| Summary: | 碩士 === 育達商業科技大學 === 資訊管理所 === 101 === The steady-state Kalman gain filter so-called an alpha-beta-gamma filter was developed for centralized measurement fusion to lessen the computational loads. In this method, the filter gain matrix is computed for gain-scheduling and obtained from three uncoupled one dimensional filters by using closed form equations in the line-of-sight Cartesian coordinate system and then transformed for use in the local inertial Cartesian coordinate system. In order to keep high tracking accuracy, adaptive filtering approaches were frequently used. There were an adaptive Kalman filter and an adaptive alpha-beta-gamma filter for tracking applications. The first approach consists of a multi-band filter (a group of parallel filters) and an on-line Bayesian classifier. Bayesian classifier is constructed based on Radial Basis Function Network (RBFN) and used to classify which output is selected among different level band filters for adapting different target maneuver turns. The second approach consists of a group of parallel alpha-beta-gamma filters and a General Probabilistic Neural Network (GPNN). By incorporating a general probabilistic formulation and Markov chain into a general regression neural network, GPNN is developed as a decision logic algorithm for on-line classification. Each activation function of GPNN is defined as Gaussian basis function whose smooth factor is a constant selected from filter’s innovation covariance matrix by utilizing the parametric method. In this thesis, we develop an adaptive gain-scheduling algorithm for improving the tracking accuracy of alpha-beta-gamma filter in transient response. In the transient period, we use the curve fitting technique for approximating the gain of decoupled Kalman filter. In the steady-state period, we use the gain of alpha-beta-gamma filter for tracking target. Finally, we verify the effectiveness of proposed algorithm including the tracking accuracy by using Monte Carlo computer simulation method.
|
|---|
