The Research Of Kalman Filter Based On Neural Network

In the signal processing area, It is a significant subject to research how to extract useful signal from the noise environment. In order to get pure signal, Wiener Filter and Kalman Filter which based on least mean-square rule are widely used to reduce noise signal. Wiener Filter is an efficient method used in one-dimensional random process, but it can't be used in multidimensional and multivariable random process, Kalman Filter use state space to describe system feature, it can be used in the environment not only stationary random process, but also multidimensional and non-stationary random process, and solved the application limited of Wiener filtering. However, the performance of Kalman Filter depends on the correct priori knowledge of noise covariance matrices and precise system model, model error may cause degradation even divergence in filtering performance, neural network is used to aided the problem of model error. This paper proposed an innovation-based neural network Kalman Filter algorithm. This algorithm not only can adaptive estimate the noise covariance matrices, but also can compensation estimate error brought by imprecise system model, is a optimal filtering algorithm.In this paper, firstly, system identification based on BP neural network is researched. Considering the influence of the noise on identification samples, this paper proposed adaptive neural network system identification both in offline and online situation, the simulations in MATLAB show system identification based on neural network are simple and feasible methods to get the system model. Secondly, Considering the impact of noise statistics to Kalman Filter performance, the innovation based Kalman filter is introduced,and the simulation in the situation of changing noise covariance shows the algorithm can adaptive estimate the noise statistical characteristics, keep a good filter effect. Thirdly, Innovation-based neural network Kalman Filter algorithm is proposed, the influence factors of system model and noise statistical characteristics are fully considered in this algorithm, the simulation results both in offline and online identification show a good filtering performance, the filtering results are significant improved and the system is more stability, and the effect on online is better than offline identification. In the last of this paper, the basic principle of target tracking is introduced. The algorithms of Innovation-based neural network Kalman filter, Interacting Multiple Model and Least Square algorithm are simulated in single target tracking, obviously, Innovation-based neural network Kalman filter has the best tracking performance,is a feasible and practical filter algorithm.