Research On Methods And Algorithms Of Optimal And Self-tuning Multisensor Measurement Fusion Filtering

With the rapid development in computer technology, communication technology, microelectonics technology, precision machine manufacture technology and control technology, all kinds of multisensor systems, suited to the complicated application backgrounds, come into being. In the multisensor systems, since the information has a lot of characters, for example, it has the various forms, the tremendous quantities, the complicated relationship, and it requires to be precise and reliable, and settled in time, in order to increase the synthesized handling ability, a new branch of science, multisensor information fusion, intersected, synthesized and expended by information subject and control subject, comes into being.Multisensor information fusion estimation is an important branch of multisensor information fusion, and its aim is how to utilize the measurement data, provided by multisensor, to make more precise estimation to the state or signal of the systems than the estimation based on the single sensor. It is widely used in many high technology fields, such as target tracking, military affairs, space flight, guidance, GPS position and robotics.This subject comes from the National Nature Science Foundation under Grant 60374026, New Theories and New Methods for Multisensor Information Fusion Optimal and Self-tuning Filtering, and Grant 60874063, Research on Self-tuning Information Fusion Filtering Theory and Its Applications.Based on modern time series analysis method, for multisensor linear stochastic systems, this paper makes the theoretical analysis primarily, the computation simulation examples secondarily, and the deep researches on the state estimation problems of optimal weighted measurement fusion and self-tuning weighted measurement fusion. The main works are as follows:First, for the multisensor systems with unknown noise statistics, applying the correlation method, the information fusion noise statistics on-line estimators, which have the strong consistence, are presented, where the nonsingular partial linear equations are selected from the sampled correlation function linear equations to be solved to obtain the local noise statistics estimators, then the fused estimators are obtained by making the weighted arithmetic average of the local estimators, and it is proved that the fused estimators can be viewed as the least squares fusers. Using the ergodicity of the correlation function, it is proved that the fused estimators converge to the true noise statistics with probability one. Since the accuracy of each fused estimator falls in between the highest and lowest accuracies of the local estimators, it has the higher reliability. For the multisensor autoregressive moving average (ARMA) systems with unknown model parameters and noise statistics, applying the existing estimation methods for ARMA model parameters, combining the information fusion on-line estimators of the unknown noise statistics, the multi-stage information fusion identification methods and the concept of the information fusion model parameter estimators are presented. The information fusion model parameter estimators are obtained by taking the average of the local model parameter estimators obtained from each sensor. In theory, the consistence of the proposed information fusion model parameter estimators is rigorously proved.Second, for the multisensor linear discrete time-invariant stochastic systems with correlated measurement noises and correlated noises, under the linear unbiased minimum variance (LUMV) criterion, applying Lagrange multiplier method and the filtering theories based on ARMA innovation model, the optimal weighted measurement fusion steady-state Kalman filtering and Wiener filtering algorithms are presented respectively, and theirs results are equal to those of the optimal weighted measurement fusion filtering algorithms based on the weighted least squares (WLS) fusion criterion, that is, they have the completely functional equivalence and the asymptotical global optimality. For the multisensor systems with different measurement matrices and correlated measurement noises, using the matrix partitioning method, combining the constraint condition of the measurement weight matrices, the reduced dimension measurement fusion algorithm is presented, which can remarkably reduce the computational burdens of computing the measurement weight matrices and the fused measurement error variance matrices. The presented comparison table of the computational burden supports this inference from the view point of quantitative analysis.Third, for the multisensor systems with unknown noise statistics, correlated measurement noises and correlated noises, by substituting the information fusion on-line estimators of unknown noise statistics into the steady-state optimal weighted measurement fusion filters, the self-tuning weighted measurement fusion Kalman filters and Wiener filters, and the self-tuning reduced dimension measurement fusion filters are presented respectively. Applying the dynamic error system analysis (DESA) method, it is proved that the proposed self-tuning fusers converge to the corresponding steady-state globally optimal fusers in a realization, that is, they have the asymptotical global optimality.Last, for the multisensor single channel autoregressive (AR) signal systems and ARMA signal systems with the common disturbance noise, unknown noise statistics and unknown model parameters, using the transformation method from ARMA model to the state space model, the signal estimation problem is transformed into the state estimation problem. By substituting the information fusion estimators of model parameters and noise statistics into the steady-state optimal weighted measurement fusion Wiener signal filters, the self-tuning weighted measurement fusion Wiener signal filters are presented respectively. By DESA method, it is rigorously proved that the proposed self-tuning weighted measurement fusion Wiener signal filters converge to the corresponding steady-state globally optimal filters in a realization, so they have the asymptotical global optimality. In the whole paper, a lot of simulation examples in the tracking system and the signal processing show the effectiveness of the proposed methods and algorithms.