Lightweight intelligent landslide displacement prediction method

Abstract

The invention provides a lightweight intelligent landslide displacement prediction method, and belongs to the field of geological disaster monitoring and prediction, and the method comprises the steps: obtaining historical monitoring displacement time sequence data of landslide mass monitoring points; performing dynamic normalization processing on the displacement time sequence data; constructing a lightweight prediction model comprising a gating enhanced ordinary differential equation network and a dynamic residual error correction module; synchronously optimizing the basic prediction and residual error correction network through a joint training module, and adopting a composite loss function comprising dynamic threshold Huber loss and L2 regularization; and based on a rolling prediction mechanism, when new monitoring data arrives, a sliding window is automatically updated, and a prediction result is calculated. According to the invention, through collaborative optimization of the gating enhanced ordinary differential equation network and the dynamic residual correction, the prediction precision is improved while the light weight of the model is ensured, and real-time and reliable technical support is provided for early warning of landslide disasters and disaster prevention and reduction decisions.

Description

Technical Field

[0001] The present invention relates to the field of public safety prediction, and in particular to a lightweight intelligent landslide displacement prediction method. Background Art

[0002] Geological disasters that occur every year cause significant casualties and property losses. Landslide geological disasters are one of the most common types of geological disasters. Displacement monitoring and prediction are one of the most important bases for their prevention and control. It has very important practical significance for reducing casualties and economic losses caused by landslides.

[0003] However, traditional machine learning models such as linear regression and random forests cannot capture the temporal dynamics of displacement. While models such as Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and Transformer models improve temporal modeling capabilities, they suffer from large parameter counts and high model complexity. Furthermore, static normalization makes these models incapable of adapting to the time-varying statistical characteristics of landslide displacement, leading to reduced prediction accuracy, delayed response times, and difficulty deploying them on edge devices at monitoring stations. Therefore, establishing a lightweight and intelligent landslide displacement prediction method that both ensures prediction accuracy and is lightweight is of great practical significance for improving prediction accuracy and response speed, and increasing emergency response time after early warnings. Summary of the Invention

[0004] This paper proposes a lightweight, intelligent landslide displacement prediction method. First, the mean μ and standard deviation σ are dynamically calculated using a sliding window, followed by real-time normalization to address the distribution shift problem associated with traditional static normalization. Second, a prediction model is constructed, consisting of a gated enhanced ordinary differential equation (ODE) network and a dynamic residual correction method. The gated enhanced ODE network generates initial states using a GRU encoder. Basic predictions are achieved through feature-state fusion, gated activation, layer normalization, and ODE solution. The dynamic residual correction module learns the prediction error using a three-layer fully connected network. Finally, a composite loss function, including a dynamic Huber loss and an adaptive regularization term, is employed to jointly optimize model parameters. This method, through a sliding window rolling prediction mechanism, improves prediction accuracy and reduces the number of model parameters, meeting the real-time requirements of geological hazard monitoring.

[0005] The present invention is achieved in that: The technical solution for achieving the purpose of the present invention is: a lightweight intelligent landslide displacement prediction method, comprising the following steps: Step 1: Obtain the historical monitoring displacement time series data of the landslide monitoring point; Step 2: using a dynamic normalization method to normalize the historical monitoring displacement sequence to obtain a standardized displacement sequence; Step 3: Enhance the ODE network through gating to obtain the basic prediction value; Step 4: Obtain the prediction results for loss calculation through the dynamic residual correction method; Step 5: Using the dynamic composite loss function as the optimization objective, and based on the prediction results used for loss calculation, jointly optimize the trainable parameters in the gated enhanced ordinary differential equation network and the residual correction method through the backpropagation algorithm; The trainable parameters include: (1) The weight matrix and bias terms in the core of the gated augmented ordinary differential equation network; (2) The fully connected layer weights and bias terms of the dynamic residual correction module; Step 6: Repeat the calculation process from step 2 to step 5 until the preset number of iterations is reached, and finally obtain the prediction model including the gated enhanced ordinary differential equation network and dynamic residual correction; Step 7: Input the latest sliding window data into the trained prediction model, calculate the final displacement prediction value, and perform denormalization on the prediction result; Step 8: When new monitoring data arrives, remove the earliest data point and add the new data. Perform dynamic normalization on the monitoring displacement sequence in the latest window according to step 2 and then re-execute step 7.

[0006] Furthermore, the gated enhanced ordinary differential equation (ODE) network is implemented by the following steps: Step S31: Dimensionality reduction temporal coding: Process the normalized displacement sequence through a gated recurrent unit, where the hidden layer dimension of the gated recurrent unit is configured as D / 2, and extract the D / 2-dimensional hidden state of the final time step, and then expand the state to D dimensions through zero padding as the initial hidden state h0, where D is the base hidden layer dimension; Step S32: performing pooling downsampling with a kernel size of k and a step size of 1 on the normalized shift sequence to obtain a downsampled sequence of length L / k, where L is the length of the normalized shift sequence and k is an integer greater than 1; Step S33: The down-sampled sequence and the initial hidden state h0 are combined to obtain a basic prediction value through a dynamic gating state evolution process.

[0007] Furthermore, the dynamic gating state evolution process is achieved by sequentially executing the following steps: S34: Feature-state fusion, which is performed by concatenating the downsampled displacement sequence with the initial hidden state of the gated recurrent unit encoder along the feature dimension and mapping it to a high-dimensional space through a linear transformation to obtain a feature-state fusion output; S35: Gated activation, the method is as follows: first, the feature-state fusion output is equally split into two sub-vectors along the last feature dimension; second, a sigmoid function with an output range of (0, 1) is applied to the first sub-vector; third, the sigmoid activation result is element-wise multiplied by the second sub-vector to obtain a gated activation output; S36: Layer normalization: normalize the gated activation output to obtain layer normalized output; S37: Calculate the hidden state residual: add the layer normalized output to λ times the initial hidden state of the gated recurrent unit encoder to obtain the residual output, λ∈[0.05,0.2]; S38: Solve ordinary differential equations (ODEs): Numerically integrate the derivatives given by the residual output to obtain the ODE output; S39: Linear decoding: The output of the ordinary differential equation is mapped to the basic prediction value through a linear transformation.

[0008] Furthermore, in step 4, the steps of the dynamic residual correction method are as follows: Step S41: flattening the normalized displacement sequence into one dimension and concatenating it with the basic prediction value of the gated enhanced ordinary differential equation network to form a modified network input; Step S42: subtracting the basic prediction value of the gated enhanced ordinary differential equation network from the actual displacement value at the time to be predicted to generate a residual label as the correction network output; Step S43: The corrected network input is passed through three fully connected layers in sequence: input layer, hidden layer, and output layer. The number of units in the input layer is one-fourth of the baseline hidden layer dimension, and the number of units in the hidden layer is one-eighth of the baseline hidden layer dimension. The input layer and hidden layer use SiLU activation function, and the output layer is a linear layer. Finally, the error estimate is output. Step S44: Add the basic prediction value and the error estimate to obtain a prediction result for loss calculation.

[0009] Furthermore, in step 5, the dynamic composite loss function is defined as L total : L total =L Huber +β×L regu Where β is the dynamic regularization coefficient, which is set dynamically according to the sliding window standard deviation σ as follows: When σ≥1.5, β=0.001 When 0.5<σ<1.5, β= 0.007-0.004σ When σ ≤ 0.5, β = 0.005 L regu is the L2 norm square sum of all trainable parameters; L Huber The Huber loss function is dynamically adjusted and is defined as: when |y- |≤δ, L Huber =0.5×(y- )² When |y- When |>δ, L Huber =δ×(|y- |-0.5×δ) Where y represents the actual displacement monitoring value, represents the prediction result used for loss calculation, δ is the threshold, and its value is 1.5 times the standard deviation σ of the sliding window.

[0010] The beneficial effects of the present invention are as follows: the present invention provides a lightweight intelligent landslide displacement prediction method. Among them, the constructed dynamic normalization processing mechanism can retain the local statistical characteristics of recent data by calculating the mean μ and standard deviation σ of the sliding window in real time, effectively solving the distribution offset problem caused by traditional static normalization; the improved liquid neural ODE basic prediction module adopts gated enhancement design and fixed step-size solver, which significantly improves the long sequence prediction accuracy while reducing the number of parameters; the dynamic residual correction module reduces the prediction error through multi-layer perception units and gradient truncation technology; the joint training module and composite loss function realize the progressive collaborative optimization of basic prediction and error correction. Through the organic combination of the above technologies, reliable technical support is provided for reducing the number of model parameters and realizing real-time monitoring and early warning of landslide displacement. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments. It should be understood that the following drawings only illustrate certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other relevant drawings can be obtained based on these drawings without paying any creative work.

[0012] Figure 1 This is a flow chart of a lightweight intelligent landslide displacement prediction method provided by an embodiment of the present invention; Figure 2 is a flow chart of a gated enhanced ordinary differential equation network provided by an embodiment of the present invention; Figure 3 It is a comparison chart of the predicted results and the measured values ​​of the method of the present invention provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0013] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in combination with the implementation cases and drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the invention for which protection is sought, but merely represents selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0014] The method of the present invention is described below by taking a specific case of real-time prediction of landslide displacement as an example.

[0015] like Figure 1 , a lightweight intelligent landslide displacement prediction method, comprising the following steps: Step 1: Obtain the historical monitoring displacement time series data of the landslide monitoring point; Step 2: using a dynamic normalization method to normalize the historical monitoring displacement sequence to obtain a standardized displacement sequence; Step 3: Enhance the ODE network through gating to obtain the basic prediction value; Step 4: Obtain the prediction results for loss calculation through the dynamic residual correction method; Step 5: Using the dynamic composite loss function as the optimization objective, and based on the prediction results used for loss calculation, jointly optimize the trainable parameters in the gated enhanced ordinary differential equation network and the residual correction method through the backpropagation algorithm; The trainable parameters include: (1) The weight matrix and bias terms in the core of the gated augmented ordinary differential equation network; (2) The fully connected layer weights and bias terms of the dynamic residual correction module.

[0016] Step 6: Repeat the calculation process from step 2 to step 5 until the preset number of iterations is reached, and finally obtain the prediction model including the gated enhanced ordinary differential equation network and dynamic residual correction; Step 7: Input the latest sliding window data into the trained prediction model, calculate the final displacement prediction value, and perform denormalization on the prediction result; Step 8: When new monitoring data arrives, remove the earliest data point and add the new data. Perform dynamic normalization on the monitoring displacement sequence in the latest window according to step 2 and then re-execute step 7.

[0017] Furthermore, the gated enhanced ordinary differential equation (ODE) network is implemented by the following steps: Step S31: Dimensionality reduction temporal coding: Process the normalized displacement sequence through a gated recurrent unit, where the hidden layer dimension of the gated recurrent unit is configured as D / 2, and extract the D / 2-dimensional hidden state of the final time step, and then expand the state to D dimensions through zero padding as the initial hidden state h0, where D is the base hidden layer dimension; By setting the hidden layer dimension of GRU directly to D / 2, the number of parameters is significantly reduced, lightweight temporal coding is achieved, and the lower-dimensional GRU output can prevent overfitting of the initial state; Step S32: performing pooling downsampling with a kernel size of k and a step size of 1 on the normalized shift sequence to obtain a downsampled sequence of length L / k, where L is the length of the normalized shift sequence and k is an integer greater than 1; Step S33: The down-sampled sequence and the initial hidden state h0 are combined to obtain a basic prediction value through a dynamic gating state evolution process.

[0018] Furthermore, the dynamic gating state evolution process is achieved by sequentially executing the following steps: S34: Feature-state fusion, which is performed by concatenating the downsampled displacement sequence with the initial hidden state of the gated recurrent unit encoder along the feature dimension and mapping it to a high-dimensional space through a linear transformation to obtain a feature-state fusion output; S35: Gated activation, the method is as follows: first, the feature-state fusion output is equally split into two sub-vectors along the last feature dimension; second, a sigmoid function with an output range of (0, 1) is applied to the first sub-vector; third, the sigmoid activation result is element-wise multiplied by the second sub-vector to obtain a gated activation output; S36: Layer normalization: normalize the gated activation output to obtain layer normalized output; S37: Calculate the hidden state residual: add the layer normalized output to λ times the initial hidden state of the gated recurrent unit encoder to obtain the residual output, λ∈[0.05,0.2]; S38: Solve ordinary differential equations (ODEs): Numerically integrate the derivatives given by the residual output to obtain the ODE output; S39: Linear decoding: The output of the ordinary differential equation is mapped to the basic prediction value through a linear transformation.

[0019] Furthermore, in step 4, the steps of the dynamic residual correction method are as follows: Step S41: flattening the normalized displacement sequence into one dimension and concatenating it with the basic prediction value of the gated enhanced ordinary differential equation network to form a modified network input; Step S42: subtracting the basic prediction value of the gated enhanced ordinary differential equation network from the actual displacement value at the time to be predicted to generate a residual label as the correction network output; Step S43: The corrected network input is passed through three fully connected layers in sequence: input layer, hidden layer, and output layer. The number of units in the input layer is one-fourth of the baseline hidden layer dimension, and the number of units in the hidden layer is one-eighth of the baseline hidden layer dimension. The input layer and hidden layer use SiLU activation function, and the output layer is a linear layer. Finally, the error estimate is output. Step S44: Add the basic prediction value and the error estimate to obtain a prediction result for loss calculation.

[0020] Furthermore, in step 5, the dynamic composite loss function is defined as L total : L total =L Huber +β×L regu； Where β is the dynamic regularization coefficient, which is set dynamically according to the sliding window standard deviation σ as follows: When σ≥1.5, β=0.001; When 0.5<σ<1.5, β= 0.007-0.004σ; When σ ≤ 0.5, β = 0.005; L regu is the L2 norm square sum of all trainable parameters; L Huber is the dynamically adjusted Huber loss function, defined as: When |y- |≤δ, L Huber =0.5×(y- )² When |y- When |>δ, L Huber =δ×(|y- |-0.5×δ) Where y represents the actual displacement monitoring value, represents the prediction result used for loss calculation, δ is the threshold, and its value is 1.5 times the standard deviation σ of the sliding window.

[0021] The experimental data comes from the online monitoring data of a tailings dam. A total of 2568 sets of monitoring sample data are obtained, of which the first 2092 sets of data are used for training the model and the last 476 sets of data are used for prediction.

[0022] The specific implementation process is as follows: like Figure 1 , a lightweight intelligent landslide displacement prediction method, comprising the following steps: Step 1: Obtain the historical monitoring displacement time series data of the landslide monitoring point; Step 2: using a dynamic normalization method to normalize the historical monitoring displacement sequence to obtain a standardized displacement sequence; Among them, the window length is selected as 120, the sliding step is 1, and the normalization method is selected as: X norm =(X-μ) / (σ+10 -6 ) Among them, X is the historical monitoring displacement sequence, X norm is the normalized displacement sequence; Step 3: Enhance the ODE network through gating to obtain the basic prediction value; The baseline hidden layer dimension is 64, the downsampling kernel size is k=3, λ=0.1, the pooling method uses mean pooling, and the ordinary differential equation uses the fourth-order Runge-Kutta method; Step 4: Obtain the prediction results for loss calculation through the dynamic residual correction method; Step 5: Using the dynamic composite loss function as the optimization objective, and based on the prediction results used for loss calculation, jointly optimize the trainable parameters in the gated enhanced ordinary differential equation network and the residual correction method through the backpropagation algorithm; The trainable parameters include: (1) The weight matrix and bias terms in the core of the gated augmented ordinary differential equation network; (2) The fully connected layer weights and bias terms of the dynamic residual correction module.

[0023] Among them, the back propagation algorithm uses the Adam optimizer with a learning rate of 0.001; Step 6: Repeat the calculation process from step 2 to step 5 until the preset number of iterations is reached, and finally obtain the prediction model including the gated enhanced ordinary differential equation network and dynamic residual correction; Among them, the maximum number of iterations is set to 300; Step 7: Input the latest sliding window data into the trained prediction model, calculate the final displacement prediction value, and perform denormalization on the prediction result; Step 8: When new monitoring data arrives, remove the earliest data point and add the new data. After performing dynamic normalization processing on the monitoring displacement sequence in the latest window according to step 2, execute step 7 to achieve rolling prediction.

[0024] Figure 3The comparison between the prediction results of the present invention and the true values ​​is shown. It can be seen that the consistency between the prediction results and the true values ​​is very high, and the landslide displacement trend can be well predicted.

[0025] To further verify the advantages of the method of the present invention, an LSTM model, a GRU model and the method of the present invention were used for comparison, wherein the hidden layer dimensions and learning rates of the LSTM and GRU were consistent with the baseline hidden layer dimensions and learning rates of the method of the present invention.

[0026] In order to fully compare the performance of each model, the mean absolute error (MAE), absolute percentage error (MAPE), root mean square error (RMSE) and goodness of fit R 2 A total of four indicators were used to quantitatively evaluate each model. Each model was calculated ten times and the average value of each indicator was obtained. Table 1 shows the comparison of the evaluation results of each model.

[0027] Table 1 Comparison of evaluation indicators of various models

[0028] As shown in Table 1, the proposed method achieves the best performance across all evaluation metrics among the three methods, demonstrating its predictive accuracy. Furthermore, it also has the fewest parameters compared to the other models, making it more lightweight and suitable for edge device deployment. In summary, this case study demonstrates that the proposed method not only has high computational accuracy but also uses fewer parameters than popular deep learning methods such as LSTM and GRU. This method can effectively improve landslide displacement prediction efficiency and edge deployment capabilities, providing a basis for landslide safety management and the prevention of major accidents.

[0029] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention are intended to be within the scope of protection of the present invention.

Claims

1. A lightweight intelligent landslide displacement prediction method, characterized in that: The following steps are involved: Step 1: Obtain the historical monitoring displacement time series data of the landslide monitoring point; Step 2: using a dynamic normalization method to normalize the historical monitoring displacement sequence to obtain a standardized displacement sequence; Step 3: Enhance the ODE network through gating to obtain the basic prediction value; Step 4: Obtain the prediction results for loss calculation through the dynamic residual correction method; Step 5: Using the dynamic composite loss function as the optimization objective, and based on the prediction results used for loss calculation, jointly optimize the trainable parameters in the gated enhanced ordinary differential equation network and the residual correction method through the backpropagation algorithm; The trainable parameters include: (1) The weight matrix and bias terms in the core of the gated augmented ordinary differential equation network; (2) The fully connected layer weights and bias terms of the dynamic residual correction module; Step 6: Repeat the calculation process from step 2 to step 5 until the preset number of iterations is reached, and finally obtain the prediction model including the gated enhanced ordinary differential equation network and dynamic residual correction; Step 7: Input the latest sliding window data into the trained prediction model, calculate the final displacement prediction value, and perform denormalization on the prediction result; Step 8: When new monitoring data arrives, remove the earliest data point and add the new data. Perform dynamic normalization on the monitoring displacement sequence in the latest window according to step 2 and then re-execute step 7.

2. A lightweight intelligent landslide displacement prediction method according to claim 1, characterized in that: In step 2, the dynamic normalization method is implemented by the following steps: Step S21: Set a sliding window of fixed length and calculate the mean μ and standard deviation σ of the monitoring displacement sequence in the window in real time; Step S22: Based on the μ and σ parameters of the current window, the monitoring displacement sequence in the window is normalized to obtain a standardized displacement sequence; The sliding window is updated as new monitoring data arrives, and the normalization calculation only uses historical data within the window.

3. A lightweight intelligent landslide displacement prediction method according to claim 1, characterized in that: In step 3, the gated enhanced ordinary differential equation network is implemented by the following steps: Step S31: Dimensionality reduction temporal coding: Process the normalized displacement sequence through a gated recurrent unit, where the hidden layer dimension of the gated recurrent unit is configured as D / 2, and extract the D / 2-dimensional hidden state of the final time step, and then expand the state to D dimensions through zero padding as the initial hidden state h0, where D is the base hidden layer dimension; Step S32: performing pooling downsampling with a kernel size of k and a step size of 1 on the normalized shift sequence to obtain a downsampled sequence of length L / k, where L is the length of the normalized shift sequence and k is an integer greater than 1; Step S33: The down-sampled sequence and the initial hidden state h0 are combined to obtain a basic prediction value through a dynamic gating state evolution process.

4. A lightweight intelligent landslide displacement prediction method according to claim 3, characterized in that: The dynamic gating state evolution process is achieved by sequentially executing the following steps: S34: Feature-state fusion, which is performed by concatenating the downsampled displacement sequence with the initial hidden state of the gated recurrent unit encoder along the feature dimension and mapping it to a high-dimensional space through a linear transformation to obtain a feature-state fusion output; S35: Gated activation, the method is as follows: first, the feature-state fusion output is equally split into two sub-vectors along the last feature dimension; second, a sigmoid function with an output range of (0, 1) is applied to the first sub-vector; third, the sigmoid activation result is element-wise multiplied by the second sub-vector to obtain a gated activation output; S36: Layer normalization: normalize the gated activation output to obtain layer normalized output; S37: Calculate the hidden state residual: add the layer normalized output to λ times the initial hidden state of the gated recurrent unit encoder to obtain the residual output, λ∈[0.05,0.2]; S38: Solve ordinary differential equations: numerically integrate the derivatives given by the residual output to obtain the ordinary differential equation output; S39: Linear decoding: The output of the ordinary differential equation is mapped to the basic prediction value through a linear transformation.

5. A lightweight intelligent landslide displacement prediction method according to claim 1, characterized in that: In step 4, the steps of the dynamic residual correction method are as follows: Step S41: flattening the normalized displacement sequence into one dimension and concatenating it with the basic prediction value of the gated enhanced ordinary differential equation network to form a modified network input; Step S42: subtracting the basic prediction value of the gated enhanced ordinary differential equation network from the actual displacement value at the time to be predicted to generate a residual label as the correction network output; Step S43: The corrected network input is passed through three fully connected layers in sequence: input layer, hidden layer, and output layer. The number of units in the input layer is one-fourth of the baseline hidden layer dimension, and the number of units in the hidden layer is one-eighth of the baseline hidden layer dimension. The input layer and hidden layer use SiLU activation function, and the output layer is a linear layer. Finally, the error estimate is output. Step S44: Add the basic prediction value and the error estimate to obtain a prediction result for loss calculation.

6. A lightweight intelligent landslide displacement prediction method according to claim 1, characterized in that: In step 5, the dynamic composite loss function is defined as L total : L total =L Huber +β×L regu ; Where β is the dynamic regularization coefficient, which is set dynamically according to the sliding window standard deviation σ as follows: When σ≥1.5, β=0.001; When 0.5<σ<1.5, β=0.007-0.004σ; When σ ≤ 0.5, β = 0.005; L regu is the L2 norm square sum of all trainable parameters; L Huber is the dynamically adjusted Huber loss function, defined as: When |y- |≤δ, L Huber =0.5×(y- )²; when |y- When |>δ, L Huber =δ×(|y- |-0.5×δ); Where y represents the actual displacement monitoring value, represents the prediction result used for loss calculation, δ is the threshold, and its value is 1.5 times the standard deviation σ of the sliding window.

Images