Structure Depiction Of Multi-agent Systems And Distributed Estimation

With the rapid development of sensor technology,network technology and computer technology,the multi-agent systems and distributed estimation have become the research focus in the fields of control theory and application.Which are attracting the attention of the academia and industry due to their potential application.The multi-agent systems can accomplish the complex tasks via the coordinated control while the single individual cannot do.They have the characteristics of fault tolerance,flexibility and robustness.The design of controllability and leader selection of multi-agent systems have been applied in engineering fields such as multi-vehicle formation,autonomous underwater vehicle,intelligent transportation and satellite attitude cooperative control.At the same time,the distributed estimation algorithms are not dependent on the center sensors.They have the advantage of scalability and can adapt to the varying environment.Up to now,the idea of distributed estimation has found wide applications in many practical fields,such as target tracking,the determination of space debris,environment pollution monitoring and camera network monitoring,etc.This dissertation mainly studies the structural controllability of networked distributed systems,cooperative observability and the design of distributed filter.On the one hand,it studies how to select the leaders such that the multi-agent systems are strongly structurally controllable.Furthermore,we consider the issue that the measurements are needed in multi-agent systems.Then the placement of sensors are discussed.On the other hand,we analyze the cooperative observability for the distributed estimation problem under the strongly connected network.Next,we present the distributed information Kalman filter and distributed unscented Kalman filter for the linear and nonlinear systems respectively.These two filters are based on accurate consensus protocols with finite iteration steps.The main analysis tools utilized in this dissertation include graph theory,combinatorial optimization,linear system theory and optimal estimate.The main contents of the dissertation are summarized as follows:1.The graph theory is used to select the leaders or measurements for multi-agent systems in the structural sense.For the multi-agent systems that can be transformed into a general linear time invariant system,the multi-agent systems will be strongly structurally controllable via selecting the leaders.We prove that the topologies of path and path-bud only need one leader to guarantee the strong structural controllability.Then we point out that any directed graph can be partitioned into disjoint paths and path-buds,and the leaders are the root agents of all paths.We also discuss the strong structural controllability of bidirectional path and grounded tree with the failure of any one node.The minimum number and position of the leaders for these two special topologies are provided.When some agents' states can be measured in multi-agent systems,we consider the case that the system can be transformed into a bilinear structured system with rank 1.This system is structurally controllable by choosing the leaders and the measurement agents.This problem is equivalent to an optimization problem on configuration design.In order to resolve this optimization problem,we introduce the directed acyclic graph decomposition,bipartite graph decomposition and dynamic graph.They are used to verify the input and output accessibility,full rank condition and coprime paths property.A method is given to provide the optimal solution to minimum cost control configuration design problem.2.The design methods of distributed estimation protocols are explored for wireless sensor networks.For the case of heterogeneous sensors,the existence of a type of distributed observer protocol is proved.The relationship between the parameters of the protocol and the network topology is analyzed when the state reconstruction is realized.For linear system with state noise and measurement noise,a distributed information Kalman filter is proposed which is based on the finite fusion of innovation vector and innovation matrix.This is the two time scale design strategy.We assume that all sensors are cooperative observable.Then the stability of this filter is proved by constructing suboptimal estimator method.When the target is modeled as a nonlinear stochastic system,the recursive estimator is designed to estimate the state of the target.If the nonlinear relationship is weaker,the distributed extended Kalman filter can be used to get the approximate state.If the nonlinear relationship is stronger,the distributed unscented Kalman filter is designed to track the plant.The consensus filtering state and covariance matrix are obtained by using the storage data of each sensor.