Study On Slope Displacement Prediction Based On Deep Belief Network

With the frequent occurrence of slope deformation accidents,people’s property is infringed and life safety is threatened.Therefore,the safety analysis of slope deformation is particularly important.Displacement is an external characteristic that reflects the internal deformation of the slope.By monitoring the obtained displacement data of the slope to establish a prediction model,the future predicted value of the slope can be effectively grasped.There are mainly curve regression methods,neural network methods and deep learning methods for establishing prediction models.Due to the advantages of deep learning that can analyze complex non-linear systems and automatically extract abstract features,it has been widely applied in many prediction fields.This article analyzes the monitored slope data based on the Deep Belief Network(DBN),and uses its ability to automatically extract data features to predict the slope displacement.The main research contents are as follows:(1)The research status of slope displacement prediction is introduced.Due to the time-varying nonlinearity and random uncertainty of slope displacement,it is difficult for traditional prediction methods to comprehensively analyze the data characteristics of slope displacement and develop deep learning prediction.At the same time,The slope displacement data acquisition experiment is introduced,and the slope displacement is monitored using a GNSS receiver to provide data support for subsequent displacement prediction.(2)The theoretical basis and training method of the DBN network are elaborated,and a slope displacement prediction model based on the DBN is established.The optimal network structure is determined by analyzing the influence of the hidden layer settings,parameter learning rate,and momentum settings on the prediction results.Experimental results show that compared with the shallow BP method,the average relative prediction error is reduced by 2%,and the root mean square error is reduced by 4%.(3)DBN slope displacement prediction model of automatic layer selection DBN is established to solve the problem that the DBN network manually determines the number of layers takes a long time and there is a certain error.The model automatically determines the number of network layers based on the reconstruction error theory,and compares it with the basic model for prediction experiments.Experimental results show that compared with the basic DBN model,the training time of the automatic layer selection DBN model is shortened by 10.05 s,and the average relative prediction error is reduced by 1.3%.(4)The slope displacement prediction model based on automatic layer selection WPA-DBN is established to solve the problems that randomly initialized DBN weights and thresholds will cause the network training to easily fall into local optimum and large prediction errors.Use WPA algorithm to optimize the initial weights and thresholds of the DBN network to further improve the feature extraction capability of the model.Experimental results show that the average relative error of the WPA-DBN model is reduced by 2.5% compared with the model of automatic layer selection DBN.