Non-Gaussian Time Series Filtering For Dynamic Deformation Monitoring Based On Particle Filter

Dynamic deformation monitoring is faced with complex and changeable environment, and multiple influence factors lead to uncertain distribution of the noise in the deformation time series. The measurement environment has to be acknowledged before filtering time series. In this paper, the distribution characteristics of time series are described by linear/nonlinear test and Gaussian/non-Gaussian test. A comprehensive test with multiple methods is used to ensure the correctness of the test results.Traditional filtering algorithms, which are blocked at the door by non-Gaussian noise, are based on the Gaussian assumption. Particle filter that under the framework of Bayesian theory is based on Monte Carlo method, sequential importance resampling, it is no need for Gaussian distribution and linear assumption. Recursive, continuous, dynamic, real-time filtering of discrete time series is completed by Bayesian estimation, with Monte Carlo method, the high dimensional integral operation in Bayesian estimation is avoided. The filtering problem of dynamic time series is solved by the non-parametric sampling test. Beyond the standard particle filter, nonlinear filter correction, which makes full use of the latest observation, is selected as proposal of particle filter, so that the true distribution of the state is approached by sampled particles. The weight degradation is eliminated by systematical resampling, and with the increase of particles diversity by Markov Chain Monte Carlo move, sample Impoverishment caused by over resampling is reduced.In this paper, the general idea about the dynamic deformation monitoring data processing based on particle filter algorithm is:first, linear/nonlinear test and Gaussian/non-Gaussian test of time series; then, according to the test results, an accurate state space model of the time series will be established; finally, filtering the time series by the appropriate particle filter suits with the test results. A particle filter application for the dynamic deformation monitoring data processing is completed with mixed programming using C# and MATLAB.According to the processing results of simulation data and dynamic deformation monitoring data, particle filter and improved particle filters showed good filtering effect, especially in non-linear, non-Gaussian time series, they are optimal filters. With the improved particle filters and sidereal filter combined, the multipath effects in GPS single epoch deformation signals were more effectively eliminated, and the standard deviations of the residual coordinate sequences on N, E directions are decreased a greater degree than U direction, particle filters improved the reliability of GPS dynamic deformation monitoring.