Professional Documents
Culture Documents
Lipschitz-continuous nonlinearity
Jan H. Richter
Mara M. Seron
Jos A. De Don
s C Re(s) <
0. For a symmetric matrix A = A
T
, the notation A 0 ( A 0)
means that all eigenvalues of A are positive (negative). Like-
wise, the notation A 0 ( A 0) means that all eigenvalues of
A are non-negative (non-positive). Unless stated otherwise, all
vector norms are Euclidean.
3. RECONFIGURATION PROBLEM FOR LURE
SYSTEMS AFTER ACTUATOR FAILURE
3.1 Nominal Lure system and controller
The nominal Lure system
P
is given by
P
:
_
_
x(t) = Ax(t) + B
v
v(t) + Bu
c
(t) + B
d
d(t)
v(t) = (C
v
x(t))
y(t) = Cx(t)
z(t) = C
z
x(t)
(1)
x(0) = x
0
, where x(t) R
n
is the state, v(t) R
s
the feedback
input, u
c
(t) R
m
the control input, d(t) R
k
is a disturbance
input, y(t) R
r
the measured output, z(t) R
q
the control-
relevant output, A is the system matrix, B
v
the feedback input
matrix, B the control input matrix, B
d
the disturbance input
matrix, C the measured output matrix, and C
z
the relevant
output matrix. All matrices are of compatible dimensions. The
feedback signal v is obtained using the nonlinear characteristic
() : R
s
R
s
.
Assumption 1. (Lipschitz continuity). The characteristic is
globally Lipschitz, namely
L R
+
s.t. x, y : j(x) (y)j Ljx yj. (2)
The Lure system (1) is controlled by means of some given
nominal controller
C
:
_
x
c
(t) = f (x
c
(t), y(t), r(t)), x
c
(0) = x
c0
u
c
(t) = h(x
c
(t), y(t), r(t))
(3)
with the internal state x
c
R
n
c
and the reference input r R
r
.
Assumption 2. (Nominal closed-loop stability). The given nom-
inal closed-loop system (1), (3) is input-to-state stable w.r.t.
the inputs (r, d) and satises given application-specic require-
ments regarding the transient and steady-state response from
(r, d) to z.
Assumption 2 can be achieved by means of controller de-
sign techniques for Lure systems such as those described in
(van de Wouw et al., 2008; de Bruin et al., 2009) and the refer-
ences therein. The nominal closed-loop system is illustrated in
Fig. 2a).
3.2 Actuator failures and faults
Although we are primarily interested in actuator failures (i.e.
blockage), we dene more general actuator faults. All methods
presented below are applicable to the following denition of
faults.
Pf
Pr
d
z
f
u
f
y
f
u
c
y y
c
=
r
d z
!
u
c
y
C
C
r
P
d z
f
u
f
y
f
Cr
r
Pf
a) b) c)
Figure 2. a) Nominal closed-loop system, b) recongured
closed-loop system with new controller, c) recongured
closed-loop system for fault-hiding.
Denition 1. (Actuator faults). An actuator fault f is an event
that changes the nominal input matrix B R
(nm)
to the faulty
input matrix B
f
R
(nm)
of the same dimensions. In this paper,
we assume that faults appear abruptly and remain eective
once they have occurred. The pair ( A, B
f
) is assumed to be
stabilizable.
Faults change the nominal Lure system (1) to the faulty Lure
system
Pf
:
_
_
x
f
(t) = Ax
f
(t) + B
v
v
f
(t) + B
f
u
f
(t) + B
d
d(t)
v
f
(t) = (C
v
x
f
(t))
y
f
(t) = Cx
f
(t)
z
f
(t) = C
z
x
f
(t),
(4)
x
f
(0) = x
0
. For example, columns of B
f
that correspond to
faulty or failed actuators are scaled or set to zero, respectively.
Zero columns of the matrix B
f
represent actuator blockage at
the operating point. To distinguish the faulty system behavior
fromthe nominal behavior, all signals that are aected by faults
are labeled by subscript f .
After the fault, the nominal controller (3) interconnected to the
faulty plant by means of the connections y = y
f
and u
c
= u
f
is
generally not suitable for controlling the faulty Lure system (4).
In particular, in the case of actuator failures, the loop is partially
opened. In the next section we seek a recongured controller
that stabilizes the faulty plant.
3.3 Lure virtual actuator
As a reconguration approach after actuator faults and failures,
we choose an extension of the virtual actuator where the recon-
gured controller is factored into the nominal controller and a
reconguration block. Keeping the nominal controller in the re-
congured closed-loop system has the practical advantage that
minimum-invasive control law adjustments can be formulated
as a reconguration goal, if so desired.
The reconguration block
R
is placed between the faulty
plant (4) and the nominal controller (3) as shown in Fig. 2c).
Together with the faulty plant (4), the reconguration block
R
forms the recongured plant
Pr
= (
Pf
,
R
) to which the
nominal controller (3) is connected via the signal pair (u
c
, y
c
)
(Fig. 2c)). To enable this idea, the reconguration block must
satisfy the following constraint.
Denition 2. (Strict fault-hiding constraint). Consider the nom-
inal system(1) and the faulty system(4). The recongured plant
Pr
satises the strict fault-hiding constraint, if there exists a
suitable particular initial condition
0
of the reconguration
block
R
such that the following relation holds:
t R
+
, d(t) = 0, u
c
(t) : y(t) y
c
(t) = 0.
SAFEPROCESS 2012
August 29-31, 2012. Mexico City, Mexico
223
The reconguration block
R
proposed in this paper is a Lure
virtual actuator
A
:
_
_
x
(t) = A
(t) + B
v
v
(t) + B
u
c
(t), x
(0) = x
0
v
(t) =
_
C
v
(x
(t) + x
f
(t))
_
_
C
v
x
f
(t)
_
u
f
(t) = Mx
(t) + Nu
c
(t)
y
c
(t) = y
f
(t) + Cx
(t)
A
A B
f
M, B
B B
f
N
(5)
(Fig. 3,
R
=
A
). The virtual actuator
A
, whose linear form
was introduced in (Steen, 2005), expresses the dierence
between nominal and recongured dynamics in its state x
(t)
_
=
_
A 0
0 A
_ _
x(t)
x
(t)
_
+
_
B
B
_
u
c
(t) +
_
B
v
v(t)
B
v
v
(t)
_
+
_
B
d
0
_
d(t) (6a)
x(0) = x
0
+ x
0
, x
(0) = x
0
(6b)
v(t) = (C
v
x(t)) (6c)
1
The preservation of stability after introducing an observer is expected but
must be analyzed separately; a generic discussion of the combination of
nonlinear virtual actuators with nonlinear observers is available in (Richter,
2011).
A
1/s C
C
z
B
d
d
B
f
x
f
v
f
u
f
z
f
y
f
Faulty plant
A
1/s C N
M
x
u
c
Lure virtual actuator
v
-
Nominal controller
r
y
c
Figure 3. Lure virtual actuator with recongured closed-loop
system.
v
(t) = (C
v
x(t)) (C
v
( x(t) x
(t))) (6d)
y
c
(t) = C x(t), z
f
(t) = C
z
x(t) C
z
x
(t). (6e)
This model shows that y
c
, the measured output made available
to the controller, depends only on the state x, which is gov-
erned by the nominal dynamics denoted by
P
(and completely
decoupled from the dierence system x
). Observing initial
conditions, y
c
(t) = y(t) follows for all t R
+
if x
0
= 0, which
proves that the Lure virtual actuator satises the strict fault-
hiding constraint. Due to Assumption 2, the interconnection
(
P
,
C
) is ISS.
The dierence system is, however, aected by the dynamics
of the state variable x through the variable v
max
(X)
min
(X)
jB
v
j
2
2
jC
v
j
2
2
X 0 (7)
is satised, where M = YX
1
.
Proof. Due to Assumption 1, the dierence system (state x
)
is governed by a continuously dierentiable and globally Lips-
chitz continuous right-hand side, uniformly in time t. Accord-
ing to (Khalil, 2002, Lemma 4.6), stability of the unforced
dierence system is sucient also for global ISS of the forced
dierence system and the control input u
c
is, hence, omitted
from the following analysis, which can therefore be limited to
the autonomous sub-dynamics obtained by dropping u
c
.
Consider a quadratic Lyapunov function candidate
V(x
) = x
T
Px
(8)
with P = P
T
0 unknown and its time derivative
V(x
) = x
T
P x
+ x
T
Px
= x
T
P( A B
f
M)x
+ x
T
( A B
f
M)
T
Px
+ x
T
PB
v
((C
v
x) (C
v
( x x
)))
+ ((C
v
x) (C
v
( x x
)))
T
B
T
v
Px
. (9)
Further estimation is possible by observing the equivalence
between the following two inequalities
0 ( )
T
P( )
T
P +
T
P
T
P +
T
P, (10)
which is straightforward to verify by expanding and reorder-
ing the rst inequality. Application of the inequality of the
form (10) to Equation (9) with the substitutions = x
and
SAFEPROCESS 2012
August 29-31, 2012. Mexico City, Mexico
224
= B
v
((C
v
x) (C
v
( x x
V(x
) = x
T
P( A B
f
M)x
+ x
T
( A B
f
M)
T
Px
+ x
T
PB
v
((C
v
x) (C
v
( x x
)))
+ ((C
v
x) (C
v
( x x
)))
T
B
T
v
Px
x
T
P( A B
f
M)x
+ x
T
( A B
f
M)
T
Px
+ x
T
Px
+ ((C
v
x) (C
v
( x x
)))
T
B
T
v
PB
v
((C
v
x) (C
v
( x x
))) (11)
x
T
P( A B
f
M)x
+ x
T
( A B
f
M)
T
Px
+ x
T
Px
+ L
2
max
(P)
min
(P)
jB
v
j
2
2
jC
v
j
2
2
x
T
Px
. (12)
Thus, a sucient condition for
V to be negative is the satisfac-
tion of the matrix inequality
P( A B
f
M) + ( A B
f
M)
T
P + P
+L
2
max
(P)
min
(P)
jB
v
j
2
2
jC
v
j
2
2
P 0. (13)
By pre- and post-multiplication with the matrix P
1
from left
and right (a congruence transformation) and substitution of X =
P
1
, Y = MX, one obtains the equivalent matrix inequality (7)
as a sucient condition for quadratic stability of the unforced
dierence system, which is easily extended to an ISS Lyapunov
function of the forced dierence system.
Remark 1. (Conditioning the solution). For all numerical pur-
poses, the inequality (7) should be complemented with an opti-
mization objective that prefers solutions X with small condition
number over solutions with large condition number, such as
minimize
t>0,s>0,X=X
T
0,Y
t s
subject to
s I X t I
X + AX + XA
T
B
f
Y Y
T
B
T
f
+
t
s
L
2
jB
v
j
2
2
jC
v
j
2
2
X 0
Remark 2. (Linear matrix inequality). The inequality (7) can
be turned into a linear matrix inequality (LMI) which is better-
suited to numerical solution by xing the variables s and t of
Remark 1 to constant values s
, t
I X t
I (15)
AX + XA
T
B
f
Y Y
T
B
T
f
+
_
1 +
t
L
2
jB
v
j
2
2
jC
v
j
2
2
_
X 0 (16)
for various t
> s
0
= t
0
= 1, and to
decrement by the rule s
i+1
s
i
/2 and increment by the rule
t
i+1
2 t
i
until feasible solutions are found.
4.2 Algorithm for Lure virtual actuator synthesis
The stabilizing design of the Lure virtual actuator (5) is summa-
rized in Algorithm 1 where, for the sake of completeness, the
models (1) and (4) are used to describe the relevant closed-loop
systems. In experiments, these models are replaced by the real
plant.
In order to implement the Lure virtual actuator, the indicated re-
quired matrix parameters and the function must be provided.
Algorithm 1 Stabilizing Lure virtual actuator synthesis
Require: A, B, B
v
, C, C
v
,
1: Initialize the nominal closed-loop system (1), (3), (5) with
B
f
= B, M = 0, N = I, x(0) = x
0
, x
(0) = 0
2: repeat
3: Run nominal closed-loop system
4: until fault f detected and isolated
5: Construct B
f
, update virtual actuator (5)
6: Set s
0
:= t
0
:= 1, i := 0
7: repeat
8: Solve feasibility problem (14)(16) with s
:= s
i
and
t
:= t
i
for X and Y
9: s
i+1
:= s
i
/2 and t
i+1
:= 2 t
i
10: until problem (14)(16) is feasible
11: Update virtual actuator (5) with M = YX
1
and arbitrary
N
12: Run recongured closed-loop system (3), (4), (5)
Result: Input-to-state stable recongured closed-loop system
Remark 1. (Convergence time). There is no a priori guarantee
for convergence of this algorithm. In other words, there might
be fault cases where problem (14)(16) is infeasible.
5. EXAMPLE
Consider the small-scale model of a mechanical drilling gear
described in (Doris, 2007; van de Wouw et al., 2008) and
illustrated in Fig 4. The gear consists of a long thin drilling shaft
with two drive motors at the upper end and a drilling head on
the lower end. The state vector x =
_
u
l
_
T
used to describe
the drilling gear consists of the relative angular displacement
between the lower and upper ends in rad, the rotation speed
u
of the upper shaft end, and the rotation speed
l
of the lower
shaft end, both measured in rad/s.
w
l
w
u
Motor
Drill-Head
Figure 4. Illustration of the drilling gear.
Around an operating point x = (0, 4.4, 4.4)
T
, the drilling gear
is well described by the Lure system
A =
_
_
0 1 1
J
u
1
J
u
(b
u
b
u
+ b
a
)
b
a
J
u
k
J
l
b
a
J
l
1
J
l
(m b
a
)
_
_
B =
_
_
0
k
m
J
u
0
0 0.8
k
m
J
u
0
_
_
T
, B
v
=
_
0 0
1
J
l
_
T
C
v
=
_
0 0 1
_
, C = I
3
, C
z
= C
SAFEPROCESS 2012
August 29-31, 2012. Mexico City, Mexico
225
15 10 5 0 5 10 15
2
1.5
1
0.5
0
0.5
1
1.5
2
)
Figure 5. Nonlinear friction characteristic.
with the parameter values k
= 0.075 Nm/rad, k
m
=
4.3228 Nm/V, J
u
= 0.4765 kgm
2
, J
l
= 0.035 kgm
2
, b
u
=
2.4245 kgm
2
/(rad s), b
u
= 0.0084 kgm
2
/(rad s), b
a
= 0.1 is a
damping coecient, and m = 0.1 Nms/rad. The characteristic
is shown in Fig. 5, with the steepest part around -1.5 modelled
by the Lipschitz constant L = 10.
The drilling gear is controlled by the nominal state-feedback
controller
K =
_
15.9 1.57 27.6
0 0 0
_
described in (Doris, 2007) that uses only the rst of the two
drive motors and achieves the control objective of stabilization
of the equilibrium without steady-state oscillations.
The considered fault scenario f is the complete failure of the
rst drive motor at t
f
= 0.5 s, which necessitates a control
reallocation to the second drive motor which is weaker and has
a smaller eect on the shaft. The fault changes the input matrix
to
B
f
=
_
0 0 0
0 0.8
k
m
J
u
0
_
T
.
The Lure virtual actuator, which has been inserted into the loop
from plant startup, is automatically redesigned as soon as a
diagnostic algorithm reports the fault at t = t
f
+
D
with the
diagnosis delay
D
= 0.5 s in this example. The LMI solutions
and the obtained gains of the Lure virtual actuator are
X =
_
_
0 0.0004 0
0.0004 0.0992 0.001
0 0.001 0
_
_
Y =
_
0 0 0
0.0002 1.8875 0.0007
_
and
M = 10
_
0 0 0
44696 6 16238
_
N =
_
0 0
1.25 1
_
.
0 0.5 1 1.5 2 2.5 3
0.5
1
1.5
2
,
r
a
d
0 0.5 1 1.5 2 2.5 3
6
4
2
0
2
u
,
r
a
d
/
s
0 0.5 1 1.5 2 2.5 3
1
0
1
2
3
l
,
r
a
d
/
s
0 0.5 1 1.5 2 2.5 3
6
4
2
0
2
u
f
1
,
u
c
1
0 0.5 1 1.5 2 2.5 3
8
6
4
2
0
2
4
Ti me t i n s
u
f
2
,
u
c
2
Figure 6. Closed-loop behavior of controlled drill systembefore
fault and after fault event and recongurations by means of
Lure virtual actuator.
The resulting closed-loop behavior before, during and after the
fault event is shown in Fig. 6, with the drill rod angle and the
top and bottom rotational speeds
u
and
l
shown on top and
the drive motor voltages on the bottom. The time t
f
= 0.5 s of
fault occurrence is marked by a red line in the gure, and the
reconguration time t
f
+
D
= 1 s is marked by a magenta line.
It is clearly visible that prior to fault occurrence (t < t
f
), the rod
is driven by the rst motor only, whereas the virtual actuator
takes the second motor into operation during reconguration
a short while after t
f
. The control commands u
c
issued by
the controller are shown in grey, and the real control input
after translation by the virtual actuator is shown in black. All
signals are stable and the angular displacement converges to
zero, while the upper and lower angular velocities converge to
their operating points.
6. CONCLUSIONS
In this paper, the Lure virtual actuator has been introduced.
Constructive and numerically checkable conditions for stabi-
lizing control reconguration based on a Lipschitz property of
the Lure nonlinearity were given and combined into an online
algorithm. The method was demonstrated by means of an ex-
ample.
Interesting topics for future work are the consideration of per-
formance recovery, explicit combination with set-invariance-
SAFEPROCESS 2012
August 29-31, 2012. Mexico City, Mexico
226
based fault diagnosis methods in the spirit of (Seron et al.,
2012a), and the extension to internal faults that modify the Lure
nonlinearity.
REFERENCES
Ashari, A.E., Sedigh, A.K., and Yazdanpanah, M.J. (2005).
Recongurable control system design using eigenstructure
assignment: static, dynamic and robust approaches. Int. J.
Control, 78(13), 100516.
Blanke, M., Kinnaert, M., Lunze, J., and Staroswiecki, J.
(2006). Diagnosis and Fault-Tolerant Control. Springer,
2nd edition.
Brogliato, B. and Heemels, W.P.M.H. (2009). Observer design
for Lure systems with multivalued mappings: a passivity
approach. IEEE Trans. Automatic Control, 54(8), 1996
2001.
Chen, S., Tao, G., and Joshi, S. (2002). On matching conditions
for adaptive state tracking control of systems with actuator
failures. IEEE Trans. Autom. Control, 47(3), 473478.
Chen, W. and Saif, M. (2007). Adaptive actuator fault detection,
isolation and accommodation in uncertain systems. Int. J.
Control, 80(1), 4563.
Ciubotaru, B., Staroswiecki, M., and Christophe, C. (2006).
Fault tolerant control of the Boeing 747 short-period mode
using the admissible model matching technique. In Proc. 6th
IFAC Sympos. Fault Detect., Superv. and Safety for Techn.
Processes. Beijing, China.
de Bruin, J.C.A., Doris, A., van de Wouw, N., Heemels,
W.P.M.H., and Nijmeijer, H. (2009). Control of mechanical
motion systems with non-collocation of actuation and fric-
tion: a Popov criterion approach for input-to-state stability
and set-valued nonlinearities. Automatica, 45, 405415.
Discenzo, F.M., Unsworth, P.J., Loparo, K.A., and Marcy, H.O.
(1999). Self-diagnosing intelligent motors: a key enabler for
nextgeneration manufacturing systems. In IEE Colloquium
on Intelligent and Self-Validating Sensors, 3/13/4. Oxford,
UK.
Doris, A. (2007). Output-feedback design for non-smooth
mechanical systems: Control synthesis and experiments. Phd
thesis, Technische Universiteit Eindhoven, Den Dolech 2,
5600 MB Eindhoven.
Gao, Z. and Antsaklis, P.J. (1992). Recongurable control
system design via perfect model following. Int. J. Control,
56(4), 783798.
Ingimundarson, A. and Snchez Pea, R.S. (2008). Using the
unfalsied control concept to achieve fault tolerance. In
Proc. 17th IFAC World Congress, 123642. Seoul, Korea.
Jonckheere, E.A. and Yu, G.R. (1999). Propulsion control of
crippled aircraft by H