Particle Filtering Algorithm Based On Improved BP Neural Network

For linear Gaussian dynamic systems,the Kalman filtering is the optimal filtering method.However,when the Kalman filter tackles with a nonlinear system,it involves the calculation of high-dimensional integrals,so the analytic estimation of the system state is infeasible.The main idea of particle filtering is to utilize discrete weighted samples to represent the posterior probability density and then to obtain an estimate of the current state.Theoretically,the particle filter algorithm can sufficiently approximate the posterior probability density when the number of particles is large enough.Nevertheless,in the resampling step,particles with large weights are more likely to be drawn,causing the variance of weights to increase with time,which will inevitably lead to the particle impoverishment.It thus needs a sufficient number of particles to ensure the accuracy.Moreover,with the dimension of the system state increasing,the computational complexity increases and the efficiency becomes lower.For specific models,the importance density function plays an important role in the performance of the particle filtering.In this thesis,the nonlinear mapping function of BP network is exploited.By adding the weight splitting step,the partial particles with smaller weights are used as input samples,the particle weights are taken as the weights of the network,and the measured values are recognized as target samples of the network.Thereafter the weights of the particles are trained multiple times to improve the diversity of the particles in the PF algorithm,alleviate the weight degeneracy and enhance the filtering performance of the PF algorithm.This thesis firstly introduces the general framework of particle filtering,the basic knowledge of BP neural network and the theoretical derivation of neural network training process,and then proposes the specific BPNNPF algorithm of two types of models: One is a four-dimensional pure azimuth radar tracking model.The other is a mixed linear/nonlinear Gaussian model.Lastly,for illustration,we employ both the standard PF algorithm and the BPNNbased PF algorithm to simulate these two types of models.the experiment results shows that the BPNN based PF algorithm has a higher estimation accuracy than the PF algorithm.