一种结合时序分解与相似分量重组的深度学习滑坡位移组合预测模型

A Deep Learning Landslide Displacement Prediction Model Combining Time Series Decomposition and Similar Component Reorganization

ES评分 0

DOI 10.14075/j.jgg.2024.04.186
刊名
Journal of Geodesy and Geodynamics
年,卷(期) 2025, 45(3)
作者
作者单位

1 长安大学地质工程与测绘学院, 西安市雁塔路126号, 710054;
2 西部矿产资源与地质工程教育部重点实验室, 西安市雁塔路126号, 710054

摘要
在对滑坡监测数据粗差进行有效处理及充分顾及滑坡监测数据自身特性的基础上,提出一种结合时序分解与相似分量重组的深度学习滑坡位移组合预测模型。首先,利用孤立森林法对滑坡时序监测数据的显著粗差进行处理,再对其平稳性、自相关性、正态性进行综合分析,确定模型预测中输入特征序列的最佳长度;其次,利用集合经验模态分解(EEMD)方法,将非稳态滑坡监测数据分解为多个平稳时间序列,再结合样本熵与K-means算法将其划分为高频、中频、低频3类时间分量;最后,通过对比不同神经网络模型的预测精度,分别构建适合于3类时间分量的预测模型,再将预测结果相叠加,实现对滑坡位移的高精度预测。实验区典型滑坡体北斗/GNSS监测数据测试表明,本文组合预测模型对含有显著粗差的滑坡监测数据具有较好的适用性,相较于单一及现有组合模型可显著提高滑坡位移预测精度。
Abstract
On the basis of effectively handling the gross errors of landslide monitoring data and fully considering the characteristics of landslide monitoring data, we develop a deep learning landslide displacement prediction model combining time series decomposition and similar component reorganization. First, we deal with the significant gross errors of landslide time series monitoring data using the isolation forest algorithm, and then comprehensively analyze its smoothness, autocorrelation, and normality to determine the optimal length of input feature sequence. Second, the non-stationary landslide monitoring data are decomposed into multiple smooth time series using the ensemble empirical mode decomposition(EEMD) method, which is then classified into three types combining the sample entropy and K-means algorithm, namely high, medium, and low frequency. Finally, comparing the prediction accuracy of different neural networks, the prediction models suitable for three types of time components are constructed respectively, and then the prediction results are superimposed to realize the high-precision prediction of landslide displacement. The testing results of Beidou/GNSS monitoring data of typical landslide body in experimental area show that the combination prediction model proposed in this paper has a better applicability to the landslide monitoring data containing significant gross errors, and can significantly improve the prediction accuracy of landslide displacement compared with single and existing combination models.
关键词
滑坡位移预测; 集合经验模态分解; 样本熵; 深度神经网络; 时间卷积网络
KeyWord
landslide displacement prediction; ensemble empirical mode decomposition(EEMD); sample entropy; deep neural network; temporal convolutional network
基金项目
页码 221-230
  • 参考文献
  • 相关文献
  • 引用本文

瞿伟, 李达, 李久元, 边子策. 一种结合时序分解与相似分量重组的深度学习滑坡位移组合预测模型 [J]. 大地测量与地球动力学. 2025; 45; (3). 221 - 230.

  • 文献评论

相关学者

相关机构