Landslide Prediction And Warning Based On Multi-model Fusion And Data Anomaly Detection

Landslide disasters are common geological hazards in China,and slope displacement prediction and landslide advance warning are effective measures for landslide prevention and mitigation.Slope monitoring systems serve as the main tools for landslide prevention and mitigation,evaluating slope stability through monitoring and analysis of slope displacement.Radar technology has been widely used in slope monitoring systems.However,radar monitoring data is often affected by high noise and strong coupling,leading to potential inaccuracies in slope displacement prediction results.Additionally,obtaining accurate hazard assessment results for slopes can be challenging due to inherent errors in radar monitoring data and prediction data.Therefore,research on multi-step prediction and landslide advance warning of slope displacement based on radar monitoring data,utilizing multi-model fusion and data anomaly detection,is of paramount importance for effective landslide prevention and mitigation.Accordingly,this study focuses on landslide prediction and advance warning based on multi-model fusion and data anomaly detection,with two main research aspects:(1)Multi-step Prediction of Radar-Monitored Slope Displacement Based on Improved Phase Space Reconstruction and Multi-Model Integration: Considering the chaotic characteristics of radar monitoring data and the need for multi-step prediction of slope displacement,this study presents a method that utilizes improved phase space reconstruction and multi-model integration for radar-monitored slope displacement prediction.The method involves several steps: First,the original data is reconstructed through resampling at a fixed sampling frequency,and the displacement sequences are decomposed into trend terms and chaotic fluctuation terms using empirical mode decomposition.Next,an autoregressive-based multi-step prediction model is designed for the trend terms.The fluctuation term displacement data is then reconstructed using the phase space reconstruction model,and the fluctuation term displacement is predicted using the extreme learning machine.The predicted results of the trend terms and fluctuation terms are fused to obtain the predicted values of slope displacement components.Finally,the predicted values of each component are integrated to obtain the final prediction of slope displacement.The proposed algorithm is applied to radar-monitored slope displacement prediction and compared with various methods.The results demonstrate that the proposed method effectively improves prediction accuracy and increases the prediction lead time.(2)Advance landslide warning based on time series data anomaly detection and multi-model fusion: In landslide warning models,the use of slope displacement prediction data can lead to early warnings,thus enabling proactive measures for landslide prevention(research aspect 1).However,prediction data may contain errors,which can reduce the accuracy of the warnings.Moreover,radar monitoring data often exhibit strong fluctuations,making direct landslide warnings prone to high false alarm rates.To address these issues,a landslide warning model based on time series data anomaly detection and multi-model fusion is proposed.The model concatenates the original data with the prediction data and inputs the concatenated data into a local outlier factor detection algorithm and the Lyapunov rule for anomaly detection.The detected anomalous values are then used as the starting point for landslide warnings using the Saito model,the inverse velocity method,and the Verhulst gray model.Finally,the warning results from multiple models are fused using weighted voting to obtain the final landslide warning result.The proposed model is validated using various types of data,and the results show that it can effectively handle noise and errors in slope displacement data,thereby improving the timeliness and accuracy of landslide warnings.This thesis includes 40 figures,4 tables,and 82 references.