Intermediate observers based robust fault estimation for multi-agent systems under communication constraints via switching scheme Authors: Yue Quan, Ruopeng Sun, Shichuan Ding, Hai Guo, Fuzhi Hu, Di Zhang, Xiaoyu Fan Published: 2023 ABSTRACT This paper addresses the problem of robust fault estimation for multi-agent systems (MASs) under communication constraints. Taking into account the possible data packet loss (DPL) in the information interaction of each subsystem, MASs are remodelled as switching systems by introducing a variable sampling strategy. Then, using the local information among agents, a novel intermediate observer design method based on switching scheme is proposed to estimate faults of MASs. Combining Lyapunov’s criterion and linear matrix inequality, sufficient conditions for the intermediate observer to be exponentially stable and have H∞ performance against bounded disturbances and the DPL are given. Finally, some simulations are provided to verify the effectiveness of the proposed method. Introduction Multi-agent systems (MASs) are distributed systems composed of a limited number of subsystems with completely independent computing capability. Due to its excellent characteristics and wide application prospects in engineering, the theory of MASs has been widely studied in many aspects and has achieved fruitful research results [1,2]. At present, MAS theory has become an important tool to solve large-scale complex system problems, and has been widely used in many fields [3,4]. With the continuous expansion of the scale of systems, the relationship between the subsystems becomes more and more complex. Different from the traditional centralized system, MASs are typical distributed systems. Due to the limitation of the communication network and the computing ability of the agent subsystem itself, the agent subsystem only exchanges and calculates the information of neighbour agent, but cannot obtain the global information of the whole system. For MASs, this distributed network structure is a double-edged sword. On the one hand, it improves the system’s strong robustness and the emergence of swarm intelligence. On the other hand, the absence of the central entity node of the monitoring system’s global state also increases the probability of system failure. Moreover, the lack of central monitoring node may also lead to a single fault spreading to the whole system through the cooperation between nodes, which will seriously threaten the safe operation of MASs. Therefore, the research on effective fault detection (FD) algorithms for MASs and the improvement of system reliability is a key problem that urgently needs to be solved in the field of MASs and limits development of MASs. According to the implementation methods, FD algorithms can be divided into data-driven FD and analytic model based FD. At present, the FD of MASs mainly improves the above two methods to meet the requirements of the distributed system structure of MASs. Dang et al. [5] and Peng et al. [6] studied the problem of FD of data-driven MASs. Using the method of mathematical statistics, it analysed the historical data of the system, extracted the fault characteristics and trained the data of each node and finally realized the FD of the system. This method requires a large amount of historical data to train each node. It increases the computational burden of the system. On the other hand, the limitations of historical data cannot fully meet the dynamic characteristics and real-time requirements of MASs. Different from the data based multi-agent FD method, the analytic model based FD does not require a lot of data training and it does not increase the computational burden of the system. Hence, the analytic model based FD can better meet the real-time requirements of the system and adapt to the dynamic characteristics of the MAS. Therefore, FD methods based on analytical models have been widely studied and applied in MASs. Among them, FD of distributed systems with unknown input observers (UIOs) is an effective method in this field, and fruitful research results have been achieved in FD of MASs. Shames et al. [7] proposed a FD framework based on UIOs for MASs, gave sufficient conditions for the existence of UIOs in an ideal topology network structure, and realized online FD of MASs. Considering the influence of uncertain factors in the MAS network, a series of improvements of results of Shames et al. have been researched. Liu et al. [8] proposed a FD method for UIOs of high-order MASs. By using UIOs, Chen et al. [9] studied the problem of FD for MASs with nonlinear disturbances, and Liang et al. [10] and Zhou et al. [11] studied the problem of FD for MASs with network delays. In addition, other aspects of FD of MASs based on UIOs are considered, such as FD of MASs in switching topology, FD of MASs in discrete systems and FD of MASs in finite time. Although many achievements have been made in FD of MASs based on UIOs, there are still some shortcomings. Firstly, the existence of UIOs requires strict matching conditions, which requires that the system parameter matrix must meet strict matrix rank conditions, increasing the difficulty of determining the unknown parameters of the system and reducing the flexibility of design. Secondly, the FD method of MAS based on UIO can only detect a single fault of the system effectively, but cannot detect multiple faults. Thirdly, the FD method of UIO can carry out fault alarm, but it cannot effectively estimate fault signals. These shortcomings limit the application of UIO in MAS FD. Comparing with the above research results, fault estimation (FE) is a more effective method to achieve FD. FE can not only alert the fault signal, but also effectively estimate the signal state, which is helpful for further processing of system faults. Intermediate observer is an effective method for distributed fault estimation (DFE) [12]. Zhu et al. [13,14] proposed a FE method based on distributed intermediate observer, which can realize FD of MASs with undirected topology network. Considering that in the actual system, the node information in the MAS may be transmitted unidirectionally, Han et al. [15] and Liu et al. [16] proposed a FE method for the distributed intermediate observer of the MASs with a directed network topology. In addition to being affected by unknown disturbance, communication delay, measurement noise and communication topology, MASs may also be affected by more complex factors, such as data packet loss (DPL), time-varying sampling period, etc. The emergence of these complex factors may make the MAS more complex. At present, the problem of MAS control under the influence of DPL has gradually attracted people’s attention [17,18]. However, in the aspect of FD of MASs, research on FD of DPL is relatively less. The objective existence of DPL may aggravate the vibration of MASs, thus increasing the difficulty of FD. Li et al. [19] proposed a FD method for MASs with DPL in the sensor controller link. In practice, DPL may occur in the sensor controller link, the controller executive link, or even both links. Zhang & Zhang [20] proposed a FD method for networked control systems with time-varying sampling and uncertain DPL rate by using δ operator method for MASs with simultaneous DPL from sensor to controller and controller to executive link. It should be pointed out that although these two methods have realized the FD of MASs with DPL, they have not realized the effective estimation of the fault signal and cannot provide a reference for the identification of future fault types. As far as we know, there are no published research results on the problem of FE of MASs with DPL. Therefore, it will be more meaningful and challenging to study the problem of FE of MASs under the influence of DPL. Based on the analysis of the above problems, this paper studies the FE problem of MASs with DPL by using the intermediate observer method. Firstly, the problem of DPL in the data transmission link is transformed into the problem of time-varying sampling period control of the system by using the input holding strategy of the controller. Furthermore, using Delta operator method, the MASs with DPL are represented as a switching system with time-varying sampling period. Secondly, using the relative state information of the MASs, an intermediate observer satisfying the switching system model is constructed to realize the FE of MASs with DPL. Finally, combined with the mean dwell time and Lyapunov stability criterion, the conditions of exponential stability for the intermediate observations designed in this paper are given. Compared with the existing achievements on multi-agent FD or FE, the main contributions of this paper can be summarized as follows: 1. The system model considers the simultaneous DPL of the sensor controller link and the controller executive link. Compared with the fault model in [12–16], the model in this paper is more consistent with the requirements of the actual system. Furthermore, the MAS with DPL is transformed into a switched system by using the method of time-varying sampling period, where the small network delay can be ignored when the network delay is smaller than the sampling period.2. Using the relative state information of MASs, an intermediate observer with switched system modes is constructed. Compared with the UIO in [8–11], this observer does not depend on the matching relationship of the parameter matrix in the system, and has strong design flexibility. Compared with the results in [15,16], this observer has better robustness to DPL. Compared with [20], the observer designed in this paper can not only realize the faulty alarm for multi-faults, but also effectively estimate the fault state signal, which can pave the way for further fault processing.3. Using the concept of average dwell time (ADT), the relationship between the stability of the intermediate observer of the switched system, the DPL rate and the number of consecutive packets is given under the condition of exponential stability. By using Lyapunov stability criterion, the matrix inequality (LMI) satisfied by the undetermined parameters of the intermediate observer is obtained. The dimension of the LMI is only related to the state dimension of a single agent, but not to state dimension of the entire system. Compared with the dimension of LMI in [12–16], the dimension of LMI in this paper is less, which can effectively reduce the computational burden of the system. 1. The system model considers the simultaneous DPL of the sensor controller link and the controller executive link. Compared with the fault model in [12–16], the model in this paper is more consistent with the requirements of the actual system. Furthermore, the MAS with DPL is transformed into a switched system by using the method of time-varying sampling period, where the small network delay can be ignored when the network delay is smaller than the sampling period. The system model considers the simultaneous DPL of the sensor controller link and the controller executive link. Compared with the fault model in [12–16], the model in this paper is more consistent with the requirements of the actual system. Furthermore, the MAS with DPL is transformed into a switched system by using the method of time-varying sampling period, where the small network delay can be ignored when the network delay is smaller than the sampling period. 2. Using the relative state information of MASs, an intermediate observer with switched system modes is constructed. Compared with the UIO in [8–11], this observer does not depend on the matching relationship of the parameter matrix in the system, and has strong design flexibility. Compared with the results in [15,16], this observer has better robustness to DPL. Compared with [20], the observer designed in this paper can not only realize the faulty alarm for multi-faults, but also effectively estimate the fault state signal, which can pave the way for further fault processing. Using the relative state information of MASs, an intermediate observer with switched system modes is constructed. Compared with the UIO in [8–11], this observer does not depend on the matching relationship of the parameter matrix in the system, and has strong design flexibility. Compared with the results in [15,16], this observer has better robustness to DPL. Compared with [20], the observer designed in this paper can not only realize the faulty alarm for multi-faults, but also effectively estimate the fault state signal, which can pave the way for further fault processing. 3. Using the concept of average dwell time (ADT), the relationship between the stability of the intermediate observer of the switched system, the DPL rate and the number of consecutive packets is given under the condition of exponential stability. By using Lyapunov stability criterion, the matrix inequality (LMI) satisfied by the undetermined parameters of the intermediate observer is obtained. The dimension of the LMI is only related to the state dimension of a single agent, but not to state dimension of the entire system. Compared with the dimension of LMI in [12–16], the dimension of LMI in this paper is less, which can effectively reduce the computational burden of the system. Using the concept of average dwell time (ADT), the relationship between the stability of the intermediate observer of the switched system, the DPL rate and the number of consecutive packets is given under the condition of exponential stability. By using Lyapunov stability criterion, the matrix inequality (LMI) satisfied by the undetermined parameters of the intermediate observer is obtained. The dimension of the LMI is only related to the state dimension of a single agent, but not to state dimension of the entire system. Compared with the dimension of LMI in [12–16], the dimension of LMI in this paper is less, which can effectively reduce the computational burden of the system. Problem description In MASs, each subsystem is composed of independent sensors, controllers, actuators, network data transmitters and receivers. These subsystems realize information interaction through the network and form a complete network distributed system. Each agent interacts with each other through the network. The topological network of MASs can be denoted as G=(V,ε,A), where V = {Vi|i = 1, 2, …, N} represents the node set, Vi represents the ith node, A = [aij] ∈ RN×N is the adjacency matrix of G and N is the number of agents in MASs. aij is determined by the edge set ε of graph G. For example, aij = 1 if (Vi,Vj)∈ε when i ≠ j, otherwise aij = 0. In the undirected topology, (Vi,Vj)∈ε and (Vj,Vi)∈ε exist at the same time, that is, aij = aji. L=[lij]∈ℜN×N is the Laplacian matrix of G, where lij=∑ j=1Naij when i = j and lij = −aij when i ≠ j. The dynamic model of the ith agent subsystem is considered as follows:2.1{x˙i(t)=A0xi(t)+B0ui(t)yi(t)=Cxi(t),i=1,2…,N,where xi(t)∈ℜn,ui(t)∈ℜnu,yi(t)∈ℜny, respectively, represent the state, control input and state output of the ith agent subsystem, A0, B0, C are known real matrices with appropriate dimensions, satisfying that (A0, B0) is controllable and (A0, C) is measurable. It is assumed that the sensor is time-driven and samples the system with a fixed sampling period. Let Tk (k = 0, 1, 2, …) represent the sampling time of the system and h=def⁡Tk−Tk−1 represent the sampling period. The structure of MASs with time-driven sensors can be shown as in figure 1. Where xit(Tk) represents the information sending to network of agent i, xir(Tk) represents the information receiving from network of agent i, xic(Tk) and ui(Tk), respectively, represent the input information and output information of the controller in agent i. Figure 1. Structure of MASs with DPL. Structure of MASs with DPL. Due to the limitation of network bandwidth, DPL may collide or communication nodes may fail to complete in the process of network transmission, resulting in the loss of packets to be transmitted. DPL may occur on the agent’s network data sending link or network data receiving link, or on two links at the same time. In order to reduce the impact of DPL on MASs, both the network data receiver and controller adopt the event driven control mode. If DPL occurs during data transmission, they will keep the last data, i.e. xir(tk)=xir(tk−1), ui(tk) = ui(tk−1) when there is DPL at Tk. The impact of DPL on system dynamics is described in figure 2. Figure 2. Control sequence of MASs with DPL. Control sequence of MASs with DPL. At T1, all data are successfully transmitted, and the output signal of the implementation controller is u(T1). At T2 and T3, DPL occurs, respectively, in data transmission and data reception. The data receiver will keep the data of the previous time, and the controller will also keep the output of the previous time, that is u(T2) = u(T1), u(T3) = u(T1). At T4, the data are successfully transmitted, and the output signal of the controller is u(T4). At T5, DPL occurs simultaneously in the sending process and the receiving process. At this time, the controller will maintain the original output, u(T5) = u(T4). According to the output characteristics of control signals, the MAS with DPL can be regarded as a sampled control system with variable sampling period. Assumption 2.1. The network induced delay is ignored, and the number of continuous DPL in the network is bounded and supposed d. Let hk = tk+1 − tk denote the variable sampling period, and tk denotes the sampling time. Based on the results of the analysis of figure 2, it can be got that hk = nk h. According the assumption 1, it is defined hk ∈ {h, …, (d + 1)h}, nk∈Zδ=def{1,…,d+1}. If the number of DPL in MASs is bounded and d, then the system (2.1) may have the following description of the control system model with non-equidistant sampling periods:2.2{x˙i(t)=A0xi(t)+B0ui(tk)yi(t)=Cxi(t),∀t∈[tk,tk−1),i=1,2,…,N. By discretizing the system (2.2) with sampling period hk, the following discrete-time system model (2.3) can be obtained:2.3xi(tk+1)=A(hk)xi(tk)+B(hk)ui(tk)andyi(tk)=Cxi(tk), } where A(hk)=eA0hk, B(hk)=∫0hkeA0τdτB. Let A1=eA0h and B1=∫0heA0τdτB, then A(hk) and B(hk) can be described as2.4A(hk)=eA0hk=eA0nkh=A1nkand2.5B(hk)=∫0hk eA0τ dτB=(∑i=0nk−1∫ih(i+1)h eA0τ dτ)B=(∑i=0nk−1(eA0h)i)∫0h eA0τ dτB=(∑i=0nk−1A1i)B1. According to (2.4) and (2.5) A(hk) and B(hk) both depend on nk, while nk can have at most d + 1 values. Furthermore, a piecewise continuous switching system signal δ(tk)∈Zσ is introduced, and let Aδ(tk)=A1δ(tk), Bδ(tk)=(∑ j=1δ(tk)A1 j−1)B1, then system (2.3) is equivalent to the following form of switching system:2.6sδ(tk)i : {xi(tk+1)=Aδ(tk)xi(tk)+Bδ(tk)ui(tk)yi(tk)=Cxi(tk), where δ(tk) = nk ∈ {1, 2, …, d + 1} represents a subsystem with continuous DPL of nk − 1. When MASs are affected by faults or external disturbance signals, the fault signals and external disturbance signals can be taken as the superposition signals of the system. Let fi(tk)∈ℜnf and di(tk)∈ℜnd, respectively, represent the fault signal vector and external disturbance of agent i. It is assumed that fi(tk) and di(tk) are bounded and satisfy fi(tk) ∈ l2(0, ∞), di(tk) ∈ l2(0, ∞). Then the dynamics of agent i affected by DPL, disturbance and faulty can be expressed as2.7sδ(tk)i : {xi(tk+1)=Aδ(tk)xi(tk)+Bδ(tk)ui(tk)+F fi(tk)+Edi(tk)yi(tk)=Cxi(tk), where F and E are known real matrices with appropriate dimensions. To simplify the analysis difficulty of the problem, the following assumption is given. Assumption 2.2. The fault signal will only affect the output of the system and will not change the dynamic parameters of the MASs. Main results Intermediate observer The intermediate variable about agent i is firstly construct to estimate the fault signal:3.1ϕi(tk)= fi(tk)−K1xi(tk),where K1 is an undetermined matrix. Then it can be got that fi (tk) = ϕi (tk) + K1xi (t). Combined with equation (2.7), it can be further obtained from equation (3.1):3.2ϕi(tk+1)= fi(tk+1)−K1xi(tk+1)= fi(tk+1)−K1(Aδ(tk)xi(tk)+Bδ(tk)ui(tk)+F fi(tk)+Edi(tk))= fi(tk+1)−(K1Aδ(tk)+K1FK1)xi(tk)−K1Bδ(tk)ui(tk)−K1Fϕi(tk)−K1Edi(tk). The following state observers are designed asx^i(tk+1)=Aδ(tk)x^i(tk)+Bδ(tk)ui(tk)+Ff^i(tk)+ρK2ςi(tk),ϕ^i(tk+1)=−K1Fϕ^i(tk+1)−(K1Aδ(tk)+K1FK1)x^i(tk) −K1Bδ(tk)ui(tk)+ρK3ςi(tk),f^i(tk)=ϕ^i(tk)+K1x^i(t)andy^i(tk)=Cx^i(tk),where x^i(tk), ϕ^i(tk) and f^i(tk) represent the system state estimation, intermediate variable estimation and fault signal estimation at tk, respectively. ςi(tk)=∑ j=1Naij[(yi(tk))−y^i(tk))−(yj(tk))−y^j(tk))] represents the system distributed output estimation error. ρ > 0, K1 and K2 are the undetermined gain matrices of the observer. Let exi(tk)=xi(tk)−x^i(tk), eϕi(tk)=ϕi(tk)−ϕ^i(tk) and efi(tk)= fi(tk)−f^i(tk) represent the state estimation error of the system, respectively. It can be further obtained that3.3exi(tk+1)=xi(tk+1)−x^i(tk+1)=Aδ(tk)xi(tk)+Bδ(tk)ui(tk)+F fi(tk)+Edi(tk)−Aδ(tk)x^i(tk)−Bδ(tk)ui(tk)−Ff^i(tk)−ρK2ςi(tk)=Aδ(tk)exi(tk)+Fefi(tk)−ρK2ςi(tk)+Edi(tk)and3.4eϕi(tk+1)=ϕi(tk+1)−ϕ^i(tk+1)= fi(tk+1)−(K1Aδ(tk)+K1FK1)xi(tk)−K1Bδ(tk)ui(tk)−K1Fϕi(tk)−K1Edi(tk)+K1Fϕ^i(tk+1)+(K1Aδ(tk)+K1FK1)x^i(tk)+K1Bδ(tk)ui(tk)−ρK2ςi(tk)=−K1Feϕi(tk)−(K1Aδ(tk)+K1FK1)exi(tk)−K1Edi(tk)+ fi(tk+1)−ρK3ςi(tk). It is noted that3.5efi(tk)= fi(tk)−f^i(tk)=ϕi(tk)+K1xi(t)−ϕ^i(tk)−K1x^i(t)=eϕi(tk)+K1exi(tk).Combining with (3.5), equation (3.3) can be rewritten as3.6exi(tk+1)=(Aδ(tk)+FK1)exi(tk)+Feϕi(tk)−ρK2ςi(tk)+Edi(tk). Let K1=εFT, where FT is transpose matrix of F and ε>0 is an undetermined constant. Combining (3.4) and (3.6), it can be got the error dynamics about agent i:3.7ei(tk+1)=A¯1 ei(tk)+E1ωi(tk)−ρK¯2∑ j=1Naij[(yi(tk)−y^i(tk))−(yj(tk)−y^j(tk))]andY i(tk)=C¯ ei(tk),} where A¯1=[Aσ(tk)+εFFT F−εFT(Aσ(tk)+εFFT) −εFTF], E1=[E O−εFTE I], K¯2=[K2K3], C¯=[In OK1 Inf] and ωi(tk)=[di(tk)fi(tk+1)] can be regarded as an interference term in error dynamics (3.7), which is bounded and satisfies ω(tk) ∈ l2[0, ∞). Let e(tk) = [(e1(tk))T, {(e2(tk))T, …, (eN(tk))T]T, ω(tk) = [(ω1(tk))T, {(ω2(tk))T, …, (ωN(tk))T]T, and Y(tk) = [(Y1(tk))T, {(Y2(tk))T, …, (YN(tk))T]T, then the error dynamic equation of the whole MASs can be obtained:3.8Sδ(tk) : {e(tk+1)=IN⊗A¯1e(tk)−(ρL⊗K¯2C)e(tk)+IN⊗E1ω(tk)Y i(tk)=IN⊗C¯e(tk). Obviously, the observer described above can effectively carry out the estimation of the state of the whole MASs when system (3.8) is convergent. Therefore, the observer design problem can be transformed into the convergence analysis of the system (3.8). Stability analysis of error dynamics Before the analysis, some definitions are given. Definition 3.1. [21] For the given switching signal δ(tk) and any tk > t0 > 0, let Nδ[t0, tk) represent the switching times of the switching signal δ(tk) at the interval [t0, tk). If there are N0 ≥ 0 and τa > 0 to satisfy Nδ[t0, tk) ≤ N0 + ((tk − t0)/τa), then τa is called the ADT of the switching signal δ(tk). In this paper, N0 = 0. Definition 3.2. [22] Considering the system (3.8) with ω(tk) = 0, if there are normal values ς and λ < 1, ∥e(tk)∥≤ςλtk−t0∥e(t0)∥ for any initiation condition e(t0)∈ℜn, then the system (3.8) is exponentially stable with exponential decay rate λ. Definition 3.3. [22] For any given γ > 0, the system (3.8) satisfies the following conditions: (1) When ω(tk) = 0, system (3.8) is exponentially stable.(2) Under zero initial condition, for any non-zero signal ω(tk), system (3.8) satisfies that ∑k=0∞eT(tk)e(tk)<∑k=0∞γ2ωT(tk)ω(tk), then it is said that the system (3.8) has the H∞ performance, where γ is the H∞ performance level. (1) When ω(tk) = 0, system (3.8) is exponentially stable. When ω(tk) = 0, system (3.8) is exponentially stable. (2) Under zero initial condition, for any non-zero signal ω(tk), system (3.8) satisfies that ∑k=0∞eT(tk)e(tk)<∑k=0∞γ2ωT(tk)ω(tk), then it is said that the system (3.8) has the H∞ performance, where γ is the H∞ performance level. Under zero initial condition, for any non-zero signal ω(tk), system (3.8) satisfies that ∑k=0∞eT(tk)e(tk)<∑k=0∞γ2ωT(tk)ω(tk), then it is said that the system (3.8) has the H∞ performance, where γ is the H∞ performance level. It is assumed that the system (3.8) resides in the subsystem Sδ(t0) (δ(t0)∈Zδ) at the initial time t0. Let tk1,tk2,…tkd indicate the switching time in the [t0, tk), which satisfies t0 1). Sδ(tk) is switched by the sequence Π : {(χ0,t0),(χ1,tk1),…,(χφ,tkφ)|χφ∈Zδ,φ=1,…,d+1}. For the convenience of analysis, let eχφ(tk) represent the output of χφth subsystems. Considering the state of the subsystem remains unchanged at the switching instant, we can get (3.9) when subsystem is switched from χφ−1 to χφ at tkφ:3.9eχφ(tkφ)=eχ(φ−1)(tkφ). Let γχφ denote the occurrence rate of DPL process of subsystem Sχφ(χφ∈Zδ), then it can be got that ∑φ=1d+1γχφ=1 and γχφ≥0. Let nχφ represent the occurrence number of DPL of subsystem Sχφ at intervals [t0, tk). According to the definition of ADT, it can be got that nχφ=γχφNδ[t0,tk)=γχφ((tk−t0)/τα), and ∑φ=1d+1nχφ=∑φ=1d+1γχφ((tk−t0)/τα)=(tk−t0)/τα. The subsystem Sχφ (φ = 2, 3, …, d + 1) is activated only when DPL occurs. Hence, DPL rate of the system can be described as3.10α=∑φ=2d+1γχφ. Theorem 3.4. For MASs with undirected connected topology, system (3.8) with ω(tk) = 0 is exponentially stable and satisfies the exponential decay rate ρ(τα,λ)=μ−(1/2τα)λ if there are scalar λ > 1, μ > 1, λχφ>1 (χφ∈Zδ,φ=1,…,d+1), positive definite matrix Pδ(tk), Pα and Pβ (α,β∈Zδ) to satisfy the following inequality:3.11Ξi=[−Pδ(tk)Pδ(tk)A¯1−ρψiPδ(tk)K¯2C)∗−λδ(tk)−2(tk+1−tk)Pδ(tk)]<0, i=1,2,…,N,3.12Pα<μPβ,(α,β∈Zδ),3.13∏φ=1d+1λχφγχφ>λ>1,and3.14τa>ln⁡μ2ln⁡λ. Proof. Choosing a candidate Lyapunov function as3.15Vδ(tk)(tk)=e T(tk)(IN⊗Pδ(tk))e(tk),when ω(tk) = 0, by (3.8) and (3.15), it follows that3.16Vδ(tk)(tk+1)−λδ(tk)−2(tk+1−tk)Vδ(tk)(tk)=e T(tk+1)Pδ(tk)e(tk+1)−λ −2(tk+1−tk)e T(tk)Pδ(tk)e (tk)=e T(tk)(IN⊗A¯1−(ρL⊗K¯2C))T(IN⊗Pδ(tk))(IN⊗A¯1−(ρL⊗K¯2C))e(tk)−λδ(tk)−2(tk+1−tk)e T(tk)(IN⊗Pδ(tk))e(tk)=e T(tk)[(IN⊗A¯1−(ρL⊗K¯2C))T(IN⊗Pδ(tk))(IN⊗A¯1−(ρL⊗K¯2C))−λδ(tk)−2(tk+1−tk)(IN⊗Pδ(tk))]e(tk).If the following equation (3.17) holds:3.17(IN⊗A¯1−(ρL⊗K¯2C))T(IN⊗Pδ(tk))(IN⊗A¯1−(ρL⊗K¯2C))−λδ(tk)−2(tk+1−tk)(IN⊗Pδ(tk))<0,it can be got that3.18Vδ(tk)(tk+1)−λ −2(tk+1−tk)Vδ(tk)(tk)<0.According to Schur complement theorem, equation (3.17) is equivalent to3.19[−IN⊗Pδ(tk)(IN⊗Pδ(tk)A¯1−ρL⊗Pδ(tk)K¯2C)∗−λδ(tk)−2(tk+1−tk)(IN⊗Pδ(tk))]<0.It is noted that LT = L in the MASs with undirected connected topology. Hence, there exists an orthogonal matrix H to satisfy L=HTΛH, where Λ=diag{ψ1,…,ψN} and ψi is the eigenvalue of L. The left multiplication and right multiplication of inequality (3.19) by [HT⊗IOOHT⊗I] and [H⊗IOOH⊗I], respectively. Equation (3.20) can be obtained from inequality (3.19) by making proper matrix transformation:3.20Ξ=diag{Ξ1,…,ΞN}<0,where Ξi=[−Pδ(tk)Pδ(tk)A¯1−ρψiPδ(tk)K¯2C∗−λδ(tk)−2(tk+1−tk)Pδ(tk)] (i = 1, 2, …, N). Obviously, inequality (3.18) holds if (3.11) holds. It means that Vδ(tk)(tk) decays exponentially along the respective subsystem trajectories. Further, the following equation (3.21) can be obtained:3.21Vδ(tι)(tk)≤λχι−2(tk−tkι)Vδ(tι)(tkι)andVδ(tφ)(tkι)≤λχφ−2(tkι−tkφ)Vδ(tφ)(tkφ).} According to (3.12) and (3.15), it can be got that3.22Vχφ(tkφ)=eχφT(tkφ)Pχφeχφ(tkφ)=eχ φ−1T(tkφ)Pχφeχφ−1(tkφ)≤ueχ φ−1T(tkφ)Pχφ−1eχφ−1(tkφ)=uVχφ−1(tkφ).Using the definition of ADT, the following recursive equation can be obtained by combining (3.21) and (3.22):3.23Vδ(tk)(tk)≤λχι−2(tk−tkι)Vχι(tkι)≤μλχι−2(tk−tkι)Vχι−1(tkι)≤μλχι−2(tk−tkι)λχι−1−2(tkι−tkι−1)Vχι−1(tkι−1)⋮≤μNδ(t0,tk)λχι−2(tk−tkι)λχι−1−2(tkι−tkι−1)⋯λχ0−2(tk1−t0)Vδ(t0)(t0)=μNδ(t0,tk)∏φ=1d+1λχφ−2ταnχφVδ(t0)(t0)=μNδ(t0,tk)∏φ=1d+1λχφ−2ταγχφNδ[t0,tk)Vδ(t0)(t0).When inequality (3.13) is established, (3.23) can be rewritten as3.24Vδ(tk)(tk)≤μNδ(t0,tk)λ−2ταNδ[t0,tk)Vδ(t0)(t0).Since Nδ[t0,tk)≤(tk−t0)/τa, it can be obtained from (3.24) that3.25Vδ(tk)(tk)≤μ(−2(tk−t0))/(−2τα)λ−2(tk−t0)Vχ0(t0).Let ρ(τα,λ)=μ−(1/2τα)λ, then (3.25) can be rewritten as3.26Vδ(tk)(tk)≤ρ−2(tk−t0)(τα,λ)Vδ(t0)(t0).Let ψ1=minδ(tk)∈Z0⁡λmin(Pδ(tk)) and ψ2=maxδ(tk)∈Z0⁡λmax(Pδ(tk)), then3.27ψ1∥e(tk)∥2≤Vδ(tk)(tk)≤ρ−2(tk−t0)(τα,λ)Vδ(t0)(t0)≤ρ−2(tk−t0)(τα,λ)ψ2‖e(t0)‖2.From (3.27), it can be got that3.28‖e(tk)‖≤ψ2ψ1ρ−(tk−t0)(τα,λ)‖e(t0)‖. Obviously, ρ(τα,λ)>1 if (3.14) holds. This means that system (3.8) with ω(tk) = 0 is exponentially stable and satisfies the exponential decay rate ρ(τα,λ), which completes the proof. ▪ Remark 3.5. Let λ¯s=min{λ2,λ3,…,λd+1}, λ¯1>λ and λ¯1>λ¯s. Since α=∑χφ=2d+1γχφ and ∑χφ=1d+1γχφ=1, it can be got that ∏φ=1d+1λχφγχφ≥λ¯1γχ 1∏φ=2d+1λχφγχφ=λ¯11−∑χφ=2d+1γχφλ¯s∑χφ=2d+1γχφ=λ¯1(1−α)λ¯sα. According to (3.13), we can get λ¯11−αλ¯sα>λ>1, which means the following equation holds:3.29α<α¯=ln⁡(λ¯1/λ)ln⁡(λ¯1/λs).Since λ=ρ2(τα,λ)μ1/τα, so it can be got that3.30α¯=ln⁡λ¯1ρ2(τα,λ)μ−(1/τα)ln⁡(λ¯1/λs)=−β1ln⁡ρ(τα,λ)+β2,where β1=(2/ln⁡(λ¯1/λs)) and β2=(ln⁡λ¯1−(1/τα)ln⁡μ)/ln⁡(λ¯1/λs). Equation (3.29) shows the upper limit of the DPL rate to ensure that system (3.8) meets exponential stability. And equation (3.30) shows that the upper bound of the DPL rate is a monotonic decreasing function of the exponential decay rate ρ(τα,λ), that is, the smaller the DPL rate, the greater the exponential decay rate of the system. Next, H∞ performance conditions of system (3.8) to bounded disturbance ω(tk) ≠ 0 will be analysed. Theorem 3.6. For MASs with undirected connected topology, if there exist scalar γ > 0, λ > 1, μ ≥ 1, λχφ>1, (χφ∈Zδ,φ=1,…,d+1), positive definite matrix Pδ(tk), Pα and Pβ (δ(tk),α,β∈Zδ) to satisfy conditions of theorem 3.4 and the following inequalities:3.31Θi=[−Pδ(tk)Pδ(tk)A¯1−ρψiPδ(tk)K¯2CPδ(tk)E∗−λδ(tk)−2(tk+1−tk)Pδ(tk)+IO∗∗−γ2I]<0,i=1,2,…,Nand3.321−μλχφ−2(tkφ+1−t φ)>0,φ=1,…,d,then the error dynamic system (3.8) with ω(tk) ≠ 0 has H∞ performance, where γ is the performance level. Proof. Let us choose the following cost function:3.33J=∑k=0∞[eT(tk)e(tk)−γ2ωT(tk)ω(tk)].According to definition 3.3, if J < 0 can be maintained, the above theorem can be proved. The same candidate Lyapunov function described as (3.15) is selected. When ω(tk) ≠ 0, combining with equation (3.8) and (3.15), it can be got that3.34Vδ(tk)(tk+1)−λδ(tk)−2(tk+1−tk)Vδ(tk)(tk)=e T(tk+1)Pδ(tk)e(tk+1)−λ −2(tk+1−tk)e T(tk)(IN⊗Pδ(tk))e (tk)=e T(tk)(IN⊗A¯1−(ρL⊗K¯2C))T(IN⊗Pδ(tk))(IN⊗A¯1−(ρL⊗K¯2C))e(tk)+e T(tk)(IN⊗A¯1−(ρL⊗K¯2C))T(IN⊗Pδ(tk))IN⊗E1ω(tk)+ωT(tk)(IN⊗E1)T(IN⊗Pδ(tk))(IN⊗A¯1−(ρL⊗K¯2C))e(tk)+ωT(tk)(IN⊗E1)T(IN⊗Pδ(tk))IN⊗E1ω(tk)−λδ(tk)−2(tk+1−tk)e T(tk)(IN⊗Pδ(tk))e(tk). Defining augmented vectors θ(tk)=[e T(tk) ωT(tk)]T, (3.34) can be rewritten as3.35Vδ(tk)(tk+1)−λδ(tk)−2(tk+1−tk)Vδ(tk)(tk)=θT(tk)Φδ(tk)θT(tk)+γ2ωT(tk)ω(tk)−eT(tk)e(tk),whereΦδ(tk)=[−λδ(tk)−2(tk+1−tk)(IN⊗Pδ(tk))Pδ(tk)+IOO−γ2I]+[(IN⊗A¯1−(ρL⊗K¯2C))T(IN⊗Pδ(tk))(IN⊗E)T(IN⊗Pδ(tk))](IN⊗Pδ(tk)−1) [(IN⊗Pδ(tk))(IN⊗A¯1−(ρL⊗K¯2C)) (IN⊗Pδ(tk))IN⊗E1].According to Schur complement theorem, when3.36[−IN⊗Pδ(tk)(IN⊗Pδ(tk))(IN⊗A¯1−(ρL⊗K¯2C))(IN⊗Pδ(tk))IN⊗E1∗−λδ(tk)−2(tk+1−tk)(IN⊗Pδ(tk))+IO∗∗−γ2I]<0holds, it can be got that θ(tk)Φδ(tk)θT(tk)<0. Then we can get3.37eT(tk)e(tk)−γ2ωT(tk)ω(tk)<λδ(tk)−2(tk+1−tk)Vδ(tk)(tk)−Vδ(tk)(tk+1). Left and right multiplication matrices [HT⊗IOO∗HT⊗IO∗∗HT⊗I] and [H⊗IOO∗H⊗IO∗∗H⊗I] of inequality (3.36), respectively, then (3.36) is equivalent toΘ=diag{Θ1,…,ΘN}<0,where Θi=[−Pδ(tk)Pδ(tk)A¯1−ρψiPδ(tk)(K¯2C)Pδ(tk)E∗−λδ(tk)−2(tk+1−tk)Pδ(tk)+IO∗∗−γ2I], i = 1, 2, …, N. Obviously, (3.36) holds if (3.31) holds. Further, the following equation can be obtained recursively from inequality (3.37):eT(tkd)e(tkd)−γ2ωT(tkd)ω(tkd)<λχd−2(tk−tkd)Vχd(tkd)−Vχd(tk)eT(tkd−1)e(tkd−1)−γ2ωT(tkd−1)ω(tkd−1)<λχd−1−2(tkd−tkd−1)Vχd−1(tkd−1)−Vχd−1(tkd)⋮eT(t0)e(t0)−γ2ωT(tk0)ω(tk0)<λχ0−2(tk1−tk0)Vχ0(t0)−Vχ0(tk1).Adding the two sides of the above inequality separately gives3.38∑t=t0tkeT(t)e(t)−∑t=t0tkγ2ωT(t)ω(t)<λχ0−2(tk1−t 0)Vχ0(t0)−(1−μλχ1−2(tk2−t k1))Vχ0(tk1)−(1−μλχ1−2(tk3−tk2))Vχ1(tk2)−⋯−Vσ(tk)(tk). If (3.32) holds, it can be obtained from (3.38) that3.39∑t=t0tkeT(t)e(t)−∑t=t0tkγ2ωT(t)ω(t)<λχ0−2(tk1−t 0)Vχ0(t0),where t0 denotes the initial moment of the system. Let t0 = 0, then it can be obtained from equation (3.39) under zero initial conditions that3.40∑t=0tkeT(t)e(t)<∑t=0tkγ2ωT(t)ω(t).Consequently, it can be got that ∑t=0∞eT(t)e(t)<∑t=0∞γ2ωT(t)ω(t) when tk → ∞, i.e. J < 0, which completes the proof. ▪ Simulation In this section, some numerical simulations are given to verify the results proposed in the paper. MASs with five nodes and undirected topology are given in figure 3. A=[0110110110110000100010000]. Parameter matrices of agent subsystem are chosen as A0=[−0.8−0.0110.1], B0=[0.40.1]. The sampling period of the system is h = 0.1. By discretizing the system with a fixed sampling period h, the parameter matrix without DPL can be obtained as Aχ1=[0.9920−0.00010.01001.0010], Bχ1=[0.00400.0010]. Assuming the maximum number of continuous DPL of the system d = 3, then the switching subsystem parameters can be obtained by discretion with a variable sampling period, which are shown as Aχ2=[0.98410.00020.01990.0020], Bχ2=[0.00790.0021], Aχ3=[0.97630.00030.02970.0030], Bχ3=[0.01190.0032], Aχ4=[0.96850.00040.03940.0040], Bχ4=[0.01570.0043]. Figure 3. Undirected topology of MASs. Undirected topology of MASs. Assume that at [0, 50h], the numbers of activations of subsystems are nχ1=44,nχ2=2,nχ3=2 and nχ4=2. The switching sequence is sχ1⋯sχ1⏞6sχ2sχ1⋯sχ1⏞6sχ3sχ1⋯sχ1⏞6sχ4 sχ1⋯sχ1⏞6sχ4sχ1⋯sχ1⏞6sχ3sχ1⋯sχ1⏞6sχ2sχ1⋯sχ1⏞6. It can be got that τα=0.42, χ0 = 0.88, γχ1=γχ2=γχ3=0.04 and α = 0.12. Let λχ1=1.44, λχ2=λχ3=1.25, λχ4=1.16, λ = 1.4, μ = 1.07, λ¯1=1.44, λs = 1.16, ρ = 0.5, ϵ=3, γ = 0.7, F=[01] and E=[c11]. Combining the inequalities in theorems 3.4 and 3.6, it can be got that K¯2=[0.00780.01010.0050.04660.00730.0571]. The disturbance signal of the system is taken as di(tk)=sin⁡(2πtk) ∗ η(tk), where η(tk) denotes the Gaussian white noise signal with amplitude of 0.1. When there is no fault in MASs, the result is as shown in figures 4–8. Figures 4 and 6 show the actual state output of MASs, figures 5 and 7 the estimation outputs of MASs. It can be seen that the observer can effectively estimate the system state when there is no fault. Figure 8 is the FE output. When there is no fault, the outputs are irregular curves near zero, which show that the observer can effectively estimate the fault signal when there is no fault. Figure 4. First-order state output of MASs without fault. Figure 5. Estimation output of first-order state of MASs without fault. Figure 6. Second-order state output of MASs without fault. Figure 7. Estimation output of second-order state of MASs without fault. Figure 8. FE output of MASs without fault. First-order state output of MASs without fault. Estimation output of first-order state of MASs without fault. Second-order state output of MASs without fault. Estimation output of second-order state of MASs without fault. FE output of MASs without fault. When there is a signal fault in MASs, suppose the fault occurs on Agent 1, and the fault signal is described as f1(tk)=2(e(tk−1)−1) when 0.5 ≤ tk ≤ 2 and f1(tk) = 0 when tk < 0.5 or tk > 2. The outputs of system are shown in figures 9 and 10. It can be seen that the observer can make a timely response to the failure of node 1 in figure 9. Since the fault only occurs in one agent, the observer’s output to other nodes is approximately zero, which also shows that the observer can carry out the fault alarm for signal fault. Figure 10 is the output of observer’s comparison between the estimated value of the fault and actual value. It can be seen that the estimator can accurately estimate the fault signal, which also verifies the effectiveness of FE for a single fault of the system. Figure 9. FE output of MASs with single fault. Figure 10. Actual fault signal and FE output of MASs with single fault. FE output of MASs with single fault. Actual fault signal and FE output of MASs with single fault. When multiple faults occur in the system MASs, it is assumed that the fault occurs on Agent 1 and Agent 2, respectively, and the fault signal is described as follows: f1(tk)=2(e(tk−1)−1) when 1 ≤ tk ≤ 2 and f1(tk) = 0 when tk < 1 or tk > 2, f2(tk)=sin⁡(0.5tk−1)+cos⁡(0.2tk−1) when tk ≥ 1.5 and f2(tk) = 0 when tk < 1.5. The output results are as shown in figures 11–13. Figure 11 shows that the observers can make a timely response to Agent 1 and Agent 2. Figures 12 and 13 shows that the estimator can effectively estimate the faults that occur in Agent 1 and Agent 2. This shows that the observer is effective in FE for multiple faults. Figure 11. FE output of MASs with multiple faults. Figure 12. Actual fault signal and FE output of Agent 1. Figure 13. Actual fault signal and FE output of Agent 2. FE output of MASs with multiple faults. Actual fault signal and FE output of Agent 1. Actual fault signal and FE output of Agent 2. Conclusion In this paper, the problem of distributed robust FE for MASs with DPL is investigated. Firstly, by decomposing and transforming the node’s information dynamic model, the problem of DPL is transformed into the problem of time-varying sampling period, and based on this, MAS switching dynamic models with time-varying sampling period are constructed. Secondly, intermediate observers independent of the matching relationship of the system parameter matrix are constructed, which can estimate the system state and multiple fault signals simultaneously. Thirdly, the output bias dynamics model of the observer is constructed, and the design problem of the observer is converted into the convergence problem of the output error dynamics. It is shown that the error dynamics satisfy exponential stability and have H∞ performance to rebounded disturbance. It is worth noting that this paper only considers the FE for undirected MASs with communication constraints. Subsequent research will further develop a robust fault estimation for directed MASs with communication constraints.