Professional Documents
Culture Documents
Abstract: This paper studies the synchronization problem of a dynamical network with event-
based communication, where each node communicates to its neighbors only when an event-
triggering condition is fulfilled. In order to achieve asymptotic synchronization as well as to
prevent the occurrence of Zeno behavior, estimators are introduced into each node to estimate
the current state of its neighbors and of its own. Then, with the assistance of these estimators,
a distributed event-triggering rule is designed, which only depends on the information that the
node can obtain, and thus can be implemented in a decentralized way. Finally, a numerical
example is given to show the effectiveness of the proposed results.
achieved. But there is no evidence that the designed event- while how to determine the time sequence {tiki }, ki ∈ Z+
triggering rule can prevent Zeno behaviors. Moreover, in when node i should communicate to its neighbors (i.e.,
Demir and Lunze (2012), estimators were introduced into samples its state and then sends the sampled value to its
each node, and were used to design a decentralized event- neighbors) such that network (1) can achieve synchroniza-
triggering rule with a fixed threshold, but only bounded tion asymptotically.
synchronization was obtained.
To reduce the unnecessary communication between inter-
In this paper, we will study asymptotic synchronization connected nodes as well as to achieve asymptotic synchro-
of a dynamical network with both event-based communi- nization, we will adopt event-triggered control, i.e., design
cation and generalized linear node dynamics by designing an event-triggering rule to determine such a time sequence
a distributed event-triggering rule. It is known that the {tiki }. In this case, Zeno behavior which is undesirable
Zeno behavior, in which two consecutive execution times in practice may occur. So it is important to exclude the
approach arbitrarily closely resulting in an accumulation occurrence of such a behavior by designing a proper control
point is undesirable in practice (Tabuada (2007)). Thus, input ui and a well-defined event-triggering rule for each
the key problem is to design a proper event-triggering rule node, and that is the main purpose of this paper.
such that the network can achieve asymptotic synchro-
nization without Zeno behaviors. Moreover, each node in We assume that the network is connected via a com-
such a network can only get information from its neighbors munication network, and each node in the network can
and also only at some discrete time instances, therefore, only access values of its neighbors at certain discrete time
it is more practical to design an event-triggering rule by instances. It is likely that these limited information is in-
only using these limited information. This in turn increases sufficient for the design purpose, in particular for networks
the difficulty of the design problem. To overcome these whose nodes will synchronize to a time-varying trajectory.
obstacles, we introduce estimators into each node. The To solve this problem, we adopt control input ui as follows
X
estimators built in each node provides the node with esti- aij x̂ij (t) − x̂ii (t) , i = 1, 2, . . . N,
ui = c (3)
mations of its neighbors by using the limited information j∈Ni
that the node can obtain. Then with the help of these
where c > 0 is the coupling strength, and aij are entries
estimators, we design a distributed event-triggering rule
such that the problem can be solved. of matrix A. x̂ij = (x̂ij1 , x̂ij2 , . . . , x̂ijn )> ∈ Rn , j ∈ N̄i are
the states of the mi estimators Oji that are built in node
Notation: We denote by R, R+ and Z+ the set of real i with the form of
numbers, non-negative numbers and non-negative integers,
x̂˙ i (t) = H x̂ij (t), t ∈ [tjkj , tjkj +1 ), j ∈ N̄i
respectively; by Rn and Rn×m the set of n-dimensional real Oji : ji + (4)
vector and n × m real matrix. In , 1n and 1n×m are the x̂j (t ) = xj (t), whenever rj (t, xj , x̂jj , ẑj ) > 0,
n-dimensional identity matrix, the n-dimensional vector
S
where N̄i = Ni {i}; Ni = {j ∈ {1, 2, . . . , N } | aij > 0} is
with all entries being 1 and the n × m matrix also with all the index set of neighbors of node i; mi is the cardinality
entries being 1, respectively. We use k · k to represent the of the set N̄i . The increasing time sequence {tjkj }, kj ∈ Z+
Euclidean norm of a vector x ∈ Rn and also the induced
when node j communicates to its neighbors, is decided by
norm of a matrix A ∈ Rn×m . “> ” is the transpose of a
the event-triggering function rj (·, ·, ·, ·) : R+ × Rn × Rn ×
vector or a matrix, “−1 ” is the inverse of a nonsingular
Rn → R to be designed, i.e.,
matrix, and “⊗” is the Kronecker product of two matrices. n o
λ(·) denotes all the eigenvalues of a square matrix, Re(·) tjkj +1 = inf t ≥ tjkj | rj (t, xj , x̂jj , ẑj ) > 0 (5)
represents the real part of a complex number. j P j j
where t0 = t0 , ẑj = l∈Nj (x̂l − x̂j ). These estimators will
2. NETWORK MODEL AND PRELIMINARIES be reinitialized as soon as node i receives the sampled state
from node j, and hence, provide node i with an estimate
Consider a network of dynamical systems interconnected of the current state of each of its neighbors as well as the
via a digital platform. The state equations of the network current state of itself.
are given as follows
Throughout the paper, we use the following assumptions:
ẋi (t) = Hxi (t) + Γui , i = 1, 2, . . . , N, (1)
where xi = (xi1 , xi2 , . . . , xin )> ∈ Rn is the state of node A1. The dynamical network (1) is connected, i.e., the
i. H ∈ Rn×n is a constant matrix describing the node outer coupling matrix A is irreducible;
dynamics. Γ ∈ Rn×n is the inner coupling matrix. ui ∈ Rn A2. There is no time delay for the sampling and sending
is the control input of node i. We suppose that the topo- executions, i.e., the time tiki represents both the ki th
logical structure of the network is given and is represented sampling time instant and the ki th time when node i
by an outer coupling matrix A = (aij )N ×N ∈ RN ×N . broadcasts its sampled value xi (tiki ) to its neighbors;
Here, we are only interested in undirected networks, i.e., A3. The communication network is under an ideal cir-
if there is a connection between nodes i and j (i 6= j), cumstance, i.e., there are no time delays or data
then aij = aji = 1; otherwise aij = aji = 0. The diagonal dropouts in communication.
entries of A satisfy Under Assumptions A2 and A3, all the estimators Oij
X X
aii = − aij = − aji , i = 1, 2, . . . , N. (2) for each j ∈ N̄i will be reinitialized simultaneously using
j i+
j∈Ni j∈Ni the value xi (tiki ) at t = ti+ i
ki , i.e., x̂i (tki ) = xi (tki ). This
j
The problem to be considered is with the given network together with (4) and t0 = t0 leads to
topology, how to design the control input ui and mean- x̂ji (t) = x̂ii (t) ∀j ∈ N̄i , t ≥ t0 ,
117
NecSys 2013
September 25-26, 2013. Koblenz, Germany
which implies that all the estimators Oij , j ∈ N̄i that are the fact Φ> Φ = IN −1 , A1N = 0, IN − ΦΦ> = IN −
built in different nodes to estimate the state of node i U = N1 1N ×N and U = ΦΦ> = IN − N1 1N ×N .
have the same state response all the time. To simplify the
Let Hi = H + Γi with Γi = cλi Γ, i = 2, 3, . . . , N , and
analysis, we will not distinguish these estimators Oij , and denote H̄ = IN −1 ⊗ H + cΛ1 ⊗ Γ = diag {H2 , H3 , . . . , HN }
use x̂i to replace x̂ji in the sequel. Therefore, network (1) and Λ̄ = cΛ1 ⊗ Γ = diag {Γ2 , Γ3 , . . . , ΓN }. Then, system
with (4) and (3) can be simplified as (10) can be simplified as
N
X ẏ = H̄y + Λ̄(Φ̄> e). (11)
ẋi (t) = Hxi (t) + c aij Γx̂j (t), i = 1, 2, . . . , N (6a)
j=1 It is shown in Liu et al. (2012) that kyk = kΦ̄> xk = |x|As .
x̂˙ i (t) = H x̂i (t), t ∈
[tiki , tiki +1 ) Therefore, if limt→∞ ky(t)k = 0, then by the definition
Oi : (6b) of synchronization, one can conclude that network (6)
x̂i (t+ ) = xi (t), whenever ri (t, xi , x̂i , ẑi ) > 0,
achieves synchronization asymptotically. We summarize
where (6a) also follows by the definition of aii in (2). this result in the following lemma.
Of course, when the communication network is not ideal,
Lemma 1. If system (11) and (6b) has the property
the above relationship will not hold. Then, the problem
becomes more complicated, deserving more attention. lim ky(t)k = 0,
t→∞
Definition 1. Let x(t; x0 ) = (x1 (t; x0 )> , x2 (t; x0 )> , . . . , then network (6) achieves synchronization asymptotically.
xN (t; x0 )> )> ∈ RnN and x̂(t; x0 ) = (x̂1 (t; x10 )> , x̂2 (t; x20 )> ,
. . . , x̂N (t; xN 0 )> )> ∈ RnN be a solution to network (6) Now, we will discuss the synchronization problem of (6) by
with the initial condition x0 = (x> > > >
10 , x20 , . . . , xN 0 ) and designing a proper event-triggering rule. In the case where
xi0 = xi (t0 ). Then we say network (6) achieves synchro- all states of network (6) can be accessed for the design
nization asymptotically, if x(t; x0 ) and x̂(t; x0 ) exist for purpose, we can get a centralized event-triggering rule.
every initial condition x0 ∈ RnN and t > t0 , and further if Theorem 2. If there exist positive definite matrices Pi ∈
lim |x(t; x0 )|As = 0, (7) Rn×n such that
t→∞
where As = x ∈ R
nN
| x1 = x2 = · · · = xN is called the Hi> Pi + Pi Hi = −2In , i = 2, 3, . . . , N, (12)
synchronization manifold of network (6), and |x|As is the then network (6) is globally asymptotically synchronized
Euclidean point-to-set distance, namely under the sampling time sequence determined by the
|x|As = d(x, As ) = inf kx − yk. (8) centralized event-triggering function r(e, x) = kΦ̄> ek −
δ >
α kΦ̄ xk, i.e.,
y∈As
118
NecSys 2013
September 25-26, 2013. Koblenz, Germany
>
the solution of (19) with initial condition φ(t0 , φ(t0 )) = ∂ ∂
φ(t0 ). Therefore, the inter-execution intervals τk can be V̇ (x, t) , V (x, t) f (x, t) + V (x, t)
∂x ∂t (27)
lower bounded by the time during which φ evolves from
≤ −α3 (kxk), whenever kxk ≥ g(t),
0 to αδ , i.e., φ(τ ∗ , 0) = αδ . Such a τ ∗ can be obtained by
solving the differential equation (19), namely and V (s, t) and V̇ (x, t), as functions of x, are equicontinu-
1
δ + αc1 − αc c0 ous on any bounded region of Rn , uniformly with respect
∗
τ = ln + >0 (20) to t, the solution x(t) of system (24) converges to the set
2ac δ + αc1 + αc a
q
(a+b)2 −a2
\ [
with c1 = a+b , c = > 0 and c0 = Q= Qa,b , (28)
a a
1 c1 −c T ≥b>0 a≥T
− 2c ln c1 +c . Thus, no Zeno behavior will occur in net-
where
work (6) under event-triggering rule (13) for all t ≥ t0 . ( )
Now, select the following Lyapunov function candidate Qa,b = x|V (x, a) ≤ sup {V (y, s)} , (29)
y∈Ω,s≥b
V = y > P y, (21)
Ω = x ∈ Rn |kxk ≤ limt→∞ g(t) ,
where P = diag{P2 , P3 , . . . , PN } and Pi , i = 2, 3, . . . , N (30)
are positive definite matrix solutions of (12). Then along and limt→∞ g(t) is the upper limit of g(t). Moreover, if
the trajectories of system (11), one has limt→∞ g(t) = 0, then system (24) is asymptotically stable.
V̇ =ẏ > P y + y > P ẏ ≤ −2kyk2 − 2kP Γ̄kkΦ̄> ek. (22) Theorem 4. If there exist positive definite matrix solutions
> Pi ∈ Rn×n to the Lyapunov equations in (12), then
The event-triggering rule (13) ensures that kΦ̄ ek ≤
δ > δ network (6) can achieve synchronization asymptotically
α kΦ̄ xk = α kyk for all t ≥ t0 . This together with (22) under the distributed event-triggering rule
makes the following inequality hold
tiki +1 = inf t ≥ tiki | ri (t, ei , ẑi ) > 0 ,
(31)
V̇ ≤ −2(1 − δ)kyk2 . (23) p −δ
2
where ri (t, ei , ẑi ) = kei k−ρ kẑi k + e −2γt ; ρ = λN (α+δ) >
Therefore, the equilibrium point y = 0 of system (11) is
asymptotically stable, i.e., limt→∞ ky(t)k = 0. Applying 0; λN < 0 is the smallest eigenvalue of A; γ is a
Lemma 1 proves the result. positive constant such that γ < λmin with λmin =
− maxi∈{2,3,...,N } {Re(λ(Hi ))}; α and δ are defined the
Remark 1. For a dynamical network whose nodes inter- same as in Theorem 2. Moreover, no Zeno behavior occurs
act with their neighbors all the time, condition (12) is a in (6) for all t ≥ t0 .
necessary and sufficient condition for asymptotic synchro-
nization (Trentelman et al. (2013)). Hence, for a network Proof. We divide the proof into two steps: 1). to show
with event-based communication (6a), it is reasonable to the existence of a lower bound on the inter-execution
require that (12) holds in Theorem 2. intervals for event-triggering rule (31), and 2). to prove
Remark 2. In practice, a centralized event-triggering rule the asymptotic convergency of network (6) to As .
(13) is usually hard to implement. For one thing, it 1). Instead of discussing event-triggering rule (31), we
may be costly and time consuming to gather the global consider the following event-triggering rule
information for the design purpose. For another, each
tik+1 = inf t ≥ tik | kei k > ρe−γt .
node normally can only get limited information from its (32)
neighbors – the sampled state value of its neighbors, which By solving differential equation (11) and using the method
can only be received at certain discrete time instants, proposed in Guinaldo et al. (2011), we can get an upper
rather than x(t) for all t > t0 . Therefore, a distributed bound on kyk, i.e.,
event-triggering rule for a given node which only relies on Z t
the information that the node can get is desirable. kyk ≤ke H̄(t−t0 )
kky(t0 )k + keH̄(t−θ) kkΛ̄kkΦ̄> e(θ)kdθ
t0
In order to design such a distributed event-triggering rule ≤k0 e−λmin (t−t0 ) ky(t0 )k
for each node i by only using the information of xi (t), Z t
x̂i (t), and x̂j (t), j ∈ Ni , we first give a lemma which will + k0 e−λmin (t−θ) kΛ̄kkΦ̄> e(θ)kdθ, (33)
be used in the proof the main result. For the proof of the t0
lemma, please refer to Zhao et al. (2012). where k0 = kUT kkUT −1 k, and UT is a nonsingular matrix
Lemma 3. (Zhao et al. (2012)). Consider the time-varying such that UT H̄UT −1 = D with D being the diagonal
nonlinear system matrix composed of the eigenvalues of H̄. The event-
ẋ = f (x, t), (24) triggering rule (32) guarantees that kei k < ρe−γt for all
where f : Rn × R+ → Rn is continuous. Suppose that for t ≥ t0 , which together with the property kΦ̄k = 1 gives
any M > 0, there exists η = ηM , which may depend on √
kΦ̄> ek ≤ kΦ̄kkek = kek < N ρe−γt .
M , such that
Putting the above inequality into (33) gives
kf (x, t)k ≤ ηM , ∀t ≥ 0 and kxk < M. (25) √
λmin t0 −λmin t k0 ρ N kΛ̄k −γt
Then, if there exist a nonnegative bounded function g(t) kyk ≤k0 ky(t0 )ke e + e . (34)
defined on R+ , a smooth function V (x, t) : Rn × R+ → R λmin − α
and K functions α1 , α2 and α3 satisfying In addition, the dynamics of ei can be rewritten as
α1 (kxk) ≤ V (x, t) ≤ α2 (kxk), (26) ėi = Hei − Γẑi , (35)
119
NecSys 2013
September 25-26, 2013. Koblenz, Germany
which gives √ !
> δ > N e−γt
kėi k ≤kHkkei k + kΓkkẑi k kΦ̄ ek < kΦ̄ xk − . (44)
α λN
≤ρkHke−γt + kΓk (k(A ⊗ In )xk + k(A ⊗ In )ek)
√ Now, select the Lyapunov candidate V = y > P y with P =
≤ρ(kHk − λN N kΓk)e−γt + kΓkkzk, (36) diag{P2 , P3 , . . . , PN }, then with (44) and kΦ̄> xk = kyk,
where z = (A ⊗ In )x, and kA ⊗ In k = −λN . Moreover, the derivative of V along system (11) satisfies
kzk2 =x> (A2 ⊗ In )x ≤ λ2N x> (U 2 ⊗ In )x V̇ ≤ − 2kyk2 + 2αkykkΦ̄> ek
(37)
=λ2N kΦ̄> xk2 = λ2N kyk2 , ≤ − 2(1 − δ)kyk2 + 2ρ1 kyke−γt
= − 2(1 − δ − δ1 )kyk2 − 2kyk δ1 kyk − ρ1 e−γt
which comes from A2 ≤ λ2N U 2 and U 2 = U = Φ̄Φ̄> .
Therefore, the inequality (36) becomes ≤ − 2(1 − δ − δ1 )kyk2 , when kyk ≥ ρ2 e−γt , (45)
√ √ √
kėi k ≤ ρ(kHk − λN N kΓk)e−γt − λN kΓkkyk. (38) where 0 < δ1 < 1 − δ, ρ1 = − δλNN
> 0, ρ2 = > 0.− δδ1 λN
N
From (38) and (34), for all t ∈ [tiki , tiki +1 ) and k ∈ Z, we −γt
Let g(t) = ρ2 e , so we have limt→∞ g(t) = 0. Applying
have Lemma 3, we get that the equilibrium point y = 0 of
i i
kėi k ≤ k1 e−λmin t + k2 e−γt ≤ k1 e−λmin tki + k2 e−γtki , system (11) is asymptotically stable, and thus the network
kΛ̄k
√ can achieve synchronization asymptotically.
where k1 = k0 kΓkky(t0 )ke−λmin t0 and k2 = k0λρminN−γ + Remark 3. Asymptotic synchronization was also investi-
√
ρ(kHk − λN N kΓk). Thus, we get gated for a similar dynamical network by using self-
Z t Z t triggered control in De Persis (2013). The main differences
kei (t)k =k ėi (θ)dθk ≤ kėi (θ)kdθ between the results proposed in De Persis (2013) and ours
tik tik are twofold. First, in De Persis (2013), each node will
i i
−λmin tik −γtik
(39) require the updated information from its neighbors when
≤ k1 e i + k2 e i (t − tiki ) the self-triggering condition is fulfilled, and therefore, bidi-
i
≤e−γtki (k1 + k2 ) τki , rectional communication is needed; whereas in our paper,
a node will broadcast its sampled state to its neighbors
where τki = t − tiki , and the second inequality follows directly as long as its own event-triggering condition is
from λmin > γ > 0. Because at t = tiki , e(t) = 0 and satisfied, and hence, only undirectional communication is
according to (31), the next event will not be triggered until required. Second, the system matrix of the node dynamics
kei (t)k = ρe−γt . Therefore, the inter-execution intervals is required to have purely imaginary eigenvalues with uni-
τk∗i ≤ tiki +1 − tiki can be lower bounded by the solution tary geometric multiplicity in De Persis (2013), and our
∗
τi∗ of the equation (k1 + k2 )τi∗ = ρe−γτi which is strictly results don’t rely on such a restriction.
positive. Please refer to Seyboth et al. (2013) for details.
4. AN EXAMPLE
Apparently, the inter-execution intervals of event-triggering
rule
p (31) are greater than those of rule (32) because This section gives an example to show the effectiveness of
kẑi k2 + e−2γt ≥ e−γt with kẑi k ≥ 0, and thus there the proposed results. Here, we adopt the example used in
also exists a positive lower bound on the inter-execution Liu et al. (2012), where the network consists of 10 nodes.
intervals for (31). The parameters of the network are as follows: c = 1,
2). The event-triggering rule (31) guarantees that 0 −0.5 0.25 0
H= , Γ= ,
p 0.5 0 −1 0.25
kei k < ρ kẑi k2 + e−2γt , ∀t ≥ t0 . (40)
−4 1 1 0 0 0 0 0 1 1
Since 1 −4 1 1 0 0 0 0 0 1
N
X √ 2 1 1 −4 1 1 0 0 0 0 0
kẑi k2 + e−2γt = kẑk2 +N e−2γt ≤ kẑk + N e−γt ,
0 1 1 −4 1 1 0 0 0 0
i=1
0 0 1 1 −4 1 1 0 0 0
A= 0 0 0 1 1 −4 1 1 0 0
.
we have
√ 0 0 0 0 1 1 −4 1 1 0
kek <ρ kẑk + N e−γt
0 0 0 0 0 1 1 −4 1 1
√ 1 0 0 0 0 0 1 1 −4 1
=ρ k(A ⊗ In )(x + e)k + N e−γt (41) 1 1 0 0 0 0 0 1 1 −4
√ Since H has two eigenvalues ±0.5i on the imaginary
≤ρ(kzk − λN kek + N e−γt ). axis, the network cannot asymptotically synchronize to its
−δ equilibrium point, but to a stable time-varying solution.
As ρ = λN (α+δ) ,
inequality (41) is equivalent to
−δ √ When all the information of network (6) is available for
kek < kzk + N e−γt . (42) the design purpose, then we can design centralized event-
αλN
triggering rule (13) to synchronize the network. Figure 1
Using (37) again leads to
√ −γt ! gives the simulation results under the designed centralized
δ > Ne event-triggering rule with δ = 0.9 and α = 2.9061,
kek < kΦ̄ xk − . (43) which shows that the network achieves synchronization
α λN
asymptotically, and the minimum inter-execution interval
Combining kΦ̄> ek ≤ kek with (42) gives is τ ∗ = 0.0186s for t from 0s to 25s.
120
NecSys 2013
September 25-26, 2013. Koblenz, Germany
−0.5
−1
−1.5
REFERENCES
0 5 10
t
15 20 25
De Persis, C. (2013). On self-triggered synchronization of
1.5 linear systems. In IFAC NecSys2013, accepted. Koblenz,
Germany.
De Persis, C., Sailer, R., and Wirth, F. (2011). On a small-
Events
1
gain approach to distributed event-triggered control. In
IFAC World Congress, 2401–2406. Milan, Italy.
0.5
0 2 4 6 8 10 Demir, O. and Lunze, J. (2012). Event-based synchronzi-
t
ation of multi-agent systems. In the 4th IFAC Conf. on
Fig. 1. Simulation for centralized event-triggering rule. Analysis and Design of Hybrid Systems, 1–6. Eindhoven,
The Netherlands.
For the network where each node can only get limited Dimarogonas, D.V. (2011). L2 gain stability analysis of
information, we can design a distributed event-triggering event-triggered agreement protocols. In IEEE Conf.
rule (31) to achieve asymptotic synchronization based Decision and Control, 2130–2135. Orlando, USA.
on Theorem 4. Figure 2 shows the state of the entire Dimarogonas, D.V. and Johansson, K.H. (2009). Event-
network and sampling times of each node, where δ = 0.9, based control for multi-agent systems. In IEEE Conf.
α = 2.9061, λN = −6.1518, λmin = 0.2972, and γ = 0.29. Decision and Control, 7131–7136. Shanghai, China.
The minimum inter-execution interval τi∗ for each node for Donkers, M.C.F. and Heemels, W.P.M.H. (2012). Output-
t from 0s to 25s is given in Table 1. based event-triggered control with guaranteed L∞ -gain
and improved and decentralised event-triggering. IEEE
1 Trans. Automatic Control, 57(6), 1362–1376.
0.5 xi1 Garcia, E. and Antsaklis, P.J. (2013). Model-based event-
0
xi2 triggered control for systems with quantization and
x(t)
−0.5
−1 time-varying network delays. IEEE Trans. Automatic
−1.5
Control, 58(2), 422–434.
0 5 10
t
15 20 25
Guinaldo, M., Dimarogonas, D.V., Johansson, K.H.,
10
Sánchez, J., and Dormido, S. (2011). Distributed event-
8
based control for interconnected linear systems. In IEEE
Conf. on Decision and Control and European Control
Events
121