Scholarlink

<p>Various deep learning models are employed to predict landslide displacement. However, existing deep learning prediction models do not take into account the rich multi-scale information of external triggering factors from multiple sources and the impact of each influencing factor on the degree of triggered landslide displacements. Therefore, this work applies the variational mode decomposition (VMD) theory to decompose the “step” displacement of a landslide into trend displacement, periodic displacement, and stochastic displacement. A polynomial function is used to predict the trend displacement, VMD is used to calculate the high-frequency and low-frequency components of periodic displacement and random displacement, and the main trigger factors are determined by the grey correlation degree. Then, a hybrid deep learning model based on multi-scale convolutional neural networks (MCNN) and squeeze-and-excitation networks (SENet) combined with the gated recurrent unit (GRU), called MCNN-SE-GRU, is proposed to predict the periodic displacement and random displacement. In the model, the MCNN block is designed to extract three convolutional kernels of different scales to form receptive fields of different sizes and obtain global and local trigger factor characteristics. The SE block builds dependencies between channels by learning global and local information. By dynamically adjusting the weights of the channels of each feature, important features are reinforced, and non-important features are suppressed. Then, the GRU block is used for feature extraction of dependencies between temporal data. Finally, the feature fusion block is used to stitch the multi-featured vectors and linear regression is used to calculate the final displacement prediction values. Taking the Baijiabao Landslide in the Three Gorges of China as an example, data from two monitoring points with the maximum and minimum displacement changes were selected to reflect the sensitivity of model predictions and a comparative analysis was conducted with four mainstream models, four latest prediction models, and three different combinations of models. The results show that the root-mean-square error (RMSE) of MCNN-SE-GRU was 2.50 mm&#xa0;and 2.33 mm, respectively. Compared with the mainstream models, the prediction performance of MCNN-SE-GRU was at least improved by 66.44 %&#xa0;and 76.05%. Compared with the latest models, the prediction performance of MCNN-SE-GRU was at least enhanced by 23.31% and 11%. It has been verified that our method can effectively suppress the random fluctuations of the prediction results during the creeping period of landslide displacement and accurately predict when a steplike deformation occurs, providing a more effective means of predicting the risk warning during the intense deformation period of landslides.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Landslide displacement prediction model based on variational mode decomposition and MCNN-SE-GRU

  • 期刊论文

  • 发表日期：

    2025年7月8日

查看全文
  • Yi Wang,
  • Yuchen Li,
  • Tianfeng Gu,
  • Jiading Wang

摘要

Various deep learning models are employed to predict landslide displacement. However, existing deep learning prediction models do not take into account the rich multi-scale information of external triggering factors from multiple sources and the impact of each influencing factor on the degree of triggered landslide displacements. Therefore, this work applies the variational mode decomposition (VMD) theory to decompose the “step” displacement of a landslide into trend displacement, periodic displacement, and stochastic displacement. A polynomial function is used to predict the trend displacement, VMD is used to calculate the high-frequency and low-frequency components of periodic displacement and random displacement, and the main trigger factors are determined by the grey correlation degree. Then, a hybrid deep learning model based on multi-scale convolutional neural networks (MCNN) and squeeze-and-excitation networks (SENet) combined with the gated recurrent unit (GRU), called MCNN-SE-GRU, is proposed to predict the periodic displacement and random displacement. In the model, the MCNN block is designed to extract three convolutional kernels of different scales to form receptive fields of different sizes and obtain global and local trigger factor characteristics. The SE block builds dependencies between channels by learning global and local information. By dynamically adjusting the weights of the channels of each feature, important features are reinforced, and non-important features are suppressed. Then, the GRU block is used for feature extraction of dependencies between temporal data. Finally, the feature fusion block is used to stitch the multi-featured vectors and linear regression is used to calculate the final displacement prediction values. Taking the Baijiabao Landslide in the Three Gorges of China as an example, data from two monitoring points with the maximum and minimum displacement changes were selected to reflect the sensitivity of model predictions and a comparative analysis was conducted with four mainstream models, four latest prediction models, and three different combinations of models. The results show that the root-mean-square error (RMSE) of MCNN-SE-GRU was 2.50 mm and 2.33 mm, respectively. Compared with the mainstream models, the prediction performance of MCNN-SE-GRU was at least improved by 66.44 % and 76.05%. Compared with the latest models, the prediction performance of MCNN-SE-GRU was at least enhanced by 23.31% and 11%. It has been verified that our method can effectively suppress the random fluctuations of the prediction results during the creeping period of landslide displacement and accurately predict when a steplike deformation occurs, providing a more effective means of predicting the risk warning during the intense deformation period of landslides.