18) does not depend on the constant
). Moreover, in the case of perturbation feedback
the
nominal control action vanishes in steady-state, which is not the
case for full output feedback
. As a result, the choice
for perturbation feedback is favorable from the point of view of
bounding the control action.
From the analysis, presented above, it can be concluded that
the SLD with active chatter controller (14) does not depend on
the chosen controller input signal. This is due to the fact that
the variable , indicating whether full output feedback
or perturbation feedback
is applied, does not appear
in the linearized equations of motion [see (18)]. Moreover it is
shown that the actuator forces, needed to stabilize the milling
process, will be zero in steady state in case of perturbation feedback whereas the actuator forces will be non-zero in steady state
in case of full output feedback. Based on the previous discussion and the fact that for an AMB it is important to limit the input
current in order not to exceed the maximum amount of carrying
force, in the remainder of this paper, perturbation feedback will
be considered. In practice, the perturbation displacements
can be obtained by using a chatter detection algorithm based on
a parametric model of the milling process, as, e.g., described in
[4].
V. MODELLING FOR ROBUST CONTROL DESIGN
The model of the milling process, discussed in Section II, can
readily be employed for stability analysis (i.e., determination of
the SLD). However, the presence of time-delay and the explicit
time-dependency of the right-hand side of the DDE (9) complicate the development of robust control synthesis techniques.
Therefore, we apply two model simplifications to construct a
finite-dimensional, time-invariant model, which will be more
suitable for controller synthesis. Moreover, the effect of these
model simplifications on the SLD is demonstrated.
This section will discuss two model simplifications to construct a finite-dimensional time-invariant model of the milling
process, which will be more suitable for controller synthesis.

VAN DIJK et al.: ROBUST ACTIVE CHATTER CONTROL IN THE HIGH-SPEED MILLING PROCESS

907

First, the discussion will focus on an autonomous approximation of the linearized nonautonomous DDE describing the linearized perturbation milling dynamics, obtained by linearising
(9) about
, given by (12)

(19)
A characteristic feature of a milling process is that the direction
. As a
of the cutting forces is a function of the rotation angle
result, time-dependent functions appear in the describing model
equations. In [17], a method is described which approximates
by means of a Fourier series expansion. The number of
harmonics to be considered for an accurate reconstruction of
depends on the immersion conditions (which indicates
the percentage of the tool diameter used during cutting) and the
number of teeth in cut. In this paper, we will consider full immersion cuts (i.e., the entire tool diameter is used for cutting).
Then, as described in [17], it is sufficient to take the average
(zero-order) component of the Fourier series expansion over one
tooth passing, i.e.
(20)
is valid only between the entry
and exit
angles
Since
of the cutter (i.e., when
), it becomes equal to the
average value of
at cutter pitch angle
(21)
, and
can be
where the integrated functions
determined analytically in case of a linear cutting model (
, see [17]) and have to be computed numerically in case of an
exponential cutting model
.
At this point we have obtained a time-invariant milling model
in which the dependency on the rotation angle
is eliminated.
Secondly, a finite-dimensional approximation of the time
delay, using a Pad approximation, is applied (see also [15]
and [16] where Pad approximation is used for controller
design in case of turning). Hereto, the delayed tool vibrations
are approximated using a Pad
approximation and the resulting approximation is denoted by
, such that
. The choice
for a suitable order of the Pad approximation depends on the

Fig. 4. Stability lobes diagram for the milling process with Pad approximation
of order
and the milling model with exact time delay.

eigenfrequencies of the spindle-toolholder-tool dynamics. In

order to accurately approximate the regenerative effect, given,
in the frequency domain by
, an approximation
function should be chosen which has magnitude 1 at the frequencies
,
and accurately approximate the phase
of
. Consequently, the order of the Pad approximation
should be chosen such that it approximates the phase of
up to at least the highest frequency of the spindle-toolholder
and tool modes which are relevant for chatter.
The milling model in (19) with cutting force averaging, defined in (21) and Pad approximation is given as shown in (22) at
the bottom of the page, where
, and
denote matrices of the state-space description of the Pad approximation. The size of these matrices
depends on the chosen order
for the Pad approximation.
Since the delayed output vector
has two elements
( - and -direction), the state-space description of the Pad approximation has two times the number of states of the Pad approximation order . The order of the Pad approximation will
be based on a desired level of accuracy regarding the predicted
chatter stability boundary using the model with Pad approximation.
In Fig. 4 the chatter stability boundary is given for the autonomous model with time-delay and for different orders
of
the Pad approximant with the parameters of the model listed
in Table I. From Fig. 4 it can be observed that, for increasing
order
of the Pad approximant, the error between the stability lobes determined using the exact delay term and the approximated delay term becomes smaller. Moreover, since the
delay is inversely proportional to the spindle speed, the approximation becomes more accurate as the spindle speed increases.

(22)

908

IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 20, NO. 4, JULY 2012

Fig. 5. Generalized plant interconnection.

In this work we focus on high-spindle speeds (i.e., above 20

krpm). Hence, it is sufficient to choose the order of the Pad approximant for
equal to be
, which will be used
throughout this paper.
VI. ROBUST CONTROLLER DESIGN
In the previous section, a model has been derived which is
suitable for robust controller design. Therefore, in this section, the actual controller design for an active chatter control
methodology, which will alter the chatter stability boundary, is
presented, such that stable operating points of higher productivity can be attained while avoiding chatter. The control goal,
as presented in Section III, can be cast into the generalized plant
framework and solved using -synthesis techniques [26]. Fig. 5
shows the configuration of this framework. The generalized
plant is a given system with three sets of inputs and three sets
of outputs. The signal pair
denote the inputs/outputs of the
uncertainty channel. The signal represents an external input in
which possible disturbances, measurement noise and reference
inputs are stacked. The signal is the control input. The output
can be considered as a performance variable while denotes
the measured outputs used for feedback. The remainder of this
section will be devoted to uncertainty modelling, in depth of cut
and spindle speed and the specification of the performance
requirement for the active chatter control problem as stated in
Section III.
A. Nominal Model
Equation (22) gives the nominal plant model used during
-synthesis. Note that, in contrast to most active chatter control methods discussed in the introduction, in this work we do
not only consider the spindle dynamics during the control design, but also take the interaction between the spindle dynamics
and the cutting forces (and therewith the regenerative effect responsible for chatter) into account. It is expected that this is a
more profound and promising method to make dedicated modifications to the chatter stability boundary by means of feedback
control.
B. Uncertainty Modelling
This section describes the modelling of the uncertainties in
the process parameters, which can be considered as a key step
in achieving the control objective defined above: robust stability

(i.e., chatter avoidance) in a predefined range of process parameters. The control design will be based on the milling model (22)
presented in the previous section.
1) Uncertainty in Process ParameterDepth of Cut :
First, the uncertainty in depth of cut
is considered which is
modelled as a parametric uncertainty. An important (practical)
aspect is that robust control design should provide stability for
small as well as (relatively) large values of the depth of cuts,
see also [16]. Hereto, the uncertain depth of cut is modelled
such that it specifies a range from zero up to a maximum value
, i.e.,
. Let us define a real scalar uncertainty set
. The uncertainty for the depth
of cut is then defined by

(23)
where is the maximal depth of cut for which stable cutting is
desired.
2) Uncertainty in Process ParameterSpindle Speed :
Next, the uncertainty model for the spindle speed is considered.
As described before, the delay is inversely proportional to the
spindle speed. Hence, uncertainty in spindle speed is modelled as an uncertainty in the delay , where
. Since
a Pad approximation is a rational function of two polynomials
in the Laplace operator and delay , modelling the interval
delay via a parametric uncertainty would result in an overall
uncertainty of very large dimensions (due to the relatively large
order of the Pad approximations).
Here an alternative approach will be used to model the delay
uncertainty. Hereto, note that for arbitrary frequency , the
value set of the frequency-domain delay operator
for all
can be represented in the complex plane as a circular
arc extending along the unit circle. This time-delay interval
can be approximated by choosing any pair of stable transfer
functions
and
such that
, with
and
, covers the uncertainty set
with
. Several alternatives exist to determine transfer
function
and
satisfying these conditions. Chen
and Knospe [16] propose to choose
and
such
that at each frequency: 1) the arc length covered by the disk
is nearly that of the delay element
, for
and 2) the area of the disk lying outside the unit circle
is minimized. Doing so results in a transfer function
which has twice the order of the chosen Pad approximation.
Since the Pad approximation needed to accurately describe the
delay term is already of a relatively high order, the generalized
plant will be of an even higher order which is not desired due
to possible computational and implementation issues for the
resulting controller. Moreover, the size of the circle covering
the circular arc of the delay uncertainty is rather large which is
due to the fact that the area of the disk lying outside the unit
circle is chosen to be minimized. This approach may therefore
give conservative results as illustrated for the milling process
in [1].
Hence, here a different approach is presented to model the
delay uncertainty. In contrast to the approach discussed above,

VAN DIJK et al.: ROBUST ACTIVE CHATTER CONTROL IN THE HIGH-SPEED MILLING PROCESS

we model the delay uncertainty based on a Pad approximation of the nominal model. The total delay uncertainty interval
is then overapproximated using a low-order transfer function
which covers the circular arc of the delay uncertainty interval
along the unit disk about the nominal delay. Hereto, consider the
linearized autonomous milling model with a delay uncertainty
only. Basically, this model can be represented by the following
state-space model:

(24)
,
,
where
and uncertainty set
. It is easy to show that (24) can
be written as a feedback interconnection between the dynamics

(25)
and uncertainty term
(26)
where

, the delay operator

is defined as
and
. The representation of
the time-domain operator in (26) can be given in the Laplace
domain as

909

Hereto, in [27] rational transfer functions

for several orders
are derived such that
. Since,
in this work, the high-speed milling process is considered, the
delay intervals will be relatively small (typically of
)
s. for typical spindle speed ranges of
rpm for spindle
speeds
rpm). Using this fact, together with the fact
that the dominant spindle dynamics resonances lie in general
between
Hz, implies that typically
and, consequently, an accurate approximation
of the frequency-dependent upper bound
is required. Moreover, from a numerical point of view, proper
transfer functions
are desired. Then, based on the results
in [28],
is chosen as
(30)
which ensures a tight over bound of
(by
) especially in the frequency region which is relevant in the case of
high-speed milling, as is described above. Hence, by using the
results presented above, the delay uncertainty is approximated
by two rational transfer functions
and
, where
is the Pad approximation of
and
,
with
as in (30), such that

(31)

(27)
where
,
the Laplace transforms of
and
,
respectively. Let
be the gain bound of the uncertainty operator (27) in the frequency domain, given as
(28)
is analytic and
Since the transfer function
bounded in the open right half of the complex plane, the
-norm of
can be determined by evaluating
the transfer function on the imaginary axis, i.e., for
.
Consequently, in order to determine a bound on
,
it should be determined for
. It can be shown, see [27],
that the upper bound
on the delay uncertainty is represented as follows:
(29)
. The frequency-dependent upper bound
where
on the delay uncertainty is not a rational function and
can therefore not readily be used during controller synthesis.

C. Performance Requirement
This section discusses the specification of a performance requirement for the active chatter control design. In essence, the
chatter control problem at hand is a robust stabilization problem
rather than a performance problem. As outlined in the problem
statement in Section III, the robust stability requirement has to
be achieved with limited control effort, since actuator forces
have to satisfy practical saturation limits (of, e.g., AMB). Therefore, the control gain should be bounded during -synthesis,
which reflects the most relevant performance requirement for
chatter control.
Limiting the control gain is done by applying an
upper bound on the control sensitivity transfer function
, where we have the
equation at the bottom of the page, which gives the transfer
function representation from
to
of the nominal plant
given by (22). Here, the control sensitivity is defined as the
transfer function from a input signal
(which can, e.g., be
interpreted as measurement noise on the measured perturbation displacements
entering the feedback loop) to the
control input
. The bound on the control sensitivity is

910

IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 20, NO. 4, JULY 2012

As often in robust control,

fashion.

is chosen in an iterative

D. Generalized Plant Formulation

Fig. 6. Block diagram of linearized approximated autonomous milling model

with performance weighting.

enforced by defining a weighting function

, which
will be described below, such that the performance output
of the generalized plant is the weighted control sensitivity
. A schematic overview of the closed-loop
approximated autonomous milling model with performance
weighting is given in Fig. 6 where the transfer function of the
spindle-actuator dynamics is denoted by
and given by

Based on the discussion on uncertainty modelling and the

specification of a performance requirement in the previous sections and the milling model for control, presented in Section V,
the control problem can now be transformed into the generalized plant framework [26].
In order to derive the generalized plant, first, consider the
following state-space descriptions of the systems
,
,
where
and
, respectively
(34)
(35)
(36)

(32)
Then, the problem in which it is aimed to find a -optimal controller
which stabilizes the milling process in the face of
modelled uncertainties (in and ) while minimizing the peak
magnitude of the weighted control sensitivity; that is a controller which achieves
.
Of course, minimizing the weighted control sensitivity actually
enforces a frequency-dependent upper bound on the magnitude
of control gain
rather than on the magnitudes of the actual control input
. Hence, by estimating the magnitude of
the inputs to the controller, i.e., of the chatter-related tool displacements in - and -direction, an appropriate bound on the
control gain can in practice be chosen such that the actuator
forces are satisfying given saturation limits. In this work, the
weighting transfer function matrix
is chosen to be diagonal, because of the 2-D nature of the control input
, i.e.,
. Moreover, its structure
is chosen such that
is a lead-lag filter with high-pass
characteristics. This means that, for frequencies below the
roll-off frequency
, the control gain is limited by a
certain value and that, for frequencies larger than
, the inputs to the controller are attenuated in order to reduce (undesired) influences on the control action, due to sensor noise and
aliasing effects due to discretization of the continuous-time controller. These (undesired) influences need to be attenuated for
any practical (digital) implementation of the controller. Moreover, in general, an actuator has a limited bandwidth in which it
can operate. Consequently, the roll-off frequency
should be
chosen such that it accounts for the actuator bandwidth limitation, high-frequent measurement noise sensor noise and respects
sampling frequency influences. Based on the discussion above,
weighting function
is written as

where the size of

depends on the order of the Pad approximation which should be chosen such that it approximates the
phase of
up to at least the highest resonance frequency
of the spindle-toolholder and tool modes which are relevant for
chatter. Next, consider the linearized autonomous milling model
as described by (22)

(37)
Then, by adding the uncertainty and performance channel
and
, respectively,
input/output, denoted by
to the system and rearranging terms, the generalized plant is
given as follows:

(38)
with the state vector
input vector
and output

,
, output vector
. The uncertainty channel input
are defined as

where the subscripts and

denote the input/output of the
delay and depth of cut uncertainty, respectively. The definition
of the state-space matrices of the generalized plant can be found
in the appendix of the paper. Combining all the sources of uncertainty as described in Section VI-B, the total uncertainty block
is given as

(33)
denotes the gain of the weighting function. A pole,
where
at frequency
(such that
), is added to obtain
a proper weighting function, necessary for implementation.

(39)
and correFrom the definition of the generalized plant
sponding uncertainty set
it becomes clear that the control

VAN DIJK et al.: ROBUST ACTIVE CHATTER CONTROL IN THE HIGH-SPEED MILLING PROCESS

problem at hand is a robust performance problem which contains structured uncertainties, i.e., the uncertainty is not a full
complex matrix but has specific elements which contain uncertainties. Hence, it is recommended to solve the problem using
-synthesis, which will be briefly discussed in the following
section.
E. Controller Synthesis
In this section, the problem of finding controllers, which satisfy the requirements as defined in Section III, will be discussed.
As discussed in the previous section, the control problem at hand
is a robust performance problem. It is well known that there
is no direct method available yet to synthesis a -optimal controller, see [26]. From the generalized plant model, presented
in the previous section, it can be concluded that we are dealing
with a so-called mixed -synthesis problem, i.e., both complex
and real uncertainties are present. Although mixed -synthesis
can be employed via D,G-K-iteration, it will in general result
in high-order controllers due to high-order fits required for the
G-scales. Moreover, as demonstrated in [13], D,G-K-iteration
does not guarantee an increase in performance. As the general
plant in this work is of relatively high order (since a relatively
high-order Pad approximation is needed to accurately approximate the time delay), the uncertainty in depth-of-cut is considered as a complex uncertainty and controller design is employed
using D-K-iteration. We accept the possible conservatism introduced by considering only complex uncertainties during the
controller design in order to avoid the design of a controller of
even higher order. The complex uncertainty set, denoted by ,
is given as
(40)
By using an upper bound on the structured singular value
,
the controller synthesis problem is transformed into an optimization problem which tries to minimize the peak value over
frequency for this upper bound, namely
(41)
denotes a (frequency-dependent) scaling matrix to
Herein,
represent an upper bound on
and
denotes the set
of functions that are analytic and bounded in the open right half
plane. The optimization problem (41) is iteratively solved for
and . For a fixed scaling transfer matrix , the problem reduces to a standard
synthesis problem, which can be turned
into a convex optimization problem. The optimization problem
for a fixed controller matrix, i.e., the problem of determining
the optimal scaling matrix
for a given frequency can also
be recast in to a convex optimization problem. Both the
as
well as the step in the D-K-iteration can be solved using algorithms from the Robust Control Toolbox of MATLAB [29].
F. Controller Order Reduction
Due to the relatively high order of the Pad approximation
needed to accurately describe the delay term, the resulting con, see [1]. As
trollers will be of relatively high order
discussed above, the spindle dynamics typically has resonances

911

TABLE II
PARAMETERS OF THE AMB MODEL

lying between
Hz, which will generally result in relatively fast controller poles which in turn require
relatively large sample frequencies in a digital implementation.
Hence, for the purpose of the feasibility of the implementation
of the proposed active chatter control methodology in practice,
controller order reduction should be applied.
Balanced truncation is an order reduction procedure which is
often applied to tackle such model reduction problems. However, balanced truncation can only be applied in case the system
to be reduced is stable. The control synthesis procedure discussed in the previous section does, however, not guarantee the
design of stable controllers. To deal with this fact, closed-loop
balanced truncation can be applied, see [30].
The controller states which do not contribute significantly to
the closed-loop input/output of the generalized plant will be removed from the controller using closed-loop balanced truncation. After that, robust performance for the closed-loop system
with the reduced-order controller is evaluated by determining
-values. The acceptable amount of reduction is defined as
the smallest controller order for which
.
As already outlined above, the robust control problem under
consideration has structured uncertainties, which will be solved
via D-K-iteration. Hence, during closed-loop balanced truncation, the D-scaling matrices obtained during controller synthesis
are absorbed into the generalized plant.
VII. RESULTS
In this section, the results for actual controller synthesis for a
realistic model of a high-speed milling machine is addressed. In
order to demonstrate the feasibility of the -synthesis approach
proposed in the previous section, control design is performed
for an illustrative example. Hereto, consider the parameters
of the milling process as given in Table I. The spindle dynamics is modelled, as before, by two decoupled subsystems
consisting of a two mass-spring-damper model in order to
capture the inherent compliance between the actuator/sensor
system (with mass
) and the cutting tool (with
mass
). The parameters of the AMB model are
listed in Table II. Moreover, a four-fluted tool is considered.
Consequently, as already discussed in Section V, a 10-th order
Pad approximation is used to approximate the time delay in
the milling model.
We aim to design controllers that stabilize milling operations
(i.e., guarantee the avoidance of chatter) for two different
spindle speed ranges, for a range of depth-of-cut
which
should be as large as possible for a given performance requirement (i.e., for a given limitation of the control gain). Hereto,
D-K-iteration is employed within a bi-section scheme.
The performance requirement, presented in the previous section, is used to limit the control forces. For an AMB it is important to limit the input current in order not to exceed the
maximum amount of carrying force. Here, we choose to limit

912

IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 20, NO. 4, JULY 2012

Fig. 7. Closed-loop
-values for reduced controllers using closed-loop balanced truncation. Black bars indicate an unstable closed-loop: (a)
rpm; (b)
rpm.

the input current to

2 A. Typical bearing displacements
, for the modelled spindle under consideration, related to onset of chatter, are of the order of
mm. Then,
an upper bound on the control gain due to physical constraints
of the actuator is set to
N/mm and consequently
in (33) is set equal to
mm/A. The
high-frequent roll-off frequency
is chosen to account for
actuator bandwidth limitation (which is typically a few kHz
in case of an AMB), high-frequent sensor noise and aliasing
effects due to discretisation of the continuous-time controller
(sampling frequencies will be typically set to a value larger than
15 kHz). Based on this
is set to
7500 Hz. The
remaining parameter of the weighting filter
is set to
Hz, the additional pole is added such that, first,
is well-posed and, second, the generalized plant fulfills
the rank conditions typically made in the
problem, see [26,
p. 354].
Controllers
are designed for two different ranges
of spindle speeds, namely a relatively small interval given
as
rpm and a relatively large interval
given as
rpm. Controller synthesis using
D-K-iteration yields a 40-th order controller for a maximal
depth of cut of
mm
for
rpm and a 36-th order controller for a maximal depth of cut of
mm
for
rpm. The difference between the controller
orders is due to a difference in the D-scales.
In Fig. 7, the
-value for different controller orders (after
reduction) are depicted for the two ranges of spindle speeds.
In general, closed-loop stability cannot be guaranteed after
controller order reduction. Therefore, before determining the
-value for a specific reduced-order controller, first stability
of the nominal closed-loop system is checked. When the
closed-loop system is stable the corresponding
-value is
determined. From these results, the lowest controller order is
selected for which robust performance can be guaranteed, i.e.,
the lowest controller order for which
. This
yields a 24-th order controller for
and
a 16-th order controller for
rpm. So, it
can be concluded that a significant reduction of the controller
order can be achieved while still guaranteeing robust stability

and performance which warrant the feasibility of the practical

implementation of such controllers in practice.
Frequency response functions (FRF) of the full- and reduced-order controllers together with the inverse of the
frequency bound imposed on the control sensitivity (i.e.,
) are given in Fig. 8. It can be seen that the resulting controllers exhibit highly dynamical characteristics
indicated by the inverse notches in the FRF. Although the
magnitude of the controllers do not exactly fulfill the imposed
bound (since the bound is imposed on the control sensitivity
and not the controller
itself), it can be seen that the magnitude is bounded.
Moreover, it can be seen that the full- and reduced-order
controller have similar FRF magnitudes. Hence, it is expected
that robust performance is maintained under controller-order
reduction. To verify whether robust performance is maintained,
stability lobes diagrams (SLDs) are determined using the
(closed-loop) linearized non-autonomous milling model (18),
as outlined in Section II-D, for the case with the reduced-order
controller and without control. Note that for the case without
control, (12) and (13) are used for the stability analysis. The resulting SLDs can be found in Fig. 9. It can be seen that the SLD
of the controlled milling are shaped such that it contains a lobe
in the desired spindle speed range. Stability is ensured up to a
depth of cut
3.055 mm (an increase of approximately
760% compared to the case without control) and
2.686 mm (an increase of approximately 241% compared to
the case without control) where controllers are designed for
rpm and
rpm, respectively. Herein,
denotes the maximal achievable depth of
cut in the SLD in the desired spindle speed range. Fig. 9 clearly
illustrates the power of the proposed approach, as the SLD is
shaped locally to be able to increase
at a specific spindle
speed (while avoiding chatter and satisfying a specified bound
on the control gain). This is contrary to the application of active
damping which lifts the SLD over the entire spindle speed
range at the cost of high required levels of actuation energy.
Whereas stability is increased at the desired spindle speeds, it
decreases significantly at other spindle speeds.
The characteristics of the controller design and its ability to
shape the SLD in a dedicated fashion can be explained by fur-

VAN DIJK et al.: ROBUST ACTIVE CHATTER CONTROL IN THE HIGH-SPEED MILLING PROCESS

913

Fig. 8. Magnitude of FRF of the full-order (black) and reduced-order (grey) controllers obtained by D-K-iteration for two different range of spindle speeds
and
rpm.

Fig. 9. Stability lobes diagrams, determined using the linearized nonautonomous milling model (18), for reduced-order controllers designed
and
for two different range of spindle speeds,
rpm.

cated at
1667 Hz. A better damped resonance around
2000 Hz is created in case of the larger spindle speed range
rpm , which lies near the edge of the range of desired tooth passing frequencies. As a matter of fact, the location
of the closed-loop resonances, which are dominant for chatter
occurrence, is closely related to the tooth excitation frequencies
for milling operations within the defined spindle
speed ranges. Hence, it can be concluded that, in order to create
a stability lobe at a certain spindle speed, the natural frequency
of the spindle dynamics should be set equal the corresponding
tooth passing excitation frequency, see also [18]. The fact that
a closed-loop spindle resonance situated at a tooth-passing excitation frequency is beneficial for avoiding chatter can be explained as follows. In the milling process the highest depth of
cut can be obtained (corresponding to a peak in the SLD) when
the dynamic chip thickness
is equal
to zero. This relation can be transformed to the frequency domain as follows:
(42)
and
are the Fourier transforms of
and
, respectively. Hence, the difference between
the tooltip displacements of the present and previous cut is
actually characterized by a filter, denoted by
, with zeros
at
. Moreover, for the milling
process, the dominant (chatter) frequency of the perturbation
vibrations lies in general close to the eigenfrequency of the
spindle dynamics. Then, by designing the controller such that
the closed-loop resonance is close to a tooth-passing frequency
and due to the filter properties of the
(in particular the
location of the zeros of
at
-related frequencies), the
dynamic chip thickness is enforced to be zero at the desired
spindle speed. This, in turn, results in a large depth of cut
within the desired spindle speed range and a peak in the SLD
at that spindle speed. So, by applying robust control design
techniques, a controller is obtained which tailors the tooltip
spindle dynamics, such that a resonance is created near a tooth
passing harmonic which in turn results in a peak in the SLD.
where

ther examining the controlled spindle dynamics. The FRF of the

closed-loop tool-tip spindle dynamics
(i.e., the FRF
to
) is given, together with the original (unconfrom
trolled) spindle dynamics, in Fig. 10. While the original (uncontrolled) spindle dynamics only has - and -components
(due to decoupled spindle dynamics), the controlled machine
dynamics also has off-diagonal components. This can be explained by the fact that controller design is performed using the
complete milling model where coupling between - and -direction is introduced by the cutting force model (resulting in a
full matrix in (21) and consequently in a full 2 2 controller
). A striking characteristic displayed in Fig. 10 is the fact
that the controller has tailored the spindle dynamics such that
the resonances are shifted. A similar conclusion was drawn in
[16] for active chatter control in case of the turning process.
For the small spindle speed range
rpm ,
a dominant weakly damped resonance can be seen which is lo-

914

IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 20, NO. 4, JULY 2012

Fig. 10. Controlled

(black) and uncontrolled (open-loop)
and
different range of spindle speeds,
spindle speed range is indicated by the grey area.

(grey) tooltip spindle dynamics for reduced-order controllers designed for the two
rpm. The interval of tooth passing excitation frequencies corresponding to the

Fig. 11. Displacements at the tooltip

for a time domain simulation performed for
nonlinear AMB model. The controller is designed for a spindle speed range of

In the final part of this section some results from time-domain simulations (TDS) will be discussed. Hereto, the nonlinear nonautonomous delay differential equations describing
the total milling model, given by (9), along with the nonlinear bearing model, given by (5), have been implemented in
MATLAB/SIMULINK. The purpose of the TDS is to further
illustrate the effectiveness of the controller design, the benefit
of perturbation feedback and investigate whether it is justified
to apply the linear AMB model for controller design.
As described before, for an AMB it is important to limit the
controller input in order to avoid actuator saturation. In order
to apply perturbation feedback, the periodic solution
with
period time has to be known. Existence of the periodic solution
in case of the nonlinear bearing model is difficult to prove. Here,
the periodic solution is approximated using the finite difference
method, as illustrated for the milling model in [23]. In practice,
the perturbation displacements
can be obtained by using a
chatter detection algorithm based on a parametric model of the
milling process, as, e.g., described in [4]. For the sake of brevity,
we do not describe these algorithms here.
In order to compare the performance of the milling process
with and without chatter control the simulation is performed
for an operating point which is originally unstable (i.e., exhibits
chatter), but is stabilized by means of control. Here, the opera-

31 000 rpm and

2 mm with and without control using the
rpm. (a) Control off. (b) Control on,
.

tion point under consideration has the process parameters

31 000 rpm and
2 mm, which is originally an unstable
working point (see Fig. 9). The results are gathered in Figs. 11
and 12. Fig. 11 presents the displacements at the tooltip with
control off and control on with perturbation feedback
.
Furthermore, the -sampled tool displacements is depicted by
dots. Fig. 12 presents the actuator currents generated by the controller for the case with control on. From the figures, it can be
seen that without control, the amplitude of the displacements
becomes relatively large (approximately 35% of the tool radius
which is chosen as 5 mm), which will result in the tool jumping
in and out of cut resulting in an inferior workpiece quality. When
the controller is switched on, the motion for the initially unstable working point is stabilized, which can be seen from the
-sampled displacements. Moreover, it can be seen that the amplitude of the displacements is considerably smaller for the case
with active chatter control as compared to the uncontrolled case.
The actuator input currents are given in Fig. 12. Due to the fact
that perturbation feedback is applied, the (steady-state) actuator
current, after some transients at the start of the simulation, are
(almost) zero. From the results of the time-domain simulations,
it can be seen that the assumptions, for which the linear AMB
model is a good approximation of the nonlinear AMB model, as
discussed in Section II-B, remain valid. Moreover, the results

VAN DIJK et al.: ROBUST ACTIVE CHATTER CONTROL IN THE HIGH-SPEED MILLING PROCESS

Fig. 12. Actuator input currents

for a time domain simulation per31 000 rpm and
2 mm where perturbation feedback
formed for
is considered. The controller is designed for a spindle speed range of
rpm.

from time-domain simulations clearly demonstrate the benefit

of applying perturbation feedback.

915

In addition, it is shown that the actuator forces, needed to stabilize the milling process, will be zero in steady state in case of
perturbation feedback (i.e., only chatter vibrations are used as a
feedback signal) whereas the actuator forces will be non-zero
in steady state in case of full output feedback. This result is
exploited for a milling model incorporating a nonlinear Active
Magnetic Bearing model, where it is important to limit the actuator input current in order to avoid actuator saturation. Results,
for illustrative examples, clearly illustrate the power of the proposed control methodology. The chatter stability boundary is
locally shaped to stabilize the desired range of working points.
This is contrary to the application of active damping which lifts
the SLD over the entire spindle speed range at the cost of high
required levels of actuation energy. By means of illustrative examples it is shown that this control strategy can render working
points of significantly higher productivity stable.
APPENDIX
Spindle Rotor Dynamics: The state-space matrices of the
spindle rotor dynamic model, given in (4), are defined as follows:

VIII. CONCLUSION
In this paper, an active chatter control design methodology for
the suppression of regenerative chatter in the high-speed milling
process has been developed. The main purpose achieved is the
suppression of chatter (i.e., stabilization of the milling process)
in an a priori specified range of process parameters (spindle
speed and depth of cut), such that working points of significantly higher productivity become feasible while avoiding undesirable chatter vibrations. Herein, the requirement for a priori
stability guarantee for a predefined range of process parameters
is cast into a robust stability requirement. Moreover, a performance requirement is imposed on the control sensitivity in order
to limit the actuator forces. Current chatter control strategies for
the milling process cannot provide such a strong guarantee of a
priori stability for a predefined range of working points. The
control problem is solved via -synthesis using D-K-iteration.

where

for
Generalized Plant: The state-space matrices of the generalized plant, given in (38), are defined as shown in the equation
at the bottom of the page.

916

IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 20, NO. 4, JULY 2012

REFERENCES
[1] N. J. M. van Dijk, N. van de Wouw, E. J. J. Doppenberg, J. A. J. Oosterling, and H. Nijmeijer, Chatter control in the high-speed milling
process using -synthesis, in Proc. Amer. Control Conf., 2010, pp.
61216126.
[2] S. Smith and T. Delio, Sensor-based chatter detection and avoidance
by spindle speed, J. Dyn. Syst., Meas. Control, vol. 114, no. 3, pp.
486492, 1992.
[3] R. E. Haber, C. R. Peres, A. Alique, S. Ros, C. Gonzalez, and J. R.
Alique, Toward intelligent machining: Hierarchical fuzzy control for
the end milling process, IEEE Trans. Control Syst. Technol., vol. 6,
no. 2, pp. 188199, Mar. 1998.
[4] N. J. M. van Dijk, E. J. J. Doppenberg, R. P. H. Faassen, N. van de
Wouw, J. A. J. Oosterling, and H. Nijmeijer, Automatic in-process
chatter avoidance in the high-speed milling process, J. Dyn. Syst.,
Meas. Control, vol. 132, no. 3, p. 031006(14), 2010.
[5] A. Yilmaz, E. AL-Regib, and J. Ni, Machine tool chatter suppression
by multi-level random spindle speed variation, J. Manuf. Sci. Eng.,
vol. 124, no. 2, pp. 208216, 2002.
[6] E. Soliman and F. Ismail, Chatter suppression by adaptive speed modulation, Int. J. Mach. Tools Manuf., vol. 37, no. 3, pp. 355369, 1997.
[7] T. Insperger and G. Stpn, Stability analysis of turning with periodic spindle speed modulation via semi-discretisation, J. Vibr. Control, vol. 10, pp. 18351855, 2004.
[8] K. J. Liu and K. E. Rouch, Optimal passive vibration control of cutting
process stability in milling, J. Mater. Process. Technol., vol. 28, no.
12, pp. 285294, 1991.
[9] Y. S. Tarng, J. Y. Kao, and E. C. Lee, Chatter suppression in
turning operations with a tuned vibration absorber, J. Mater. Process.
Technol., vol. 105, no. 1, pp. 5560, 2000.
[10] J. L. Dohner, J. P. Lauffer, T. D. Hinnerichs, N. Shankar, M. E. Regelbrugge, C. M. Kwan, R. Xu, B. Winterbauer, and K. Bridger, Mitigation of chatter instabilities in milling by active structural control, J.
Sound Vibr., vol. 269, no. 12, pp. 197211, 2004.
[11] S. Kern, C. Ehmann, R. Nordmann, M. Roth, A. Schiffler, and E. Abele,
Active damping of chatter vibrations with an active magnetic bearing
in a motor spindle using -synthesis and an adaptive filter, presented
at the 8th Int. Conf. Motion Vibr. Control, Daejeon, Korea, 2006.
[12] S. Kern, A. Schiffler, R. Nordmann, and E. Abele, Modelling and active damping of a motor spindle with speed-dependent dynamics, presented at the 9th Int. Conf. Vibr. Rotat. Mach., Exeter, Great Britain,
2008.
[13] R. L. Fittro, C. R. Knospe, and L. S. Stephens, -synthesis applied
to the compliance minimization of an active magnetic bearing HSM
spindles thrust axis, Machin. Sci. Technol., vol. 7, no. 1, pp. 1951,
2003.
[14] Y. Zhang and N. D. Sims, Milling workpiece chatter avoidance using
piezoelectric active damping: A feasibility study, Smart Mater. Structures, vol. 14, no. 6, pp. N65N70, 2005.
[15] M. Shiraishi, K. Yamanaka, and H. Fujita, Optimal control of chatter
in turning, Int. J. Mach. Tools Manuf., vol. 31, no. 1, pp. 3143, 1991.
[16] M. Chen and C. R. Knospe, Control approaches to the suppression of
machining chatter using active magnetic bearings, IEEE Trans. Control Syst. Technol., vol. 15, no. 2, pp. 220232, Mar. 2007.
[17] Y. Altintas, Manufacturing Automation. Cambridge, U.K.: Cambridge Univ. Press, 2000.
[18] Y. Altintas and M. Weck, Chatter stability of metal cutting and
grinding, CIRP AnnalsManuf. Technol., vol. 53, no. 2, pp.
619642, 2004.
[19] R. Faassen, Chatter prediction and control for high-speed milling:
Modelling and experiments, Ph.D. dissertation, Dept. Mech. Eng.,
Eindhoven Univ. Technol., Eindhoven, The Netherlands, 2007.
[20] G. Stpn, Modelling nonlinear regenerative effects in metal cutting,
Philosophical Trans. Royal Soc. A: Math., Phys., Eng. Sci., vol. 359,
no. 1781, pp. 739757, 2001.
[21] R. P. H. Faassen, N. van de Wouw, J. A. J. Oosterling, and H. Nijmeijer, Prediction of regenerative chatter by modelling and analysis
of high-speed milling, Int. J. Mach. Tools Manuf., vol. 43, no. 14, pp.
14371446, 2003.
[22] G. Schweitzer, H. Bleuler, and A. Traxler, Active Magnetic Bearings. Zrich: Verlag der Fachvereine, 1994.

[23] R. P. H. Faassen, N. van de Wouw, H. Nijmeijer, and J. A. J. Oosterling, An improved tool path model including periodic delay for chatter
prediction in milling, J. Comput. Nonlinear Dyn., vol. 2, no. 2, pp.
167179, 2007.
[24] M. Farkas, Periodic Motions. Berlin: Springer-Verlag, 1994.
[25] T. Insperger and G. Stpn, Updated semi-discretization method for
periodic delay-differential equations with discrete delay, Int. J. for
Numer. Methods Eng., vol. 61, no. 1, pp. 117141, 2004.
[26] S. Skogestad and I. Postlethwaite, Multivariable Feedback Control:
Analysis and Design, 2nd ed. Chichester, U.K.: Wiley, 2005.
[27] Y.-P. Huang and K. Zhou, Robust stability of uncertain time-delay
systems, IEEE Trans. Autom. Control, vol. 45, no. 11, pp. 21692173,
Nov. 2000.
[28] Z.-Q. Wang, P. Lundstrm, and S. Skogestad, Representation of uncertain time delays in the H1 framework, Int. J. Control, vol. 59, no.
3, pp. 627638, 1994.
[29] The MathWorks, Inc., Natick, MA, Documentation, 2010. [Online].
Available: http://www.mathworks.com
[30] C. Ceton, P. M. R. Wortelboer, and O. H. Bosgra, Frequency weighted
closed-loop balanced reduction, in Proc. 2nd Euro. Control Conf.,
1993, pp. 697701.

Niels van Dijk received the M.Sc. and Ph.D. degrees

in mechanical engineering from the Eindhoven University of Technology, Eindhoven, The Netherlands,
in 2006 and 2011, respectively.
Currently he is with the Mechatronics Technologies Group, Philips Innovation Services, Eindhoven,
The Netherlands. His interests include modelling and
control of manufacturing and high-precision systems.

Nathan van de Wouw received the M.Sc. (with

honors) and Ph.D. degrees in mechanical engineering from the Eindhoven University of
Technology, Eindhoven, The Netherlands, in 1994
and 1999, respectively.
From 1999 to 2010, he has been an Associate
Professor with the Department of Mechanical Engineering, Eindhoven University of Technology. He
was with Philips Applied Technologies, Eindhoven,
The Netherlands, in 2000 and at the Netherlands
Organization for Applied Scientific Research, The
Netherlands, in 2001. He was a visiting Professor with the University of
California Santa Barbara, in 2006/2007 and the University of Melbourne,
Australia, in 2009/2010. He has published a large number of journal and
conference papers and the books Uniform Output Regulation of Nonlinear
Systems: A convergent Dynamics Approach with A.V. Pavlov and H. Nijmeijer
(Birkhauser, 2005) and Stability and Convergence of Mechanical Systems with
Unilateral Constraints with R.I. Leine (Springer-Verlag, 2008).

Ed J. J. Doppenberg received the B.S. degree in

control engineering from the Rotterdam University,
the Netherlands, in 1974 and the M.Sc. Ing-degree
in control engineering from the Delft University of
Technology, Delft, the Netherlands, in 1989.
He has been working from 1987 until 2002
in the development of active noise and vibration
adaptive control system for the transportation sector
(aerospace and automotive applications) at TNO
Science and Industry, Delft, the Netherlands. He was
involved from 2004 to 2010 with the manufacturing
technology development at TNO Science and Technology. In 2011 he joined
the Department of Precision Motion Systems, TNO. He has published several
papers on adaptive (MIMO) control systems and on machining technology,
chatter control and micro milling.

VAN DIJK et al.: ROBUST ACTIVE CHATTER CONTROL IN THE HIGH-SPEED MILLING PROCESS

Han Oosterling received the M.Sc. degree in mechanical engineering from the Eindhoven University
of Technology, Eindhoven, The Netherlands, in 1981.
From 1981 to 1992, he has been with Manufacturing Technology Development, Dutch Navy,
Philips Tool and Die Shop, and Philips Research. In
1991, he started with TNO Science and Technology,
Delft, The Netherlands, as a Research Leader of the
Machining Technology Research Group. In 2011, he
joined the Department Space and Science, TNO. He
has published several papers on coating technology,
machining technology, chatter control, and micro milling.

Henk Nijmeijer (F99) received the M.Sc. and

Ph.D. degrees in mathematics from the University
of Groningen, Groningen, The Netherlands, in 1979
and 1983, respectively.
From 1983 to 2000, he was with the Department
of Applied Mathematics, University of Twente,
Enschede, The Netherlands. Since 1997, he was
also part-time affiliated with the Department of
Mechanical Engineering, Eindhoven University of
Technology, Eindhoven, The Netherlands, where
since 2000, he has been a Full Professor with and

917

chairs the Dynamics and Control section. He has published a large number
of journal and conference papers and several books, including the classical Nonlinear Dynamical Control Systems (Springer, 1990) coauthored
with A. J. van der Schaft, Synchronization of Mechanical Systems (World
Scientific, 2003) coauthored with A. Rodriguez, Dynamics and Bifurcations
of Non-Smooth Mechanical Systems (Springer-Verlag, 2004) coauthored with
R. I. Leine, and Uniform Output Regulation of Nonlinear Systems (Birkhauser
2005) coauthored with A.Pavlov and N. van de Wouw.
Dr. Nijmeijer is Editor-In-Chief of the Journal of Applied Mathematics, corresponding editor of the SIAM Journal on Control and Optimization, and board
member of the International Journal of Control, Automatica, Journal of Dynamical Control Systems, International Journal of Bifurcation and Chaos, International Journal of Robust and Nonlinear Control, Journal of Nonlinear Dynamics, and the Journal of Applied Mathematics and Computer Science. He was
a recipient of the IEE Heaviside premium in 1990. In the 2008 research evaluation of the Dutch Mechanical Engineering Departments the Dynamics and
Control Group was evaluated as excellent regarding all aspects (quality, productivity, relevance, and viability).