A Multi-Time-Scale Displacement Prediction Method for Landslides Based on Temporal and Physical Dual-Flow Structures

By employing a temporal and physical dual-stream structure approach, combined with signal decomposition, convolutional networks, long short-term memory networks, and a common attention mechanism, the problems of single-scale and insufficient information fusion in landslide displacement prediction were solved, achieving high-precision and stable prediction of the entire landslide process.

CN121636922BActive Publication Date: 2026-11-10HUAZHONG UNIV OF SCI & TECH +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202511954707.0
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2025-12-23
Publication Date
2026-11-10
Estimated Expiration
2045-12-23

AI Technical Summary

Technical Problem

Existing landslide displacement prediction methods suffer from problems such as a single prediction scale, lack of physical mechanisms, and insufficient fusion of multimodal information. As a result, the models can reflect trends in the short term but cannot take into account medium and long-term trends, failing to meet the needs of real-time early warning and long-term planning, and lacking the ability to collaboratively analyze features at different time scales.

Method used

A method based on temporal and physical dual-stream structure is adopted. Multi-scale signal decomposition and noise reduction are performed by ensemble empirical mode decomposition. Global temporal dependency features are extracted by combining temporal convolutional network and bidirectional long short-term memory network. Physical feature embedding is performed by gradient boosting tree model and feature fusion is performed by co-attention mechanism. Finally, multi-time-scale landslide displacement prediction values ​​are generated in multiple parallel regressors.

Benefits of technology

It improves the accuracy and stability of landslide displacement prediction, realizes dynamic intelligent prediction of the entire landslide process, can quickly respond to displacement changes in the short term and maintain trend consistency in the medium and long term, and enhances the model's generalization ability and physical interpretability.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN121636922B_ABST
    Figure CN121636922B_ABST
Patent Text Reader

Abstract

This application belongs to the field of geological disaster monitoring technology, specifically disclosing a landslide multi-timescale displacement prediction method based on temporal and physical dual-flow structure. The method includes: preprocessing the original displacement sequence, and using ensemble empirical mode decomposition for multi-scale signal decomposition and denoising signal reconstruction to obtain a denoised displacement sequence; inputting the denoised displacement sequence into a prediction model to obtain global temporal dependency features; determining physical feature quantities based on the denoised displacement sequence, inputting the physical feature quantities into a gradient boosting tree model, performing nonlinear mapping and feature embedding to obtain physical constraint features; and using a common attention mechanism to perform bidirectional interaction and weighted fusion of the global temporal dependency features and physical constraint features, inputting the fused representation vector into multiple parallel regressors to obtain multi-timescale landslide displacement prediction values. This application can improve the model's generalization ability and prediction accuracy.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application belongs to the field of geological disaster monitoring technology, and more specifically, relates to a method for predicting landslide displacement at multiple time scales based on a time-series and physical dual-flow structure. Background Technology

[0002] Landslides, as a common and highly destructive geological hazard, seriously threaten the lives and property of people in mountainous areas, the operation of major projects, and regional economic development. Therefore, achieving accurate prediction of landslide displacement to provide a scientific basis for disaster early warning and risk management has become a core issue in the fields of geological engineering and disaster prevention and mitigation.

[0003] Existing landslide displacement prediction methods mainly rely on single time-series data-driven models, such as deep learning structures like LSTM and TCN, or empirical models based on statistical regularities. While these methods can reflect landslide displacement trends in the short term, they generally suffer from the following problems: First, the prediction scale is singular. Most models only model the short-term stage and cannot take into account the medium and long-term trend prediction. This results in an incomplete characterization of the risk evolution of the entire landslide process and an inability to take into account the immediate early warning needs such as personnel evacuation and equipment transfer in the short-term prediction as well as the pre-decision decision-making for prevention and control planning and resource allocation in the medium and long-term prediction. Second, multimodal physical information has not been effectively integrated. Existing models have not yet been able to establish a dynamic correlation between time series signals and multi-source physical parameters, and lack the ability to collaboratively analyze features at different time scales, thus limiting the accuracy and applicability of the prediction results.

[0004] Therefore, overcoming the technical shortcomings of existing landslide displacement prediction methods, such as single prediction scale, lack of physical mechanisms, and insufficient fusion of multimodal information, and improving the generalization ability and prediction accuracy of the models, is a technical problem that urgently needs to be solved. Summary of the Invention

[0005] In view of the shortcomings of the existing technology, the purpose of this application is to overcome the technical defects of existing landslide displacement prediction methods, such as single prediction scale, lack of physical mechanism and insufficient fusion of multimodal information, and improve the generalization ability and prediction accuracy of the model.

[0006] To achieve the above objectives, in a first aspect, this application provides a landslide multi-timescale displacement prediction method based on temporal and physical dual-flow structures, comprising: The system receives the original displacement sequence obtained from monitoring, performs data preprocessing on the original displacement sequence, and uses ensemble empirical mode decomposition to perform multi-scale signal decomposition and noise reduction signal reconstruction to obtain the noise-reduced displacement sequence. Based on the denoised displacement sequence, physical feature quantities are determined, including velocity, acceleration, and improved tangential angle. The denoised displacement sequence is input into the prediction model to obtain global temporal dependency features; the prediction model is based on a concatenated temporal convolutional network and a bidirectional long short-term memory network. The physical feature quantities are input into the gradient boosting tree model, and nonlinear mapping and feature embedding are performed to obtain the physical constraint features; By utilizing the co-attention mechanism, the global temporal dependency features and the physical constraint features are bidirectionally interacted and weighted to generate a unified fusion representation vector. The fused representation vector is input into multiple parallel regressors to obtain landslide displacement predictions at multiple time scales.

[0007] Optionally, the method for obtaining the denoised displacement sequence includes: Multiple independent white noise sequences are superimposed on the original displacement sequence; Empirical mode decomposition is performed on the sequences after each superposition to obtain multiple sets of eigenmode function sequences; Averaging all the eigenmode function sequences of the same order obtained from the decomposition yields stable average eigenmode function components; The denoised displacement sequence is reconstructed based on the average intrinsic mode function components and residual terms.

[0008] Optionally, the method for obtaining the physical characteristic quantity includes: The first-order and second-order differences are performed on the denoised displacement sequence to obtain the velocity and acceleration, respectively. The cumulative displacement curve is normalized by the displacement rate during the constant velocity phase to obtain the normalized curve. Based on the target time, the time before the target time, and the corresponding normalized curves, the slope angle of the tangent line of the normalized curve at each monitoring time point is calculated and used as the improved tangent angle.

[0009] Optionally, the method for obtaining the global temporal dependency features includes: The local fluctuation features of the denoised displacement sequence are captured using a temporal convolutional network; the temporal convolutional network is established using a one-dimensional causal dilated convolutional structure with residual connections. A bidirectional long short-term memory network is used to model the forward and backward global dependencies of the output of the temporal convolutional network, resulting in a concatenated latent vector. The hidden states are weighted and pooled using a temporal attention mechanism to obtain the global temporal dependency features.

[0010] Optionally, the method for obtaining the physical constraint features includes: The physical feature quantities are input into the pre-trained gradient boosting tree model to obtain the physical constraint features; The training method for the gradient boosting tree model includes: The sum of the m regression trees for each physical feature is used as the predicted output value, and the target displacement is used as the true label value. The squared error loss is determined based on the predicted output value and the true label value. The squared error loss is then combined with the summation result of the tree structure regularization term to construct the overall loss function of the gradient boosting tree model. A second-order Taylor expansion is used to approximate the overall loss function. The sample residuals are fitted and the tree model parameters are updated step by step to obtain a pre-trained gradient boosting tree model.

[0011] Optionally, the method for obtaining the fused representation vector includes: The time-trainable parameter matrix corresponding to the global time-series dependency features is determined, and the physical trainable parameter matrix corresponding to the physical constraint features is determined; the time-series trainable parameter matrix includes a time-series query matrix, a time-series key matrix, and a time-series value matrix; the physical trainable parameter matrix includes a physical query matrix, a physical key matrix, and a physical value matrix. The attention distribution of the global temporal dependency features on the physical constraint features is calculated based on the temporal query matrix and the temporal key matrix, and the physical constraint features are interactively fused based on the attention distribution and the temporal value matrix to obtain the first interactive feature; The anti-attention response of the physical constraint feature to the global temporal dependency feature is calculated based on the physical query matrix and the physical key matrix. The temporal constraint feature is then interactively fused based on the anti-attention response and the physical value matrix to obtain the second interactive feature. The first interaction feature, the second interaction feature, the original global temporal dependency feature, and the physical constraint feature are pooled, concatenated, and nonlinearly mapped to obtain the fused representation vector.

[0012] Optionally, the method for obtaining the predicted landslide displacement includes: Determine multiple time scales, including weekly, monthly, and quarterly scales; The fused representation vector is input into multiple regressors, and linear transformation is performed according to the weight parameters and bias terms corresponding to each time scale to obtain the predicted output values ​​corresponding to multiple time scales. Based on the predicted output value and the true label value corresponding to each time scale, the loss weight, regularization coefficient and the set of trainable parameters for each time scale, the joint loss function is determined. An adaptive optimization algorithm is used to optimize the joint loss function, and the learning rate and early stopping strategy are dynamically adjusted according to the validation set error to obtain the trained two-stream network model. The landslide displacement prediction values ​​at multiple time scales were obtained using the trained dual-stream network model.

[0013] Secondly, this application provides a landslide multi-timescale displacement prediction system based on temporal and physical dual-flow structure, comprising: The input preprocessing module is used to receive the original displacement sequence obtained from monitoring, perform data preprocessing on the original displacement sequence, and use ensemble empirical mode decomposition to perform multi-scale signal decomposition and noise reduction signal reconstruction to obtain the noise-reduced displacement sequence. The temporal flow modeling module is used to input the denoised displacement sequence into the prediction model to obtain global temporal dependency features; the prediction model is based on a temporal convolutional network and a bidirectional long short-term memory network connected in series. The physical flow modeling module determines physical feature quantities based on the denoised displacement sequence, inputs the physical feature quantities into the gradient boosting tree model, performs nonlinear mapping and feature embedding, and obtains physical constraint features; the physical feature quantities include velocity, acceleration and improved tangential angle; The dual-stream fusion module is used to perform bidirectional interaction and weighted fusion of the global temporal dependency features and the physical constraint features using a common attention mechanism to generate a unified fusion representation vector. The multi-timescale prediction module is used to input the fused representation vector into multiple parallel regressors to obtain landslide displacement prediction values ​​at multiple timescales.

[0014] Thirdly, this application provides an electronic device, comprising: at least one memory for storing a program; and at least one processor for executing the program stored in the memory, wherein when the program stored in the memory is executed, the processor is configured to execute the method described in the first aspect or any possible implementation thereof.

[0015] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when run on a processor, causes the processor to perform the method described in the first aspect or any possible implementation thereof.

[0016] Fifthly, this application provides a computer program product that, when run on a processor, causes the processor to perform the method described in the first aspect or any possible implementation thereof.

[0017] It is understood that the beneficial effects of the second to fifth aspects mentioned above can be found in the relevant descriptions in the first aspect mentioned above, and will not be repeated here.

[0018] Overall, the technical solutions conceived in this application have the following beneficial effects compared with the prior art: This application explicitly embeds the dynamic relationship between velocity, acceleration, and improved tangential angle into the physical flow aspect of the model, enabling the deep learning system to possess interpretability and robustness of physical constraints, as well as the ability to learn and fit medium- to long-term trends. The temporal flow aspect achieves collaborative modeling of short-term features through a cascaded structure of temporal convolutional networks and bidirectional long short-term memory networks. A common attention fusion mechanism further strengthens the correlation between the two, forming an intelligent prediction framework adaptable to the entire landslide prevention cycle. The method provided in this application effectively improves the accuracy, stability, and generalization ability of the model in multi-timescale landslide displacement prediction, solving the problems of single prediction scale and lack of mechanism in traditional pure data-driven models under complex geological environments, and realizing dynamic intelligent prediction of the entire landslide process. Attached Figure Description

[0019] Figure 1 This is one of the flowcharts illustrating the landslide multi-timescale displacement prediction method based on temporal and physical dual-flow structure provided in this application embodiment; Figure 2 This is a flowchart of the EEMD decomposition and reconstruction algorithm according to an embodiment of this application; Figure 3 This is a schematic diagram of the landslide multi-timescale displacement prediction system based on temporal and physical dual-flow structure provided in the embodiments of this application; Figure 4 This is the second flowchart of the landslide multi-timescale displacement prediction method based on temporal and physical dual-flow structure provided in the embodiments of this application; Figure 5 This is a flowchart of the time-series TCN-BiLSTM model algorithm according to an embodiment of this application; Figure 6 This is a comparison chart of the prediction results of the single-stream model and the dual-stream model in this application embodiment; Figure 7(a) shows the training results on a 1-day timescale; Figure 7(b) shows the training results on a 7-day timescale; Figure 7(c) shows the training results on a 30-day timescale; Figure 7(d) shows the training results on a 90-day timescale. Figure 8(a) shows the test results of the new dual-flow model on a 1-day timescale; Figure 8(b) shows the test results of the new dual-flow model on a 7-day timescale; Figure 8(c) shows the test results of the new dual-flow model on a 30-day timescale; Figure 8(d) shows the test results of the new dual-flow model on a 90-day timescale. Figure 9 This is a multi-model comparison diagram at a 1-day timescale according to an embodiment of this application; Figure 10 This is a 7-day multi-model comparison chart on a weekly scale according to an embodiment of this application; Figure 11 This is a multi-model comparison chart on a monthly scale over 30 days, based on an embodiment of this application. Figure 12 This is a multi-model comparison chart for a 90-day quarterly scale according to an embodiment of this application; Figure 13 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0021] In this article, the term "and / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. The symbol " / " in this article indicates that the related objects are in an "or" relationship; for example, A / B means A or B.

[0022] The terms "first" and "second," etc., used in the specification and claims herein are used to distinguish different objects, not to describe a specific order of objects. For example, "first response message" and "second response message," etc., are used to distinguish different response messages, not to describe a specific order of response messages.

[0023] In the embodiments of this application, the terms "exemplary" or "for example" are used to indicate that something is an example, illustration, or description. Any embodiment or design that is described as "exemplary" or "for example" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design. Specifically, the use of the terms "exemplary" or "for example" is intended to present the relevant concepts in a specific manner.

[0024] In the description of the embodiments of this application, unless otherwise stated, "multiple" means two or more, for example, multiple processing units means two or more processing units, multiple elements means two or more elements, etc.

[0025] The embodiments of this application are described below with reference to the accompanying drawings.

[0026] Reference Figure 1 This application provides a landslide multi-timescale displacement prediction method based on temporal and physical dual-flow structure, including: S101. Receive the original displacement sequence obtained from monitoring, perform data preprocessing on the original displacement sequence, and use ensemble empirical mode decomposition to perform multi-scale signal decomposition and noise reduction signal reconstruction to obtain the noise-reduced displacement sequence. S102. Input the denoised displacement sequence into the prediction model to obtain global temporal dependency features; the prediction model is based on a concatenated temporal convolutional network and a bidirectional long short-term memory network; S103. Determine physical feature quantities based on the denoised displacement sequence; input the physical feature quantities into the gradient boosting tree model, perform nonlinear mapping and feature embedding, and obtain physical constraint features; the physical feature quantities include velocity, acceleration, and improved tangential angle; S104 utilizes a co-attention mechanism to perform bidirectional interaction and weighted fusion of the global temporal dependency features and the physical constraint features, generating a unified fusion representation vector; S105. Input the fused representation vector into multiple parallel regressors to obtain landslide displacement prediction values ​​at multiple time scales.

[0027] Specifically, in step S101, the raw data obtained from monitoring is first subjected to strict preprocessing, including outlier removal, missing value filling and smoothing, etc. At the same time, the sequence is smoothed using methods such as moving average to suppress some high-frequency random noise.

[0028] Based on preprocessing, this embodiment employs the Ensemble Empirical Mode Decomposition (EEMD) method for in-depth analysis and reconstruction of the displacement sequence. EEMD effectively overcomes the mode aliasing problem of traditional EMD methods by repeatedly superimposing white noise with different phases into the original signal and performing ensemble averaging on the decomposition results each time. This embodiment removes the high-frequency IMF components that mainly represent noise and reconstructs the remaining low-frequency IMFs and residual terms, ultimately obtaining a denoised displacement sequence that clearly reflects the main movement trend of the landslide while effectively suppressing random disturbances, providing a stable and reliable input for subsequent modeling.

[0029] Secondly, in step S102, the prediction model used in this step is a cascaded two-stream structure model, which consists of two parts: a temporal stream and a physical stream. The temporal stream takes the displacement sequence after EEMD denoising as input and uses a cascaded structure of a Temporal Convolutional Network (TCN) and a Bidirectional Long Short-Term Memory Network (BiLSTM) to extract time-dependent features. The TCN module adopts a one-dimensional causal dilated convolutional structure, which achieves multi-scale receptive field expansion through progressively increasing dilation rate to capture the local fluctuation features of landslide displacement.

[0030] Subsequently, these local features are fed into a bidirectional long short-term memory network (BiLSTM). BiLSTM scans the entire sequence in both forward and backward directions, and its unique gating mechanism enables it to learn and memorize temporal dependencies across long distances, such as the persistence of landslide acceleration phases and the lag effect of rainfall.

[0031] Finally, the hidden states output by BiLSTM are aggregated through an attention-weighted pooling layer that adaptively emphasizes time-step information that is more important to the current prediction task, thereby generating a comprehensive global temporal dependency feature that contains both local details and integrates global long-term dependencies.

[0032] Further, in step S103, key physical parameters that can directly characterize the kinematic state of the landslide are extracted from the processed displacement data, including velocity, acceleration, and improved tangential angle. Then, a vector composed of these three physical quantities (velocity, acceleration, and improved tangential angle) is input into the Gradient Boosting Tree (XGBoost) model. The XGBoost model is used to achieve nonlinear mapping and physical feature embedding, ultimately outputting physical constraint features.

[0033] In step S105, the global temporal dependency feature matrix output from the temporal stream and the physical constraint feature matrix output from the physical stream are simultaneously input into a common attention module. This module first generates corresponding query (key) and value vectors for the two types of features through learnable linear transformations, thus constructing a bidirectional interaction channel.

[0034] The original temporal features, physical features, and two features enhanced by cross-attention are concatenated, pooled, and nonlinearly transformed to generate a unified and fused representation vector that can comprehensively and evenly reflect the current state and evolution of the landslide.

[0035] Finally, the fused representation vector generated in the previous step is fed into a set of parallel linear regressors via S105. Each regressor is independently responsible for predicting a specific future time scale (e.g., 1 day, 3 days, 7 days, 30 days, 90 days), performing a linear transformation operation to output the displacement prediction value at that scale.

[0036] Optionally, the method for obtaining the denoised displacement sequence includes: Multiple independent white noise sequences are superimposed on the original displacement sequence; Empirical mode decomposition is performed on the sequences after each superposition to obtain multiple sets of eigenmode function sequences; Averaging all the eigenmode function sequences of the same order obtained from the decomposition yields stable average eigenmode function components; The denoised displacement sequence is reconstructed based on the average intrinsic mode function components and residual terms.

[0037] The details are as follows, please refer to... Figure 2 , Figure 2 This is a flowchart of the EEMD decomposition and reconstruction algorithm according to an embodiment of this application. First, the original displacement sequence obtained by monitoring is processed... After outlier removal, missing value imputation, and smoothing, ensemble empirical mode decomposition (EEMD) is used to achieve multi-scale signal decomposition. EEMD introduces the concept of white noise ensembles into traditional empirical mode decomposition (EMD). The specific process is as follows: in the original sequence... Several independent Gaussian white noise sequences are superimposed on the middle. Perform EMD decomposition on each of them to obtain the first... Sequence of intrinsic mode functions (IMFs) from the next decomposition Then, the average of the results is taken to obtain a stable IMF component:

[0038] in, For the average intrinsic mode function components, denoted as the number of sets of intrinsic mode function sequences.

[0039] Based on this, the denoised signal is reconstructed to obtain the denoised displacement sequence:

[0040] in, For the residual term, This is the denoised displacement sequence. After EEMD processing, high-frequency disturbances in the signal are effectively suppressed, while low-frequency trend terms are preserved, providing a stable input for deep network modeling.

[0041] Optionally, the method for obtaining the physical characteristic quantity includes: The first-order and second-order differences are performed on the denoised displacement sequence to obtain the velocity and acceleration, respectively. The cumulative displacement curve is normalized by the displacement rate during the constant velocity phase to obtain the normalized curve. Based on the target time, the time before the target time, and the corresponding normalized curves, the slope angle of the tangent line of the normalized curve at each monitoring time point is calculated and used as the improved tangent angle.

[0042] Specifically, based on the denoised displacement sequence described above, three physical characteristic quantities derived from the displacement sequence are calculated, namely velocity. acceleration and improved tangential angle Velocity and acceleration are obtained through first-order and second-order differences, respectively:

[0043] To more accurately reflect the evolution characteristics of a landslide from deceleration to acceleration, an improved tangential angle is introduced. Displacement-time curve displacement rate in constant velocity phase Normalization is performed based on the baseline to construct the cumulative displacement curve for the constant velocity phase. The improved tangential angle is defined as the curve at the monitoring time. The slope angle of the tangent line, that is:

[0044] in, For the current moment The slope angle of the tangent line, For the target time, The previous moment before the target moment. The normalized curve at the target time. This is the normalized curve of the previous time step before the target time step.

[0045] These three types of physical quantities constitute the physical flow input feature matrix, which is used to characterize the dynamic evolution of landslide displacement.

[0046] Optionally, the method for obtaining the global temporal dependency features includes: The local fluctuation features of the denoised displacement sequence are captured using a temporal convolutional network; the temporal convolutional network is established using a one-dimensional causal dilated convolutional structure with residual connections. A bidirectional long short-term memory network is used to model the forward and backward global dependencies of the output of the temporal convolutional network, resulting in a concatenated latent vector. The hidden states are weighted and pooled using a temporal attention mechanism to obtain the global temporal dependency features.

[0047] Specifically, the dual-flow structure model in this embodiment consists of two parts: a temporal flow and a physical flow. The temporal flow takes the displacement sequence after EEMD denoising as input and uses a cascaded structure of a Temporal Convolutional Network (TCN) and a Bidirectional Long Short-Term Memory Network (BiLSTM) to extract time-dependent features. The TCN module uses a one-dimensional causal dilated convolutional structure, which expands the receptive field at multiple scales through progressively increasing the dilation rate to capture the local fluctuation features of landslide displacement. Its calculation expression is as follows:

[0048] in, Indicates the void ratio causal convolution, For batch normalization, This is a non-linear activation function. This structure can achieve stable information propagation in deep networks through multiple layers of residual connections while maintaining temporal causality. Subsequently, the BiLSTM module performs global dependency modeling on the TCN output, with its forward and backward states as follows:

[0049]

[0050] The concatenation yields the latent vector. .

[0051] Global temporal representation is computed using attention-weighted pooling:

[0052]

[0053] Timing stream output It combines short-term volatility characteristics with long-term dependence.

[0054] Optionally, the method for obtaining the physical constraint features includes: The physical feature quantities are input into the pre-trained gradient boosting tree model to obtain the physical constraint features; The training method for the gradient boosting tree model includes: The sum of the m regression trees for each physical feature is used as the predicted output value, and the target displacement is used as the true label value. The squared error loss is determined based on the predicted output value and the true label value. The squared error loss is then combined with the summation result of the tree structure regularization term to construct the overall loss function of the gradient boosting tree model. A second-order Taylor expansion is used to approximate the overall loss function. The sample residuals are fitted and the tree model parameters are updated step by step to obtain a pre-trained gradient boosting tree model.

[0055] Specifically, in this embodiment, the physical flow is represented by three types of physical quantities: velocity. acceleration and improved tangential angle As input, the Gradient Boosting Tree (XGBoost) model is used to implement nonlinear mapping and physical feature embedding. The XGBoost model is built based on the additive tree ensemble principle, and its expression is:

[0056] in, Predict the output value. For the input vector, Indicates the first A tree of return, This is a set of CART decision trees. The model is optimized by minimizing the following objective function:

[0057] in, The actual label value. For the squared error loss, This is a tree structure regularization term used to prevent overfitting. In each iteration, XGBoost uses a second-order Taylor expansion to approximate the overall loss function, fitting the sample residuals and progressively updating the tree model parameters. The model ultimately outputs a physical constraint vector. Each dimension corresponds to the nonlinear mapping result of the physical evolution characteristics of the landslide body in the latent space.

[0058] In this physical flow, velocity characteristics Describes the instantaneous trend of displacement change, acceleration characteristics Characterizing the dynamic response during the acceleration phase, while improving the tangential angle This reflects local deviations in the displacement growth rate and potential instability. After being combined using a tree structure in XGBoost, these three elements form a physical representation with strong interpretability and nonlinear coupling, effectively revealing the evolution mechanism of landslides from stable to accelerating stages. The vector output by the physical flow... Features of time-series stream output They will jointly participate in modeling during the subsequent integration phase.

[0059] Optionally, the method for obtaining the fused representation vector includes: The time-trainable parameter matrix corresponding to the global time-series dependency features is determined, and the physical trainable parameter matrix corresponding to the physical constraint features is determined; the time-series trainable parameter matrix includes a time-series query matrix, a time-series key matrix, and a time-series value matrix; the physical trainable parameter matrix includes a physical query matrix, a physical key matrix, and a physical value matrix. The attention distribution of the global temporal dependency features on the physical constraint features is calculated based on the temporal query matrix and the temporal key matrix, and the physical constraint features are interactively fused based on the attention distribution and the temporal value matrix to obtain the first interactive feature; The anti-attention response of the physical constraint feature to the global temporal dependency feature is calculated based on the physical query matrix and the physical key matrix. The temporal constraint feature is then interactively fused based on the anti-attention response and the physical value matrix to obtain the second interactive feature. The first interaction feature, the second interaction feature, the original global temporal dependency feature, and the physical constraint feature are pooled, concatenated, and nonlinearly mapped to obtain the fused representation vector.

[0060] Specifically, in the dual-stream fusion stage, to achieve information exchange between the temporal and physical streams, a feature interaction structure based on a co-attention mechanism is adopted. This mechanism generates bidirectional attention weights by calculating the similarity matrix between temporal and physical features, thereby achieving dynamic fusion in the feature space. Let the temporal stream output matrix be... The physical flow output matrix is ,in, For time steps, For time-series feature dimensions, For physical characteristics dimension.

[0061] Common attention is calculated as follows:

[0062]

[0063] in, For a trainable parameter matrix, These are the time-series query matrix, the time-series key matrix, and the time-series value matrix, respectively. The physical query matrix, physical key matrix, and physical value matrix are... It is a scaling factor, usually the square root of the feature dimension, used to stabilize the gradient. This represents the attention distribution of temporal features on physical features. The enhanced feature obtained by focusing on the physical flow of the temporal stream is the first interaction feature mentioned above. This represents the inverse attention response of physical features to temporal features. The enhanced features obtained by focusing on the temporal flow of the physical flow are the second interactive features mentioned above.

[0064] Through the common attention mechanism, the model can adaptively enhance the response of displacement signals to changes in physical quantities during the acceleration phase, so that the model can maintain adaptive attention to the dominant factors under different geological stages (stable-transition-acceleration).

[0065] The interacting features are pooled and concatenated to obtain a fused representation vector:

[0066] in, The nonlinear mapping function includes: fully connected layers, activation, and normalized combination. This fusion vector... It also preserves temporal dependence and physical constraint information, serving as a unified input for multi-timescale prediction.

[0067] Optionally, the method for obtaining the predicted landslide displacement includes: Determine multiple time scales, including short-term time scales, medium-term time scales, long-term time scales, and ultra-long-term time scales; The fused representation vector is input into multiple regressors, and linear transformation is performed according to the weight parameters and bias terms corresponding to each time scale to obtain the predicted output values ​​corresponding to multiple time scales. Based on the predicted output value and the true label value corresponding to each time scale, the loss weight, regularization coefficient and the set of trainable parameters for each time scale, the joint loss function is determined. An adaptive optimization algorithm is used to optimize the joint loss function, and the learning rate and early stopping strategy are dynamically adjusted according to the validation set error to obtain the trained two-stream network model. The landslide displacement prediction values ​​at multiple time scales were obtained using the trained dual-stream network model.

[0068] Specifically, in this embodiment, the multi-timescale prediction model consists of multiple parallel regression heads, corresponding to prediction tasks of 1, 3, 7, 30 and 90 days, respectively, corresponding to the weekly scale (1, 3, 7), monthly scale (30) and quarterly scale (90).

[0069] Each prediction head is structured as a single-layer fully connected linear regressor:

[0070] The model uses a joint loss function for multi-scale optimization:

[0071] in, For the loss weights at each time scale, The regularization coefficient is . The model uses a complete set of trainable parameters. The training process employs the Adam adaptive optimization algorithm, dynamically adjusting the learning rate and early stopping strategy based on the validation set error. Once trained, the model can simultaneously output displacement predictions across multiple time scales in a single forward propagation.

[0072] Reference Figure 3 This application provides a landslide multi-timescale displacement prediction system based on temporal and physical dual-flow structure, comprising: The input preprocessing module 310 is used to receive the original displacement sequence obtained by monitoring, perform data preprocessing on the original displacement sequence, and use ensemble empirical mode decomposition to perform multi-scale signal decomposition and noise reduction signal reconstruction to obtain the noise-reduced displacement sequence. The temporal flow modeling module 320 is used to input the denoised displacement sequence into the prediction model to obtain global temporal dependency features; the prediction model is based on a temporal convolutional network and a bidirectional long short-term memory network connected in series. The physical flow modeling module 330 determines physical feature quantities based on the denoised displacement sequence, inputs the physical feature quantities into the gradient boosting tree model, performs nonlinear mapping and feature embedding, and obtains physical constraint features; the physical feature quantities include velocity, acceleration and improved tangential angle; The dual-stream fusion module 340 is used to perform bidirectional interaction and weighted fusion of the global temporal dependency features and the physical constraint features using a common attention mechanism to generate a unified fusion representation vector. The multi-timescale prediction module 350 is used to input the fused representation vector into multiple parallel regressors to obtain landslide displacement prediction values ​​at multiple timescales.

[0073] Reference Figure 4 , Figure 4 This is a complete flowchart of an embodiment of this application, including the following steps: The system receives the original sequence of monitored displacement data.

[0074] The raw data enters the preprocessing module and undergoes three key processes: Remove outliers, fill in missing values, and smooth the data.

[0075] The preprocessed data is then fed into a signal decomposition module for noise reduction. The denoised signal is then split into two independent analysis channels for further processing. Physical feature analysis: Physical feature quantities are calculated from the denoised signal, mainly including velocity, acceleration and improved tangential angle.

[0076] These physical features are combined into a feature matrix, which is then input into a machine learning model to generate analysis results of the physical flow.

[0077] Time series trend analysis: The denoised signal is directly input into the time series modeling module, which consists of two advanced neural networks to learn the deep patterns of data changes over time and generate analysis results of the data stream.

[0078] Dual-stream fusion: Deeply integrates the analysis results from two sets of data streams, namely physical stream and data stream. Through feature interaction, co-attention mechanism, pooling, feature concatenation, and non-linear mapping; Multi-scale prediction: The fused vector is input into the multi-timescale prediction module. This module contains multiple parallel regression heads, each unit specifically responsible for predicting the displacement after a specific time length.

[0079] Reference Figure 5 , Figure 5 This is a flowchart of the temporal flow TCN-BiLSTM model algorithm of this application.

[0080] The following detailed description of the solution in this application is provided with specific examples: Using a landslide in a certain area as a case study, the temporal-physical dual-flow structure model proposed in this application is validated. The landslide slopes from north to south, has a long, tongue-shaped planform, a main sliding direction of 207°, a length of approximately 340m, a width of 45–81m, an average thickness of approximately 12.0m, and a landslide area of ​​approximately 2.26 × 10⁻⁶ m². 4 m², volume approximately 20.34 × 10 4 The landslide, with a depth of m³, is classified as a medium-sized sedimentary landslide. The elevation difference between the front and rear edges is approximately 131 m, with an overall slope of about 21°. The landslide mass is primarily composed of fragmented rock and soil, with the slip zone consisting of gravelly soil interbedded with silty clay (approximately 0.2 m thick). The underlying slip bed is sericite-quartz schist from the Doushantuo Formation (Z2d) of the Sinian System. Under prolonged rainfall infiltration, the landslide mass has softened significantly, exhibiting a typical "creep-acceleration-restabilization" staged deformation process. The landslide is currently in a creeping state, with deformation concentrated at the central and front edges and the free face. Cracks are well-developed at the front edge, with significant tension and displacement; the central terraced zone is a concentrated area of ​​arc-shaped cracks, which gradually extend laterally. To achieve continuous monitoring, five GNSS displacement monitoring points (G1–G5) were deployed along the main profile of the landslide mass, spaced approximately 50–80 m apart, with a monitoring cycle of one day and an observation period of approximately 30 months. An automatic rain gauge was installed in the landslide front platform area to synchronously record daily rainfall. In this embodiment, data G2–G5 are used as the training set, and G1 is used as the test set and model generalization validation point. G1 is located in the leading edge main sliding zone and is the most displacement-sensitive point, which can reflect the response characteristics of the overall acceleration stage of the landslide.

[0081] The test results, conducted on a 3-day timescale, compared the predictions of G1 monitoring points by a two-stream structure model incorporating physical information and a single-stream model relying solely on deep learning for time series prediction. Figure 6 As shown: Error metrics for the EEMD+TCN+BiLSTM (with physical information embedding) test set: 1. Root Mean Squared Error (RMSE): 4.1061; 2. Mean Absolute Error (MAE): 2.9963; 3. Mean Relative Percentage Error (MAPE): 1.9568%; 4. Coefficient of Determination (R²) 2 ): 0.99733.

[0082] Error metrics for test sets without physical data: 1. Root Mean Squared Error (RMSE): 13.8; 2. Mean Absolute Error (MAE): 12.8088; 3. Mean Relative Percentage Error (MAPE): 6.8493%; 4. Coefficient of Determination (R²)2 ): 0.96989.

[0083] Four time scales (1 day, 7 days, 30 days, and 90 days) were selected for model training, and the following models were compared: EEMD+TCN+BiLSTM (dual-stream XGBoost physical information), EEMD+TCN+BiLSTM (single-stream), EEMD+TCN (dual-stream), EEMD+BiLSTM (dual-stream), and TCN+BiLSTM (dual-stream). The training results for the 1, 7, 30, and 90-day time scales are shown in Figure 7, and the test results of the new dual-stream model for the 1, 7, 30, and 90-day time scales are shown in Figure 8.

[0084] The error indices for multiple time scales are shown in the table below: Table 1 Error Indicators for Multiple Time Scales

[0085] As can be seen, the time-series and physical dual-flow structure model of this application exhibits stability across short-term and medium-to-long-term timescales, demonstrating high prediction accuracy and strong generalization ability. In the short-term phase (1 day), the model responds sensitively to abrupt displacement changes, with an RMSE of only 2.35 and an R² of 0.999. In the medium-to-long-term phases (7, 30, and 90 days), the model maintains good consistency in predicting displacement evolution trends, with RMSEs of 4.83, 3.31, and 3.69, respectively, and R² exceeding 0.996 for all three phases. Therefore, this model can function stably across different timescales, achieving unified prediction of landslide displacement from short-term early warning to medium-to-long-term trend assessment.

[0086] Reference Figures 9-12 The following is a comparison of error metrics for multiple models at different time scales: Referring to Table 2, which shows a comparison of multiple models over one day: Table 2. Multi-model comparison over 1 day

[0087] Referring to Table 3, which shows the comparison of multiple models over 7 days: Table 3. Multi-model comparison over 7 days

[0088] Refer to Table 4, which shows a 30-day comparison of multiple models: Table 4. Multi-model comparison over 30 days

[0089] Referring to Table 5, which shows a 90-day comparison of multiple models: Table 5. Multi-model comparison over 90 days

[0090] The comparison of the four time scale indicators shows that the single-flow model without embedded physical information degrades significantly with increasing prediction step size: the error accumulates and expands over time, drifting and oscillations occur in the long term, stability is insufficient, and R² drops to around 0.991. Traditional TCN or BiLSTM single models are sensitive to non-stationary signals. Although they can partially fit fluctuations in the short term, they show trend deviations or delayed responses in medium- and long-term predictions.

[0091] This application introduces three physical constraints—velocity, acceleration, and improved tangential angle—into the time-series data and employs a common attention mechanism to dynamically fuse temporal and physical characteristics. This enables the model to quickly capture displacement abrupt changes in the short term and maintain trend consistency in the medium to long term, effectively suppressing error accumulation and drift. Overall, the model performs stably in both the short and medium-to-long term, improving prediction accuracy and significantly enhancing its generalization and physical interpretability, achieving highly stable, multi-timescale prediction of the entire landslide process.

[0092] Reference Figure 13 Based on the methods in the above embodiments, this application provides an electronic device that may include: a processor 121, a communications interface 122, a memory 123, and a communication bus 124. The processor 121, communications interface 122, and memory 123 communicate with each other via the communication bus 124. The processor 121 can call logical instructions stored in the memory 123 to execute the methods in the above embodiments.

[0093] Furthermore, the logical instructions in the aforementioned memory 123 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application.

[0094] Based on the methods in the above embodiments, this application provides a computer-readable storage medium storing a computer program that, when run on a processor, causes the processor to execute the methods in the above embodiments.

[0095] Based on the methods in the above embodiments, this application provides a computer program product that, when run on a processor, causes the processor to execute the methods in the above embodiments.

[0096] It is understood that the processor in the embodiments of this application can be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. A general-purpose processor can be a microprocessor or any conventional processor.

[0097] The method steps in this application embodiment can be implemented in hardware or by a processor executing software instructions. The software instructions can consist of corresponding software modules, which can be stored in random access memory (RAM), flash memory, read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), registers, hard disks, portable hard disks, CD-ROMs, or any other form of storage medium known in the art. An exemplary storage medium is coupled to the processor, enabling the processor to read information from and write information to the storage medium. Of course, the storage medium can also be a component of the processor. The processor and the storage medium can reside in an ASIC.

[0098] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially as a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted through the computer-readable storage medium. The computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state disk (SSD)).

[0099] It is understood that the various numerical designations used in the embodiments of this application are merely for the convenience of description and are not intended to limit the scope of the embodiments of this application.

[0100] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A landslide multi-timescale displacement prediction method based on temporal and physical dual-flow structure, characterized in that, include: The system receives the original displacement sequence obtained from monitoring, performs data preprocessing on the original displacement sequence, and uses ensemble empirical mode decomposition to perform multi-scale signal decomposition and noise reduction signal reconstruction to obtain the noise-reduced displacement sequence. The denoised displacement sequence is input into the prediction model to obtain global temporal dependency features; The prediction model is based on a concatenation of a temporal convolutional network and a bidirectional long short-term memory network. Based on the denoised displacement sequence, physical feature quantities are determined, and these physical feature quantities are input into a gradient boosting tree model for nonlinear mapping and feature embedding to obtain physical constraint features; the physical feature quantities include velocity, acceleration, and improved tangential angle. By utilizing the co-attention mechanism, the global temporal dependency features and the physical constraint features are bidirectionally interacted and weighted to generate a unified fusion representation vector. The fused representation vector is input into multiple parallel regressors to obtain landslide displacement predictions at multiple time scales.

2. The landslide multi-timescale displacement prediction method based on temporal and physical dual-flow structure according to claim 1, characterized in that, The method for obtaining the denoised displacement sequence includes: Multiple independent white noise sequences are superimposed on the original displacement sequence; Empirical mode decomposition is performed on the sequences after each superposition to obtain multiple sets of eigenmode function sequences; Averaging all the eigenmode function sequences of the same order obtained from the decomposition yields stable average eigenmode function components; The denoised displacement sequence is reconstructed based on the average intrinsic mode function components and residual terms.

3. The landslide multi-timescale displacement prediction method based on temporal and physical dual-flow structure according to claim 1, characterized in that, The method for obtaining the physical characteristic quantities includes: The first-order and second-order differences are performed on the denoised displacement sequence to obtain the velocity and acceleration, respectively. The cumulative displacement curve is normalized by the displacement rate during the constant velocity phase to obtain the normalized curve. Based on the target time, the time before the target time, and the corresponding normalized curves, the slope angle of the tangent line of the normalized curve at each monitoring time point is calculated and used as the improved tangent angle.

4. The landslide multi-timescale displacement prediction method based on temporal and physical dual-flow structure according to claim 1, characterized in that, The method for obtaining the global temporal dependency features includes: The local fluctuation features of the denoised displacement sequence are captured using a temporal convolutional network; the temporal convolutional network is established using a one-dimensional causal dilated convolutional structure with residual connections. A bidirectional long short-term memory network is used to model the forward and backward global dependencies of the output of the temporal convolutional network, resulting in a concatenated latent vector. The hidden states are weighted and pooled using a temporal attention mechanism to obtain the global temporal dependency features.

5. The landslide multi-timescale displacement prediction method based on temporal and physical dual-flow structure according to claim 1, characterized in that, The method for obtaining the physical constraint features includes: The physical feature quantities are input into the pre-trained gradient boosting tree model to obtain the physical constraint features; The training method for the gradient boosting tree model includes: The sum of the m regression trees for each physical feature is used as the predicted output value, and the target displacement is used as the true label value. The squared error loss is determined based on the predicted output value and the true label value. The squared error loss is then combined with the summation result of the tree structure regularization term to construct the overall loss function of the gradient boosting tree model. A second-order Taylor expansion is used to approximate the overall loss function. The sample residuals are fitted and the tree model parameters are updated step by step to obtain a pre-trained gradient boosting tree model.

6. The landslide multi-timescale displacement prediction method based on temporal and physical dual-flow structure according to claim 1, characterized in that, The method for obtaining the fusion representation vector includes: The time-trainable parameter matrix corresponding to the global time-series dependency features is determined, and the physical trainable parameter matrix corresponding to the physical constraint features is determined; the time-series trainable parameter matrix includes a time-series query matrix, a time-series key matrix, and a time-series value matrix; the physical trainable parameter matrix includes a physical query matrix, a physical key matrix, and a physical value matrix. The attention distribution of the global temporal dependency features on the physical constraint features is calculated based on the temporal query matrix and the temporal key matrix, and the physical constraint features are interactively fused based on the attention distribution and the temporal value matrix to obtain the first interactive feature; The anti-attention response of the physical constraint feature to the global temporal dependency feature is calculated based on the physical query matrix and the physical key matrix. The temporal constraint feature is then interactively fused based on the anti-attention response and the physical value matrix to obtain the second interactive feature. The first interaction feature, the second interaction feature, the original global temporal dependency feature, and the physical constraint feature are pooled, concatenated, and nonlinearly mapped to obtain the fused representation vector.

7. The landslide multi-timescale displacement prediction method based on temporal and physical dual-flow structure according to claim 1, characterized in that, The method for obtaining the predicted landslide displacement values ​​includes: Determine multiple time scales, including weekly, monthly, and quarterly scales; The fused representation vector is input into multiple regressors, and linear transformation is performed according to the weight parameters and bias terms corresponding to each time scale to obtain the predicted output values ​​corresponding to multiple time scales. Based on the predicted output value and the true label value corresponding to each time scale, the loss weight, regularization coefficient and the set of trainable parameters for each time scale, the joint loss function is determined. An adaptive optimization algorithm is used to optimize the joint loss function, and the learning rate and early stopping strategy are dynamically adjusted according to the validation set error to obtain the trained two-stream network model. The landslide displacement prediction values ​​at multiple time scales were obtained using the trained dual-stream network model.

8. A landslide multi-timescale displacement prediction system based on temporal and physical dual-flow structure, characterized in that, include: The input preprocessing module is used to receive the original displacement sequence obtained from monitoring, perform data preprocessing on the original displacement sequence, and use ensemble empirical mode decomposition to perform multi-scale signal decomposition and noise reduction signal reconstruction to obtain the noise-reduced displacement sequence. The temporal flow modeling module is used to input the denoised displacement sequence into the prediction model to obtain global temporal dependency features; The prediction model is based on a concatenation of a temporal convolutional network and a bidirectional long short-term memory network. The physical flow modeling module is used to determine physical feature quantities based on the denoised displacement sequence, input the physical feature quantities into the gradient boosting tree model, perform nonlinear mapping and feature embedding, and obtain physical constraint features; the physical feature quantities include velocity, acceleration and improved tangential angle; The dual-stream fusion module is used to perform bidirectional interaction and weighted fusion of the global temporal dependency features and the physical constraint features using a common attention mechanism to generate a unified fusion representation vector. The multi-timescale prediction module is used to input the fused representation vector into multiple parallel regressors to obtain landslide displacement prediction values ​​at multiple timescales.

9. An electronic device, characterized in that, include: At least one memory for storing computer programs; At least one processor is configured to execute a program stored in the memory, wherein when the program stored in the memory is executed, the processor is configured to perform the method as described in any one of claims 1-7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is run on the processor, it causes the processor to perform the method as described in any one of claims 1-7.