J Syst Sci Complex (2023) 36: 1809–1829

Observer-Based Intra-Cluster Lag Consensus of

Multi-Agent Systems via Intermittent Adaptive Pinning
Control∗
ZHANG Yunlong · WEN Guoguang · RAHMANI Ahmed · PENG Zhaoxia
· WEN Shiping

DOI: 10.1007/s11424-023-1222-0
Received: 25 June 2021 / Revised: 5 September 2021
c The Editorial Office of JSSC & Springer-Verlag GmbH Germany 2023

Abstract The intra-cluster lag consensus means that the agents in the same cluster can achieve lag
consensus asymptotically while the agents in different clusters can achieve different consensus. In this
paper, the authors investigate the observer-based intra-cluster lag consensus problems of multi-agent
systems (MASs) with general linear dynamics and nonlinear dynamics via intermittent adaptive pinning
control. The interaction network is considered to be weakly connected, i.e., it is not necessary to be
strongly connected, in-degree balanced or contain a directed spanning tree. To realise the intra-cluster
lag consensus, a class of observers is designed to estimate the states of followers. Then a class of
observer-based intermittent adaptive pinning control protocols is proposed according to the difference
that the agents receive information source. Moreover, the pinning gains are designed to be intermittent
adaptive and have an exponential convergence rate, which will effectively reduce communication costs,
ZHANG Yunlong
CRIStAL, UMR CNRS 9189, Centrale Lille, Villeneuve d’Ascq 59651, France.
Email: 15121529@bjtu.edu.cn.
WEN Guoguang (Corresponding author)
Department of Mathematics, Beijing Jiaotong University, Beijing 100044, China;
Email: guoguang.wen@bjtu.edu.cn.
RAHMANI Ahmed
CRIStAL, UMR CNRS 9189, Centrale Lille, Villeneuve d’Ascq 59651, France.
Email: ahmed.rahmani@centralelille.fr.
PENG Zhaoxia
School of Transportation Science and Engineering, Beihang University, Beijing 100191, China.
Email: pengzhaoxia@buaa.edu.cn.
WEN Shiping
Centre for Artificial Intelligence (CAI), University of Technology Sydney, Australia.
Email: Shiping.Wen@uts.edu.au.
∗ This work was supported in part by the National Natural Science Foundation of China under Grant Nos.

61977004 and 61503016, and in part by the Fundamental Research Funds for the Central Universities under
Grant Nos. YWF20-BJ-J-634 and YWF-19-BJ-J-259.
⋄ This paper was recommended for publication by Editor FU Minyue.
1810 ZHANG YUNLONG, et al.

avoid the pinning gains being larger than those needed in real applications and guarantee the pinning
gains quickly converge to steady value. Correspondingly, some sufficient consensus criteria are derived
and rigorous proofs are given based on matrix theory and Lyapunov stability theory. Finally, the
effectiveness for the proposed intermittent adaptive pinning control strategy is validated by a numerical
simulation.
Keywords Lag consensus, multi-agent systems, intermittent adaptive pinning control, state observer.

1 Introduction
In recent decades, much attention has been paid to the research of multi-agent systems
(MASs)[1–3] , the main reason is that MASs can be potentially applied in various areas such
as search and rescue missions[4] , spacecraft formation flying[5, 6] , cooperative surveillance[7] .
Moreover, different from individual agent which has limited sensing and computing ability, the
MASs can perform or solve complex tasks in a coordinated fashion with improving scalability,
increasing flexibility, reducing cost, and avoiding a single-point of failure. As an important
and critical issue arising from coordination of MASs, consensus problem is aimed at design-
ing appropriate distributed control protocol by negotiating with their neighbors to reach an
agreement, one can see [8–10] and references therein.
Notice that the relevant works on consensus problem for MASs mainly focus on complete
consensus. However, in many physical applications, due to some unanticipated situations such
as the changes of cooperative tasks, external environment or event-time, the agents usually are
required to be divided into several clusters. Therefore, an extended concept about complete
consensus is proposed to deal with such complex phenomena, that is cluster consensus, which
means the agents in the same cluster share a common state while there is no consensus behavior
among different clusters. In fact, cluster consensus is appealing in various areas, for examples,
in the large scale complex networks, due to the specific requirements, it is easier to analysis
and design by decomposing the large scale complex network into several smaller clusters; in
nature, the birds, fish and bacteria colonies are often emerge several clusters; particularly in
human society, some opinion formation models reveal that agents in the same cluster eventually
form the same opinion. Up to now, numerous results about cluster consensus have been derived
from various aspects such as system dynamics, network topologies, time-delays[11–13] . In [11],
the authors investigated the cluster synchronization of a network system with nonidentical
nodes, where the dynamics of agents are either general linear or nonlinear. In [12], the authors
addressed the cluster consensus problem for a second-order MAS by pinning leader-follower
approach, where the communication topology is assumed to be weakly connected, i.e., it is
not necessary to be strongly connected, in-degree balanced or contain a directed spanning
tree. Moreover, as we all know, time delays exist widely in many situations and it can cause
oscillations or instabilities. Hence, [13] considered the lag consensus of nonlinear MASs with
two types of time-delays.
It is well known that the control techniques are especially important for MASs[14] to reach
objective state. In the most of prior works, the communication among all the agents is com-
 CLUSTER LAG CONSENSUS OF MULTI-AGENT SYSTEM 1811

monly assumed to be continuous. However, in real engineering, due to the presence of obstacles,
the limitation of computing and communication ability, or some other factors, the informa-
tion communication may be discontinuous. To mitigate these issues, the intermittent control
strategy has been introduced and widely applied in transportation, manufacturing and com-
munication. Reference [15] analysed the cluster lag synchronization issue for heterogeneous
complex networks by employing an intermittent control to reduce control cost and information
communication. In [16], an intermittent sampled data control mechanism was employed, which
not only shortens the working time, but also decreases the load of controllers’ updating rate.
In addition, pinning control is more economical and more effective for the large-scale leader-
following MASs due to it pins only a small fraction of followers rather than all followers, which
has been attracted much attention. In [17], the authors applied pinning control scheme to het-
erogeneous MASs for solving the group consensus with reference. In [18], the authors presented
a distributed pinning controller for robotic networks to realize cluster consensus, where the
system dynamics are characterized by Euler Lagrangian equations. Nevertheless, it is noted
that the disadvantage for pinning control in existing literature is that the pinning strength is
usually fixed and given in advance for avoiding the appearance of the worst situation, which is
unreasonable in real applications due to the worst case seldom happens. To overcome the draw-
back, many researchers adopt adaptive control approach to adjust the pinning weight during
the process of dynamics evolution. For example, [19] investigated cluster consensus problem of
second-order nonlinear MASs by employing an adaptive pinning control method. Based on the
above discussions about control techniques, the intermittent adaptive pinning scheme is more
general, practical and economical due to it combines the advantages of the above three control
strategies. To the authors’ best knowledge, the intra-cluster lag consensus problem for general
linear MASs and nonlinear MASs via intermittent adaptive pinning control under a weakly
connected graph is still open issue.
Furthermore, due to the states of many actual systems are not always available or measured
directly in expensive cost, it is quite necessary to design an observer to estimate the agent’s
real-time state. Up to now, various issues are addressed by observer-based method. [20] inves-
tigated the formation tracking problem, where suitable Luenberger observers are used for the
agents to estimate linear and angular velocities. [21] studied the containment control issue of
general linear MASs with exogenous disturbances, where the disturbances are estimated by dis-
turbance observer-based control scheme. In order to estimate the state of leader and deal with
fixed-time stability problem, [22] considered the consensus problem for MASs with high-order
integrator dynamics by designing a distributed consensus observer. To our best knowledge, so
far the observer-based intra-cluster lag consensus of MASs has not been investigated. More-
over, compared with a single system or integrator MASs, the observer-based intra-cluster lag
consensus of MASs with general linear dynamics and nonlinear dynamics is more challenging.
Inspired by the aforementioned these considerations, this paper investigates the observer-
based intra-cluster lag consensus of general linear MASs and nonlinear MASs via intermittent
adaptive pinning control, where each cluster has a virtual leader whose state is available to
only a small part of followers, and the pinning gains are changeable because of communication
1812 ZHANG YUNLONG, et al.

constraints. The so-called intra-cluster lag consensus means that the followers in the same
cluster can achieve lag consensus asymptotically while the followers in different clusters can
achieve different agreements. The main contributions of this work are listed as follows.

• Firstly, considering the relevant full state information of each follower is not always avail-
able, the Luenberger observers are designed for general linear and nonlinear MASs re-
spectively to estimate the states of followers. Hence, the control protocols in this paper
only use the observed states information of followers instead of their real-time states
information.

• Secondly, to realize the intra-cluster lag consensus, the followers in each cluster are clas-
sified into three types. Specifically, the followers can only receive information from the
followers in their own cluster, the ones receive information from the followers in other
clusters, and the others cannot receive information from any followers. Then, the dis-
tributed intermittent adaptive pinning control protocols are further designed according
to the different categories of followers, which make sure that the followers in the same
cluster achieve intra-cluster lag consensus whereas the followers in the different clusters
achieve different consensus.

• Thirdly, compared with the most of existing results about cluster consensus where each
block unit in the weighted adjacency matrix of network topology is assumed to be a
zero-row-sum matrix or an equal-row-sum matrix, which is relative conservative in actual
applications, in this paper a weakly connected topology is considered and all the coupling
weights of network topology are positive, which is more universal in practice.

• Fourthly, different from continuous control protocols in [12] and [17], the pinning gains
in proposed control protocols are designed to be intermittent adaptive and with an ex-
ponential convergence rate, which can effectively reduce communication costs, avoid the
pinning gains being larger than those needed in practice. Meanwhile, it guarantees the
pinning gains quickly converge to steady value.

Notations Throughout this paper, the adopted notations are presented as follows. N
denotes the positive integer. Rn stands for the Euclidean space with dimension n. Rn×m
represents the set of entire n × m real matrices. Particularly, IN is the identity matrix. A−1
and AT are used to represent respectively the inverse and transpose of a real matrix A. A > 0
means that A is a positive definite matrix. For a matrix A, its ith eigenvalue is denoted as
λi (A) and correspondingly, the maximum and minimum eigenvalues are represented as λmax (A)
and λmin (A) respectively. The diag{· · · } is used to indicate a block-diagonal matrix. ∥ · ∥ not
only refers to Euclidean vector norm, but also spectral matrix norm. · and ⊗ denote product
T
and Kronecker product respectively. Denote A = A+A 2 , Al represents the minor matrix of A
by removing first l row and column.
 CLUSTER LAG CONSENSUS OF MULTI-AGENT SYSTEM 1813

2 Preliminaries
2.1 Graph Theory
A triplet G=(V, E, A) is used to characterize the interaction among follower agents, where
V = {v1 , v2 , · · · , vN } is the set of N follower agents and edge set E ⊆ V × V represents informa-
tion exchange of MASs, A = [aij ]N ×N is the weighted adjacency matrix. Besides, a directed
edge eij = (vi , vj ) ∈ E indicates that the information flows can be transformed from follower
j to follower i. Correspondingly, aij > 0 implies eji ∈ E; aij = 0, otherwise. The Laplacian
ΣN
matrix L = [lij ]N ×N of topology G is denoted as follows: lij = −aij if i /= j; lij = j=1,j/=i aij ,
ΣN
otherwise. Apparently, the Laplacian matrix holds the diffusion property, that is j=1 lij = 0.
ΣN ΣN
degin (vi ) = j=1 aij and degout (vi ) = j=1 aji denote the ith follower’s in-degree and out-
degree respectively. Moreover, for a directed topology G, a directed path indicates a finite
ordered nodes sequence vi1 , · · · , vik such that (vil , vil+1 ) ∈ E, where l = 1, · · · , k − 1. As a com-
parison, a weak path is that there exists a finite ordered nodes sequence vi1 , · · · , vik such that
(vil , vil+1 ) or (vil+1 , vil ) ∈ E. Furthermore, a directed topology G is called weakly connected
if there exists at least one weak path for any two distinct agents, which means its correspond-
ing undirected graph is connected and each follower can either obtain information from other
followers or send information to other followers. Based on the above discussions, in order to
reduce the restriction, the communication graph of followers agents is supposed to be weakly
connected throughout this paper.
Suppose that the MASs contains m clusters and every cluster has just only one virtual
leader. Let Vp be the set of pth cluster, one has V = {v1 , v2 , · · · , vN }=V1 ∪ V2 ∪ · · · ∪ Vm
and Vp ∩ Vq = ∅ with p /= q (p, q = 1, 2, · · · , m). For simplicity, the subscript set of virtual
leaders and followers are denoted as Ω = {1, 2, · · · , m} and V = {1, 2, · · · , N }, respectively.
For convenience of expression, there is no difference between the two descriptions i ∈ V and
vi ∈ V. Specially, the subscript of the jth leader and jth cluster is identical. Moreover, vi ∈ Vi
means that the ith follower is in the ith cluster, where i is the subscript of the cluster. In the
same cluster, the followers have different property. Refer to [13] for the definition of inter-act
followers and intra-act followers. Vi represents the subset of Vi and the agents in Vi means that
they can receive information from the agents in other cluster, that is, for any vi ∈ Vi , there
exists at least one vj ∈ Vi \ Vi such that aij /= 0. Furtherly, if vi ∈ Vi , the node vi is termed as
the inter-act follower, otherwise, if vi ∈ Vi \ Vi , vi is termed as the intra-act follower.

2.2 Some Useful Lemmas

Lemma 2.1 (see [23]) For the matrices P1 , P2 , P3 , and P4 with appropriate dimensions,
the following equalities are established:

(P1 ⊗ P2 )(P3 ⊗ P4 ) = (P1 P3 ) ⊗ (P2 P4 ),

(P1 ⊗ P2 )T = P1T ⊗ P2T ,
(P1 + P2 ) ⊗ P3 = P1 ⊗ P3 + P2 ⊗ P3 .

Lemma 2.2 (see [24]) G ∈ RN ×N is a symmetric matrix and D = diag{d1 , · · · , dl , 0 · · · ,

1814 ZHANG YUNLONG, et al.

( )
R−D S
0}N ×N is a diagonal matrix with d1 , · · · , dl > 0, 1 ≤ l ≤ N . Let G − D = S T Gl
, where
D = diag{d1 , · · · , dl }, matrices R and S have the appropriate dimensions, Gl is obtained by
removing the first l row(s) and column(s). If di > λmax (R − SG−1 T
l S ), then G − D < 0 is
identical with Gl < 0.

3 Main Results
3.1 Observer-Based Intra-Cluster Lag Consensus of MASs with General Linear
Dynamics via Intermittent Adaptive Pinning Control
Consider a general linear MAS containing N followers and m virtual leaders (N > m ≥ 2).
The ith follower’s dynamics is
{
ẋi (t) =Axi (t) + Bui (t),
i ∈ V, (1)
yi (t) =Cxi (t),

where xi ∈ Rn , ui ∈ Rm , yi ∈ Rq denote the state, control protocol and output measurement

of ith follower, respectively. The constant matrices A ∈ Rn×n , B ∈ Rn×m and C ∈ Rq×n have
appropriate dimensions. Suppose that the matrix triple (A, B, C) is stabilizable and detectable.
The jth virtual leader’s dynamics is modeled by
{
ṡj (t) = Asj (t),
j ∈ Ω, (2)
yj (t) = Csj (t),

where sj (t) ∈ Rn , yj ∈ Rq denote the state and output measurement, respectively.

Definition 3.1 The MASs (1)–(2) achieve the intra-cluster lag consensus if limt→∞ ||ηi (t)||
= 0 and limt→∞ ||sl (t) − sk (t)|| /= 0, where l /= k, l, k ∈ Ω , ηi (t) = xi (t) − si (t − τi ), i ∈ V ,
i ∈ Ω is the subscript of the leader that the ith follower will track with, τi denotes the time
delay in i cluster between the ith follower and its own leader.
Remark 3.2 Note that the trajectories of all virtual leaders are determined by system
matrix A owing to sj (t) = sj (0)eAt . If A is a Hurwitz matrix, then the states of all leaders will
asymptotically reach zeros, which goes against our objective. Therefore, in this paper, we as-
sume the matrix A is not Hurwitz. By giving different initial values for homogeneous system (2),
the states of virtual leaders will be different, that is, for any l /= k, limt→∞ ||sl (t) − sk (t)|| /= 0,
where l, k ∈ Ω .
Remark 3.3 Note that the existence of time-delays in many practical systems is ubiqui-
tous and inevitable due to communication, calculation, actuation. Therefore, more and more
researchers have focused their attention on the consensus problem of MASs with time delays.
Lag consensus, including consensus as a special case, which means the followers’ states are
consistent with the delayed states of leaders. For example, proper time delays between different
vehicles in the way can keep the road safe and orderly. Otherwise, congestion often occurs.
Taking into account lag consensus can mitigate network or traffic congestion problem, in this
paper, we consider the intra-cluster lag consensus of MASs, which means that the followers
 CLUSTER LAG CONSENSUS OF MULTI-AGENT SYSTEM 1815

in the same cluster can achieve lag consensus asymptotically while the followers in different
clusters can achieve different agreements.
Before proceeding, some explanations are presented for the reason why we propose the
following control protocol. In the former works about MASs[25] , the cluster lag consensus
was achieved based on two common assumptions. Firstly, the local relative state information
among all agents is available. Secondly, information between the leaders and followers can
be transmitted continuously. However, in many circumstances, full state measurements and
continuous communication are not always available due to physical constrain. Therefore, in
this paper, assuming that each follower can receive the relative output information instead of
its neighbors’ state information, the following observer-based control protocol for the follower
i in (1) is proposed:
⎧ N
⎪ Σ
⎪
⎪ aij (xj (t) − xi (t)) − di (t)(xi (t) − si (t − τi ))
⎪
⎪ K
⎪
⎪
⎪
⎪ j=1
⎪
⎪
⎪
⎪ ΣN
⎪
⎪ lij sj (t − τj ), if vi ∈ Vi ,
⎪
⎪ + K
⎪
⎪
⎪
⎨ j=1

ΣN
ui (t) = (3)
⎪
⎪ K aij (xj (t) − xi (t)) − di (t)(xi (t) − si (t − τi )),
⎪
⎪
⎪
⎪ j=1
⎪
⎪
⎪
⎪
⎪
⎪ if vi ∈ Vi \ Vi and deg(vi )in = 0,
⎪
⎪
⎪
⎪ N
⎪
⎪ Σ
⎪
⎪ K aij (xj (t) − xi (t)), otherwise,
⎩
j=1

where xi (t) ∈ Rn is the observer state, i and j are the subscripts of the leaders, τi and τj are
the time delays, K represents the feedback control gain matrix. The observer state xi (t) and
adaptive law di (t) are defined as follows:
{
x˙ i (t) = Axi (t) + Bui (t) + F (yi (t) − yi (t)),
(4)
yi (t) = C xi (t),
{ σt
σe 1 (xi (t) − si (t − τi ))T Γ (xi (t) − si (t − τi )), if t ∈ [mT, mT + δ),
d˙i (t) = (5)
0, if t ∈ [mT + δ, (m + 1)T ),

where matrices F and Γ will be determined later, σ and σ1 are positive constants, T > 0
and δ ∈ (0, T ) represent the control period and control width respectively, [mT, mT + δ) and
[mT + δ, (m + 1)T ) are called as work time interval and rest time interval.
Remark 3.4 Note that the followers in each cluster can be divided into three types.
Correspondingly, the controller (3) is designed according to the special property of each follower.
ΣN
For all the followers, The term K j=1 aij (xj (t) − xi (t)) in the control protocol (3) is necessary,
which reflects the interaction between agent i and its neighbours. For the followers which
can receive information from other clusters or can not receive information from any followers,
1816 ZHANG YUNLONG, et al.

the term di (t)(xi (t) − si (t − τi )) is introduced to obtain the information of leaders and the
time delay τi can be different among different clusters. Besides, for the followers that can
ΣN
receive information from other clusters, the term K j=1 lij sj (t − τj ) is used to counteract the
influence among clusters. The purpose of the controller (3) is to make sure that the followers
in the same cluster achieve intra-cluster lag consensus whereas the followers in the different
clusters achieve different consensus.
Remark 3.5 The design of the controller (3) is partly inspired by [12, 13], where the
pinning control gain di (t) is a fixed constant all the time or in every time interval. However,
due to the change of external environment or in the consideration of reducing control cost, it
is more reasonable and effective to design the pinning control gain to be intermittent adaptive.
Furthermore, in order to make the pinning control gain quickly converge to steady value, the
pinning control gain di (t) possesses an exponential convergence rate. Based on the above
analysis, the derivative of di (t) is designed to be the form of the equation (5) if ith follower is
pinned; di (t) = 0 otherwise.
ΣN
Lemma 3.6 If vi is the follower that belongs to Vi \ Vi , j=1 lij sj (t − τj ) = 0.
Proof From the definition of Vi , if vi ∈ Vi \ Vi , one has lij = 0 for ∀vj ∈ / Vi . Besides, if
ΣN
vj ∈ Vi , one has sj (t − τj ) = si (t − τi ). Combining the fact j=1 lij = 0, it follows that
N
Σ Σ Σ
lij sj (t − τj ) = lij sj (t − τj ) + lij sj (t − τj )
j=1 vj ∈Vi vj ∈V
/ i
Σ
= lij si (t − τi )
vj ∈Vi
( Σ Σ ) (6)
= lij + lij si (t − τi )
vj ∈Vi vj ∈V
/ i
N
Σ
= lij si (t − τi ) = 0.
j=1

For convenience, let Vi1 be the set of the followers that belong to Vi , Vi2 be the set of the
followers that belong to Vi \ Vi and deg(vi )in = 0, and Vi3 be the set of remaining followers.
Combining the definition ei (t) = xi (t) − si (t − τi ) and the formulas (2)–(5), one has

⎧ N
⎪
⎪ Σ
⎪
⎪ Aei (t) − BK lij ej (t) − di (t)BKei (t) + F (yi (t) − yi (t)), if vi ∈ Vi1 ∪ Vi2 ,
⎪
⎨ j=1
ėi (t) = (7)
⎪
⎪ N
Σ
⎪
⎪
⎪
⎩ Ae i (t) − BK lij ej (t) + F (yi (t) − yi (t)), if vi ∈ Vi3 ,
j=1
{ σ1 t T
σe ei (t) Γ ei (t), if t ∈ [mT, mT + δ),
d˙i (t) = (8)
0, if t ∈ [mT + δ, (m + 1)T ).

For convenience, let the first l (l = 1, 2, · · · , N ) followers be pinned by rearranging the followers’
 CLUSTER LAG CONSENSUS OF MULTI-AGENT SYSTEM 1817

order. Due to the fact that the derivative of di (t) satisfies the equation (5), that is, for any
time t, di (t) > 0 if the i-th follower is pinned; di (t) = 0 otherwise. When vi ∈ Vi3 , which means
di (t) = 0, then the error system (7) can be reformulated as
N
Σ
ėi (t) =Aei (t) − BK lij ej (t) − di (t)BKei (t)
j=1 (9)
+ F (yi (t) − yi (t)), vi ∈ Vi1 ∪ Vi2 ∪ Vi3 .

The proof is finished.

Theorem 3.7 Suppose that the matrix triple (A, B, C) is stabilizable and detectable, there
exist a diagonal matrix D, a positive constant μ1 and a positive definite matrix P = P T > 0
such that the following equations:

L + D > 0, (10)
μ1 · λmin (2L + 2D) > 1, (11)
AT P + P A − μ1 P BB T P λmin (2L + 2D) < 0, (12)
T
where L = L+L 2 , D = diag{d1 , · · · , dl , 0 · · · , 0}N ×N is a diagonal matrix with d1 , · · · , dl >
0, 1 ≤ l ≤ N . Then letting K = μ1 B T P , Γ = 2μ1 P BB T P , and designing a matrix F such that
A + F C is Hurwitz matrix, the observer-based intra-cluster lag consensus of MASs (1)–(2) can
be reached by control input (3) with the following condition:

−α1 δ + β1 (T − δ) < 0, (13)

−λ1 T T
where α1 = min{ λmax (P ) , σ1 }, λ1 = λmax (A P + P A − μ1 P BB P λmin (2L + 2D)), β1 =
λ2
max{ λmin (P ) , σ1 }, λ2 = max{λmax (IN ⊗ (A P + P A) − 2μ1 L ⊗ P BB T P ), ε}, σ, σ1 and ε are
T

positive constants.
Proof Theorem 3.7 can be proven by two steps: The first step is to show the effectiveness of
state observer, i.e., limt→∞ ||xi (t) − xi (t)|| = 0, and the second step is to prove the asymptotic
stability of error system (9), that is, limt→∞ ||ei (t)|| = 0.
Step 1 Let ξi (t) = xi (t) − xi (t), one has

ξ˙i (t) =Axi (t) + Bui (t) − Axi (t) − Bui (t) − F (yi (t) − yi (t))
=A(xi (t) − xi (t)) + F C(xi (t) − xi (t)) (14)
=(A + F C)ξi (t).

Based on the condition in Theorem 3.7: A + F C is Hurwitz matrix, we have limt→∞ ||ξi (t)|| =
lim ||xi (t) − xi (t)|| = 0, which indicates the state observer is feasible. Furthermore, one has
t→∞
F (yi (t) − yi (t)) = F Cξi (t) = 0 as t → ∞. In the equation (9), since ξi (t) is decoupled from the
error ei (t), the stability of (9) is equal to the stability of the following system:
N
Σ
ėi (t) = Aei (t) − BK lij ej (t) − di (t)BKei (t), vi ∈ Vi1 ∪ Vi2 ∪ Vi3 . (15)
j=1
1818 ZHANG YUNLONG, et al.

Step 2 Construct the Lyapunov function for the equation (15):

N
Σ N
1 Σ −σ1 t (di (t) − di )2
V (t) = eT
i (t)P ei (t) + e , (16)
i=1
2 i=1 σ

where matrix P and the parameters σ, di , σ1 are defined in Theorem 3.7. Evidently, V (t) ≥ 0.
1) When t ∈ [mT, mT + δ), differentiating V (t) along (15) yields
N
Σ ( N
Σ )
V̇ (t) =2 eT
i (t)P Aei (t) − BK lij ej (t) − di (t)BKei (t)
i=1 j=1
(17)
N
Σ ΣN
1 (di (t) − di )2
+ (−σ1 ) e−σ1 t + (di (t) − di ) eT
i (t)Γ ei (t).
2 i=1
σ i=1

Let D(t) = diag{d1 (t), · · · , dl (t), 0, · · · , 0}N ×N , e(t) = (eT T T T

1 (t), e2 (t), · · · , eN (t)) . With
K = μ1 B T P and Γ = 2μ1 P BB T P , the derivative of V (t) can be expressed as

V̇ (t) =eT (t)(IN ⊗ AT P − LT ⊗ K T B T P − D(t) ⊗ K T B T P )e(t)

+ eT (t)(IN ⊗ P A − L ⊗ P BK − D(t) ⊗ P BK)e(t)
N
1 Σ −σ1 t (di (t) − di )2
+ (−σ1 ) · e + eT (t)(D(t) − D) ⊗ (2μ1 P BB T P )e(t)
2 i=1 σ
=eT (t)(IN ⊗ (AT P + P A) − 2μ1 L ⊗ P BB T P − 2μ1 · D(t) ⊗ P BB T P )e(t)
N
1 Σ −σ1 t (di (t) − di )2
+ (−σ1 ) · e + eT (t)(D(t) − D) ⊗ (2μ1 P BB T P )e(t)
2 i=1 σ
=eT (t)(IN ⊗ (AT P + P A) − μ1 (2L + 2D) ⊗ P BB T P )e(t) (18)
N
1 Σ −σ1 t (di (t) − di )2
+ (−σ1 ) · e
2 i=1 σ
N
1 Σ −σ1 t (di (t) − di )2
≤λ1 eT (t)e(t) + (−σ1 ) · e
2 i=1 σ
N
λ1 1 Σ −σ1 t (di (t) − di )2
≤ eT (t)(IN ⊗ P )e(t) + (−σ1 ) · e
λmax (P ) 2 i=1 σ
≤ − α1 V (t),

−λ1
where λ1 = λmax (AT P + P A − μ1 P BB T P λmin (2L + 2D)), α1 = min{ λmax (P ) , σ1 } > 0.
2) When t ∈ [mT + δ, (m + 1)T ), the derivative of (16) can be obtained:
N
Σ N
Σ ΣN
1 (di (t) − di )2
V̇ (t) =2 eT
i (t)P (Aei (t) − BK lij ej (t)) + (−σ1 ) e−σ1 t
i=1 j=1
2 i=1
σ
 CLUSTER LAG CONSENSUS OF MULTI-AGENT SYSTEM 1819

N
1 Σ −σ1 t (di (t) − di )2
=eT (t)(IN ⊗ (AT P + P A) − μ1 (2L) ⊗ P BB T P )e(t) − σ1 e
2 i=1 σ
N
1 Σ −σ1 t (di (t) − di )2
≤λmax (IN ⊗ (AT P + P A) − 2μ1 L ⊗ P BB T P )eT (t)e(t) + σ1 · e
2 i=1 σ
N
1 Σ −σ1 t (di (t) − di )2
≤λ2 eT (t)e(t) + σ1 · e
2 i=1 σ
N
λ2 1 Σ −σ1 t (di (t) − di )2
≤ eT (t)(IN ⊗ P )e(t) + σ1 · e
λmin (P ) 2 i=1 σ
≤β1 V (t),
λ2
where λ2 = max{λmax (IN ⊗ (AT P + P A) − 2μ1 L ⊗ P BB T P ), ε}, β1 = max{ λmin (P ) , σ1 } > 0.
Accordingly, when t ∈ [0, δ), one has

V (t) ≤ V (0)e−α1 t , (19)

when t ∈ [δ, T ), one has

V (t) ≤ V (δ)eβ1 (t−δ) ≤ V (0)e−α1 δ+β1 (t−δ) . (20)

By induction, one obtains, when t ∈ [mT, mT + δ),

V (t) ≤V (0)e−α1 (mδ+t−mT )+β1 m(T −δ)

≤V (0)e−α1 mδ+β1 m(T −δ) (21)
m(−α1 δ+β1 (T −δ))
=V (0)e .

Similarly, when t ∈ [mT + δ, (m + 1)T ),

V (t) ≤V (0)e−α1 (m+1)δ+β1 (m(T −δ)+t−T2k+1 )

≤V (0)e−α1 (m+1)δ+β1 (m(T −δ)+(T −δ)) (22)
(m+1)(−α1 δ+β1 (T −δ))
=V (0)e .

According to the equation (13) in Theorem 3.7: −α1 δ + β1 (T − δ) < 0, one has V (t) → 0
and limt→∞ ||ei (t)|| = 0 as m → ∞. Combining the first step and the second step, one has
limt→∞ ||ηi (t)|| = limt→∞ ||xi (t) − xi (t) + xi (t) − si (t − τi )||=0. It means that intra-cluster lag
consensus of general linear MASs via intermittent adaptive pinning control can be reached.
This proof is completed.
Remark 3.8 The proof of Theorem 3.7 is divided into two parts. The first step is to
show that the state observer is valid by only using the relevant output information of the
system (1). In the second step, when t ∈ [mT, mT + δ), the derivative of V (t) is negative due
to the effect of adaptive pinning gain di (t), which indicates the limt→∞ ||ei (t)|| = 0. However,
when t ∈ [mT + δ, (m + 1)T ), the derivative of V (t) cannot be determined to be negative, while
it is only obtained that V̇ (t) ≤ β1 V (t) by choosing the appropriate positive constants ε, λ2 , β1 .
1820 ZHANG YUNLONG, et al.

Hence, the equation (13) is critical to ensure V ((m + 1)T ) → 0 as m → ∞. To make it more
δ
clear, define the control rate θ = T −δ , where δ is the control width, T − δ is the rest width,
then the equation (13) can be rewritten as: θ > αβ11 . It can be seen that the control rate θ plays
an essential role in realizing intra-cluster lag consensus. In consideration of this fact, in real
applications, we can select the control period T according to our request.
Remark 3.9 The condition (12) is equivalent to the following Riccati equation condition:
A P + P A − μ1 P BB T P λmin (2L + 2D) + Q = 0, where Q is any positive definite matrix.
T

The purpose for the form (12) is to make a comparison with the following condition (32) in
Theorem 3.14.
The general criteria are presented in Theorem 3.7 to reach intra-cluster lag consensus of
MASs. However, how to construct an appropriate gain matrix D to satisfy the condition (10):
L + D > 0 is not obvious. Hence, we give Theorem 3.10 to solve this problem.
Theorem 3.10 The matrix L + D is positive definite with the following prerequisites

λmax (−Ll ) < 0, (23)

−1 T
di > λmax (R − S(−Ll ) S ). (24)
( )
R−D S
Proof In Lemma 2.2, let G = −L, one has −L − D = S T −Ll
, where D = diag{d1 , · · · ,
dl }, −Ll is obtained by removing the first l row and column, R and S represent the matrices
with appropriate dimensions. Thanks to Lemma 2.2 and combine with (23), (24), we have
−L − D < 0, that is, L + D is a positive definite matrix. This completes the proof.
Remark 3.11 In [11, 26, 27], in order to achieve cluster synchronization or group con-
sensus, the topology contains negative weight couplings aij < 0 and is required to satisfy the
Σ Σ
following assumption: j∈Vk aij = 0, or j∈Vk aij = β, ∀k /= i, which is too conservative
in practical applications. In order to remove the restrictions, [12] and [13] employed pinning
leader-following control scheme to reach cluster consensus of MASs under weakly connected
graph. Inspired by the works [12] and [13], in this paper, the intermittent adaptive pinning
control is designed for general linear MASs and Lipschitz nonlinear MASs under weakly con-
nected graph, in which the Laplace matrix is just supposed to satisfy the general diffusion
ΣN
property, that is, j=1 lij = 0. Moreover, by choosing an appropriate low bound pinning gain
di , one can get L + D > 0. For an undirected connected graph, L + D > 0 can hold naturally
due to the Laplace matrix is a real symmetry. Therefore, the MASs with undirected connected
graph can be regarded as a special case only by choosing the inter-act followers to be pinned.

3.2 Observer-Based Intra-Cluster Lag Consensus of MASs with Nonlinear Dy-

namics via Intermittent Adaptive Pinning Control
In practice, many physical systems are nonlinear. In this section, the observer-based intra-
cluster lag consensus of MASs with nonlinear dynamics via intermittent adaptive pinning control
is further considered, which is more challenging due to the complexity of the system structure.
 CLUSTER LAG CONSENSUS OF MULTI-AGENT SYSTEM 1821

The dynamics of the ith follower is characterized by

{
ẋi (t) = Axi (t) + Bui (t) + f (xi (t)),
i ∈ V, (25)
yi (t) = Cxi (t),
and the virtual leader j’s dynamics is described by
{
ṡj (t) = Asj (t) + f (sj (t)),
j ∈ Ω, (26)
yj (t) = Csj (t),
where the nonlinear function f (·) holds the following Assumption 3.12.
Assumption 3.12 It is assumed that the nonlinear function f (·) satisfies the Lipschiz
condition, i.e., for any vectors x, y ∈ Rm , there exists a constant γ > 0 such that ∥f (x) −
f (y)∥ ≤γ∥x − y∥.
Remark 3.13 The Lipschiz condition assumption is widely used in the analysis of non-
linear systems. With the help of Lipschiz condition assumption, the existence and uniqueness
of nonlinear systems’ solutions can be guaranteed. Refer to literatures [28] and [29] for more
details.
Before going on, the state observer for the ith nonlinear follower agent is designed as follows
{
x˙ i (t) = Axi (t) + Bui (t) + f (xi (t)) + F (yi (t) − yi (t)),
i ∈ V. (27)
yi (t) = C xi (t),
In this section, let’s consider control protocol (3) with (5), taking the control protocol (3) for
the system (25)–(26) and combining to the definition of ei (t), one has
⎧ N
⎪
⎪ Σ
⎪
⎪ Ae (t) − BK lij ej (t) − di (t)BKei (t) + Fi (t) + F (yi (t) − yi (t)), if vi ∈ Vi1 ∪ Vi2 ,
⎪
⎨
i
j=1
ėi (t) =
⎪
⎪ ΣN
⎪
⎪
⎪
⎩ Ae i (t) − BK lij ej (t) + Fi (t) + F (yi (t) − yi (t)), if i ∈ Vi3 ,
j=1

{ (28)
σeσ1 t ei (t)T Γ ei (t), if t ∈ [mT, mT + δ),
d˙i (t) = (29)
0, if t ∈ [mT + δ, (m + 1)T ),
where Fi (t) = f (xi (t)) − f (si (t)).
Theorem 3.14 Suppose that the matrix triple (A, B, C) is stabilizable and detectable,
there exist a diagonal matrix D, a positive constant μ2 and a positive definite matrix P =
P T = Q−1 > 0 such that the following equations:

L + D > 0, (30)
μ2 · λmin (2L + 2D) > τ, (31)
⎛ ⎞
AQ + QAT − τ BB T + γ 2 I Q
⎝ ⎠ < 0, (32)
QT −I
1822 ZHANG YUNLONG, et al.

T
where L = L+L 2 , D = diag{d1 , · · · , dl , 0 · · · , 0}N ×N is a diagonal matrix with d1 , · · · , dl >
0, 1 ≤ l ≤ N . Then letting K = μ2 B T P , Γ = 2μ2 P BB T P , and designing a matrix F such that
A+γI +F C is Hurwitz matrix, the observer-based intra-cluster lag consensus of MASs (25)–(26)
can be reached by control input (3) with the following condition:

−α2 δ + β2 (T − δ) < 0, (33)

−1 2
where α2 = min{−λ3 · λmin (P ) T 2 2 T
λmax (P ) , σ1 }, λ3 = λmax (AQ + QA + γ I + Q − τ BB ), β2 =
λ4 T 2 2 T
max{ λmin (P ) , σ1 }, λ4 = max{λmax (IN ⊗ (A P + P A + I + γ P ) − 2μ2 L ⊗ P BB P ), ε}, σ,
σ1 and ε are positive constants.
Proof The proof is still divided into two parts: The first part is to show the effective-
ness of state observer, i.e., limt→∞ ||xi (t) − xi (t)|| = 0. And the second part is to prove
limt→∞ ||ei (t)|| = 0.
The first step Let ξi (t) = xi (t) − xi (t), one has

ξ˙i (t) =Axi (t) + Bui (t) − Axi (t) − Bui (t) + f (xi (t)) − f (xi (t)) − F (yi (t) − yi (t))
=A(xi (t) − xi (t)) + F C(xi (t) − xi (t)) + f (xi (t)) − f (xi (t)) (34)
=(A + F C)ξi (t) + f (xi (t)) − f (xi (t)).

Construct the following Lyapunov function candidate:

N
1Σ T
V (t) = ξ (t)ξi (t). (35)
2 i=1 i

Taking the time derivative of (35) yields

N
Σ N
Σ
V̇ (t) = ξiT (t)((A + F C)ξi (t) + f (xi (t)) − f (xi (t))) ≤ ξiT (t)(A + F C + γI)ξi (t). (36)
i=1 i=1

Based on the condition A+F C+γI is Hurwitz matrix, one has V̇ (t) < 0. Thus, limt→∞ ||ξi (t)||
= limt→∞ ||xi (t) − xi (t)|| = 0, as well F (yi (t) − yi (t)) = F Cξi (t) = 0. In the equation (28),
owing to ξi (t) is decoupled from the error ei (t), the stability of (28) is equal to the stability of
the following system:
⎧ N
⎪
⎪ Σ
⎪ Aei (t) − BK
⎪ lij ej (t) − di (t)BKei (t) + Fi (t), if vi ∈ Vi1 ∪ Vi2 ,
⎪
⎨ j=1
ėi (t) = (37)
⎪
⎪ ΣN
⎪
⎪ 3
⎩ Aei (t) − BK
⎪ lij ej (t) + Fi (t), if vi ∈ Vi .
j=1

The second step Construct the Lyapunov function candidate:

N
Σ N
1 Σ −σ1 t (di (t) − di )2
V (t) = eT
i (t)P ei (t) + e . (38)
i=1
2 i=1 σ
 CLUSTER LAG CONSENSUS OF MULTI-AGENT SYSTEM 1823

1) When t ∈ [mT, mT + δ), the derivative of (38) is

N
Σ ( N
Σ )
V̇ (t) =2 eT
i (t)P Ae i (t) − BK l e
ij j (t) − di (t)BKe i (t) + f (xi (t)) − f (s i (t))
i=1 j=1
(39)
N
Σ 2 ΣN
1 −σ1 t (di (t) − di )
+ (−σ1 ) e + (di (t) − di ) eT
i (t)Γ ei (t).
2 i=1
σ i=1

Denote F (t) = (f T (x1 (t)) − f T (s1 (t)), f T (x2 (t)) − f T (s2 (t)), · · · , f T (xN (t)) − f T (sN (t)))T .
Based on Assumption 3.12, one has

2eT T 2 2
i (t)P (f (xi (t)) − f (si (t)) ≤ 2γ||P ei (t)|| · ||ei (t)|| ≤ ei (t)(I + γ P )ei (t). (40)

Combining (40) and the conditions in Theorem 3.14, V̇ (t) can be written as
V̇ (t)
=eT (t)(IN ⊗ AT P − LT ⊗ K T B T P − D(t) ⊗ K T B T P )e(t)
+ eT (t)(IN ⊗ P A − L ⊗ P BK − D(t) ⊗ P BK)e(t) + 2eT (t)(IN ⊗ P )F (t)
N
1 Σ −σ1 t (di (t) − di )2
+ (−σ1 ) e + eT (t)(D(t) − D) ⊗ (2μ2 P BB T P )e(t)
2 i=1 σ

≤eT (t)(IN ⊗ (AT P + P A + I + γ 2 P 2 ) − 2μ2 (L) ⊗ P BB T P − 2μ2 D(t) ⊗ P BB T P ))e(t) (41)

N
1 Σ −σ1 t (di (t) − di )2
+ (−σ1 ) e + eT (t)(D(t) − D) ⊗ (2μ2 P BB T P )e(t)
2 i=1 σ

≤eT (t)(IN ⊗ (AT P + P A + I + γ 2 P 2 ) − μ2 (2L + 2D) ⊗ P BB T P ))e(t)

N
1 Σ −σ1 t (di (t) − di )2
+ (−σ1 ) e .
2 i=1 σ

Based on the condition in Theorem 3.14, the matrix 2L + 2D > 0, there exists a constant μ2
satisfies μ2 λmin (2L + 2D) > τ . Let Q = P −1 , η(t) = (IN ⊗ P )e(t), one has
V̇ (t) ≤η T (t)(IN ⊗ (AQ + QAT + γ 2 I + Q2 ) − IN ⊗ τ BB T )η(t)
N
1 Σ −σ1 t (di (t) − di )2 (42)
+ (−σ1 ) e .
2 i=1 σ
( T
)
BB T +γ 2 I
Furthermore, since the conditions in Theorem 3.14: AQ+QA −τ QT
Q
−I
< 0, and with
the help of Schur complement lemma , one has AQ + QA + γ I + Q − τ BB T < 0. Denote
[30] T 2 2

λ3 = λmax (AQ + QAT + γ 2 I + Q2 − τ BB T ), it follows

N
1 Σ −σ1 t (di (t) − di )2
V̇ (t) ≤ λ3 η T (t)η(t) + (−σ1 ) · e
2 i=1 σ
N
T 1 Σ −σ1 t (di (t) − di )2 (43)
≤ λ3 λmin (P )e (t)(IN ⊗ P )e(t) + (−σ1 ) · e
2 i=1 σ
≤ −α2 V (t),
1824 ZHANG YUNLONG, et al.

where α2 = min{−λ3 λmin (P ), σ1 } > 0.

2) When t ∈ [mT + δ, (m + 1)T ), the derivative of (38) follows
N
Σ ( N
Σ ) N
Σ
V̇ (t) =2 eT
i (t)P Aei (t) − BK lij ej (t) + 2 eT
i (t)P (f (xi (t)) − f (si (t))
i=1 j=1 i=1
N
Σ 2
1 (di (t) − di )
+ (−σ1 ) e−σ1 t
2 i=1
σ
≤eT (t)(IN ⊗ (AT P + P A + I + γ 2 P 2 ) − μ2 (LT + L) ⊗ P BB T P )e(t)
N
1 Σ −σ1 t (di (t) − di )2
− σ1 e (44)
2 i=1 σ
N
1 Σ −σ1 t (di (t) − di )2
≤λ4 eT (t)e(t) + σ1 · e
2 i=1 σ
N
λ4 1 Σ −σ1 t (di (t) − di )2
≤ eT (t)(IN ⊗ P )e(t) + σ1 · e
λmin (P ) 2 i=1 σ
≤β2 V (t),

where λ4 = max{λmax IN ⊗ (AT P + P A + I + γ 2 P 2 ) − μ2 (LT + L) ⊗ P BB T P , ε} > 0, β2 =

λ4
max{ λmin (P ) , σ1 } > 0.
Next, by the same operations with (19)–(22) and combining to the condition in Theo-
rem 3.14: −α2 δ +β2 (T −δ) < 0, one has V (t) → 0 and limt→∞ ||ei (t)|| = 0 as m → ∞. Combin-
ing the first step and the second step, one obtains limt→∞ ||ηi (t)|| = limt→∞ ||xi (t)− xi (t)+ xi (t)
− si (t − τi )||=0. It means that intra-cluster lag consensus of MASs with nonlinear dynamics
via intermittent adaptive pinning control can be reached. This completes the proof.

4 Simulation Results
For conveniently, we just present the simulation result of Theorem 3.14. Consider a nonlinear
MAS composed of three clusters, where the virtual leaders are denoted as S1 , S2 and S3 , and
the followers are denoted as 1 to 9. Figure 1 is the communication topology. In Figure 1,
followers 1 and 2 can receive information from other clusters, followers 3 and 4 can not receive
information from any followers. Therefore, according to the control strategy (3), followers 1, 2,
3, 4 should be pinned. In addition, from the graph theory, the matrices A and L can be derived
as follows: ⎛ ⎞ ⎛ ⎞
0 0 0 0 0 0 0 1 0 1 0 0 0 0 0 0 −1 0
⎜0 0 0 0 0 0 0 0 1⎟ ⎜ 0 1 0 0 0 0 0 0 −1 ⎟
⎜ ⎟ ⎜ ⎟
⎜0 0 0 0 0 0 0 0 0⎟ ⎜ 0 0 0 0 0 0 0 0 0 ⎟
⎜ ⎟ ⎜ ⎟
⎜0 0 0 0 0 0 0 0 0⎟ ⎜ 0 0 0 0 0 0 0 0 0 ⎟
⎜ ⎟ ⎜ ⎟
A = ⎜1 0 0 0 0 0 0 0 0⎟ and L = ⎜ −1 0 0 0 1 0 0 0 0 ⎟.
⎜0 ⎟ ⎜ 0 ⎟
⎜ 1 0 0 0 0 0 0 0⎟ ⎜ −1 0 0 0 1 0 0 0 ⎟
⎜0 ⎟ ⎜ 0 ⎟
⎜ 0 0 0 0 1 0 0 0⎟ ⎜ 0 0 0 0 −1 1 0 0 ⎟
⎝0 0 1 0 0 0 0 0 0⎠ ⎝ 0 0 −1 0 0 0 0 1 0 ⎠
0 0 0 1 0 0 0 1 0 0 0 0 −1 0 0 0 −1 2
 CLUSTER LAG CONSENSUS OF MULTI-AGENT SYSTEM 1825

Figure 1 The interaction topology of MASs with s1 , s2 , s3 being

leaders and 1 to 9 being followers
( 0 −2 )
Furthermore, (
for the
) dynamics of nonlinear MASs (25) and (26), we choose A = 1 0 ,
B = ( 01 ) , C = 1 1 , and f (x) = 0.1 cos(x). Meanwhile, by calculating, the conditions
( 3.9453 −3.2213 ) ( )
in Theorem 3.14 are shown as follows: P = Q−1 = −3.2213 , F = −4
−2 , K = 2 ·
( ) 6.5765

−3.2213 6.5765 , di = 4, i = 1, 2, 3, 4, τ = 1.5521, μ2 = 2, σ = 1, σ1 = 0.1, ε = 0.01. T = 1

s, δ = 0.96. Specially, the time delays are given as: τ1 = 0.1, τ2 = 0.15, τ3 = 0.12. Figures 2,
3 and 4 are state trajectories of leader and followers in each cluster respectively. Figure 5 is
the time evolution of di (t). Figure 6 is the observer error ξi (t). From the above figures, we
can conclude that the intra-cluster lag consensus of MASs is realised via intermittent adaptive
pinning control.

2
(t)

1
1j
x i(t) and s

-1

-2

-3

-4

-5
0 2 4 6 8 10 12 14 16 18 20
time t

Figure 2 The state trajectories of leader S1 and followers in the

first cluster, where j = 1, 2; i = 1, 5
1826 ZHANG YUNLONG, et al.

2
(t) 2j
x i(t) and s

-2

-4

-6
0 2 4 6 8 10 12 14 16 18 20
time t

Figure 3 The state trajectories of leader S2 and followers in the

second cluster, where j = 1, 2; i = 2, 6, 7

10

4
(t)

2
3j
x i(t) and s

-2

-4

-6

-8

-10
0 2 4 6 8 10 12 14 16 18 20
time t

Figure 4 The state trajectories of leader S3 and followers in the

third cluster, where j = 1, 2; i = 3, 4, 8, 9

4
d i(t)

d1
3
d2
d3
2
d
4

0
0 2 4 6 8 10 12 14 16 18 20
time t

Figure 5 The time evolution of di (t), where i = 1, 2, 3, 4

 CLUSTER LAG CONSENSUS OF MULTI-AGENT SYSTEM 1827

1.5

0.5
(t)
i

-0.5

-1

-1.5
0 2 4 6 8 10 12 14 16 18 20
time t

Figure 6 The observer error ξi (t), where i = 1, 2, · · · , 9

5 Conclusions
In this paper, the observer-based intra-cluster lag consensus problem for general linear MASs
and nonlinear MASs under a weakly connected graph via intermittent adaptive pinning control
has been investigated. Considering the relevant full state information of each follower is not
always available, a class of observers has been designed to estimate the states of followers. Then
a class of observer-based intermittent adaptive pinning control protocols is proposed according
to the difference that the agents receive information source. Moreover, the pinning gains have
been designed to be intermittent adaptive and with an exponential convergence rate, which
will effectively reduce communication costs, avoid the pinning gains being larger than those
needed in practice and guarantees the pinning gains quickly converge to steady value. Finally,
rigorous proofs and numerical simulations have been provided to guarantee the correctness of
the theoretical results. Future works will focus on furtherly reducing communication among
agents by employing intermittent event-triggered control mechanism.

References

[1] Li C and Liu G, Data-driven leader-follower output synchronization for networked non-linear
multi-agent systems with switching topology and time-varying delays, Journal of Systems Science
& Complexity, 2018, 31(1): 87–102.
[2] Wang F, Wen G, Peng Z, et al., Event-triggered consensus of general linear multiagent sys-
tems with data sampling and random packet losses, IEEE Transactions on Systems, Man, and
Cybernetics: Systems, 2021, 51(2): 1313–1321.
[3] Xiong G, Wen G, Peng Z, et al., Pull-based event-triggered containment control for multiagent
systems with active leaders via aperiodic sampled-data transmission, IEEE Transactions on Sys-
tems, Man, and Cybernetics: Systems, 2022, 52(1): 214–225.
1828 ZHANG YUNLONG, et al.

[4] George K, Sanjiv S, Ronald P, et al., Distributed search and rescue with robot and sensor teams,
Field and Service Robotics, Springer, Berlin, Heidelberg, 2003, 529–538.
[5] Ren W, Formation keeping and attitude alignment for multiple spacecraft through local interac-
tions, Journal of Guidance, Control, and Dynamics, 2007, 30(2): 633–638.
[6] Li B, Wen G, Peng Z, et al., Time-varying formation control of general linear multi-agent systems
under Markovian switching topologies and communication noises, IEEE Transactions on Circuits
and Systems II: Express Briefs, 2020, 68(4): 1303–1307.
[7] Olfati-Saber R, Flocking for multi-agent dynamic systems: Algorithms and theory, IEEE Trans-
actions on Automatic Control, 2006, 51(3): 401–420.
[8] Wen G, Zhai X, Peng Z, et al., Fault-Tolerant secure consensus tracking of delayed nonlin-
ear multi-agent systems with deception attacks and uncertain parameters via impulsive control,
Communications in Nonlinear Science and Numerical Simulation, 2020, 82: 105043.
[9] Hu W, Wen G, Rahmani A, et al., Distributed consensus tracking of unknown nonlinear chaotic
delayed fractional-order multi-agent systems with external disturbances based on ABC algorithm,
Communications in Nonlinear Science and Numerical Simulation, 2019, 71: 101–117.
[10] Rahmani R, Toshani H, and Mobayen S, Consensus tracking of multi-agent systems using con-
strained neural-optimiser-based sliding mode control, International Journal of Systems Science,
2020, 51(14): 2653–2674.
[11] Kang Y, Qin J, Ma Q, et al., Cluster synchronization for interacting clusters of nonidentical nodes
via intermittent pinning control, IEEE Transactions on Neural Networks and Learning Systems,
2018, 29(5): 1747–1759.
[12] Ma Q, Wang Z, and Miao G, Second-order group consensus for multi-agent systems via pinning
leader-following approach, Journal of the Franklin Institute, 2014, 351(3): 1288–1300.
[13] Huang D, Jiang H, Yu Z, et al., Cluster-delay consensus in multi-agent systems via pinning
leader-following approach with intermittent effect, International Journal of Control, 2018, 91(10):
2261–2272.
[14] Wang J, Shi L, and Guan X, Semi-global leaderless consensus of linear multi-agent systems with
actuator and communication constraints, Journal of Systems Science & Complexity, 2020, 33(4):
882–902.
[15] Yang F, Li H, Chen G, et al., Cluster lag synchronization of delayed heterogeneous complex
dynamical networks via intermittent pinning control, Neural Computing and Applications, 2019,
31(11): 7945–7961.
[16] Yu Z, Jiang H, and Cheng H, Second-order consensus for multi-agent systems via intermittent
sampled data control, IEEE Transactions on Systems Man & Cybernetics Systems, 2017, 48(11):
1986–2002.
[17] Huang J, Wen G, Peng Z, et al., Group-consensus with reference states for heterogeneous multi-
agent systems via pinning control, International Journal of Control, Automation and Systems,
2019, 17(5): 1096–1106.
[18] Wang Z and Li H, Cluster consensus in multiple Lagrangian systems under pinning control, IEEE
Access, 2017, 5: 11291–11297.
[19] Xu C, Zheng Y, Su H, et al., Cluster consensus for second-order mobile multi-agent systems via
distributed adaptive pinning control under directed topology, Nonlinear Dynamics, 2016, 83(4):
1975–1985.
[20] Rosaldo-Serrano M, Santiaguillo-Salinas J, and Aranda-Bricaire E, Observer-based time-varying
 CLUSTER LAG CONSENSUS OF MULTI-AGENT SYSTEM 1829

backstepping control for a quadrotor multi-agent system, Nonlinear Dynamics, 2019, 93(1–2):
135–150.
[21] Han T, Li J, Guan Z, et al., Containment control of multi-agent systems via a disturbance
observer-based approach, Journal of the Franklin Institute, 2019, 356(5): 2919–2933.
[22] Zuo Z, Defoort M, Tian B, et al., Distributed consensus observer for multi-agent systems with
high-order integrator dynamics, IEEE Transactions on Automatic Control, 2019, 65(4): 1771–
1778.
[23] Roger H and Johnson C, Topics in Matrix Analysis, Cambridge University Press, Cambridge,
1994.
[24] Song Q and Cao J, On pinning synchronization of directed and undirected complex dynamical
networks, IEEE Transactions on Circuits and Systems I: Regular Papers, 2009, 57(3): 672–680.
[25] Zhang Y, Wen G, Peng Z, et al., Group multiple lags consensus of fractional-order nonlinear
leader-following multi-agent systems via adaptive control, Transactions of the Institute of Mea-
surement and Control, 2019, 41(5): 1313–1322.
[26] Cai S, Jia Q, and Liu Z, Cluster synchronization for directed heterogeneous dynamical networks
via decentralized adaptive intermittent pinning control, Nonlinear Dynamics, 2015, 82(1): 689–
702.
[27] Tan C, Liu G, and Duan G, Group consensus of networked multi-agent systems with directed
topology, IFAC Proceedings Volumes, 2011, 44(1): 8878–8883.
[28] Wen G, Duan Z, Chen G, et al., Consensus tracking of multi-agent systems with Lipschitz-type
node dynamics and switching topologies, IEEE Transactions on Circuits and Systems I: Regular
Papers, 2013, 61(2): 499–511.
[29] Ekramian M, Observer-based controller for Lipschitz nonlinear systems, International Journal of
Systems Science, 2017, 48(16): 3411–3418.
[30] Boyd S, El Ghaoui L, Feron E, et al., Linear Matrix Inequalities in System and Control Theory,
SIAM, Philadelphia, 1994.