| Summary: | 碩士 === 國立海洋大學 === 導航與通訊系碩士班 === 91 === Every navigation system has its drawbacks and this will limit its application. Integration of GPS and INS gains benefit at the technique and price than alone navigation system was used in particular missions and can produce a system performance superior to either one acting alone. In general, the navigation information has been processed by Kalman filter. The noise covariance matrix must be precise for obtaining optimal (minimum RMS error) solution. In filtering process, the rough process and measurement covariance matrix would lead to degraded result. The adaptation performance of conventional adaptive Kalman filtering approach depended on the window sizes of innovation. When the length of window sizes is too short, the estimation of covariance will be noisy. On the other hand, much time is needed to obtain the estimation to fit in with the correct covariance. Back-Propagation Neural Network (BPNN) and Fuzzy Neural Network (FNN) are used to adapt the covariance matrix for Kalman filtering.
BPNN is the artificial learning algorithm that has been widely used in many experimental works. The learning ability of the neural network is capable of nonlinear mapping when the network is trained by appropriate training pattern. The network that has been trained is used to estimate the covariance matrix in the filtering process. The estimation ability of BPNN does not relate to the system model and the network must be integrated with the navigation system because the innovations are the inputs of BPNN. When the navigation system is operating, the covariance matrix is estimated by BPNN simultaneously, and then the performance of navigation system will be improved.
BPNN has two drawbacks in estimation process. The result of estimation does not satisfy if the numbers of input nodes in input layer are too little; on the other hand, the time of estimation would increase. To fix the problem, fuzzy logic will be added to BPNN and attempts to improve the accuracy of estimation and reduces the input number. Numerical simulations show that the performance based on the proposed approach is substantially improved.
|
|---|
