Distributed State Estimation Algorithm For Multi-Agent Systems In Complex Environment

Recently,scholars in different fields have achieved a lot of results for distributed state estimation problem in multi-agent system.It has also been successfully applied in biology,sociology and physics.For multi-agent networks,agent’s classification and information interaction are the key points,which require agents to cooperate and jointly process the collected information.Agents are designed to transfer the collected information to all other nodes or to the fusion center for joint processing.Since there is an upper limit on the available resources and data bandwidth of each node,it is generally not feasible to transfer the original measurements to other nodes.An effective processing method is to use local information to process the collected data in a distributed fashion.Then,the local information can be fused with the information collected from adjacent nodes,which reduces the communication and computing requirements.In addition,in the process of multi-agent collecting data,data missing or wrong data is a common problem in practical engineering.Also,in some processes,one needs to estimate the states or parameters through sparse data and noisy data,which is also an important problem.With the improvement of the complexity of tasks in the real environment,the performance requirements of state estimator are also enhanced,especially in improving the reliability of task execution.How to maintain the coordination and perfect integration of multi-sensor measurements,how to ensure the accuracy of system model establishment and the reliability of communication network,and how to ensure the accuracy,timeliness,ease of operation,anti-interference,robustness and other performance requirements of state estimation algorithm have become a meaningful and urgent research topic.For the state estimation problems of state space model in complex environment,this dissertation carries out the research on distributed state estimation algorithm,distributed Kalman filter in state space model with faulty sensors,and distributed data-driven(Gaussian process)state estimation for trajectory learning in multi-agent system.The main considered problems and contributions are summarized as follows:1)Joint estimation algorithm for unknown structural parameters and states.For linear systems,the maximum likelihood(ML)criterion,the expectation maximum(EM)algorithm,Kalman filter and Kalman smoother are used to estimate the unknown structural parameters.The results of the previous step are used to perform filtering and smoothing,and then the EM algorithm is performed based on the results of the current filter and smoother.For nonlinear system,particle filter,particle smoother,and variational Bayesian(VB)inference are employed to perform joint estimation of states and parameters.Finally,simulation examples are provided to verify the proposed algorithms.2)A trust based fusion strategy is proposed to resist sensor faults and corresponding distributed Kalman filter is presented.First,the sensors collect measurements,and then update local state estimations and estimation error covariance matrices.Then,the sensors exchange information with adjacent sensors(state estimations and estimation error covariance matrices).After obtaining the estimation information from adjacent sensors,an iterative clustering algorithm is proposed,which includes three steps(initialization step,assignment step and update step).The collected estimation information can be divided into two clusters(trusted cluster and untrusted cluster).Third,the fused state and error covariance matrix are calculated by the average fusion algorithm(geometric average fusion on matrices/vectors and Wasserstein average fusion on probabilities).Finally,the time update is performed based on the fusion information.The stability and convergence of the proposed distributed filter are analyzed.Target tracking simulation examples are provided to verify the effectiveness of the proposed distributed filter.3)Data-driven distributed state estimation problem of multi-agent systems is studied.Data driven methods are widely used in state space models because they are suitable for tracking target motion in complex uncertain and unknown system models.Gaussian process model is a typical data-driven method,which allows to train system model and perform tracking trajectory in complex uncertain environment,and can be used to learn and train data collected from agents.Then,a distributed Gaussian process prediction method is proposed to solve the trajectory tracking problem,in which multiple nodes cooperate without central coordination to estimate a common Gaussian trajectory function based on local measurements and adjacent data.The considered trajectory can be generated by continuous and discrete systems.For continuous systems,a control Lyapunov function is designed for training.For the discrete system,the method based on distributed model predictive control(MPC)is used to solve it.Then,the local prediction results are fused by performing Kullback-Leibler average consensus.Simulation and actual trajectory tracking experiments verify the performance of the proposed method.