Professional Documents
Culture Documents
1 s2.0 S1474667016361055 Main
1 s2.0 S1474667016361055 Main
554
does not accommodate the unique behavior of pre- 1.1 Coulomb Friction
sliding faithfully. In fact, the friction force shows The classical Coloumb model of friction is described
hysteresis behavior with nonlocal memory in that by a discontinuous relation between the friction force
regime (Prajogo, 1999, Swevers et al., 2000). The and the relative velocity between the rubbing
Leuven model succeeded in including this type of surfaces. In this model, when the mass that is
hysteresis, but introduced some modelling subjected to friction is slipping, the friction force will
complexities and difficulties (Swevers et al., 2000, remain constant until the motion is reversed.
Lampaert et al., 2002). Finally, the GMS model
manages to overcome those difficulties by modeling 1.2 Stribeck Friction
friction as a Maxwell-Slip model where the slip The Stribeck friction consists of (i) a function s(v)
elements satisfy a certain, new state equation. As a that is decreasing in the velocity and bounded by an
result, the GMS model is able to predict not only the upper limit at zero velocity equal to the static friction
friction behavior in pre-sliding regime with nonlocal force Fs, and a lower limit equal to the Coulomb
memory hysteresis, but also the friction force in force Fc., and (ii) a viscous friction part. In this
sliding regime, which behaves in a similar way to approach, the constant portion of the Coulomb model
that in the LuGre model. is replaced by Stribeck function. Moreover, in order
to overcome the jump discontinuity of the Coloumb
An important feature in the friction identification
model, at v=0, that jump is replaced by a line of finite
procedure in this paper is that we will use only a
slope, up to a very small threshold , as shown in
single set of experiments to identify all the unknown
Figure 1.
parameters together, using a suitable optimization
method, namely the Nelder-Mead Simplex algorithm.
Once the friction models have been optimized,
position control incorporating friction compensation
is performed. For this purpose, the inertial force and
friction behavior are compensated for using a
feedforward control, while a simple (PID) feedback
part is included to track set-point changes and to
suppress unmeasured disturbances.
Figure 1. Friction curve of Stribeck models
In the following, section 2 formulates the DC motor
torque balance, and outlines the friction models used
in this investigation. Section 3 describes the 1.3 LuGre Friction
experimental apparatus, while section 4 discusses the The above two methods have proven to be
identification procedure. The identification results impractical for friction compensation at motion stop
are discussed in section 5. Thereafter, the friction and inversion, where the worst effects due to friction,
compensation scheme is sketched in section 6, while namely stick-slip motion, could arise. Motion never
the compensation results are discussed in section 7. starts or stops abruptly and pre-sliding displacements
Finally, appropriate conclusions are drawn in Section are actually observed (Prajogo, 1999, Swevers et al.,
8. 2000). Consequently, two different friction regimes
can be distinguished, i.e. the pre-sliding and the gross
sliding regimes, as explained in section 1. The LuGre
2. MOTOR TORQUES model (Canudas de Wit et al., 1995) was the first
In general, a DC motor can be viewed as a black box formulation that could effect a smooth transition
with two inputs: current and load torque, and an between those two regimes, i.e. without recourse to
output angular displacement (or velocity). Torque switching functions. It, furthermore, accounts for
balance for a DC motor can be written as: other friction characteristics such as the breakaway
force and its dynamics. The model achieves this by
τm = τi + τf +τ0 , (1) introducing a state variable, representing the average
where τm is the motor torque generated by the deflection of elastic bristles (representing surface
asperities) under the action of a tangential force,
amplifier current, τi is the inertial torque from motor
together with a state equation, governing this
armature and shaft, τf is the friction torque and τ0 is
variable’s dynamics and friction equation.
the load torque.
The LuGre model is very popular in the domain of
Due to the limitation of the amplifier current in the
control and simulation of friction due to its simplicity
motor, it can be shown that the motor torque τm = Km and the integration it affords of pre-sliding and
(i,isat); and the inertial torque τ i = J mθ . Motor sliding into one model. However, it has been
torque is bounded due to the current saturation limit subjected to important criticism (Swevers et al.,
of the servo amplifier. The saturation function is 2000) in regard to its failure to model pre-sliding/pre-
represented by (i,isat), where it has a constant rolling hysteresis with nonlocal memory. The latter
authors proposed an extension in form of the Leuven
value for |i| > isat; and it has a slope of 1 for |i| ≤ isat.
model and a subsequent improvement (Lampaert et
As for the friction torque τf in equation (1), which is al., 2002), however, not without introducing further
our main concern, several models were used as difficulties.
described below.
555
1.4 GMS Friction 4. IDENTIFICATION OF FRICTION MODELS
The Generalized Maxwell-Slip (GMS) friction model The identification experiment was carried out by
(Al-Bender et al., 2004, Lampaert et al., 2003) is a applying an input velocity signal. A special input
qualitatively new formulation of the rate-state command was designed for this purpose. This
approach of the LuGre and the Leuven models. The command signal is composed of a band limited
GMS model retains the original Maxwell-slip model random signal with cutoff frequency of 4 Hz, which
structure (see e.g., Iwan, 1966), which is a parallel is enveloped by a certain signal. This envelope signal
connection of different elementary slip-blocks and is composed of the three first non-zero terms of the
springs, but replaces the simple Coulomb law Fourier series of a rectangular signal, which has
governing each block, by another state equation to period of 10 seconds. The purpose of enveloping this
account for sliding dynamics. Thus, the friction force input signal is to emphasize the behavior of
is given as the summation of the outputs of the N presliding friction in the experiment, which
elementary state models. The dynamic behavior of corresponds to the friction at low velocity and low
each elementary block is represented by one of two displacement. The signal and its corresponding
equations (in which Fi is the elementary friction envelope signal are shown in Figure 3. These signals
force, ki is the elementary spring constant, and v is were applied through a dSPACE-1104 acquisition
the velocity): card to the servo amplifier. The real current input to
• If the model is sticking: the motor, which is assumed to correspond to the
dFi torque was measured and recorded using the same
= ki v ; (2)
dt acquisition unit.
• If the model is slipping: 30
dFi F
= sgn(v) ⋅ C ⋅ α i + i (3)
dt s(v) 20
10
Each elementary model (or block) corresponds to a
velocity (rad/sec)
approaches zero value. When the asperity is sticking Figure 3. Velocity input signal
the elementary model will act as a spring with
stiffness of ki.
5. IDENTIFICATION RESULTS
3. EXPERIMENTAL SETUP
Fifty thousand points, at one millisecond time
Experimental identification of friction in a DC motor sampling, were collected for each test. The first 10
was performed on ABB motor type M19-S, with seconds were used as a training set and the remaining
maximum rated torque of 0.49 Nm/amp and armature 40 seconds were used as testing set. The Nelder-
inertia of 0.001 kgm2. To allow a straightforward Mead Simplex algorithm was used for optimizing the
position and velocity data collection on the motor, an parameters in each model.
incremental angular encoder was connected to the
Since the system being identified here is (highly)
shaft through a toothed belt and pulleys. The pulleys
nonlinear, error quantification techniques pertaining
give a reduction ratio of 1:3 to increase the sensitivity
to linear systems, such cross-correlation between
of the encoder, whose resolution is 5000 pulses per
input, output and residuals, etc., cannot be used.
revolution. Figure 2 shows a schematic of the
experimental setup. Instead, as a measure of performance, the normalized
mean square error was used, which is defined by:
N
MSE ( yˆ ) = 100 ( yˆi − yi )2 (4)
Nσ 2y
i =1
where y is the output (in this case is the friction
force), σy2 is its variance and the caret denotes an
estimated quantity.
The performance values can be seen in Table 1. The
values between brackets represent the peak-to-peak
errors that are normalized by the RMS value of the
actual torques. In this table the performances are
divided into two groups, one for high velocities and
Figure 2. Schematic of the setup the other for low velocities. Each group is calculated
based on its envelope signal. High velocity
556
performance values are calculated at the instances respectively. It can be clearly seen that the
when the envelope velocity is high, and low velocity identification results of the classical friction models
performance values are calculated when the envelope fall far from the actual value in some regions,
velocity is low. especially in the low velocity regions (approximately
before time 3 seconds), which have lower angular
The estimated torques of the modeling technique and
displacement, i.e. more pre-sliding regime portion.
the measured torque are shown in Figure 4, for
Coulomb, Stribeck-Coulomb, LuGre and GMS (with
4 elements), from left to right and top to bottom
Coloum b Stribeck-Coloum b
5 5
Torque (Nm)
0 0
-5 -5
2 2.2 2.4 2.6 2.8 3 2 2.2 2.4 2.6 2.8 3
LuGre GMS
4
5 5
Torque (Nm)
0 0
-5 -5
2 2.2 2.4 2.6 2.8 3 2 2.2 2.4 2.6 2.8 3
time (s) time (s)
Figure 4. Measured and estimated torques for different models. Dotted lines represent the measured torques, and
solid lines are the estimated torques.
Error of Coloumb Error of Stribeck-Coloumb sliding regime is shown as the superiority of the
5 5 GMS model, while it does not lose its ability to
Error (Nm)
557
included in the feedforward loop. The control Low Stroke Random Reference Signal
scheme is depicted in Figure 7. 0.02
Amplitude (rad)
0
The feedback loop, which is required to track set-
point changes and to suppress unmeasured -0.02
disturbances, is chosen to be a PID controller with
gains Kp = 200 Nm/rad, Ki = 0 Nm/rad.s, and Kd = -0.04
0 5 10 15 20 25
x1000
0
compensation, and used with the same values for all 0
x1000
0 0
-2 -2
-4 -4
0 5 10 15 20 25 0 5 10 15 20 25
0 0
are given in Table 2. -2 -2
-4 -4
Friction compensation based on the Coloumb -6 -6
0 5 10 15 20 25 0 5 10 15 20 25
model yields large errors, and is almost identical to
LuGre GMS 4
the scheme that has no feedforward compensation. 6 6
Tracking Error (rad)
4 4
Since the reference signal has a very small stroke
2 2
x1000
558
random signal was set to π rad (or half of the the hysteresis behavior with non-local memory
shaft’s rotation). In this case, the pre-sliding friction for the presliding regime friction, while
is not expected to dominate the overall behavior. modeling the gross sliding also.
559