Research On The Methodology And Application Of State Estimation For Uncertain Dynamic Systems

The complexity and uncertainty of dynamic systems in modern industry are increasing.Accurate state estimation is the key to realize intelligent control,fault diagnosis and target tracking.Uncertain dynamic systems are characterized by the uncertainties of model parameter,gross error of observed data,initial state information,state model,and noise distribution.The existence of these uncertainties makes the state estimation of dynamic system by using general state estimation method deviate greatly,and affects the accuracy of the state estimation.Therefore,accurate state estimation for uncertain dynamic systems is the focus and difficulty of the current research.Aiming at the problems of state estimation for the uncertainty in model parameters and noise distribution,the uncertainty in gross errors of observed data and noise distribution,the uncertainty in initial states,the uncertainty in state model and noise distribution of dynamic systems,the corresponding state estimation methods are studied.The main research contents include:(1)For the problem of state estimation of the dynamic system with uncertain model parameters and unknown noise distribution an evolutionary algorithm for state and parameter estimation of dynamic process system(AE-CRPF)is proposed.In this algorithm,the estimability of parameters is analyzed,and the method for initial parameter estimation is given.Combined with artificial evolution and cost reference particle filter algorithm,the model parameters and the state vectors of dynamic system can be estimated simultaneously under the condition of unknown noise distribution and model parameters.Cost reference particle filter(CRPF)is a variant of particle filter(PF),which is simpler,more robust and more flexible than the general PF.In particular,it does not require any distribution information of state noise and observation noise in the application.The AE-CRPF method is suitable for joint state and parameter estimation of nonlinear dynamic systems with unknown model parameters and noise distribution.(2)For the problem of state estimation of dynamic system with uncertain gross error of observation data and unknown noise distribution,a state estimation method based on dynamic data reconciliation(DDR)and gross error detection(GED)is proposed.This method combine cost reference particle filter(CRPF)and boxplot detection,so it is called Box CRPF-DDRGED.Dynamic data reconciliation(DDR)and gross error detection(GED)are the keys to achieve accurate state estimation,because gross error will seriously affect the result of state estimation.The detection of gross errors in observational data in the existing techniques is too dependent on the assumption that the distribution of observational errors is known prior and is generally considered to be white Gaussian noise distribution.Although general particle filtering(GPF)can be used to detect gross errors and to reconcile dynamic data,it relies on the pre-known noise distribution.However,in practical application,the observed data of dynamic system have more or less uncertain gross errors,and the noise distribution is unknown,so the Box CRPF-DDRGED method is proposed.According to the DDR results of CRPF,the boxplot detection method is used to detect gross errors.Considering that the type of gross error is uncertain,a method of identifying and compensating gross errors by using GED results and the residuals is proposed.The results of gross error compensation can significantly improve the accuracy of state estimation.This method is suitable for DDR,GED and state estimation of nonlinear dynamic systems with gross errors in observed data and unknown noise distribution.(3)For the problem of obtaining the initial state and state estimation of dynamic systems with uncertain initial state,a robust state estimation method based on PF and data reconciliation is proposed.In this method,the measurement test criterion to obtain the reliable initial value of state under the condition of sufficient observation information and the data reconciliation method to increase the observed data information in sequence are proposed.The method iteratively improves the initial value of state through the information interaction between PF and data reconciliation,accurate state estimation results are obtained.This method is suitable for state estimation of nonlinear dynamic system with unknown initial state and for tracking maneuvering target with unknown initiation of trajectory.(4)For the problem of state estimation of dynamic systems under the uncertainty of state model and unknown noise distribution,an interactive multi-model cost reference particle filter(IMMCRPF)based state estimation method is proposed,which combines interactive multi-model estimation with cost reference particle filter.In the IMMCRPF method,a group of CRPF filters running in parallel are used to estimate the states of the system and reconcile the observations,each filter applies a specific model to estimate the state variables and reconcile the observations.The interaction between multiple filters allows the fusion interaction of state estimates,and the transition probability of switching between different models uses Markov chains.The cost increment in each model filter is regarded as the likelihood of each model,then a joint model exponential function is defined,and a rule is selected to adaptively determine which model is most suitable for process state in real time.This method is suitable for state estimation of nonlinear dynamic system with uncertain state model and maneuvering target tracking with uncertain motion model.