Abstract: In the field of time-frequency decomposition, the Local Mean Decomposition(LMD) method is applied in settlement monitoring, but the phenomenon of mode mixing can appear during the application, which results in inaccurate deformation signal extraction.The Ensemble Local Mean Decomposition(ELMD) method can be used to improve the mode of mixing the local mean decomposition by adding auxiliary noise to the original signal, and also can use the statistical characteristics of auxiliary noise to remove the mode mixing. This paper uses simulation data to analyze the model error in the ELMD method and presents a parallel combination prediction method based on ELMD. In the case of high speed railway bridge monitoring data, it divides a series of discrete nonlinear and unstable signal into three product function(PF) components and one remaining component. The method takes advantage of the support vector machine and Kalman filter algorithms to predict these components, and analyses the superiority of ELMD in the case of mode mixing and overall feasibility empirically. The results indicate that: the parallel combination model, based on ensemble local mean decomposition (ELMD), can eliminate the mode mixing problem in the local mean decomposition (LMD) method very well and extracts the deformation signal accurately. In terms of prediction precision, the mean relative error can reach 8.3%, and may provide areference for prediction of deformation monitoring.