Research On Landslide Geological Disaster Prediction Method Based On Deep Learning

There are a large number of landslide disasters in the world every year,and in result,a great many of casualties and property losses are caused,and the prediction of landslides is extremely important.With the increasing variety of landslide fixed-point monitoring sensor devices,and more and more researchers introducing deep learning technology into the field of landslide prediction,which provides an important data and technical basis for landslide prediction.However,the traditional landslide prediction model has problems such as the information between historical data cannot be well used,the correlation between multiple variables cannot be well mined,and its prediction accuracy is also insufficient.In order to solve these problems,this thesis studies the landslide prediction of Maliu Village along the Lancheng-Chongqing Pipeline as an example,and constructs a landslide prediction model.The work and conclusions carried out in thesis are as follows:1.The landslide monitoring scheme is designed for the overview of the study area,and the monitoring data is removed from outliers,missing value filling,data noise reduction,data standardization and data alignment pre-processing,of which the monitoring data mainly include pipeline strain,deep part shift,anti-slip pile tilt and rainfall data.The pretreatment work provides preparation and foundation for subsequent experiments.2.This thesis proposes a two-layer LSTM landslide prediction model based on attention mechanism(Attention-Bi-LSTM Model)for single-sequence landslide data,and the attention mechanism can well excavate the characteristics and change laws between landslide data time series,and the double-layer LSTM solves the problem that backward time information cannot be well used in single-layer LSTM for historical data.Compared with the landslide prediction model of RNN,double-layer LSTM and single-layer LSTM based on attention mechanism,this thesis uses root mean square error(RMSE)and mean absolute percentage error(MAPE)as the evaluation indicators of the prediction model.Experimental results show that the RMSE value of Attention-Bi-LSTM Model was increased by 9.8392,8.6775 and 3.7590,respectively,MAPE values increased by4.2267%,0.1880% and 2.6510%,respectively.3.This thesis proposes a convolutional neural network double-layer LSTM landslide prediction model based on dual-attention mechanism(CNN-Attention-Bi-LSTMAttention Model)for multi-feature landslide data,which uses convolutional neural networks to extract the features between the data of various variables,and introduces an attention mechanism to better mine the information between variables.Experimental comparisons were performed with the Attention-Bi-LSTM Model,and experimental results show that the RMSE value and MAPE value increased by 0.8492 and 0.1512%,respectively.4.A landslide data correction and prediction early warning system is designed by combining the data preprocessing process method and the multi-feature landslide prediction model in this thesis,which provides the function of outlier removal,missing value filling function,data prediction function and landslide early warning function,and visualizes the system.