Beruflich Dokumente
Kultur Dokumente
a r t i c l e i n f o
Keywords:
State estimation
Kalman lters
Multibody dynamics
Non-linear ltering
abstract
The aim of this work is to provide a thorough research on the implementation of some non-linear
Kalman lters (KF) using multibody (MB) models and to compare their performances in terms
of accuracy and computational cost. The lters considered in this study are the extended KF (EKF) in
its continuous form, the unscented KF (UKF) and the spherical simplex unscented KF (SSUKF).
The MB formulation taken into consideration to convert the differential algebraic equations (DAE) of
the MB model into the ordinary differential equations (ODE) required by the lters is a state-space
reduction method known as projection matrix-R method. Additionally, both implicit and explicit
integration schemes are used to evaluate the impact of explicit integrators over implicit integrators in
terms of accuracy, stability and computational cost. However, state estimation through KFs is a closedloop estimation correcting the model drift according to the difference between the predicted
measurement and the actual measurement, what limits the interest in using implicit integrators
despite being commonly employed in MB simulations. Performance comparisons of all the aforementioned non-linear observers have been carried out in simulation on a 5-bar linkage. The mechanism
parameters have been obtained from an experimental 5-bar linkage and the sensor characteristics from
off-the-shelf sensors to reproduce a realistic simulation. The results should highlight useful clues for
the choice of the most suitable lters and integration schemes for the aforementioned MB formulation.
& 2013 Elsevier Ltd. All rights reserved.
1. Introduction
In recent years, the interest in state estimation in mechanical
systems, and in particular in MB systems, has increased. A state
observer is a dynamical system that employs the model of a real
system in order to provide estimates of its internal states, given
measurements of the inputs and outputs of the real system. It is
typically a computer-implemented mathematical model. This allows
to extract otherwise unmeasurable variables of the real system [1]. In
practice, the knowledge of the system state allows synthesizing
effective state controllers, replacing expensive sensors with virtual
sensors, improving reliability by making the controlled system robust
to sensor fault. On the one hand, many works in the literature address
the synthesis of optimal observers for linear mechanical systems
through the KF [2,3]. On the other hand, when non-linear mechanical
systems are considered, only sub-optimal approaches based on the
LKF (linearized KF) have usually been adopted to ensure highfrequency and hard real-time estimation [46]. Indeed, up to now,
the use of other types of non-linear observers in MB systems has only
0020-7462/$ - see front matter & 2013 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.ijnonlinmec.2013.01.016
Please cite this article as: R. Pastorino, et al., State estimation using multibody models and non-linear Kalman lters, International
Journal of Non-Linear Mechanics (2013), http://dx.doi.org/10.1016/j.ijnonlinmec.2013.01.016i
wt N 0,Qt,
vt N 0,Rt,
1
2
_
where x(t) represents the states of the system, xt
is the time
derivative of the state vector, f(t) is the dynamics system function,
y(t) contains the measurements and h(t) is the measurement
sensitivity matrix. The differential equation for the recursive
estimation of the states is presented as follows:
^
^
x^_ Efxt,t
wt Ktytyt
^ Ktytyt
^
Fxt
^ Ehxt,t
^
^
yt
vt Hxt
where G(t) is the coupling matrix between the process noises and
the states of the system. The case in which f(t) is a non-linear
function will be treated in Section 2.2.
The estimation problem can also be formulated in discretetime form which is the most suitable form for hardware implementation as the estimation is divided in two parts depending on
the availability of sensor information. The system nominal and
measurement models, which are now linear difference equations
with independent states, are shown in Eqs. (7) and (8). The
aforementioned assumptions have also been made for the process
and measurement noises
xk 1 /k xk wk ,
yk hk xk vk ,
wk N 0,Qk
vk N 0,Rk
7
8
T
T
T
^
^
P
k 1 Exk 1 x k 1 xk 1 x k 1 Uk Pk Uk Gk Q k Gk
10
11
Pk 1 P
k1
12
x^ k 1 x^ k 1 x^ k 1 K k 1 yk 1 y^ k 1
13
Pk 1 Pk 1 P
k 1 K k 1 Hk 1 Pk 1
14
^ are
where E is the expectation, Kt is the Kalman gain matrix, yt
the predicted measurements and F(t) is the linearization of f(t)
(here Ft ft as the latter is linear). The Kalman gain and the
predicted measurements are calculated using the following equations:
K k 1 Exk 1 x^ k 1 yk 1 y^ k 1 T
Kt PtHT tR1 t
y^ k 1 Ehk 1 x^ k 1 vk Hk 1 x^ k 1
15
16
Please cite this article as: R. Pastorino, et al., State estimation using multibody models and non-linear Kalman lters, International
Journal of Non-Linear Mechanics (2013), http://dx.doi.org/10.1016/j.ijnonlinmec.2013.01.016i
17
^ Ehxt,t
^
^
yt
vt C hxt
18
vk 1 /k vk
19
^
^
^
^
x^_ Efxt,t
wt Ktytyt
Cfxt
Ktytyt
P
i 1, . . . ,L
vk i
k
k i
20
>
: x^ gp
P
i L 1, . . . ,2L
k
k i
p
where vk i is the i-th sigma-point, g L l, l a2 L k, a and
p
k are user-dened tuning parameters, is the matrix squareroot using lower triangular Cholesky decomposition and i
represents the i-th column. 0 o a r 1 is a scaling factor dening
the extension of the spread of the sigma-points around the mean
of the estimates. k is another scaling factor, usually set to 0. Fig. 1
exemplies the sigma-point set for a 2-dimensional Gaussian
random variable (nsp 5). The size of the points represents their
possible weights.
21
nX
sp 1
wm
i vk 1 i
22
i0
P^ k 1 Exk 1 x^ k 1 xk 1 x^ k 1 T
nX
sp 1
T
^
^
wci v
k 1 ix k 1 vk 1 ix k 1
23
i0
c
m
2
c
m
where wm
0 l=L l, w0 w0 1a b, wi wi 1=2L l
for i 1 . . . nsp 1 and b is a scaling factor used to control the
weighting of the zeroth sigma-point. Then, if no information is
available from the sensors, Eqs. (11) and (12) are used.
If some information is available, the measurement-update
equations, which also differ substantially from the ones of the
EKF, are employed. The Kalman gain is obtained using the
weighted covariances as shown below
K k 1 Exk 1 x^ k 1 yk 1 y^ k 1 T
nX
sp 1
T
^
^
wci v
k 1 ix k 1 Y k 1 iy k 1
i0
nX
sp 1
T
^
^
wci Y
k 1 iy k 1 Y k 1 iy k 1
24
i0
nX
sp 1
wci Y
k 1 i
25
i0
Y
k 1 hk 1 vk 1
26
T
Pk 1 P
k 1 K k 1 Pyk yk K k 1
27
Please cite this article as: R. Pastorino, et al., State estimation using multibody models and non-linear Kalman lters, International
Journal of Non-Linear Mechanics (2013), http://dx.doi.org/10.1016/j.ijnonlinmec.2013.01.016i
v11 1= 2w1 and v12 1= 2w1 . All the points (apart from the
zeroth
point) lie on a hypersphere, the radius of which is
p
L=1w0 , as it is exemplied in Fig. 2. The second difference
with respect to the UKF is related to the weights used in (22)(24)
where only one set of weights is employed: wi for i 0, . . . ,nsp 1.
30
Uq q_ 0
31
The second base equation, Eq. (32), is the time-derivative of Eq. (30)
q Rz R_ z_
32
33
Mq UTq l Q
U0
29
Dt
2
Dt
2
v_ k v_ k 1
34
k1 k2
35
with
k1 ftk ,vk ,
_
and f(t) is a system of rst-order ODE
where v t qtqt
equations coming from reshaping the second-order ODE equations
obtained with the matrix-R method and using variable duplication. Eq. (36) demonstrates how f(t) is calculated
_
q F t,q, q
_
) vt
"
_
wt
wt
"
#
F t,q,w
fvt,t
q_
36
Please cite this article as: R. Pastorino, et al., State estimation using multibody models and non-linear Kalman lters, International
Journal of Non-Linear Mechanics (2013), http://dx.doi.org/10.1016/j.ijnonlinmec.2013.01.016i
# "
# "
#
wt
_
wt
zt
fxt,t
1
_
_
RT MR1 RT Q MR_ z
wt
M Q
38
Taking into account the chosen state vector, Eq. (17) can be
written as shown in Eq. (39)
# " z#
" # "
^
w
z^_
K
^
39
T
T
1
w yy
^
R MR R Q MR_ w
^_
w
K
z
40
z^ k 1 z^ k 1 z^ k 1 K k 1 yk 1 y^ k 1
41
Once the state vector has been dened, the lters equations can
be applied. First, the set of sigma-point is calculated using Eq. (20)
for the UKF or Eq. (28) for the SSUKF. Then, each sigma-point is
propagated through the non-linear discrete-time system function
(Eq. (21)), in other words, a function evaluation of the MB model
is performed for each sigma-point. The posterior independent
acceleration vector z^ k 1 is obtained using Eq. (22). Then, it is used
to assess the posterior independent position (zk 1 ) and velocity
(z_ k 1 ) vectors by means of the TR. If no information from the
sensors is available, the covariance matrix is obtained through
Eq. (10) and the time-update equations (11) and (12) are applied.
If any sensor data are available, the measurement-update equations (24)(27) and (41) are applied.
qT xB yB xC yC xD yD f1 f2
42
Please cite this article as: R. Pastorino, et al., State estimation using multibody models and non-linear Kalman lters, International
Journal of Non-Linear Mechanics (2013), http://dx.doi.org/10.1016/j.ijnonlinmec.2013.01.016i
Velocity (rad/s)
4
2
0
2
4
6
0
0.5
1.5
Time (s)
2.5
Fig. 6. Angular velocity of the left crank. (For interpretation of the references to
color in this gure legend, the reader is referred to the web version of this article.)
20
Acceleration (rad/s2)
10
0
10
20
0
1.5
Time (s)
2.5
Fig. 7. Angular acceleration of the left crank. (For interpretation of the references
to color in this gure legend, the reader is referred to the web version of this
article.)
Position (rad)
0.5
Table 1
Color code for the gures of the 5-bar linkage.
0
1
2
3
0
0.5
1.5
Time (s)
2.5
Fig. 5. Angle of the left crank. (For interpretation of the references to color in this
gure legend, the reader is referred to the web version of this article.)
observers. The color code for the gures is shown in Table 1. The
behavior discrepancies between the real mechanism and the
system nominal model without observer (green) are clearly
demonstrated. It is worth pointing out that all the observers
follow the motion of the real mechanism with accuracy.
5.2. Comparison of observer performances
As previously mentioned, the motion of the real mechanism is
completely known as it is simulated. To provide a clear evaluation
of the observer accuracy, the estimation errors and their rst and
second time-derivatives are compared. The sensors for all the
lters are the two crank encoders. Figs. 810 show the errors in
positions, velocities and accelerations of the left crank for an
integration time step Dti of 2 ms and an update time step of the
sensors Dts of 2 ms. This means that data from each sensor
is available at each integration time step. In a similar way,
Figs. 1113 show the same errors but for an update time step of
Color
Filter
MB formu.
Integ.
Note
Gray
Green
Black
Red
Brown
Blue
No
EKF
UKF
SSUKF
SSUKF
Matrix-R
Matrix-R
Matrix-R
Matrix-R
Matrix-R
TR
TR
TR
TR
RK2
Encoder noise
System model
the sensors Dts of 6 ms. The CPU time for all the mentioned lters
with respect to the lter with the lower computational cost is
presented in Table 2. The RMSE (root-mean-square error), where
the error is the difference between the predicted measurements
and the actual measurements, is also presented in Table 2. It is
worth pointing out that, when the update time step of the sensors
is equal to the integration time step, the most accurate lter is
the EKF [10], with an RMSE slightly superior to the noise standard
deviation (103 rad). The rest of lters present also very similar
results, while correcting only the independent acceleration
vector. However, when the update time step of the sensors is
greater than the integration time step, the RMSE of the EKF in its
continuous form increases dramatically compared to the RMSE of
the other lters. When the update time step of the sensors
increases, the corrections of the lter are less frequent meaning
that less information on the behavior of the real mechanism is
available. Therefore, the model of the lter has more difculties to
follow accurately the motion of the real mechanism and the RMSE
increases. The bad behavior of the continuous-form EKF in
multi-rate situations is in part due to the assumptions made in
Section 2.2. Indeed, in the theory of the EKF in continuous form, it
Please cite this article as: R. Pastorino, et al., State estimation using multibody models and non-linear Kalman lters, International
Journal of Non-Linear Mechanics (2013), http://dx.doi.org/10.1016/j.ijnonlinmec.2013.01.016i
0.1
x 103
Error (rad/s)
0.05
Error (rad)
2
0
0
0.05
0.1
4
0
0.5
1.5
Time (s)
2.5
0.5
1.5
Time (s)
2.5
Fig. 12. Errors of the lters for the angular velocity of the left crank.
Fig. 8. Errors of the lters for the angle of the left crank.
1
0.5
Error (rad/s2)
Error (rad/s)
0.05
0
0.5
0.05
0.5
1.5
Time (s)
2.5
1.5
Time (s)
2.5
Table 2
CPU times and RMSE for Dti 2 ms.
0.5
Error (rad/s2)
0.5
Fig. 13. Errors of the lters for the angular acceleration of the left crank.
Fig. 9. Errors of the lters for the angular velocity of the left crank.
0.5
0
0.5
1.5
Time (s)
2.5
0.01
0.005
0
0.005
0.01
0.015
0
0.5
1.5
Time (s)
2.5
Filter
MB formu.
Integ.
CPU (%)
RMSE
Dts ms
EKF
Matrix-R
TR
SSUKF
Matrix-R
RK2
SSUKF
Matrix-R
TR
UKF
Matrix-R
TR
100
103
115
109
176
169
209
203
1.36 10 3
9.22 10 3
2.16 10 3
3.17 10 3
2.18 10 3
5.35 10 3
1.98 10 3
3.1 10 3
2
6
2
6
2
6
2
6
Fig. 10. Errors of the lters for the angular acceleration of the left crank.
Error (rad)
Fig. 11. Errors of the lters for the angle of the left crank.
Please cite this article as: R. Pastorino, et al., State estimation using multibody models and non-linear Kalman lters, International
Journal of Non-Linear Mechanics (2013), http://dx.doi.org/10.1016/j.ijnonlinmec.2013.01.016i
6. Conclusions
This work presents the theoretical development and practical
implementation of non-linear state observers using MB models.
Several types of Kalman lters have been considered: the EKF in
its continuous form and some SPKFs (UKF, SSUKF). Although these
SPKFs outperform the accuracy of the EKF in its continuous form
by using better approximations of system non-linearities, the
sigma-points have brought about an additional computational
cost. However, they present several advantages over the latter:
easy implementation, parallel computing structure that helps to
reach real-time. Implicit integrators have proven to improve
observer accuracy insignicantly when used in closed-loop state
estimation so explicit integrators are used to help reducing the
computational costs. Indeed, the difference between the predicted
measurement and the actual measurement (i.e. the innovation) is
adopted to correct the estimates and allows the prevention of
numerical energy loss due to the integration. In light of the
results, the choice of the most suitable set of lter, MB formulation
and integrator depends on the application requirements and is a
trade-off between estimation accuracy and computational efciency. Future works will be devoted to deeper investigations on
wider classes of MB formulations and lters, and to the experimental validation of the proposed observers.
Acknowledgments
The authors would like to thank the Spanish Ministry of Science
and Innovation and ERDF funds through the Grant TRA2009-09314
for its support in this research.
References
[1] A. Radke, Z. Gao, A survey of state and disturbance observers for practitioners,
in: Proceedings of the American Control Conference, Minneapolis, Minnesota,
USA, 2006, pp. 51835188.
[2] R.E. Kalman, A new approach to linear ltering and prediction problems,
Transactions of the ASMEJournal of Basic Engineering 82 (Series D) (1960)
3545.
[3] M.S. Grewal, A.P. Andrews, Kalman FilteringTheory and Practice Using
MATLAB, 3rd ed., John Wiley & Sons, Inc, 2008.
[4] R. Caracciolo, D. Richiedei, A. Trevisani, Experimental validation of a modelbased robust controller for multi-body mechanisms with exible links,
Multibody System Dynamics 20 (2) (2008) 129145, http://dx.doi.org/
10.1007/s11044-008-9113-7.
[5] R. Caracciolo, D. Richiedei, A. Trevisani, Robust piecewise-linear state observers for exible link mechanisms, Journal of Dynamic Systems, Measurement
and Control 130(3) (2008), 031011, http://dx.doi.org/10.1115/1.2909600.
[6] R. Caracciolo, D. Richiedei, A. Trevisani, Design and experimental validation of
piecewise-linear state observers for exible link mechanisms, Meccanica 41
(2006) 623637.
[7] A. Barreiro, E. Delgado, J. Cuadrado, D. Dopico, Extended-Kalman-lter
observers for multibody dynamical systems, in: Sixth EUROMECH Nonlinear
Dynamics Conference ENOC, Saint Petersburg, Russia, 2008.
[8] J. Cuadrado, D. Dopico, A. Barreiro, E. Delgado, Real-time state observers
based on multibody models and the extended Kalman lter, in: Proceedings
of the Fourth Asian Conference on Multibody Dynamics, no. 127, Jeju, Korea,
2008.
[9] J. Cuadrado, D. Dopico, J.A. Perez, R. Pastorino, Inuence of the sensored
magnitude in the performance of observers based on multibody models and
the extended Kalman lter, in: Proceedings of the Multibody Dynamics
ECCOMAS Thematic Conference, Warsaw, Poland, 2009.
[10] J. Cuadrado, D. Dopico, A. Barreiro, E. Delgado, Real-time state observers based
on multibody models and the extended Kalman lter, Journal of Mechanical
Science and Technology 23 (4) (2009) 894900, http://dx.doi.org/10.1007/
s12206-009-0308-5.
[11] R.E. Kalman, R.S. Bucy, New results in linear ltering and prediction theory,
Transactions of the ASME. Series D, Journal of Basic Engineering 83 (1961)
95107.
[12] S.J. Julier, J.K. Uhlmann, H.F. Durrant-Whyte, A new approach for ltering
nonlinear systems, in: Proceedings of the American Control Conference, vol.
3, 1995, pp. 16281632, http://dx.doi.org/10.1109/ACC.1995.529783.
[13] S.J. Julier, J.K. Uhlmann, A new extension of the Kalman lter to nonlinear
systems, in: The Eleventh International Symposium on Aerospace/Defense
Sensing, Simulation and Controls, Orlando, FL, USA, 1997, pp. 182193.
[14] S.J. Julier, J.K. Uhlmann, Unscented ltering and nonlinear estimation, in:
Proceedings of the IEEE, vol. 92, 2004, pp. 401422.
[15] M. Nrgaard, N.K. Poulsen, O. Ravn, New developments in state estimation for
nonlinear systems, Automatica 36 (11) (2000) 16271638, http://dx.doi.org/
10.1016/S0005-1098(00)00089-3.
[16] S. Julier, J. Uhlmann, Reduced sigma point lters for the propagation of means
and covariances through nonlinear transformations, in: Proceedings of the
American Control Conference, vol. 2, 2002, pp. 887892.
[17] S.J. Julier, The spherical simplex unscented transformation, in: Proceedings of
the American Control Conference, vol. 3, 2003, pp. 24302434.
[18] R. Van der Merwe, E. Wan, The square-root unscented Kalman lter for state
and parameter-estimation, in: IEEE International Conference on Acoustics,
Speech, and Signal Processing, vol. 6, Salt Lake City, USA, 2001, pp. 34613464.
[19] R. Van der Merwe, E.A. Wan, S. Julier, Sigma-point Kalman lters for
nonlinear estimation and sensor-fusion-applications to integrated navigation, in: AIAA Guidance, Navigation, and Control Conference and Exhibit,
Rhode Island, USA, 2004.
[20] J. Garca de Jalon, E. Bayo, Kinematic and Dynamic Simulation of Multibody
Systems: The Real-Time Challenge, Springer-Verlag, 1994.
[21] J. Cuadrado, D. Dopico, M. A. Naya, R. Pastorino, Automotive observers based
on multibody models and the extended Kalman lter, in: The First Joint
International Conference on Multibody System Dynamics, Lappeenranta,
Finland, 2010.
[22] J. Cuadrado, D. Dopico, J.A. Perez, R. Pastorino, Automotive observers based
on multibody models and the extended Kalman lter, Multibody System
Dynamics 27 (1) (2011) 319, http://dx.doi.org/10.1007/s11044-011-9251-1.
Please cite this article as: R. Pastorino, et al., State estimation using multibody models and non-linear Kalman lters, International
Journal of Non-Linear Mechanics (2013), http://dx.doi.org/10.1016/j.ijnonlinmec.2013.01.016i