Vous êtes sur la page 1sur 4

ISA Transactions 71 (2017) 21–24

Contents lists available at ScienceDirect

ISA Transactions
journal homepage: www.elsevier.com/locate/isatrans

Research article

A further result on consensus problems of second-order multi-agent


systems with directed graphs, a moving mode and multiple delays$
Xue Li a, Kai Gao b, Peng Lin c,n, Lipo Mo d
a
School of Astronautics and Aeronautic, University of Electronic Science and Technology of China, Chengdu 611731, PR China
b
Jiangsu Automation Research Institute, Lianyungang 222061, PR China
c
School of Information Science and Engineering, Central South University, Changsha 410083, PR China
d
School of Science, Beijing Technology And Business University, Beijing 100048, PR China

art ic l e i nf o a b s t r a c t

Article history: This paper considers a consensus problem of a class of second-order multi-agent systems with a moving
Received 23 November 2016 mode and multiple delays on directed graphs. Using local information, a distributed algorithm is adopted
Received in revised form to make all agents reach a consensus while moving together with a constant velocity in the presence of
16 May 2017
delays. To study the effects of the coexistence of the moving mode and delays on the consensus con-
Accepted 17 June 2017
Available online 3 July 2017
vergence, a frequency domain approach is employed through analyzing the relationship between the
components of the eigenvector associated with the eigenvalue on imaginary axis. Then based on the
Keywords: continuity of the system function, an upper bound for the delays is given to ensure the consensus
Consensus convergence of the system. A numerical example is included to illustrate the obtained theoretical results.
Delays
& 2017 ISA. Published by Elsevier Ltd. All rights reserved.
Moving mode
Frequency domain analysis

1. Introduction problem of second-order multi-agent systems was studied where


all agents finally converge to a common point and move together
As a fundamental problem in distributed control of multi-agent with a constant velocity while in [21–24], distributed circular
systems, consensus problem has received a great deal of attention control problems were studied where all agents finally form a
from the control community [1–24]. The studies on consensus formation and circle around a common point. However, there are
problem mainly concentrate on the communication topologies [1– rare works concerned about the case where the final consensus
3], the information delays [4–14], the constraints on the states and point of the agents is moving with a nonzero velocity. Articles [8]
inputs [17–20]. In this paper, our objective is to study a consensus and [9] studied the consensus problems for second-order multi-
problem of a class of second-order multi-agent systems with a agent systems with time-delay and a moving mode, but the ob-
moving mode and multiple delays. Current works about consensus tained results are limited to either the case where the time-delays
problem with delays mainly concentrated on the case where all are all equal or the case where the graph is undirected.
agents finally converge to a static point. For example, articles [4,5] Founded on the works of [8], we study the consensus problem
investigated the consensus problems of first-order multi-agent of a class of second-order multi-agent systems with a moving
systems with identical and nonidentical delays, where all agents mode and multiple delays on directed graphs. Using local in-
finally converge to a static point, while articles [6,7] studied the formation, a distributed algorithm is adopted to make all agents
consensus problems of second-order multi-agent systems with reach a consensus while moving together with a constant velocity
nonuniform delays, where all agents also finally reach a consensus in the presence of delays. To study the effects of the coexistence of
at a static point. In some practical applications, agents not only the moving mode and delays on the consensus convergence, a
need to reach a consensus but also need to move together with a frequency domain approach is employed through analyzing the
constant or time-varying velocity. For example, in [1], a consensus
relationship between the components of the eigenvector asso-
ciated with the eigenvalue on imaginary axis. Then based on the

This work was supported by the Foundation for Innovative Research Groups of continuity of the system function, an upper bound for the delays is
the National Natural Science Foundation of China (61321003), the 111 Project given to ensure the consensus convergence of the system.
(B17048) the National Natural Science Foundation of China (61573082, 61203080), The following notations will be used throughout this paper. Rm
the National Program 973 of China (613237201506) and the Innovation-driven Plan
denotes the set of all m dimensional real column vector;  m de-
at Central South University.
n
Corresponding author. notes the set of all m dimensional complex column vectors; Im
E-mail addresses: lin_peng@csu.edu.cn, lin_peng0103@sohu.com (P. Lin). denotes the m dimensional unit matrix; ⊗ denotes the kronecker

http://dx.doi.org/10.1016/j.isatra.2017.06.013
0019-0578/& 2017 ISA. Published by Elsevier Ltd. All rights reserved.
22 X. Li et al. / ISA Transactions 71 (2017) 21–24

product; (·)T denotes the transpose of vectors or matrixes and (·)* balanced graphs are taken into account, the relationship between
denotes the conjugate transpose of them. 1 represents [1, 1, … , 1]T the delays and the eigenvalues of the system matrix becomes too
with compatible dimensions (sometimes, we use 1n to denote 1 complicated to analyze and hence the analysis approaches in [8]
with dimension n); 0 denotes zero value or zero matrix with ap- and [9] are no longer valid for the system (1) with (3).
propriate dimensions; ∥·∥ refers to the standard Euclidean norm of For convenience of discussion, let 0 ≤ τ1 < τ2⋯ < τM denote all
vectors. different delays of the system (1) with (3). Let
ξ(t ) = [r1(t ) , v1(t ) , … , rn(t ) , vn(t )]T . Writing the system (1) with (3) in
a compact form, we have
2. Graph theory and model M
ξ (̇ t ) = (In ⊗ E )ξ (t ) − ∑ (L m ⊗ F )ξ (t − τm),
Let  ( ,  ,  ) be a directed graph, where  is the set of m=1 (4)
agents,  is the set of edges, and  = [aik ] is a weighted adjacency where
matrix. The agent indexes belong to a finite index set
 = {1, 2, … , n}. An edge of  is denoted by eik = (i, k ), which in- ⎡ 0 1⎤ ⎡ 0 0⎤
E=⎢ ,F=⎢ ,
dicates that agent i can receive information from agent k. The ⎣ 0 0⎥⎦ ⎣ 1 1⎥⎦
adjacency matrix is defined as aii = 0 and aik > 0 when eik ∈  . The
and Lm denotes the component of L corresponding to the delay τm
set of neighbors of agent i is denoted by Ni = {k ∈  : (i, k ) ∈ }. M
for m = 1,…,M. It is clear that L m1 = 0 and L = ∑m = 1 L m .
The Laplacian of the graph  is defined as L = [lik ], where
n
lii = ∑k = 1 aik and lik = − aik , i ≠ k . A path is a sequence of ordered
edges of the form (i1, i2) , (i2, i3) , ⋯, where ik ∈  . A directed graph
4. Main theorem
is strongly connected if there is a directed path from any node vi to
any other node vk. A directed graph is balanced if
n n
∑k = 1 aik = ∑k = 1 aki for all i ∈  . Moreover, a graph is undirected if Lemma 2 ([8]). Suppose that the directed graph  is strongly con-
aik = aki for all i, k . An undirected graph is connected if there is an nected and balanced, the matrix In ⊗ E − L ⊗ F has an eigenvalue at
edge between every pair of different agents. [25] 0 with multiplicity 2 and the other eigenvalues have negative real
parts.
Lemma 1 ([25]). If the graph  is undirected and connected, then its
Laplacian L has properties as follows: Since the graph  is directed and balanced, we need first study
L + LT
the matrix 2
before presenting the main results. Since the
(1) 0 is a simple eigenvalue with associated eigenvector 1
L + LT
(2) all its other n − 1 eigenvalues are positive and real. graph  is balanced and strongly connected, 2 can be regarded
as the Laplacian of a connected undirected graph. From Lemma 1,
L + LT
all the eigenvalues of 2
are real and nonnegative. Let λmax be the
3. Model L + LT
largest eigenvalue of 2
. We have the following theorem.
Consider a multi-agent system with n agents. Each agent is Theorem 1. Suppose that the directed graph  is strongly connected
regarded as a node in a directed graph  . Suppose that each agent and balanced, and τm = τ for some constant τ > 0 and all m. Using (3)
has the following dynamics: for (2),
ri̇(t ) = vi(t ), lim [ri(t ) − rk(t )] = 0,
t →+∞
vi̇ (t ) = ui (t ) (1) n

where ri ∈  and vi ∈  denote the position and the velocity of


lim vi(t ) =
t →+∞
∑ vi(0),
i=1
agent i for i = 1, … , n and ui(t ) ∈  is the control input. It is as-
sumed that the initial condition ri(s) and vi(s) satisfy the dynamics if one of the following two conditions holds:
of (1) for all s ≤ 0.
arctanω¯ λ
The goal of the agents is to use local information to reach a (1) 0 ≤ τ < ω¯ λ
where ω̄ = 1 −max
λmax
when λ max < 1;.
arctan2 max
consensus and move together with a constant velocity under di- (2) 0 ≤ τ < 2λmax
when λ max ≥ 1.
rected balanced communication graphs, i.e.,
Proof. Consider the frequency transformation of the system (4). It
lim [ri(t ) − rk(t )] = 0, follows that ξ(s ) = φ−1(s )ξ(0),
t →+∞ M
where φ(s ) = sI2n − In ⊗ E+∑m = 1 (L m ⊗ F )e−sτm .
lim vi(t ) = v0,
t →+∞ (2)
When s = 0, φ(s ) = − In ⊗ E + L ⊗ F . According to Lemma 2, it
for all i, k ∈  and some constant v0 when multiple delays are can be obtained that φ(s ) has an eigenvalue at 0 with multiplicity
involved. To this end, we adopt the following algorithm introduced 2. Moreover, φ(0)(1n ⊗ [1, 0]T ) = 0 and φ(0)2(1n ⊗ [a, b]T ) = 0,
in [8]: where a and b are two constants such that ab ≠ 0. If all other ei-
genvalues of φ(s ) have negative real parts, from Lemma 4 in [8], all
ui (t ) = ∑ aik(rk(t − τik ) − ri(t − τik )) agents will finally converge to the space spanned by 1n ⊗ [1, 0]T
k ∈ Ni
and 1n ⊗ [a, b]T . That is, consensus can be reached.
+ ∑ aik(vk(t − τik ) − vi(t − τik )), To show that all the other eigenvalues have negative real parts,
k ∈ Ni (3) we consider the delays such that φ(s ) = 0 has nonzero roots on the
imaginary axis. Suppose that s = jw ≠ 0 is a root of φ(s ) = 0 on the
where τik is the communication delay from agent k to i. In [8] and
imaginary axis. Let q ∈  2n be a vector such that ∥ q ∥ = 1 and
[9], consensus of second-order multi-agent systems with a moving
mode and delays was studied and conditions were given for the ⎡ M ⎤
system consensus by finding the explicit relationship between the ⎢ jωI2n − In ⊗ E + ∑ (L m ⊗ F )e−jωτm⎥q = 0
⎢⎣ m=1
⎥⎦ (5)
delays and the eigenvalues of the system matrix. When directed
X. Li et al. / ISA Transactions 71 (2017) 21–24 23

Since the system (4) is a real system, if jω is a root of the system λ max ≥ 1. From the continuity of functions, the system (4) has one
(4), then −jω is also a root of the system (4). Without loss of eigenvalue at 0 with multiplicity 2 and all its other eigenvalues
generality, suppose that ω > 0. For simplicity of expression, let have negative real parts. Hence, as discussed previously, all agents
q = q1 ⊗ [1, 0]T + q2 ⊗ [0, 1]T , where q1, q2 ∈  n . From the odd row can reach a consensus as t → + ∞ on position and velocity states
of (5), we have respectively. That is, limt →+∞(ri(t ) − rk(t )) = limt →+∞(vi(t ) − vk(t )) = 0
n
for all i, k . Since the graph  is balanced, ∑i = 1 ui(t ) = 0, i.e.,
jωq1 = q2. (6) n n n
∑i = 1 vi̇ (t ) = 0, for all t ≥ 0. Hence, ∑i = 1 vi(t ) = ∑i = 1 vi(0) for all t ≥ 0
Note that ∥ q ∥2 = ∥ q1 ∥2 + ∥ q2 ∥2 = 1. It follows that n
and limt →+∞vi(t ) = (1/n) ∑i = 1 vi(0) for all i.
1 Remark 1. In [8] and [9], the consensus problems for second-or-
q1*q1=
ω2 + 1 der multi-agent systems with time-delay and a moving mode were
ω2 considered, but the obtained results are limited to either the case
q2*q2= 2
ω +1 (7) where the time-delays are all equal or the case where the graph is
undirected. In contrast to [8] and [9], in this paper, we do not only
From (5), we have q*⎡⎣ jωI2n − In ⊗ E + ∑m = 1 (L m ⊗ F )e−jωτm⎤⎦q = 0. It
M
consider the balanced directed graphs but also consider the non-
follows that uniform delays by analyzing the system stability with the eigen-
L + LT
M M
value of 2 .
jωq1*q1 + ∑ q2*L me−jωτmq1 − q1*q2 + jωq2*q2 + ∑ q2*L me−jωτmq2 Remark 2. In this paper, we only studies the case where the
m=1 m=1
graphs are balanced and strongly connected and the agents finally
= 0. (8)
move together in a line. Our future work would be directed to-
Note that τm = τ for all m, from (6), (7) and (8), it follows that wards the case of the general directed graphs and the case where
all agents move in a circle.
(ω2 + 1)q1*Lq1sinωτ = ω3 ∥ q1 ∥2
Theorem 2. Suppose that the graph  is undirected and connected.
(ω2 + 1)q1*Lq1cosωτ = ω2 ∥ q1 ∥2 (9) Using (3) for (2),
L + LT lim [ri(t ) − rk(t )] = 0,
Note that q1*Lq1/∥ q1 ∥2 = q1*Lq1/q1*q1 = q1* 2
q1/q1*q1 ≤ λ max , where t →+∞
L + LT n
λmax denotes the largest eigenvalue of . Since |sinωτ| ≤ 1, we
2 lim vi(t ) = ∑ vi(0),
have t →+∞
i=1

ω3 ≤ λ max(ω2 + 1), (10) if one of the following two conditions holds:

Similarly, we have (1) 0 ≤ τm < arctanω¯


for all m where ω̄ =
λmax
when λ max < 1;.
ω¯ 1 − λmax
ω2 ≤ λ max(ω2 + 1) (11) (2) 0 ≤ τm <
arctan2λmax
for all m when λ max ≥ 1.
2λmax

If λ max < 1, from (11), we have (1 − λ max )ω2 ≤ λ max . Hence, Proof. Following the lines of the proof of Theorem 1, this theorem
can be proved. But it should be emphasized that the Eq. (9) would
λ max
ω≤ , when λ max < 1 be replaced with the equation
1 − λ max
M
ω3 (ω2 + 1)q1*L mq1sinωτ = ω3 ∥ q1 ∥2
If λ max ≥ 1 and ω < λ max , from (10),
ω2 + 1
=ω 1− ( 1
ω2 + 1 ) ≤ λ max . ∑
m=1

If λ max ≥ 1 and ω ≥ λ max , from (10),


ω
2
≤ω 1− ( 1
ω2 + 1 )≤λ max , we

M
(ω2 + 1)q1*L mq1cosωτ = ω2 ∥ q1 ∥2 .
have ω ≤ 2λ max . As a result, we have m=1 (12)
ω ≤ 2λ max , when λ max ≥ 1 and the inequality
M
∑m = 1 (ω2 2
+ 1)q1*L mq1/∥ q1 ∥ ≤ λ max would be
From (9), we have tan(ωτ ) = ω , i.e., τ = arctanω
. By simple cal- used.
ω
d arctanω
culations, we have dω ω < 0 when ω > 0. That is, τ is a de-
creasing function of ω when ω > 0. Since ω is bounded, we can 5. One numerical example
π
adopt τ sufficiently small such that ωτ ≤ 2 . Specifically, when
λmax λmax
λ max < 1, since ω ≤ 1 − λmax
, τ can be selected as τ ≤ 2π / 1 − λmax
. In this section, we give one numerical example to illustrate the
π obtained results. Consider a multi-agent system with 8 agents.
When λ max ≥ 1, since ω ≤ 2λ max , τ can be selected as τ ≤ 4λ .
max Fig. 1 shows the directed communication graph, where each edge
Moreover, from the monotonicity of τ with respect to ω, we have weight is 1. By calculations according to Theorem 1, an upper
arctanω¯ λmax arctan2λmax
τ≥ ω¯
with ω̄ = 1 − λmax
when λ max < 1 and τ ≥ 2λmax
when
λ max ≥ 1. Summarizing the above analysis, when λ max < 1 and
arctanω¯ arctan2λmax
0≤τ< ω¯
hold, or λ max ≥ 1 and 0 ≤ τ ≤ 2λmax
hold, the
equations (4) do not hold and hence f(t) has no eigenvalues at the
imaginary axis.
Let τ = 0 in (4), from Lemma 2, the system (4) has one eigen-
value at 0 with multiplicity 2 and all its other eigenvalues have
negative real parts. Note that the system (4) has no other ima-
ginary eigenvalues except the zero eigenvalue of multiplicity
2 when τ ∈ 0, ( arctanω¯
ω¯ ) for λ max < 1 and τ ∈ ( 0, arctan2λmax
2λmax ) for Fig. 1. The communication topology for the multi-agent system.
24 X. Li et al. / ISA Transactions 71 (2017) 21–24

18 delays is adopted to drive all agents to reach a consensus while


moving together with a constant velocity. Based on a frequency
16
domain approach and the continuity of the system function, an
14 upper bound for the delays is given to ensure the consensus
convergence of the system.
12

10
References
Position

8
[1] Ren W, Atkins E. Distributed multi-vehicle coordinated control via local in-
6 formation exchange. Int J Robust Nonlinear Control 2007;17(10–11):1002–33.
[2] Hong Y, Gao L, Cheng D, Hu J. Lyapunov-based approach to multiagent systems
4 with switching jointly connected interconnection. IEEE Trans Autom Control
2007;52(5):943–8.
2 [3] Xie G, Wang L, Consensus control for a class of networks of dynamic agents:
fixed topology. In: Proceedings of the 44th IEEE Conference on decision and
0 control, and the European Control conference; 2005. p. 96–101.
[4] Bliman PA, Ferrari-Trecate G. Average consensus problems in networks of
−2 agents with delayed communications. Automatica 2008;44(8):1985–95.
0 5 10 15 20 25 [5] Tian Y, Liu C. Consensus of multi-agent systems with diverse input and
Time communication delays. IEEE Trans Autom Control 2008;53(9):2122–8.
[6] Lin P, Jia Y. Consensus of second-order discrete-time multi-agent systems with
nonuniform time-delays and dynamically changing topologies. Automatica
2009;45(9):2154–8.
5 [7] Tian Y, Liu C. Robust consensus of multi-agent systems with diverse input
delays and asymmetric interconnection perturbations. Automatica 2009;45
(5):1347–53.
4 [8] Lin P, Jia Y, Du J, Yuan S. Distributed consensus control for second-order agents
with fixed topology and time-delay. In: Proceedings of the Chinese control
3 conference; 2007. p. 577–581.
[9] Lin P, Dai M, Song Y. Consensus stability of a class of second-order multi-agent
systems with nonuniform time-delays. J Frankl Inst 2014;351(3):1571–6.
2 [10] Münz U, Papachristodoulou A, Allgöwer F. Delay robustness in consensus
problems. Automatica 2010;46(8):1252–65.
Velocity

[11] Xiao F, Wang L. State consensus for multi-agent systems with switching
1 topologies and time-varying delays. Int J Control 2006;79(10):1277–84.
[12] Liu C, Tian Y. Formation control of multi-agent systems with heterogeneous
communication delays. Int J Syst Sci 2009;40:627–36.
0
[13] Wang L, Qian W, Wang Q. Exponential synchronization in complex networks
with a single coupling delay. J Frankl Inst 2013;350(6):1406–23.
−1 [14] Su H, Chen X, Chen MZQ, Wang L. Distributed estimation and control for
mobile sensor networks with coupling delays. ISA Trans 2016;64:141–50.
[15] Wang L, Dai H, Sun Y. Synchronization criteria for a generalized complex
−2 delayed dynamical network model. Phys A: Stat Mech Appl 2007;383(2):703–
13.
[16] Lin P, Ren W. Constrained consensus in unbalanced networks with commu-
−3
0 5 10 15 20 25 nication delays. IEEE Trans Autom Control 2014;59(3):775–81.
[17] Nedić A, Ozdaglar A, Parrilo P-A. Constrained consensus and optimization in
Time
multi-agent networks. IEEE Trans Autom Control 2010;55(4):922–38.
[18] Lin P, Ren W, Song Y. Distributed multi-agent optimization subject to non-
identical constraints and communication delays. Automatica 2016;65(3):120–
Fig. 2. State trajectories of all agents. 31.
[19] Lin P, Ren W, and Gao H, Distributed velocity-constrained consensus of dis-
crete-time multi-agent systems with nonconvex constraints, switching
bound for the delays is 0.149. Take the τ21 = τ32 = τ43 = τ54= topologies, and delays, IEEE Trans Autom Control, 2017 accepted and available
τ65 = τ76 = τ87 = τ18 = 0.14 , τ14 = τ45 = τ56 = τ67 = τ78 = τ81 = 0.1, online. 〈http://dx.doi.org/10.1109/TAC.2016.2637005〉.
τ12 = τ23 = τ34 = τ41 = 0.05, where τij denotes the delay corre- [20] Lin P, Ren W, Farrell JA. Distributed continuous-time optimization: Nonuni-
sponding to the edge (i,j). Fig. 2(a) and (b) show the simulation form gradient gains, finite-time convergence, and convex constraint set. IEEE
Trans Autom Contol 2017;62(5):2239–53.
results. It is clear that all agents reach a consensus and move to- [21] Lin P, Jia Y. Distributed rotating formation control of multi-agent systems. Syst
gether in a constant velocity. This is consistent with Theorem 1. Control Lett 2010;59(10):587–95.
[22] Sepulchre R, Paley D, Leonard NE. Stabilization of planar collective motion
with limited communication. IEEE Trans Autom Control 2007;53(3):706–19.
[23] Hernandez S, Paley D. Three-dimensional motion coordination in a time-in-
6. Conclusion variant flowfield. In: Proceedings of the 48th IEEE conference on decision and
control; 2009. p. 7042–7048.
[24] Pavone M, Frazzoli E. Decentralized policies for geometric pattern formation
This paper studies the consensus problem of second-order
and path coverage. ASME J Dyn Syst Meas Control 2007;129(5):633–43.
multi-agent systems with a moving mode and multiple delays on [25] Godsil C, Royle GF. Algebraic graph theory. Vol. 207. Springer Science &
directed graphs. A distributed control algorithm with multiple Business Media, 2013.