Research On The Optimization Of Neural Networks Method And Its Application For Integrated Navigation

Artificial neural network is a rapid development theory and technology in the research field of computer intelligence.It has the ability of learning the knowledge from the environment and adapting the environment in a way of interacting similar creature.The relationship between input and output of neural networks can be established without knowing the accurate system model.Therefore,the neural network method is used to predict the parameters of INS/GNSS integrated navigation system during the satellites signals are interrupted.However,the existing neural network methods have the following defects:the training efficiency of learning algorithm is low and rely on training samples heavily when the satellite signal is available;the prediction accuracy and generalization ability of network is poor during the satellite signal is interrupted.Therefore,in order to improve the performance of integrated navigation system,it is necessary to intense the research on the optimization of neural networks method and its application for integrated navigation.The research results have important academic meaning and practical applied value.In order to improve the efficiency of the INS/GNSS integrated navigation parameters prediction and correction by neural network technique during the satellites signals are interrupted,the improved learning algorithms of BP neural network and structure optimization algorithms of RBF neural network are researched deeply in this paper,and then a set of neural network optimization methods are presented for INS/GNSS integrated navigation system.The main research work and innovative contributions are as follows.(1)A novel neural network learning algorithm has been presented based on model prediction filter in this paper.Firstly,the state space equation was set up in the forward transfer process of the network while the weights of network were used as state variables.The local minimum problem generated by gradient descent method effectively could be avoided in this learning algorithm.Then,with considering the influence of sample noise on the network model identification,the model error of network was modified by prediction filter for making up the influence of model error on weights updating.Thereby,the trained network has stronger adaptability for the new test samples.The simulation results show that the proposed learning algorithm was superior to traditional BP algorithm and EKF algorithm in the two aspects of the training speed and generalization ability,thus the navigation precision with the INS work alone is improved.(2)Based on the study of BP neural network and fading Kalman filter,a novel neural network learning algorithm is presented on the basis of fading Kalman filter in this paper.The state space equation was built in the linear part of neuron in this learning algorithm.The measurement value was obtained through the study of inverse operation of sigmoid function for target output.Then,the filtering gain and the updated error covariance matrix is calculated by the recursion equation of fading Kalman filter with considering the influence of the latest measurement value fully.Finally,the weights of network were updated accoding to the obtained measurement value.The simulation results show that the learning speed of neural network could be improved by this proposed algorithm without loss in training accuracy.(3)A novel neural network learning algorithm based on fading UKF was designed in this paper.Unlike fading Kalman filter learning algorithm,the state space equation was built in the nonlinear part of neuron in fading UKF learning algorithm.Therefore,the measurement value could be acquired directly from training samples and thus the measurement error caused by inverse operation of sigmoid function in neuron will be avoid.Besides,in fading UKF algorithm,the length of the historical data could be limit by selecting the appropriate fading factor,and thus the present measurement information is utilized effectively.Finally,the covariance matrix of current measurement prediction was adjusted and the weights of network were corrected.The proposed algorithm was applied to identify the nonlinear function containing the random noise.The simulation result demonstrates the effectiveness of the proposed algorithm in the aspect of improving the network generalization ability.The same viewpoint is confirmed by the application in INS/GNSS integrated navigation system.(4)Aimed at the problem of structure design of RBF neural network,a new structure optimization algorithm for RBF neural network is presented based on variance significance in output sensitivity in this paper.First of all,the variance of the output sensitivity of hidden layer nodes on the samlple set is used as significance measurement and the calculation method of statistical variance significance is given.Then,the significance of nodes of hidden layer are determined by comparing the relationship of size between the calculation results of variance statistical and examined variable.Thus,the nodes are judged whether to split or delete based on the relationship of size.Finally,the parameters of trained network are updated for improving the prediction accuracy and convergence of network.The validity of the proposed algorithm is proved by the application in nonlinear function approximation and error prediction of integrated navigation.Compared with other structure optimization algorithm of RBF network,the determined network structure of the proposed algorithm is more simple and the generalizing ability is better.(5)Aimed at the problem of interruption of satellite communication due to the satellite navigation system are susceptible to interference,the layout schemes of pseudolite positioning independently has been studied in this paper.Based on the analysis on the impact of satellite number on the geometric dilution of precision(GDOP)from the definition of GDOP,the schemes are comfirmed with six pseudolite positioning in near space airship.Then the influences of altitude and azimuth angle of pseudolite to GDOP are analyzed and the new scheme of six pseudolite in near space airship is designed in this paper.Finally,for comparing the performance between the proposed schemes and the existing schemes,the simulation test are made.The simulation results show that the proposed schemes not only increase the range of the position area of users,but also decrease the GDOP and improve the positioning accuracy.