Research On Landslide Evolution State Prediction And Control Based On Multi-task Learning

Landslide is a highly destructive geological disaster.Landslide disasters frequently occur worldwide,posing a serious threat to the safety of people's lives and property.Therefore,the in-depth study of the prediction and control of the landslide evolution state has significant practical and economic significance to reduce the impact of landslide disasters based on the basic data provided by the landslide monitoring system.As a nonlinear dynamic system,landslides have complex and unknown evolutionary deformation mechanisms,and the actual amount of available data is relatively small.Long and short-term memory networks are suitable for modelling dynamic systems,and multitask learning can alleviate the problem of few-shot to a certain extent with the help of information sharing.In this thesis,multi-task learning is introduced into the study of landslide evolution state prediction and control,and models and methods are verified on the Baishuihe and Shiliushubao landslides.Firstly,we propose a landslide evolution state level classification prediction problem and define two landslide evolution states.Specifically,we use Gaussian mixture model(GMM)to reconstruct labeled data sets and establish a landslide evolution state level prediction model based on Multi-Task Learning-Stacked Long-Short Term Memory(MTL-SLSTM),and then use task weight rules to design multitask losses for network training.The high-precision parallel multi-step prediction of the evolution state of single monitoring point is achieved by the above works.Secondly,considering the spatiotemporal correlation of different monitoring points on the same landslide,we analyze spatiotemporal data to construct spatiotemporal features,and design a multi-task correlation learning mechanism combined multi-task weight learning and multi-task relationship learning methods to construct spatiotemporal relation.The landslide multi-point prediction model based on Multi-Task Correlation Learning-Stacked LongShort Term Memory(MTCL-SLSTM)achieves a single-step prediction of the evolution state of multiple monitoring points with high accuracy.Finally,according to the idea of neural direct inverse control,we propose a landslide down-level control method based on the prediction of the landslide evolution state.We build an interval prediction network based on bootstrap method and model selection strategies(S-BS-PIs),and then safe rainfall interval predictors are trained offline.Moreover,we design an online down-level control process combined the landslide evolution state level predictor,which realizes single-step control of the dangerous landslide at a single-point.This thesis focuses on the two key issues of landslide,that are prediction and control.Using neural networks and multi-task learning methods to predict the evolution state of landslide,and further obtaining effective control variables through data-driven control methods,which provide early warning information and reference plans for landslide prevention.