Research On Deformation Analysis And Forecast Method Based On Model Combined By Kalman Filter And BP Neural Network

Research on deformation analysis method is one of the important contents of project health monitoring and assessment. Currently, there’re many deformation analysis and forecast models, including:trend analysis method, time series method, grey model method GM(1,1), neural network model method etc. Based on the advantage of Kalman filter in dynamic real-time forecast and the strong nonlinear mapping ability of BP neural network, discusses the fusion algorithm of the two models and its application in deformation analysis and forecast. In the progress of study, the main research results are as follow:1.Deduces the mathematical model Kalman filter; Studies maximum posterior estimation adaptive Kalman filter and variance component estimation adaptive Kalman filter which can limit the divergence problem of traditional Kalman. Studies the principle and training procedure of BP neural network detailedly, provides BP neural network program flow based on Batch sample learning.2.Studies Kalman and BP optimal weighted combined model(OWC Kalman-BP) which considers Kalman filter and BP neural network as single model. According to the Residual of every single model, OWC Kalman-BP weighted combines forecast results. The analysis shows that OWC Kalman-BP reduces the uncertainty influence of every single model and improves forecast accuracy effectively.3.Studies BP neural network based on Kalman learning algorithm(KLA Kalman-BP) which considers BP neural network as the core model, Kalman filter is used as the learning algorithm to adjust the weights and thresholds of the network. KLA Kalman-BP greatly improves network convergence rate and avoids local optimal problem of the traditional BP. The analysis shows that efficiency and accuracy of KLA Kalman-BP is superior to the traditional BP and OWC Kalman-BP.4. According to engineering experience, this article divides project deformation trend into four typical types:"stationary", "gradual change", "mutation" and "high frequency stack extension period". Based on the "real-time tunnel three dimensional displacement monitoring data", simulation analysis the four typical types of deformation trend and discusses the accuracy and applicability of two Kalman-BP combined models.