A landslide catastrophe trend prediction method based on machine learning

Abstract

The application provides a landslide disaster trend prediction method based on machine learning, and relates to the field of trend prediction. The method comprises the following steps: constructing a landslide disaster data set, data preprocessing, proposing a landslide disaster trend prediction model, training and testing. The landslide disaster trend prediction model is composed of a landslide disaster space feature extraction module, a landslide disaster multi-level time sequence modeling module, a time sequence adaptive attention mechanism and a trend-mutation collaborative modeling and prediction module. The space feature extraction module adopts an adaptive convolution kernel mechanism, a larger convolution kernel is used for long-time dependent features, and a smaller convolution kernel is used for short-time dependent features; the multi-level time sequence modeling module fuses Bi-LSTM and Bi-GRU, and dynamically adjusts the output feature proportion through a gating mechanism; the time sequence adaptive attention mechanism calculates adaptive attention weights; and the trend-mutation collaborative modeling module combines trend prediction and mutation detection, and captures chronic risk accumulation and sudden landslide signals.

Description

Technical Field

[0001] This invention belongs to the field of trend prediction, specifically relating to a method for predicting landslide disaster trends based on machine learning. Background Technology

[0002] Landslides are a common geological hazard, characterized by their suddenness, destructiveness, and wide-ranging impact, seriously threatening people's lives and property and the regional ecological environment. Their occurrence is usually closely related to multiple factors such as rainfall, earthquakes, geological structures, soil and rock properties, and human engineering activities, exhibiting complex triggering mechanisms and highly nonlinear evolution processes. With the increasing frequency of extreme weather events and the intensification of development in mountainous areas, the frequency and intensity of landslide disasters have risen significantly, placing higher demands on the effective prediction and trend analysis of landslide disasters.

[0003] Traditional methods rely heavily on geological surveys, remote sensing, and hazard assessments, but suffer from limitations such as insufficient data fusion, poor real-time performance, and limited accuracy. To improve prediction accuracy, researchers have recently explored the application of deep learning and big data analytics in landslide prediction. For example, monitoring systems based on ground-based synthetic aperture radar (S-SAR) technology can monitor slope deformation 24 / 7, over a wide area, and with high precision, overcoming the shortcomings of traditional methods. Furthermore, researchers have proposed methods based on geological hazard risk assessment, predicting potential disaster risks by analyzing geological environmental conditions and engineering activities. However, these methods still face challenges in data processing and model fusion, requiring further research and improvement.

[0004] This invention proposes a machine learning-based method for predicting landslide disaster trends. It ensures data comprehensiveness by constructing a multi-source heterogeneous dataset containing geological and geomorphological data, meteorological and hydrological data, engineering activities, historical landslide disaster information, and real-time sensor monitoring data. In the data preprocessing stage, data cleaning methods are used to handle missing values, outliers, and noisy data, improving data quality. The proposed landslide disaster trend prediction model combines spatial feature extraction, temporal modeling, a temporal adaptive attention mechanism, and a trend-abrupt change co-modeling prediction module to effectively capture the spatial characteristics and temporal variation patterns of landslide disasters. During training, hyperparameters such as learning rate, batch size, optimizer, regularization, gradient pruning, and the number of training epochs are appropriately set to ensure model stability and generalization ability. Validation on the test set ensures the model has high-precision predictive capabilities. This method has advantages in data fusion, feature extraction, and modeling strategies, significantly improving the accuracy and reliability of landslide disaster trend prediction. Summary of the Invention

[0005] The main objective of this invention is to provide a landslide disaster trend prediction method based on machine learning. The method aims to construct a landslide disaster trend prediction model, which consists of a landslide disaster spatial feature extraction module, a landslide disaster multi-level temporal modeling module, a temporal adaptive attention mechanism, and a trend-change collaborative modeling prediction module. Specifically, the landslide disaster spatial feature extraction module introduces an adaptive convolutional kernel size mechanism; for features that are dependent on landslide disasters over a long period, the convolutional kernel size is increased, while for features that are dependent on landslide disasters over a short period, the convolutional kernel size is decreased. The landslide disaster multi-level temporal modeling module... The sequential modeling module introduces a hybrid architecture of Bi-LSTM and Bi-GRU, and introduces a gating mechanism to dynamically determine the fusion ratio of the output features of Bi-LSTM and Bi-GRU according to different time steps; the temporal adaptive attention mechanism introduces a similarity metric parameter to compare the similarity between the hidden states of the current time step t and the historical time step t-1, thereby obtaining the temporal adaptive attention weight for each time step t; the trend-mutation co-modeling and prediction module constructs two branches: trend prediction and mutation detection, which respectively capture the landslide disaster trend of chronic risk accumulation and the signal of sudden landslide disaster.

[0006] To achieve the above objectives, the technical solution of the present invention is: a method for predicting landslide disaster trends based on machine learning, the method comprising:

[0007] S1. Construct a multi-source heterogeneous landslide disaster dataset, the features of which include: geological and geomorphological features, meteorological and hydrological features, engineering activities, historical landslide disaster information and real-time sensor monitoring data;

[0008] S2. Preprocess the landslide disaster dataset. Clean the landslide disaster dataset and divide the cleaned dataset into training and testing sets in a 7:3 ratio.

[0009] S3. Propose a landslide disaster trend prediction model, including the following methods:

[0010] S31. Construct a spatial feature extraction module for landslide disasters. By introducing an adaptive convolution kernel size mechanism, a larger convolution kernel is used for long-term dependent features in landslide disasters, and a smaller convolution kernel is used for short-term dependent features.

[0011] S32. Construct a multi-level temporal modeling module for landslide disasters, integrate Bi-LSTM and Bi-GRU architectures, and introduce a gating mechanism to dynamically adjust the fusion ratio of Bi-LSTM and Bi-GRU outputs;

[0012] S33. Construct a temporal adaptive attention mechanism, introduce a similarity metric to compare the similarity between the hidden states of the current time step and the historical time steps, and dynamically calculate the temporal adaptive attention weights.

[0013] S34. Construct a trend-mutation collaborative modeling and prediction module, and construct two branches of loss: trend prediction and mutation detection, to capture the chronic accumulation risk and sudden signals of landslide disasters, respectively.

[0014] S4. Train the landslide disaster trend prediction model by using the training set to train the landslide disaster trend prediction model.

[0015] S5. Test the landslide disaster trend prediction model. Input the test set into the landslide disaster trend prediction model and output the landslide disaster probability value to achieve landslide disaster trend prediction.

[0016] Furthermore, in step S31, an adjustable hyperparameter is introduced into the convolutional layer. This hyperparameter dynamically adjusts the size of the convolutional kernel during training. For features that are dependent on landslide disasters over a long period, the convolutional kernel can be enlarged to capture a wider range of spatial features; for features that are dependent on landslide disasters over a short period, the convolutional kernel is reduced to extract more refined local features of landslide disasters. The mathematical model for the dynamic convolution operation is as follows:

[0017]

[0018] In the formula, For dynamically adjustable convolutional kernel size, X t For input data, Z t This represents the convolution result at time t;

[0019] To ensure the appropriateness of the convolution kernel size, a loss function is introduced to constrain the change in the convolution kernel size, avoiding excessive expansion or contraction. The mathematical model of the loss function is as follows:

[0020]

[0021] In the formula, k0 is the default initial size of the convolution kernel, k t Let L be the size of the convolution kernel at time t, λ be the regularization coefficient, and L be the value of the kernel. size To limit the deviation of the convolution kernel size, L... size The size of k is constantly adjusted t The size of the convolution kernel is adjusted until the most suitable kernel size for the data features of the current task is found.

[0022] The convolution results of different sizes are concatenated along the channel direction. The concatenated feature map contains multi-scale information and can preserve feature representations under different receptive fields. The mathematical model for the concatenation operation is as follows:

[0023] Z fused =Concat(Z1, Z2, ..., Z T );

[0024] In the formula, Concat(·) is the concatenation operation, and Zfused The feature map is the fused feature map, which contains features from different convolutional kernels.

[0025] Furthermore, in step S32, firstly, the input landslide disaster time series data is processed through a Bi-LSTM layer. A forward LSTM is used to process past dependencies, and a backward LSTM is used to process future dependencies, capturing long-term dependencies and trends. The mathematical model of Bi-LSTM is:

[0026]

[0027] In the formula, Z fused (t,:) represents the input data, which comes from the output of the landslide disaster spatial feature extraction module. This represents the output of the Bi-LSTM at time t.

[0028] Subsequently, the features output by the Bi-LSTM are fed into the Bi-GRU layer as input. The Bi-GRU captures short-term time dependencies, enabling the model to quickly respond to instantaneous changes in the time series. The mathematical model of Bi-GRU is as follows:

[0029]

[0030] In the formula, This represents the output of the Bi-LSTM at time t.

[0031] A gating mechanism is introduced to finely control the fusion method of Bi-LSTM and Bi-GRU outputs. The fusion ratio of the Bi-LSTM and Bi-GRU output features is dynamically determined according to different time steps. The model can select an appropriate time series model for modeling based on the data characteristics of the current time step. The gating mechanism outputs a value g through a sigmoid activation function. t g t The mathematical model is as follows:

[0032]

[0033] In the formula, This represents the positive output of the Bi-LSTM at time t. b is the positive output of the Bi-GRU at time t. g As a paranoid term, W g The weights are then used to weight the inverse outputs of Bi-LSTM and Bi-GRU. The mathematical model for this fusion is as follows:

[0034]

[0035] In the formula, This represents the inverse output of the Bi-LSTM at time t. h represents the inverse output of the Bi-GRU at time t. t The result is the fusion of the outputs of Bi-LSTM and Bi-GRU.

[0036] Furthermore, in step S33, the similarity metric parameter calculates the similarity between the hidden states of the current time step t and the historical time step t-1 by comparing them. This application uses a learnable weight matrix W. t The mathematical model for calculating similarity measurement parameters is as follows:

[0037]

[0038] In the formula, h t h is the feature vector at time step t, which is the output of the landslide disaster multi-level time series modeling module. t-1 Let W be the eigenvector at time step t-1. h W t This is a learnable weight matrix, applied to the features of the current time step and the previous time step, respectively. The weight matrix W t The transpose operation, sim(h) t h t-1 Let be the similarity measure between time step t and time step t-1;

[0039] After calculating the relevance at each time step based on the similarity metric parameters, the temporal adaptive attention weights are calculated using the Softmax function. These weights control the importance of features at each time step in the weighted fusion. The Softmax function transforms the similarity measure into a probabilistic form, normalizing the attention weights at each time step across all time steps, resulting in a sum of 1. This is the temporally adaptive attention weight. The mathematical model is as follows:

[0040]

[0041] In the formula, Let be the temporal adaptive attention weights at time step t. To normalize the similarity measure across all time steps, ensure that the sum of all weights is 1;

[0042] Based on the temporal adaptive attention weights, the hidden states at all time steps are weighted and summed to obtain the fused feature representation. The mathematical model of the fused feature representation is as follows:

[0043]

[0044] In the formula, z t This is the final weighted fusion feature representation. h represents the temporal adaptive attention weights at time step i. i Let be the hidden state at time step i.

[0045] Furthermore, in step S34, loss functions for two branches, trend prediction and mutation detection, are introduced. The trend prediction loss penalizes the difference between the predicted value and the true trend value, enabling the model to identify and fit slow but dangerous evolutionary signals. The mutation detection loss is used to capture sudden signals before a landslide occurs. The method involves using the output z of the time-adaptive attention mechanism... t Two prediction branches are fed in, and the mathematical models for the two branches are as follows:

[0046]

[0047] In the formula, MLP trend MLP burst These are two MLP neural network layers, The trend-based landslide risk value of the MLP prediction layer at time step t. To determine the mutation risk probability of the MLP prediction layer at time t for mutation detection, the magnitude of the loss function is then calculated. The mathematical model for the trend prediction loss function is as follows:

[0048]

[0049] In the formula, T is the total time step, and L is the total time step. trend To optimize the regression loss of the trend prediction branch, Let be the true trend value at time step t; the mathematical model for the mutation detection loss function is:

[0050]

[0051] In the formula, L burst To optimize the binary cross-entropy loss for the mutation detection branch, Let t be the landslide mutation marker at time step t, where 0 indicates no mutation and 1 indicates the presence of a mutation;

[0052] Finally, the two prediction results are merged into a single landslide disaster risk output. The fusion mathematical model is as follows:

[0053]

[0054] In the formula, γ is the weighting parameter. The final landslide disaster risk prediction value is obtained from the trend-mutation co-modeling prediction module.

[0055] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0056] This invention proposes a landslide disaster trend prediction model, which consists of a landslide disaster spatial feature extraction module, a landslide disaster multi-level temporal modeling module, a temporal adaptive attention mechanism, and a trend-abrupt change collaborative modeling prediction module. Specifically, the landslide disaster spatial feature extraction module introduces an adaptive convolutional kernel size mechanism; for features that are dependent on landslide disasters over a long period, the convolutional kernel size is increased, while for features that are dependent on landslide disasters over a short period, the convolutional kernel size is decreased. The landslide disaster multi-level temporal modeling module introduces a hybrid architecture of Bi-LSTM and Bi-GRU, and incorporates a gating mechanism to dynamically determine the fusion ratio of the output features of Bi-LSTM and Bi-GRU according to different time steps. The sequential adaptive attention mechanism introduces a similarity metric parameter to compare the similarity between the hidden states at the current time step t and the historical time step t-1, thereby obtaining the temporal adaptive attention weight for each time step t. The trend-mutation co-modeling prediction module constructs two branches: trend prediction and mutation detection, which respectively capture the landslide disaster trend of chronic risk accumulation and the signal of sudden landslide disaster. Compared with traditional landslide prediction methods, the landslide disaster trend prediction model proposed in this invention integrates multi-source heterogeneous data, adaptive convolution kernels, a hybrid architecture of Bi-LSTM and Bi-GRU, and trend-mutation co-modeling, which can more comprehensively capture the complex temporal characteristics of landslide disasters and improve prediction accuracy, stability, and real-time response capability. Attached Figure Description

[0057] Figure 1 This is a flowchart illustrating the steps of a machine learning-based method for predicting landslide disaster trends.

[0058] Figure 2 This is a flowchart illustrating the steps involved in a landslide disaster trend prediction model.

[0059] Figure 3 Internal structure diagram of Bi-LSTM in the multi-level time series modeling module for landslide disasters.

[0060] Figure 4 Internal structure diagram of Bi-GRU in the multi-level time series modeling module for landslide disasters.

[0061] Figure 5 This is a comparison between the predicted values ​​and the actual values ​​of the landslide disaster trend prediction model. Detailed Implementation

[0062] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0063] Please see Figures 1-5 The present invention provides a technical solution: a landslide disaster trend prediction method based on machine learning. The method includes the steps of constructing a multi-source heterogeneous landslide disaster dataset, preprocessing the landslide disaster dataset, proposing a landslide disaster trend prediction model, training the landslide disaster trend prediction model, and testing the landslide disaster trend prediction model.

[0064] Please refer to Figure 1 As shown in the figure, a landslide disaster trend prediction method based on machine learning in this application embodiment has the following specific steps:

[0065] S1. Construct a multi-source heterogeneous landslide disaster dataset, the features of which include: geological and geomorphological features, meteorological and hydrological features, engineering activities, historical landslide disaster information and real-time sensor monitoring data;

[0066] Furthermore, in step S1, the geological and geomorphological characteristics include slope, aspect, landform type, lithological distribution, geological structure, fault distribution, soil type, shear strength, and void ratio; the meteorological and hydrological characteristics include hourly and daily rainfall, cumulative rainfall, air temperature, relative humidity, wind speed, surface runoff, groundwater level, and evaporation; the engineering activity characteristics include the distribution of roads, buildings, tunnels, and mining areas, changes in land use types, and the layout of drainage facilities and the distribution of retaining structures; historical landslide disaster information includes landslide location, occurrence time, landslide volume, sliding direction, triggering factors, and sliding frequency; real-time monitoring data includes surface and deep displacement data, tilt angle, crack width, groundwater pore pressure, soil moisture content, rainfall intensity, and cumulative rainfall; this multi-source heterogeneous landslide disaster dataset contains a total of 7,000 landslide disaster sample points.

[0067] S2. Preprocess the landslide disaster dataset. Clean the landslide disaster dataset and divide the cleaned dataset into training and test sets in a 7:3 ratio.

[0068] Furthermore, in step S2, the landslide disaster dataset is first cleaned by using mean imputation to fill missing values ​​and outliers to be identified and removed through statistical analysis and visualization. Secondly, data standardization and normalization are performed to ensure that all features have the same scale and avoid the influence between different units. Z-score standardization is used to transform continuous variables, and one-hot encoding is used to convert categorical variables into numerical features. Simultaneously, time-series data is stabilized by using differencing to eliminate trend and periodic effects, ensuring data stability. For the integration of multi-source heterogeneous landslide disaster datasets, data fusion technology is used to unify data from different sources (geology, meteorology, hydrology, engineering activities) into a unified data format and time scale to ensure comparability and consistency between features. Finally, feature selection is performed on the landslide disaster dataset, using correlation analysis and principal component analysis to remove redundant features, reduce computational complexity, and ensure that the final model can efficiently process and learn key information. Finally, the cleaned dataset is divided into training and testing sets in a 7:3 ratio.

[0069] S3. A landslide disaster trend prediction model is proposed. Specific methods include: constructing a landslide disaster spatial feature extraction module, constructing a landslide disaster multi-level temporal series modeling module, constructing a temporal adaptive attention mechanism, and constructing a trend-abrupt change collaborative modeling and prediction module. The flowchart of these steps is shown below. Figure 2 As shown.

[0070] S31. Construct a spatial feature extraction module for landslide disasters. The spatial feature extraction module for landslide disasters introduces an adaptive convolution kernel size mechanism. For features that landslide disasters depend on for a long time, the convolution kernel size is increased, and for features that landslide disasters depend on for a short time, the convolution kernel size is decreased.

[0071] Furthermore, in step S31, an adjustable hyperparameter is introduced into the convolutional layer. This hyperparameter dynamically adjusts the size of the convolutional kernel during training. For features that are dependent on landslide disasters over a long period, the convolutional kernel can be enlarged to capture a wider range of spatial features; for features that are dependent on landslide disasters over a short period, the convolutional kernel is reduced to extract more refined local features of landslide disasters. The mathematical model for the dynamic convolution operation is as follows:

[0072]

[0073] In the formula, For dynamically adjustable convolutional kernel size, the initial value is set to 3, X t For input data, Z t This represents the convolution result at time t;

[0074] To ensure the appropriateness of the convolution kernel size, a loss function is introduced to constrain the change in the convolution kernel size, avoiding excessive expansion or contraction. The mathematical model of the loss function is as follows:

[0075]

[0076] In the formula, k0 is the default initial size of the convolution kernel, which is set to 5. t Let L be the size of the convolution kernel at time t, initialized to 3, and λ be the regularization coefficient. size To limit the deviation of the convolution kernel size, L... size The size of k is constantly adjusted t The size of the convolution kernel is adjusted until the most suitable kernel size for the data features of the current task is found.

[0077] The convolution results of different sizes are concatenated along the channel direction. The concatenated feature map contains multi-scale information and retains feature representations under different receptive fields. The mathematical model for the concatenation operation is as follows:

[0078] Z fused =Concat(Z1, Z2, ..., Z) T );

[0079] In the formula, Concat(·) is the concatenation operation, and Z fused The feature map is the fused feature map, which contains features from different convolutional kernels.

[0080] Furthermore, in step S31, the use of adaptive convolutional kernel size and loss function optimized for convolutional kernel size, along with convolutional results of different sizes being concatenated by channel, can significantly improve the multi-scale feature extraction capability of the landslide disaster trend prediction model. Adaptive convolutional kernel size enables the model to dynamically adjust its receptive field, thereby flexibly capturing local and global information in landslide data. Optimized loss function helps the convolutional kernel size to self-adjust, improving the model's adaptability to landslide disasters and prediction stability. By concatenating features of different scales, the model can comprehensively consider multiple factors, enhance feature fusion capabilities, and improve prediction accuracy.

[0081] S32. Construct a multi-level time series modeling module for landslide disasters. The multi-level time series modeling module for landslide disasters introduces a hybrid architecture of Bi-LSTM and Bi-GRU, and introduces a gating mechanism to dynamically determine the fusion ratio of the output features of Bi-LSTM and Bi-GRU according to different time steps.

[0082] Furthermore, in step S32, firstly, the input landslide disaster time series data is processed through a Bi-LSTM layer, the Bi-LSTM structure being as follows: Figure 3As shown, the forward LSTM processes past dependencies, and the backward LSTM processes future dependencies, capturing long-term dependencies and trends. The mathematical model of Bi-LSTM is:

[0083]

[0084] In the formula, Z fused (t,:) represents the input data, which comes from the output of the landslide disaster spatial feature extraction module. This represents the output of the Bi-LSTM at time t.

[0085] Subsequently, the features output by the Bi-LSTM are fed into the Bi-GRU layer as input. The Bi-GRU captures short-term temporal dependencies, and its structure is as follows: Figure 4 As shown, to enable the model to respond quickly to instantaneous changes in a time series, the mathematical model of Bi-GRU is:

[0086]

[0087]

[0088] In the formula, This represents the output of the Bi-LSTM at time t.

[0089] A gating mechanism is introduced to finely control the fusion method of Bi-LSTM and Bi-GRU outputs. The fusion ratio of the Bi-LSTM and Bi-GRU output features is dynamically determined according to different time steps. The model can select an appropriate time series model for modeling based on the data characteristics of the current time step. The gating mechanism outputs a value g through a sigmoid activation function. t g t The mathematical model is as follows:

[0090]

[0091] In the formula, This represents the positive output of the Bi-LSTM at time t. b is the positive output of the Bi-GRU at time t. g As a paranoid term, W g The weights are then used to weight the inverse outputs of Bi-LSTM and Bi-GRU. The mathematical model for this fusion is as follows:

[0092]

[0093] In the formula, This represents the inverse output of the Bi-LSTM at time t. h represents the inverse output of the Bi-GRU at time t. tThe result is the fusion of the outputs of Bi-LSTM and Bi-GRU.

[0094] Furthermore, in step S32, the Bi-LSTM and Bi-GRU hybrid architecture with a gating mechanism can effectively combine the long-term dependency capture capability of LSTM and the short-term dependency capture capability of GRU, flexibly handling long-term trends and sudden changes in landslide disasters; the gating mechanism dynamically adjusts the fusion ratio of the outputs of Bi-LSTM and Bi-GRU, enabling the model to adaptively select appropriate dependency modes according to different time steps, improving prediction accuracy and flexibility; the Bi-LSTM and Bi-GRU hybrid architecture with a gating mechanism improves computational efficiency, enhances the model's generalization ability, and captures information at different time scales simultaneously through multi-level modeling, thereby providing more accurate and stable results in landslide disaster prediction.

[0095] S33. Construct a temporal adaptive attention mechanism. The temporal adaptive attention mechanism introduces a similarity metric parameter to compare the similarity between the hidden states of the current time step t and the historical time step t-1, thereby obtaining the temporal adaptive attention weight for each time step t.

[0096] Furthermore, in step S33, the similarity metric parameter calculates the similarity between the hidden states of the current time step t and the historical time step t-1 by comparing them. This application uses a learnable weight matrix W. t To calculate the similarity metric parameters, the mathematical model for the similarity metric parameters is as follows:

[0097]

[0098] In the formula, h t h is the feature vector at time step t, which is the output of the landslide disaster multi-level time series modeling module. t-1 Let W be the eigenvector at time step t-1. h W t This is a learnable weight matrix, applied to the features of the current time step and the previous time step, respectively. The weight matrix W t The transpose operation, sim(h) t h t-1 Let be the similarity measure between time step t and time step t-1;

[0099] After calculating the relevance at each time step based on the similarity metric parameters, the temporal adaptive attention weights are calculated using the Softmax function. These weights control the importance of features at each time step in the weighted fusion. The Softmax function transforms the similarity measure into a probabilistic form, normalizing the attention weights at each time step across all time steps, resulting in a sum of 1. This is the temporally adaptive attention weight. The mathematical model is as follows:

[0100]

[0101] In the formula, Let be the temporal adaptive attention weights at time step t. To normalize the similarity measure across all time steps, ensure that the sum of all weights is 1;

[0102] Based on the temporal adaptive attention weights, the hidden states at all time steps are weighted and summed to obtain the fused feature representation. The mathematical model of the fused feature representation is as follows:

[0103]

[0104] In the formula, z t This is the final weighted fusion feature representation. h represents the temporal adaptive attention weights at time step i. i Let be the hidden state at time step i.

[0105] Furthermore, in step S33, a time-adaptive attention mechanism based on similarity measurement is introduced in the landslide disaster trend prediction. This mechanism can dynamically calculate the attention weight of each time step by comparing the similarity between the hidden states of the current time step and the previous time step, thereby highlighting the modeling of key moments of trend continuation and abrupt change. The introduction of the time-adaptive attention mechanism based on similarity measurement enables the model to adaptively identify change nodes and sudden risk signals in the long-term accumulation process, accurately focus on the most valuable moment information for prediction, enhance the perception of landslide evolution, and significantly improve the accuracy, robustness and adaptability to complex time-series structures of the prediction.

[0106] S34. Construct a trend-mutation co-modeling prediction module. The trend-mutation co-modeling prediction module constructs loss functions for two branches: trend prediction and mutation detection, which respectively capture the landslide disaster trend of chronic risk accumulation and the signal of sudden landslide disaster.

[0107] Furthermore, in step S34, loss functions for two branches, trend prediction and mutation detection, are introduced. The trend prediction loss penalizes the difference between the predicted value and the true trend value, enabling the model to identify and fit slow but dangerous evolutionary signals. The mutation detection loss is used to capture sudden signals before a landslide occurs. This is achieved by using the output z of the time-adaptive attention mechanism. tTwo prediction branches are fed in, and the mathematical models for the two branches are as follows:

[0108]

[0109] In the formula, MLP trend MLP burst These are two MLP neural network layers, The trend-based landslide risk value of the MLP prediction layer at time step t. To determine the mutation risk probability of the MLP prediction layer at time t for mutation detection, the magnitude of the loss function is then calculated. The mathematical model for the trend prediction loss function is as follows:

[0110]

[0111] In the formula, T is the total time step, and L is the total time step. trend To optimize the regression loss of the trend prediction branch, Let be the true trend value at time step t; the mathematical model for the mutation detection loss function is:

[0112]

[0113] In the formula, L burst To optimize the binary cross-entropy loss for the mutation detection branch, Let t be the landslide mutation marker at time step t, where 0 indicates no mutation and 1 indicates the presence of a mutation;

[0114] Finally, the two prediction results are merged into a single landslide disaster risk output. The fusion mathematical model is as follows:

[0115]

[0116] In the formula, γ is a weighting parameter with an initial value of 0.5 and a range of [0,1]. The final landslide disaster risk prediction value is obtained from the trend-mutation co-modeling prediction module.

[0117] Furthermore, in step S34, a trend prediction loss function and a mutation detection loss function are introduced to model the long-term evolution and short-term sudden signals in landslide disasters, respectively, ensuring that the model can simultaneously capture the potential risk accumulation and sudden changes of landslides. The trend branch focuses on the slowly changing accumulated risk, while the mutation branch is used to identify the rapid changes that may trigger landslides in the short term. Through this branch design, the model can avoid logical conflicts between different signals, improve the prediction accuracy and sensitivity of landslide disasters, ensure that the early warning system can take into account both long-term trends and sudden events, and improve the reliability and accuracy of the overall risk assessment.

[0118] S4. Train the landslide disaster trend prediction model by using the training set to train the landslide disaster trend prediction model.

[0119] Furthermore, in step S4, hyperparameter settings are required during the training of the landslide disaster trend prediction model. The learning rate controls the update step size, and is set to 0.001. The batch size determines the number of samples updated each time, and is set to 64. The optimizer selection affects training efficiency and stability, and is set to Adam. Regularization prevents overfitting, and is set to 0.00001. Gradient clipping ensures training stability, and is set to a gradient clipping threshold of 5.0. The number of training epochs sets the training duration, and is set to 1000. By adjusting these hyperparameters, it is possible to ensure efficient convergence of the model in the landslide disaster prediction task, avoid overfitting, and improve prediction accuracy.

[0120] S5. Test the landslide disaster trend prediction model. Input the test set into the landslide disaster trend prediction model and output the landslide disaster probability value to achieve landslide disaster trend prediction.

[0121] Furthermore, in step S5, the landslide disaster trend prediction model trained in step S4 is tested, and the prediction results are as follows: Figure 5 As shown in the figure, the actual values ​​of landslide disaster data, the predicted values ​​of the prediction model proposed in this application, and the predicted values ​​of LSTM and Transform prediction models were also tested. It can be seen from the figure that the predicted values ​​of the prediction model proposed in this application are closest to the actual values ​​and can predict the trend of landslide disasters with high accuracy.

Claims

1. A method for predicting landslide disaster trends based on machine learning, characterized in that, Includes the following steps: S1. Construct a multi-source heterogeneous landslide disaster dataset, the features of which include: geological and geomorphological features, meteorological and hydrological features, engineering activities, historical landslide disaster information and real-time sensor monitoring data; S2. Preprocess the landslide disaster dataset. Clean the landslide disaster dataset and divide the cleaned dataset into training and testing sets in a 7:3 ratio. S3. Propose a landslide disaster trend prediction model, including the following methods: S31. Construct a spatial feature extraction module for landslide disasters. By introducing an adaptive convolution kernel size mechanism, a larger convolution kernel is used for long-term dependent features in landslide disasters, and a smaller convolution kernel is used for short-term dependent features. S32. Construct a multi-level time series modeling module for landslide disasters, integrate Bi-LSTM and Bi-GRU architectures, and introduce a gating mechanism to dynamically adjust the fusion ratio of Bi-LSTM and Bi-GRU outputs; S33. Construct a temporal adaptive attention mechanism, introduce a similarity metric to compare the similarity between the hidden states of the current time step and the historical time steps, and dynamically calculate the temporal adaptive attention weights. S34. Construct a trend-mutation collaborative modeling and prediction module, and construct two branches of loss: trend prediction and mutation detection, to capture the chronic accumulation risk and sudden signals of landslide disasters, respectively. Trend prediction loss penalizes the difference between predicted and true trend values, enabling the model to identify and fit slow but dangerous evolutionary signals. Mutation detection loss is used to capture sudden signals preceding landslides. The method involves applying the output of a time-adaptive attention mechanism... Two prediction branches are fed in, and the mathematical models for the two branches are as follows: ； ； In the formula, , These are two MLP neural network layers, The trend-based landslide risk value of the MLP prediction layer at time step t. To determine the mutation risk probability of the MLP prediction layer at time t for mutation detection, the magnitude of the loss function is then calculated. The mathematical model for the trend prediction loss function is as follows: ； In the formula, T is the total time step. To optimize the regression loss of the trend prediction branch, Let be the true trend value at time step t; the mathematical model for the mutation detection loss function is: ； In the formula, To optimize the binary cross-entropy loss for the mutation detection branch, Let t be the landslide mutation marker at time step t, where 0 indicates no mutation and 1 indicates the presence of a mutation; Finally, the two prediction results are merged into a single landslide disaster risk output. The fusion mathematical model is as follows: ； In the formula, For weight parameters, The final landslide disaster risk prediction value is obtained for the trend-abrupt change co-modeling prediction module; S4. Train the landslide disaster trend prediction model by using the training set to train the landslide disaster trend prediction model. S5. Test the landslide disaster trend prediction model. Input the test set into the landslide disaster trend prediction model and output the landslide disaster probability value to achieve landslide disaster trend prediction.

2. The landslide disaster trend prediction method based on machine learning according to claim 1, characterized in that, In step S1, the geological and geomorphological characteristics include slope, aspect, landform type, lithological distribution, geological structure, fault distribution, soil type, shear strength and void ratio; the meteorological and hydrological characteristics include hourly and daily rainfall, cumulative rainfall, air temperature, relative humidity, wind speed, surface runoff, groundwater level and evaporation; the engineering activity characteristics include the distribution of roads, buildings, tunnels, mining areas, changes in land use types, layout of drainage facilities and distribution of retaining structures; the historical landslide disaster information includes landslide location, occurrence time, landslide volume, sliding direction, triggering factors and sliding frequency; the real-time monitoring data includes surface and deep displacement data, tilt angle, crack width, groundwater pore pressure, soil moisture content, rainfall intensity and cumulative rainfall.

3. The landslide disaster trend prediction method based on machine learning according to claim 2, characterized in that, In step S31, an adjustable hyperparameter is introduced into the convolutional layer. This hyperparameter dynamically adjusts the size of the convolutional kernel during training. For features that depend on landslide disasters over a long period, the convolutional kernel can be enlarged to capture a wider range of spatial features. For features that depend on landslide disasters over a short period, the convolutional kernel is reduced to extract more refined local features of landslide disasters. The mathematical model for the dynamic convolution operation is as follows: ； In the formula, For dynamically adjusted convolutional kernel size, For input data, This represents the convolution result at time t; To ensure the appropriateness of the convolution kernel size, a loss function is introduced to constrain the change in the convolution kernel size, avoiding excessive expansion or contraction. The mathematical model of the loss function is as follows: ； In the formula, This is the default initial size for the convolutional kernel. Let be the size of the convolution kernel at time t. The regularization coefficient is . To limit the deviation of the convolution kernel size, by Size constantly adjusted The size of the convolution kernel is adjusted until the most suitable kernel size for the data features of the current task is found. The convolution results of different sizes are concatenated along the channel direction. The concatenated feature map contains multi-scale information and retains feature representations under different receptive fields. The mathematical model for the concatenation operation is as follows: ； In the formula, For splicing operations, The feature map is the fused feature map, which contains features from different convolutional kernels.

4. The landslide disaster trend prediction method based on machine learning according to claim 3, characterized in that, In step S32, the input landslide disaster time series data is processed through a Bi-LSTM layer. A forward LSTM processes past dependencies, and a backward LSTM processes future dependencies. The mathematical model of Bi-LSTM is: ； ； In the formula, For input data, , This represents the output of the Bi-LSTM at time t. The features output by Bi-LSTM are fed into the Bi-GRU layer as input. Bi-GRU captures short-term time dependencies. The mathematical model of Bi-GRU is: ； ； In the formula, , This represents the output of the Bi-LSTM at time t. A gating mechanism is introduced to finely control the fusion method of the Bi-LSTM and Bi-GRU outputs. The fusion ratio of the Bi-LSTM and Bi-GRU output features is dynamically determined according to different time steps. The gating mechanism outputs a value through a sigmoid activation function. , The mathematical model is as follows: ； In the formula, This represents the positive output of the Bi-LSTM at time t. This represents the positive output of the Bi-GRU at time t. For paranoia, The weights are then used to weight the inverse outputs of Bi-LSTM and Bi-GRU. The mathematical model for this fusion is as follows: ； In the formula, This represents the inverse output of the Bi-LSTM at time t. This represents the inverse output of the Bi-GRU at time t. The result is the fusion of the outputs of Bi-LSTM and Bi-GRU.

5. The landslide disaster trend prediction method based on machine learning according to claim 4, characterized in that, In step S33, the similarity measurement parameter is obtained by comparing the current time step t with the historical time step. The similarity between them is calculated from their hidden states using a learnable weight matrix. To calculate the similarity metric parameters, the mathematical model for the similarity metric parameters is as follows: ； In the formula, This is the feature vector at time step t, which is the output of the multi-level temporal series modeling module for landslide disasters. For the first The feature vector at each time step , The weight matrix is ​​a learnable matrix. Weight matrix The transpose operation, This is a similarity measure between time step t and time step t-1; Based on the correlation calculated at each time step using the similarity metric parameters, the temporal adaptive attention weights are calculated using the Softmax function. To control the importance of features at each time step in the weighted fusion, the Softmax function transforms the similarity metric into a probabilistic form, ensuring that the attention weights at each time step are normalized across all time steps, summing to 1. This results in temporally adaptive attention weights. The mathematical model is as follows: ； In the formula, Let be the temporal adaptive attention weights at time step t. To normalize the similarity measure for all time steps, ensure that the sum of all weights is 1; Based on the temporal adaptive attention weights, the hidden states at all time steps are weighted and summed to obtain the fused feature representation. The mathematical model of the fused feature representation is as follows: ； In the formula, This is the final weighted fusion feature representation. Let be the temporal adaptive attention weights at time step i. Let be the hidden state at time step i.

Images