Research On The Application Of Wavelet Transformation And Neural Network In GPS/DR Integrated Navigation System

Global position system (GPS) is of superior long-term performance in position, but of poor short-term performance, while dead reckon (DR.) system has good position precision in short-term, but poor in long-term. GPS/DR integration provides position data for vehicle navigation system (VNS ) with high precision and reliability. Through analyzing the virtues and deficiencies of Kalman filter (KF) data fusion algorithm for the VNS, some new data fusion algorithms are presented based on wavelet transformation and neural network. The studies are as follows:(1) Through analyzing the dynamic filter property of vehicle GPS navigation system with KF, a new dynamic filter algorithm is presented based on discrete wavelet transformation (DWT). Firstly, the algorithm decomposes GPS navigation signal with discrete wavelet function. Secondly, the signal's observation gross error is detected and removed at different resolution levels with statistic 3σtheory, and random noise is de-noised at different resolution levels with DWT model square root soft-threshold de-noising algorithm. Finally, the signal is reconstructed and provides precise position data for the VNS.(2) It is difficult to establish precise mathematical model in extended kalman filter (EKF) data fusion algorithm for vehicle GPS/DR integrated navigation system , an improved EKF data fusion algorithm is put forward based on stationary wavelet transformation (SWT). Firstly, GPS, odometer and gyroscope signals are de-noised with SWT model square root soft-threshold de-noising algorithm. Finally, precise position data is provided for the VNS with EKF.(3) It is difficult to do 'zero update' for DR system's position in case of GPS signal blockage, a new position algorithm is put forward based on adaptive linear neural (ADLINE) network. Firstly, GPS, DR position data is de-noised with DWT model square root soft-threshold de-noising algorithm during the presence of GPS signal. Secondly, the algorithm compares the two position data and acquires DR position error data. Thirdly, an ADLINE network model to mimic DR position error property is trained with least mean square (LMS) algorithm. Finally, DR position error data is predicted with the model, and the algorithm provides precise position data through DR position error data updating DR position data during GPS outages.(4) The ADLINE network model is available and effective for DR position error prediction in simple vehicle running trajectory; while the model isn't in complex one, because ADLINE network is the only linear fitting and classifying ability; so another new position algorithm is put forward based on back-propagation (BP) neural network. Firstly, DR position data is de-noised with SWT model square root soft-threshold de-noising algorithm, and vehicle precise position data is acquired with EKF data fusion algorithm based on SWT during the presence of GPS signal. The algorithm compares the two position data at different SWT decomposition level, reconstructs the SWT coefficients and DR position error data is acquired. A BP network model to mimic DR position error property is trained with back-propagation algorithm. To improve generalization and training speed of the BP network, the back-propagation algorithm is improved by Bayesian regularization (BR) theory and Levenberg-Marquardt (LM) algorithm in this paper. During GPS outages, DR position error data is predicted with the model, and the algorithm provides precise position for the VNS through DR position error data updating DR position data.All the presented data fusion algorithms are also validated by computer simulation experiment in this paper.