Research On Neural Network And Kalman Filter

This paper researches on two kinds of combination of Neural Network (NN) and Kalman Filter. NN is applied to lots of non-liner problems due to its map-approaching and self-learning abilities. Meanwhile Kalman Filter is one of the most popular algorithms in the fields like information and navigation systems which can implement the optimal estimation of the system states. But these two methods have flaws respectively: though NN does not need specific mathematic model, it can't work without lots of correct learning samples (or training samples). And noise learning samples can lead to an error learning result to make NN invalid. In the mean time, its low learning speed and poor generalization ability restrict its application in engineering. On the other hand, Kalman Filter works badly without specific mathematic model of system and probability character of noise information. Its accuracy is low which makes it fade easily. It also has problems on large computation and dimension issue.This paper presents two improved methods for the learning problems of NN: one algorithm based on DFP and the other based on Kalman Filter. The former magnifies updating parameters and adds noise data so as to meliorate the stability of NN. It also solves the overflow problem and enhances its practicability. While the latter updates the estimation state parameters by other learning algorithm in the time updating process, and then modify the formula of Kalman Gain by these results. In this way, a suit of new time update and measure update formulas come up. This new method solves the problems of dimension disaster and large computation. And it enhances the robust of NN and improves the learning ability by batch learning.In allusion to shortcomings of Kalman Filter, this paper researches on the other combination between NN and Kalman Filter: based on Kalman Filter, two NN are applied to learning the estimated data and true data of Kalman Filter so as to correct the respective results in time update process and measure update process. This method solves these problems of Kalman Filter on bad stability and low accuracy.Plenty of experiments on the improved methods are presented in this paper. Through the comparison with former methods, the new methods show their advantages. The combination of NN and Kalman Filter is one of the hotspots in research of information fuse. And the combination of these two methods make them learn from the other's strong points to offset its weakness and make the algorithm more practicable by improving the accuracy and anti-jamming ability.