A landslide surge height prediction method and system based on a gated recurrent unit

Abstract

This invention relates to the field of landslide surge and data-driven technology, and discloses a method and system for predicting landslide surge height based on gated cyclic units (GRUs). The method includes collecting relevant data from large-scale physical model experiments in landslide surge areas, performing data standardization and normalization preprocessing on the feature data, and using the preprocessed feature data as input to a GRU model to obtain the surge height; constructing a GRU network using tensor flow in a Python compilation environment; training the GRU network model using historical landslide data to adjust model parameters; and deploying the trained GRU model into a real-world landslide surge prediction system for real-time monitoring. The system includes a data preprocessing module, a model building module, and a model training module. This invention forms a highly efficient and accurate prediction system.

Description

Technical Field

[0001] This invention relates to the field of landslide surge and data-driven technology, and in particular to a method and system for predicting landslide surge height based on a gated cyclic unit. Background Technology

[0002] In the field of landslide surge disaster prevention and control, the movement of soil and rock masses during a landslide is a complex dynamic process that leads to rapid changes in topography and subsequently triggers surges in the landslide area or downstream water bodies. Accurate prediction of surge height is crucial for early warning and prevention of landslide disasters. With the development of deep learning technology, Gated Recurrent Units (GRUs), as an effective time-series data processing method, have shown broad application prospects in landslide surge height prediction. GRUs can capture the time-series characteristics of the landslide process and accurately predict the surge height triggered by landslides by learning the patterns of historical landslide data. Compared with traditional methods, GRU-based landslide surge height prediction methods have the following advantages: First, GRU models can effectively handle long-term dependencies and capture the dynamic changes during the landslide process; second, GRUs are trained using experimental data, which better reflects the physical characteristics of the landslide process and improves the model's prediction accuracy; third, GRU models can automatically learn and optimize parameters, avoiding the problem of parameter overfitting in traditional methods.

[0003] Existing technology 1, Chinese patent application number 202410069549.X, discloses a coupling method and system for the near-field and far-field propagation evolution of landslide surges. The method includes the following steps: simulating the near-field motion of a granular landslide using particle discrete element method and smoothed particle hydrodynamics; calculating the far-field motion of the granular landslide using the simulated near-field motion as initial boundary conditions; and completing the near-field and far-field propagation evolution of the granular landslide during its movement and entry into water based on drag and buoyancy. While detailed numerical simulations of key areas can more accurately predict the propagation path, velocity, and impact range of landslide surges, and can also consider topography, geology, and hydrology in detail, improving the realism of landslide surge simulations, and by simulating different scenarios, the potential impact of landslide surges on coastal structures, residential areas, and critical infrastructure can be assessed, providing hazard analysis for rescue and planning, the method itself lacks integration with mechanical theory. This can easily lead to parameter overfitting or even the appearance of physically meaningless parameters, making it difficult to analyze the landslide surge process and related risk prevention and control.

[0004] Prior art two, Chinese patent application number 202410014296.6, discloses a method for judging the type of swell induced by landslides and predicting wave height. This method includes establishing a dynamic model of slope instability, obtaining the landslide velocity u(t) and displacement s(t), calculating dimensionless parameters Fr and Sz, identifying the swell wave type, and using oscillating wave / single wave formulas to calculate the wave height Am of the corresponding wave type. While classifying and identifying landslide swell waves and proposing wave height prediction methods for different swell waves provides reliable technical support for determining relevant disaster prevention and mitigation measures and formulating emergency plans, it mainly relies on numerical simulations or field monitoring data, resulting in insufficient calibration of dynamic parameters and inadequate handling of their uncertainties.

[0005] Prior art three, Chinese patent application number 202311122865.0, discloses a method for calculating the exceedance probability of landslide surges using random sampling. The method includes: analyzing the range of landslide parameters in the target area; processing the landslide input parameters required for constructing a numerical model to generate random variables; performing Latin hypercube sampling on the random variables to generate random samples; inputting the random samples into the numerical model; extracting the surge propagation height corresponding to the random variables; determining the surge propagation height threshold based on the affected bodies in the target area; and calculating the exceedance probability of the surge propagation height exceeding the threshold. While this method quantifies the uncertainty of landslide input parameters and improves the accuracy of landslide-surge disaster prediction by performing Latin hypercube sampling on selected random variables to generate random samples and then processing the random samples through the model, it is limited by numerous assumptions and boundary condition processing, and cannot reflect the true situation of landslide surges.

[0006] Currently, existing technologies 1, 2, and 3 lack integration with mechanical theory, easily leading to parameter overfitting or even physically meaningless parameters. They are insufficient for analyzing landslide surge processes and related risks, relying primarily on numerical simulations or field monitoring data. These methods suffer from inadequate calibration of dynamic parameters and handling of uncertainties, and are limited by numerous assumptions and boundary conditions, failing to reflect the true situation of landslide surges. Therefore, this invention proposes a landslide surge height prediction method based on GRU, combining experimental data and deep learning algorithms to significantly improve the accuracy and reliability of predictions, providing more effective technical support for risk prevention and control of landslide surge disaster chains. Summary of the Invention

[0007] The main objective of this invention is to provide a method and system for predicting landslide surge height based on a gated cyclic unit, in order to solve the problems in the prior art that the calibration of dynamic parameters and the handling of their uncertainties are insufficient, and that the prior art is limited by many assumptions and boundary conditions, and therefore cannot reflect the true situation of landslide surge.

[0008] To achieve the above objectives, the present invention provides the following technical solution:

[0009] A method for predicting landslide surge height based on a gated cyclic unit, the method comprising:

[0010] Data from large-scale physical model tests in the landslide surge area were collected, and the characteristic data were preprocessed by data standardization and normalization. The preprocessed characteristic data were used as input to the gated recurrent unit (GRU) model to obtain the surge height.

[0011] In a Python compilation environment, a gated recurrent unit (GRU) network is built using tensor flow. Update gates and reset gates are introduced to control the flow of information. The update gate determines how much information from the previous state should be retained in the current state, while the reset gate controls how much information from the previous state should be forgotten.

[0012] A gated recurrent unit (GRU) network model was trained using historical landslide data to adjust the model parameters. The trained GRU model was then deployed to an actual landslide surge prediction system for real-time monitoring.

[0013] As a further improvement of the present invention, the relevant data includes landslide volume, landslide velocity, slope angle, and characteristic data of similar materials used to simulate the properties of real soil and rock materials.

[0014] As a further improvement to the present invention, a gated recurrent unit (GRU) network is constructed, including:

[0015] Design the input layer, GRU layer, fully connected layer, and output layer of a gated recurrent unit (GRU) network;

[0016] Mean squared error is used as the loss function for the regression task, and the learning rate is dynamically adjusted according to gradient changes.

[0017] When new data is input, the weights are updated through the backpropagation algorithm. When different environmental conditions change, the functions of resetting and updating the gates are also adjusted through feedback.

[0018] As a further improvement of the present invention, the input layer is used to input feature vectors, the GRU layer captures long-term and short-term dependencies in time series data through a gating mechanism, the fully connected layer maps the output of the GRU layer to a new feature space, and combines the time series information extracted from the GRU layer to generate a higher-level feature representation, and the output layer maps the output of the fully connected layer to the specific prediction result, the height of the landslide surge.

[0019] As a further improvement of the present invention, the input layer sets the input feature vector x. t for:

[0020] x t =[x t,1 ,x t,2 ,…,x t,n ]

[0021] In the formula, n represents the number of features, including landslide-related parameters such as landslide volume, rate, and slope; environmental information such as rainfall, soil moisture, and temperature; and data on historical landslide events.

[0022] The input data needs to be organized as a three-dimensional tensor:

[0023]

[0024] In the formula, m is the number of samples, T is the time step, and n is the number of features at each time step;

[0025] GRU layer reset gate r t It is an important gating mechanism used to control how much historical information should be forgotten when calculating a new hidden state. Its calculation formula is:

[0026] r t =σ(W r ·[h t-1 ,x t ]+b r )

[0027] In the formula, W r h represents the weight matrix of the reset gate. t-1 It represents the hidden state of the previous time step, σ(·) is the sigmoid activation function, and its output range is [0,1]. r This resets the door's bias.

[0028] Update Gate Z t It is used to control the degree to which the previous hidden state retains the current state, and its calculation formula is as follows:

[0029] z t =σ(W z ·[h t-1 ,x t]+b z )

[0030] In the formula, W z It is the weight matrix of the updated gate, b z It updates the bias term of the gate;

[0031] Candidate hidden state It is generated when the decision is made to update the current hidden state, and its calculation formula is:

[0032]

[0033] In the formula, ⊙ denotes element-wise multiplication, W is the weight matrix of the candidate state, and b is the bias term of the candidate state.

[0034] Current hidden state h t The update formula is:

[0035]

[0036] By stacking multiple GRU layers, the GRU units in subsequent layers not only obtain information from the previous layer, but also combine the hidden states of the previous layers in their computation, forming a residual connection mechanism, i.e.:

[0037]

[0038] In the formula, This represents the current hidden state of the l-th layer, which is calculated based on the hidden states of the previous layer and the hidden states of the two preceding layers. This indicates the current hidden state of the (l-1)th layer, providing information about the previous layer. This indicates the current hidden state of layer l-2, providing information about earlier layers;

[0039] The fully connected layer will process the output h of the GRU layer. t Input a fully connected layer and perform a nonlinear mapping using the ReLU activation function:

[0040] h fc =ReLU(W fc ·h t +b fc )

[0041] In the formula, h fc W represents the output of the fully connected layer, specifically the activation value after the output from the GRU layer is passed to the fully connected layer. fc This represents the weight matrix of the fully connected layer, which determines the contribution of the input features to the output features. fc The bias term of the fully connected layer is used to adjust the output; the ReLU activation function maps negative values ​​to zero, i.e., ReLU(x) = max(0,x), introducing non-linear characteristics;

[0042] The output layer maps features to predicted values ​​through a fully connected layer, using a linear activation function.

[0043] y pred =W out ·h fc +b out

[0044] In the formula, y pred This is the final predicted surge height; W out The weight matrix of the output layer is used to map the features of the fully connected layer to the output values, b out The bias term of the output layer.

[0045] As a further improvement to the present invention, the loss function is:

[0046]

[0047] In the formula, L is the mean squared error calculated by the loss function, which measures the difference between the wave height predicted by the model and the actual value, and N is the total number of samples used to calculate the average loss. The actual surge height of the i-th sample The model predicts the surge height for the i-th sample.

[0048] As a further improvement of the present invention, the reset gate and the update gate adopt a dynamic coupling strategy, including:

[0049] The interaction between the reset gate and the update gate is calculated. When calculating the reset gate at the current time step, the combined information of the previous hidden state and the current input is considered. The update gate from the previous time step is introduced, which simultaneously refers to the current input and the previous hidden state, and the amount of information retained at the current time step.

[0050] Historical information is integrated during the generation of candidate hidden states. Through comprehensive analysis of historical information and new inputs, the candidate states reflect the characteristics of the current moment.

[0051] The hidden state update process and its dynamic coupling effect: the update gate is responsible for determining how much information from the previous hidden state should be retained in the current hidden state.

[0052] As a further improvement of the present invention, the process of generating candidate hidden states includes the interaction between reset gate and update gate, and the weight of reset gate affects the contribution of the previous hidden state in the candidate hidden state.

[0053] As a further improvement of the present invention, through a dynamic coupling mechanism, the reset gate and the update gate jointly determine the flow and storage of information; when updating the hidden state at the current moment, the reset gate determines the degree to which the previous state needs to be forgotten.

[0054] To achieve the above objectives, the present invention also provides the following technical solution:

[0055] A landslide surge height prediction system based on a gated cyclic unit is provided, which is applied to the aforementioned landslide surge height prediction method based on a gated cyclic unit. The landslide surge height prediction system based on a gated cyclic unit includes:

[0056] The data preprocessing module is used to collect relevant data from large-scale physical model tests in the landslide surge area, and to perform data standardization and normalization preprocessing on the feature data. The preprocessed feature data is used as input to the gated recurrent unit (GRU) model to obtain the surge height.

[0057] The model building module is used to build a gated recurrent unit (GRU) network using tensor flow in a Python compilation environment. Update gates and reset gates are introduced to control the flow of information. The update gate determines how much information from the previous state should be retained in the current state, while the reset gate controls how much information from the previous state should be forgotten.

[0058] The model training module is used to train the gated recurrent unit (GRU) network model. It uses historical landslide data to train the GRU network model and adjust the model parameters. The trained GRU model is then deployed to the actual landslide surge prediction system for real-time monitoring.

[0059] This invention addresses the problems of insufficient feature extraction and low prediction accuracy in existing technologies for landslide surge height prediction. The implementation steps are as follows: acquiring relevant feature data of landslide surges; building a GRU network using TensorFlow in a Python compilation environment; controlling the flow of information by introducing an update gate and a reset gate. The update gate determines how much information from the previous state should be retained in the current state, while the reset gate controls how much previous state information should be forgotten; training the GRU network using historical landslide data; and using the trained GRU network to predict the height of landslide surges. This invention constructs a GRU-based landslide surge height prediction model that can effectively capture the long-term and short-term time dependencies in the landslide surge process, improving prediction accuracy and efficiency. Attached Figure Description

[0060] Figure 1 This is a flowchart illustrating the steps of an embodiment of the landslide surge height prediction method based on a gated cyclic unit according to the present invention.

[0061] Figure 2 This is a schematic diagram of an embodiment of the landslide surge height prediction method based on a gated cyclic unit according to the present invention;

[0062] Figure 3 This is a schematic diagram of a gated loop unit, representing an embodiment of the landslide surge height prediction method based on a gated loop unit according to the present invention.

[0063] Figure 4 This is a schematic diagram of the preprocessing steps for data standardization and normalization of feature data in an embodiment of the landslide surge height prediction method based on gated cyclic units of the present invention.

[0064] Figure 5 This is a schematic diagram illustrating the steps of building a gated recurrent unit (GRU) network in one embodiment of the landslide surge height prediction method based on gated recurrent units of the present invention.

[0065] Figure 6 This is a schematic diagram of the steps of a gated cyclic unit (GRU) network model in an embodiment of the landslide surge height prediction method based on gated cyclic units according to the present invention.

[0066] Figure 7 This is a functional module diagram of an embodiment of the landslide surge height prediction system based on a gated loop unit of the present invention;

[0067] Figure 8 This is a schematic diagram of the structure of an embodiment of the electronic device of the present invention;

[0068] Figure 9 This is a schematic diagram of the structure of a storage medium according to an embodiment of the present invention;

[0069] Figure 10 This is a comparison chart of the experimental data of the physical model of this invention and the predicted values ​​of the gated loop unit method;

[0070] Figure 11 This is a graph showing the K-fold cross-validation results of the present invention (K=5). Detailed Implementation

[0071] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0072] The terms "first," "second," and "third" used in this invention are for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined as "first," "second," or "third" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified. All directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of this invention are only used to explain the relative positional relationships and movements between components in a specific orientation (as shown in the figures). If the specific orientation changes, the directional indications also change accordingly. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or devices.

[0073] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a mutually exclusive, independent, or alternative embodiment. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0074] like Figure 1 As shown, this embodiment provides an example of a landslide surge height prediction method based on a gated cyclic unit. In this embodiment, the landslide surge height prediction method based on a gated cyclic unit specifically includes the following steps:

[0075] Step S1: Collect relevant data from large-scale physical model tests in the landslide surge area. The relevant data includes landslide volume, sliding velocity, slope angle, and characteristic data of similar materials used to simulate real soil and rock materials. The characteristic data is preprocessed by data standardization and normalization. The preprocessed characteristic data is used as input to the gated recurrent unit (GRU) model to obtain the surge height.

[0076] Step S2: Build a gated recurrent unit (GRU) network using TensorFlow in the Python compilation environment; introduce update gate and reset gate to control the flow of information. The update gate determines how much information from the previous state should be retained in the current state, while the reset gate controls how much information from the previous state should be forgotten.

[0077] Step S3: Train the gated recurrent unit (GRU) network model. Use historical landslide data to train the GRU network model and adjust the model parameters. Deploy the trained GRU model into the actual landslide surge prediction system for real-time monitoring.

[0078] Preferably, step S1 of this embodiment, data collection and preprocessing, focuses on collecting relevant feature data within the landslide surge area, including landslide volume, velocity, slope angle, and similar material properties. The collected feature data is standardized and normalized to eliminate the influence of different feature dimensions and ranges, ensuring the stability and accuracy of model training. The significance is that preprocessing the feature data ensures higher quality data input to the GRU model, helping to improve the model's prediction accuracy. Standardization and normalization can accelerate model convergence, reduce training time, and avoid gradient vanishing or exploding problems caused by differences in feature values. Step S2 involves building the GRU network. The GRU network is constructed in a Python environment using the TensorFlow framework, with update and reset gates introduced to control information flow. This allows the model to dynamically adjust its state based on historical information; the update gate can determine how much information from the previous state to retain based on the current input data, while the reset gate effectively removes old information. The significance is that introducing a gating mechanism enables the GRU network to better capture long-term and short-term dependencies in time-series data. This is an improvement over traditional RNNs, making the network more efficient in processing time-series data. Through dynamic information management, the model can more flexibly adapt to changes in input data, thereby improving its predictive ability for complex systems (such as landslide surges). Step S3 trains the network and performs real-time monitoring. Historical landslide data is used to train the GRU model, and model parameters are adjusted to optimize prediction performance. The trained GRU model is then deployed to an actual landslide surge prediction system for real-time data monitoring and prediction. The significance is that through the training process, the model learns the complex relationship between landslide surge height and historical data, thus providing more accurate results in actual predictions. Real-time monitoring and prediction can provide relevant agencies with timely information, helping to make faster decisions and reduce the safety risks that landslide surges may bring. Furthermore, it can help optimize emergency response mechanisms, ensuring people's safety and property protection (see appendix for details). Figure 2 Gated Loop Unit Reference Appendix Figure 3 ).

[0079] In summary, this embodiment, through steps such as data collection and preprocessing, network construction, model training, and real-time application, forms an efficient and accurate prediction system; it helps to build an effective natural disaster prediction and early warning system under complex environmental conditions.

[0080] This embodiment of the landslide surge height prediction method based on gated loop units has the following significant advantages: (1) It processes the monitoring data of landslide surge physical model tests, combines landslide surge control factors and deep learning algorithms to dynamically evaluate surge height, and this method can effectively handle long-term dependencies, and can still maintain efficient operation even in the face of a large amount of monitoring data. (2) It can establish a deep learning model based on deep learning algorithms and experimental data, which significantly improves the accuracy of surge height prediction. (3) It combines deep learning theory and mathematical models based on actual large-scale physical model test data to achieve accurate prediction of landslide surge height, forming a new landslide disaster prevention and control evaluation method, and providing new ideas for related projects. This embodiment can quantitatively analyze the changes in mechanical parameters during the landslide process, deepen the engineering understanding of the landslide surge process, and achieve better prediction results; compared with existing numerical simulation and simple statistical methods, it combines deep learning and mechanical mechanisms, realizes dual innovation of data-driven and theory-driven approaches, significantly improves the accuracy and reliability of landslide surge height prediction, and provides more reliable and efficient technical support for landslide disaster prevention and control.

[0081] Furthermore, such as Figure 4 As shown, the preprocessing of feature data standardization and normalization in step S1 specifically includes the following steps:

[0082] Step S11: Calculate the mean and standard deviation of each feature data point, and use the statistical information of the mean and standard deviation to readjust each data point so that the feature data can be compared on the same scale; for the value range of each feature data, scale the feature data to the [0,1] interval, find the maximum and minimum values ​​of each feature, and adjust the feature data through linear mapping;

[0083] Step S12: Obtain the normalized and adjusted feature data, sort all feature data according to timestamps, and use a dynamic segmentation strategy to divide the training set, validation set and test set. Use 70% of the feature data for the training set, 15% of the feature data for the validation set and 15% of the feature data for the test set.

[0084] Step S13: Starting from the beginning position of the feature data, place the window at the current starting point and collect 10 data points within the window as a sample; move the window by 1 data point each time and repeat the above steps until enough samples are generated; for each sample generated by the sliding window, determine the corresponding label, i.e., the surge height at the next moment.

[0085] Preferably, in step S11 of this embodiment, the standardization and normalization of feature data involves calculating the mean and standard deviation of each feature and using this statistical information to adjust the data. This allows the feature data to be compared on the same scale, eliminating biases caused by different value ranges between different features and improving the effectiveness of the model. Scaling the feature data to the [0,1] interval ensures that each feature is within the same range, avoiding the influence of feature value magnitude on the model learning process, which is especially important when using distance-based algorithms (such as KNN). Significance: Standardization and normalization make optimization algorithms more efficient in updating parameters, reduce model volatility during training, and help accelerate convergence. Standardized data makes the model more generalizable, thereby improving the prediction accuracy for new data, especially when dealing with high-dimensional features. Step S12, Data Sorting and Dynamic Segmentation, sorts the feature data according to timestamps to ensure the continuity of training, validation, and test sets in time. This is a crucial step in processing time-series data, avoiding data leakage issues, as test data is not used during training. A dynamic segmentation strategy is adopted, dividing the data into 70% training set, 15% validation set, and 15% test set, ensuring that each dataset is representative and covers the changing trends of the entire dataset. Significance: By confirming the order of time-series data, the validation and test sets can accurately reflect the model's prediction performance in real-world environments, improving the credibility of model evaluation; ensuring that the training set can learn sufficient and diverse data features helps it perform well on the validation and test sets, further enhancing the model's practicality and reliability. Step S13, Sliding Window Sample Generation, uses the sliding window method to extract continuous data points within a window from the time-series data to generate new training samples. This makes data utilization more efficient, ensuring that the model obtains enough samples to learn dynamic trends; each sample generated through the sliding window is assigned a clear label, namely the surge height at the next moment, ensuring that the model can learn the direct relationship between the data and the target output. Significance: By using the sliding window method, the model can better understand the changes in feature values ​​at different times, thus effectively capturing the dynamic characteristics in the time series; by creating more samples, especially when data is scarce, the training set size of the model can be significantly increased, making the model more capable of learning and reasoning.

[0086] In summary, this embodiment achieves the standardization and normalization of feature data, ensuring improved data quality and effective model training. Secondly, by sorting and dynamically segmenting the data, the rationality and security of the time series data are ensured, enhancing the reliability of model evaluation. Finally, the sliding window technique generates effective samples, enhancing the diversity and accuracy of model learning. Overall, these processing steps complement each other, helping to improve the model's predictive performance and generalization ability, laying a solid foundation for accurate prediction of landslide surge height.

[0087] Furthermore, the process of dynamically dividing the training set, validation set, and test set using the segmentation strategy in step S12 specifically includes the following steps:

[0088] Step S121: Set a critical event threshold when the wind speed or tidal surge is detected to reach the standard in the weather report, automatically trigger data segmentation, and generate training, validation and test sets based on the data samples before and after the event; define the flow of time (i.e. the continuity of fluctuations), create a flow sample pool in each time period, and fill it according to the continuous data fluctuations. The deep learning framework is used to realize the automatic loading of data from the flow sample pool.

[0089] Step S122: Define a fuzzy range for the division of a specific time period, including the data of the time period as well as the data of several time points before and after it; by introducing fuzzy logic, more contextual information of adjacent time periods can be received during training.

[0090] Step S123: The division of the training set and validation set is determined by the actual influence of the input features through feature selection; the time period corresponding to the feature with high importance score is given priority to enter the training set, while the feature with lower score is entered into the validation set.

[0091] Preferably, step S121, the key event threshold segmentation and creation of the flow sample pool in this embodiment, automatically identifies time periods that significantly impact environmental prediction by setting key event thresholds for wind speed or tidal surge, and incorporates their data into the training set; this reduces manual intervention and improves the real-time performance of the segmentation process; defining time flow and creating a flow sample pool based on data volatility ensures that contextual information across different time periods can be recognized and utilized by the system; the flow sample pool will contain known and practical data samples, facilitating rapid loading and model training. Significance: The automatic segmentation mechanism ensures the timeliness and relevance of model training, enabling the model to quickly adapt to sudden changes in the environment; it improves the timeliness of training data, ensuring that the model can obtain real-time data when high-risk events (such as extreme weather) occur, thereby improving predictive capabilities. Step S122, the introduction of fuzzy time period segmentation, sets a fuzzy range for a specific time period, allowing the training set to include data within the time period while also integrating data from preceding and following time points; this better captures potential trends and patterns in the time series; through the setting of the fuzzy range, the model can receive contextual information from adjacent time periods during training, enhancing the model's comprehensive understanding of short-term fluctuations and long-term trends. Significance: By integrating data from adjacent time periods, the model improves its comprehensive understanding of time series features, enabling more accurate predictions, especially under extreme climate change conditions. It reduces reliance on rigid time points, making the model more flexible in the face of nonlinear changes and contributing to improved robustness. Steps S123: Feature Selection and Partitioning. Through feature selection techniques, the system can identify and weigh the actual influence of input features, selecting data samples crucial to the prediction results. The training and validation sets are partitioned based on feature importance, ensuring the training data contains higher-quality feature information. Assigning features with high importance scores to the training set and sending features with low scores to the validation set helps the model focus on the most influential features, thereby improving learning efficiency. Significance: Partitioning based on feature importance ensures more efficient model training, effectively utilizes data resources, and avoids potential feature noise interference. It improves model interpretability, clearly identifying which factors are most critical in the prediction, which aids in result analysis and subsequent decision-making.

[0092] In summary, the dynamic segmentation strategy in this embodiment ensures the flexibility, timeliness, and efficiency of dataset partitioning. Its application in dynamic environments effectively improves the model's prediction accuracy and adaptability, especially when dealing with complex climate change or environmental data, enabling better capture and learning of motion features. By reducing human intervention and enhancing automated processing capabilities, this method ensures high-quality input of datasets into deep learning frameworks, thereby promoting research and application progress.

[0093] Furthermore, such as Figure 5As shown, the process of building the gated recurrent unit (GRU) network in step S2 specifically includes the following steps:

[0094] Step S21: Design the input layer, GRU layer, fully connected layer, and output layer of the Gated Recurrent Unit (GRU) network; the input layer is used to input feature vectors, the GRU layer captures long-term and short-term dependencies in time series data through a gating mechanism (reset gate and update gate), the fully connected layer maps the output of the GRU layer to a new feature space, and combines the time series information extracted from the GRU layer to generate a higher-level feature representation, and the output layer maps the output of the fully connected layer to specific prediction results, such as the height of landslide surges;

[0095] The input layer is defined as the input feature vector x. t for:

[0096] x t =[x t,1 ,x t,2 ,…,x t,n ]

[0097] In the formula, n represents the number of features, including landslide-related parameters (such as landslide volume, rate, and slope), environmental information (such as rainfall, soil moisture, and temperature), and data on historical landslide events.

[0098] The input data needs to be organized as a three-dimensional tensor:

[0099]

[0100] In the formula, m is the number of samples, T is the time step, and n is the number of features at each time step;

[0101] GRU layer reset gate r t It is an important gating mechanism used to control how much historical information should be forgotten when calculating a new hidden state. Its calculation formula is:

[0102] r t =σ(W r ·[h t-1 ,x t ]+b r )

[0103] In the formula, W r h represents the weight matrix of the reset gate. t-1 It represents the hidden state of the previous time step, σ(·) is the sigmoid activation function, and its output range is [0,1]. r This resets the door's bias.

[0104] Update Gate Z tIt is used to control the degree to which the previous hidden state retains the current state, and its calculation formula is as follows:

[0105] z t =σ(W z ·[h t-1 ,x t ]+b z )

[0106] In the formula, W z It is the weight matrix of the updated gate, b z It updates the bias term of the gate;

[0107] Candidate hidden state It is generated when the decision is made to update the current hidden state, and its calculation formula is:

[0108]

[0109] In the formula, ⊙ denotes element-wise multiplication, W is the weight matrix of the candidate state, and b is the bias term of the candidate state.

[0110] Current hidden state h t The update formula is:

[0111]

[0112] By stacking multiple GRU layers, the GRU units in subsequent layers not only obtain information from the previous layer, but also combine the hidden states of previous layers in their computation, forming a residual connection mechanism (inspired by ResNet).

[0113]

[0114] In the formula, This represents the current hidden state of the l-th layer, which is calculated based on the hidden states of the previous layer and the hidden states of the two preceding layers. This indicates the current hidden state of the (l-1)th layer, providing information about the previous layer. This indicates the current hidden state of layer l-2, providing information about earlier layers;

[0115] The fully connected layer will process the output h of the GRU layer. t Input a fully connected layer and perform a nonlinear mapping using the ReLU activation function:

[0116] h fc =ReLU(W fc ·h t +b fc )

[0117] In the formula, h fcW represents the output of the fully connected layer, specifically the activation value after the output from the GRU layer is passed to the fully connected layer. fc This represents the weight matrix of the fully connected layer, which determines the contribution of the input features to the output features. fc The bias term of the fully connected layer is used to adjust the output; the ReLU activation function maps negative values ​​to zero (i.e., (ReLU(x) = max(0,x))), introducing non-linear characteristics;

[0118] The output layer maps features to predicted values ​​through a fully connected layer, using a linear activation function.

[0119] y pred =W out ·h fc +b out

[0120] In the formula, y pred This is the final predicted surge height; W out The weight matrix of the output layer is used to map the features of the fully connected layer to the output values, b out The bias term of the output layer;

[0121] Step S22: Use mean squared error as the loss function for the regression task, and dynamically adjust the learning rate according to the gradient change;

[0122]

[0123] In the formula, L represents the mean squared error (MSE) calculated by the loss function, which measures the difference between the wave height predicted by the model and the actual value; N represents the total number of samples used to calculate the average loss. The actual surge height of the i-th sample The model predicts the surge height for the i-th sample;

[0124] Step S23: When new data is input, the weights are updated through the backpropagation algorithm. When different environmental conditions change, the functions of resetting and updating the gates are also adjusted through feedback.

[0125] Preferably, in step S21 of this embodiment, the GRU network structure design includes an input layer that receives landslide-related feature vectors. The GRU layer captures long-term and short-term dependencies in time-series data through reset and update gates. A fully connected layer maps the GRU output to a new feature space, generating a higher-level feature representation. The output layer ultimately maps to specific prediction results, such as the height of landslide surges. This effectively processes time-series data, utilizes the GRU's gating mechanism to prevent gradient vanishing, and enables the model to learn historical information and temporal features. Through the mapping of the fully connected layer, the model can better understand and summarize the complex relationships of the input features. Significance: Achieving accurate prediction of landslide surge height enhances the early warning capability for natural disasters and helps reduce risks. In step S22, the loss function is set using Mean Squared Error (MSE) as the loss function for the regression task. The learning rate is dynamically adjusted according to gradient changes. The MSE loss function provides a quantitative standard to evaluate the difference between the prediction and the true value, allowing the model to optimize based on actual performance. Dynamic learning rate adjustment helps improve the convergence speed and stability of training. Significance: By optimizing the loss function, the model's prediction accuracy is improved, thereby increasing the model's reliability and adaptability. Step S23, weight update and gating mechanism feedback, updates the network weights through backpropagation when new data is input. When environmental conditions change, the functions of the reset and update gates are adjusted through feedback. The backpropagation algorithm enables automatic optimization of model parameters, thereby improving adaptability to new data. The dynamic adjustment of the gating mechanism ensures that the GRU network can effectively capture the temporal characteristics of data under different environmental conditions. Significance: It improves the model's resistance to interference in changing environments, making it more sensitive to real-time input data and possessing stronger generalization ability.

[0126] In summary, the GRU network construction steps in this embodiment form a complete closed loop from feature input to model output, enabling accurate prediction of landslide surge height. In practical applications, it can be used in multiple fields such as natural disaster monitoring, environmental protection, and engineering safety, not only improving scientific prediction capabilities but also facilitating timely response measures in emergencies, thereby mitigating disaster losses.

[0127] Furthermore, the process of using a dynamic coupling strategy for resetting and updating the gate in step S21 specifically includes the following steps:

[0128] Step S211: Calculate the interaction between the reset gate and the update gate. When calculating the reset gate at the current time step, consider the combined information of the previous hidden state and the current input. Introduce the update gate from the previous time step, which will simultaneously refer to the current input and the previous hidden state, and retain the amount of information at the current time step.

[0129] Step S212: In the process of generating candidate hidden states, historical information is integrated. The generation process of candidate hidden states includes the interaction between the reset gate and the update gate. The weight of the reset gate affects the contribution of the previous hidden state in the candidate hidden state. Through the comprehensive analysis of historical information and new input, the candidate state reflects the characteristics of the current moment.

[0130] Step S213: The hidden state update process and its dynamic coupling effect. Through the dynamic coupling mechanism, the reset gate and the update gate jointly determine the flow and storage of information. When updating the hidden state at the current moment, the reset gate determines the degree to which the previous state needs to be forgotten. The update gate is responsible for determining how much information from the previous hidden state should be retained in the current hidden state. According to the dynamic coupling adjustment, when the update gate is high, past information is retained. When the update gate is low, the influence of historical information is suppressed, and the introduction of new information is allowed.

[0131] Preferably, step S211 of this embodiment calculates the interaction between the reset gate and the update gate. By combining the previous hidden state and the current input, the reset gate and the update gate enable the model to dynamically adjust the information flow according to the context. This operation increases the model's flexibility, allowing it to adaptively change its dependence on historical information based on different current inputs. Introducing the influence of the update gate from the previous time step on the reset gate enables the model to intelligently retain or forget information when processing continuous time series. For example, in some cases, the model may need to pay more attention to historical data, while in other cases it needs to accept new information more quickly. The significance is that through this step, the model can adjust its utilization of historical information under different input scenarios, thereby improving its adaptability to complex time series data patterns. This mechanism makes the model's predictions more accurate because it can dynamically balance the weights between old and new information. Step S212 integrates historical information during the generation of candidate hidden states. The interaction between the reset gate and the update gate during this process means that the previous state and the current input information are dynamically fused through the weights of the reset gate, improving the accuracy and representativeness of the candidate states. By integrating historical information with the current input, the candidate states can more comprehensively reflect the characteristics of the current moment, possessing better contextual understanding and thus better capturing changes in the time series. The significance is that the model can more meticulously capture short-term and long-term dependencies in the time series when updating states, improving the model's ability to process complex data. The accurate generation of candidate states makes information transmission in the recurrent neural network more efficient, reducing unnecessary information interference and improving model efficiency. Step S213 involves the hidden state update process and its dynamic coupling effect. The dynamic coupling mechanism used by the reset gate allows the model to determine the degree of information retention from the previous state at the current moment, effectively managing the flow and storage of information. Through dynamic coupling, the update gate is responsible for determining how much historical information needs to be retained, enabling the model to adapt to the current input and flexibly introduce or suppress the influence of historical information. Significance achieved: The interaction between the reset gate and the update gate enhances the model's memory capacity when processing information, allowing the model to more reasonably forget and retain information, thereby improving the understanding of time series; This dynamic update method makes the model more adaptable to the environment, and can still maintain high accuracy and robustness under constantly changing input conditions.

[0132] In summary, the dynamic coupling strategy employed in the reset and update gates of this embodiment achieves intelligent management of information flow and storage through three key steps. These steps, individually and collectively, enhance the model's flexibility, adaptability, and predictive ability, enabling GRU to achieve better performance when handling complex time series tasks. The dynamic coupling strategy not only optimizes the model's structural design but also significantly improves its performance in practical applications.

[0133] Furthermore, the process of comprehensively analyzing historical information and new inputs in step S212 specifically includes the following steps:

[0134] Step S2121: If the value of the reset gate is close to 0, it is assumed that historical information will be forgotten and the current calculation depends more on the new input than on the previous state; if the value of the reset gate is close to 1, the historical state is preserved.

[0135] Step S2122: Process the previous hidden state by weighting certain parts of the previous hidden state through the reset gate to obtain the modified historical information; integrate the historical information according to the current situation through the reset gate.

[0136] Step S2123: The current input is introduced. In the generation of candidate hidden states, the current input is combined with the modified historical information to form a new candidate state that reflects the characteristics of the current moment. This is a comprehensive processing of historical information and new input.

[0137] Preferably, the value of the reset gate in step S2121 of this embodiment has an impact. When the value of the reset gate is close to 0, the system effectively ignores past hidden states and mainly relies on the current input data, enabling the network to quickly adapt to sudden events or input scenarios with large changes. When the value of the reset gate is close to 1, the system can retain more historical information. In this case, the model will rely more on previous knowledge and incorporate the features of the previous state into the current calculation. The significance achieved is that the dynamic forgetting and retention mechanism makes the model more flexible in adaptability and resource utilization, and can flexibly adjust its dependence on historical information according to the characteristics of the data. In sequence data processing, it effectively improves the model's performance and enhances the stability and accuracy of the model when processing long sequence data. Step S2122, the weighted processing of historical information, is a weighted processing step for the previous hidden state. By using the reset gate, historical information is filtered and integrated to form modified historical information. Irrelevant or unimportant information can be removed, and only historical information that is helpful to the current task can be retained. The model can selectively retain and utilize previous states in various contexts, thereby providing a state representation that is more in line with the current input for subsequent calculations. Significance Achieved: Through this targeted processing method, the model can more accurately reflect the characteristics of the input data when generating candidate states; it avoids redundancy and noise in historical information, making the candidate states more informative, thereby helping the model better understand the current input and make corresponding decisions. Step S2123: Combining the current input with historical information, and combining the current input with modified historical information, forms a new candidate state, allowing the candidate state to fully reflect the characteristics of the current moment; it not only integrates historical information but also ensures the immediacy and relevance of the new input; the candidate state contains important information from the past and is also directly affected by the current input, thus achieving a dynamic state update. Significance Achieved: Effectively combining historical and current information enables the model to better capture time dependencies in sequence learning tasks. It gives the model stronger contextual understanding capabilities, enabling it to make more accurate predictions and judgments; this comprehensive processing method is the key to the model learning the essence of time series data, thereby improving the overall performance of tasks such as language modeling and time series prediction.

[0138] In summary, this embodiment constructs a flexible and adaptable model framework by dynamically integrating historical information with current input. This enables the model to fully utilize historical experience and real-time information when processing complex sequence data, effectively improving the performance of deep learning models in practical applications and enhancing the model's robustness and accuracy across various data environments.

[0139] Furthermore, such as Figure 6 As shown, the process of training the Gated Recurrent Unit (GRU) network model in step S3 specifically includes the following steps:

[0140] Step S31: Extract a batch of data samples from the training set for training the GRU network model, select a loss function and optimization algorithm to adjust the parameters of the GRU network model, and minimize the prediction error;

[0141] Step S32: In each iteration, the input data is provided to the GRU network model for forward propagation, and the prediction result is calculated; with a random sampling probability u (in a uniform distribution [0,1]), the acceptance probability of the current parameter update is determined according to the acceptance probability of the GRU network model parameters.

[0142] If u is less than the acceptance probability, accept the current parameter update (i.e., m). i =m * );

[0143] Otherwise, retain the model parameters from the previous step (i.e., m). i =m (i-1) );

[0144] Step S33: Repeat the above parameter update steps until the loss function converges to the minimum value or the preset number of training iterations is reached; evaluate the performance of the GRU network model using validation set data, and adjust the hyperparameters based on the performance of the GRU network model on the validation set; perform generalization tests by changing the datasets of the training set and validation set; evaluate the trained GRU network model using test set data.

[0145] Preferably, in step S31 of this embodiment, data sample extraction and parameter adjustment involves extracting a batch of data samples from the training set to ensure the diversity and representativeness of the input data, which is crucial for training the model. A loss function (such as mean squared error, cross-entropy, etc.) and optimization algorithm (such as Adam, SGD, etc.) are selected, and tools are used to define the model's learning objective and optimization method. Significance: Data sample extraction enables the model to learn the latent patterns in the data, while batch processing improves training efficiency. Choosing appropriate loss functions and optimization algorithms can guide the model to converge more effectively, reducing prediction errors and thus improving the model's accuracy and generalization ability. In step S32, forward propagation and parameter update decision-making, in each iteration, the input data undergoes forward propagation through the GRU model. The model evaluates its current performance by calculating the error between the network's output and the true label. Based on the probability u of random sampling and the acceptance probability (based on the evaluation of the current parameters), a decision is made on whether to update the parameters, introducing a random search mechanism to optimize the model parameters. Significance: The forward propagation process is the core of model training. Through this step, the model can learn and adjust its weights to better fit the training data. Introducing a randomized parameter update mechanism (such as the MCMC method) can explore the parameter space, avoid getting trapped in local optima, improve the model's exploration ability, make parameter updates more flexible, and enhance the model's robustness. Step S33 Iteration and Model Evaluation: Repeat the update steps to allow the model to gradually converge. Training stops when the loss function reaches its minimum or the preset number of training iterations are reached. Use the validation set to evaluate model performance, monitor the training effect of the model in real time, and provide feedback for hyperparameter tuning. Significance: Through multiple iterations, the model can be gradually optimized, improving its performance on the training data while controlling the risk of overfitting. The use of the validation set ensures the model's generalization ability. Through hyperparameter tuning, the model can be further optimized to improve its performance on unseen data. Generalization testing provides additional signals to verify the model's performance under various data conditions (different distributions, different features, etc.), which helps to enhance the model's confidence and reliability.

[0146] In summary, this embodiment, through the organic combination of the above steps, achieves the following key objectives in training the Gated Recurrent Unit (GRU) network model: using appropriate data and loss functions, the model extracts meaningful patterns and features from the data; through precise parameter update strategies, the model's performance on the test set is improved, ensuring not only good performance on the training set but also reliability in real-world applications; through rigorous validation and generalization testing, the model's robustness is ensured, improving performance under various practical conditions. This ensures the scientific and efficient training process, enabling the final GRU model to perform excellently in the task of predicting landslide surge height.

[0147] Furthermore, the process of minimizing the prediction error in step S31 specifically includes the following steps:

[0148] Step S311: After a batch of data samples is selected from the training set, the batch of data samples is input into the GRU model for processing; each sample is a fixed-length sequence, the sequence is represented as a vector, and word embeddings are applied to generate the corresponding vector representation for each time step;

[0149] Step S312: The update gate and candidate states are combined to generate new hidden states. The update gate determines how much of the previous state is retained and how much new information is introduced. The last hidden state is used to generate the final output.

[0150] Step S313: Use the loss function to evaluate the model's performance, calculate the current loss value, and calculate the error between the current model output and the actual label using the loss function;

[0151] The adaptive weighted cross-entropy loss function is expressed as follows:

[0152]

[0153] In the formula, L represents the loss value, and y true The one-hot encoded vector of the real label, y pred The probability distribution predicted by the model, the total number of n samples, the total number of C classes, w i The adaptive weights of the i-th sample are dynamically adjusted according to the characteristics of the sample. λ controls the weight hyperparameter of the regularization term. Regularization(w): Regularization function, which can be L1 or L2 regularization of the weights, used to prevent overfitting.

[0154] Preferably, in step S311 of this embodiment, the input data is fed into the GRU model by selecting a batch of data samples from the training set and converting these samples into fixed-length sequences, enabling the model to accept structured input. This allows GRU to effectively process time series or sequence data (such as text). Word embedding techniques (such as Word2Vec, GloVe, or BERT) are applied to generate corresponding vector representations for each time step. These vectors help express the semantic features of words, thus providing the model with richer contextual information. Through vectorization, the model processes input data more efficiently, helping to capture the relationships and semantics between words. The significance achieved is that by representing time series data as vector input, the GRU model can capture the interdependencies between different time steps in the sequence; this lays the foundation for subsequent information processing, enabling the model to better understand the dynamic changes in the data; word embedding converts discrete text data into vectors in a high-dimensional continuous space, allowing the model to mathematically perform linear combinations and nonlinear transformations, enhancing the model's expressive power. Step S312 combines the update gate with the candidate state to generate a new hidden state. Through the update gate mechanism, the model can selectively retain information from the previous hidden state while introducing new information. This allows GRU to dynamically adjust the retention and discarding of information flow according to the specific needs of the current input. By calculating the current candidate state and combining it with the output of the update gate, the model generates a new hidden state. When processing different types of inputs, GRU can adaptively adjust its internal state. Significance: GRU, through its unique gating design, addresses long-term dependency problems, overcoming the shortcomings of traditional RNNs that are prone to gradient vanishing or exploding on long sequences. Therefore, GRU can retain and utilize important information within a longer context. The dynamically updated hidden state can synthesize and integrate information across multiple time steps, thereby generating a deep understanding of sequence data, which is crucial for many sequence prediction tasks (such as language modeling and sentiment analysis). Step S313 uses a loss function to evaluate model performance. The model uses the loss function to assess the error between its predicted results and the actual labels. Common loss functions such as mean squared error (MSE) and cross-entropy loss fully reflect the model's performance and prediction accuracy. The loss value provides feedback signals to the model during training, enabling subsequent gradient updates to effectively reduce errors. The significance is that the calculation of the loss function guides the direction and extent of model optimization. Through backpropagation, the decrease in the loss value directly drives model parameter adjustment, ultimately making the model perform better on the target task. The loss function is not only used for gradient updates during training but also for monitoring model performance, including overfitting and underfitting. It provides a basis for further model adjustments (such as hyperparameter tuning and regularization selection).

[0155] In summary, this embodiment creates an efficient closed loop for data input processing, dynamic information integration, and predictive performance evaluation. Through these techniques, the GRU model can maintain flexibility when dealing with complex time series data, improving its accuracy and generalization ability, thereby minimizing prediction errors.

[0156] Furthermore, the process of performing generalization testing by changing the datasets of the training and validation sets in step S33 specifically includes the following steps:

[0157] Step S331: After the initial training, record the performance of each sample on the validation set, calculate the misclassification rate and cumulative loss value of each sample, and calculate its F1-score on the validation set; assign weights to samples based on their misclassification rates, and set a threshold to increase the weight of samples with a misclassification rate greater than 0.2 by 1.5 times.

[0158] Step S332: In the next round of training, the weights are directly applied to the loss function; adversarial examples are generated, the gradient of the loss function with respect to the input is calculated, and several adversarial examples are generated; the generated adversarial examples are incorporated into the validation set to re-evaluate the accuracy and robustness of the model.

[0159] Step S333: Define the main task (such as time series prediction) and related supplementary tasks (such as trend detection or anomaly detection) separately; design an independent sub-network for each task, sharing the bottom feature extraction layer of GRU; adopt alternating training during training, that is, update the shared layer of all tasks periodically, and update the supplementary task once every 5 epochs.

[0160] Step S334: Evaluate the performance of each task after each epoch and plot the loss change curve between different tasks; use TensorBoard to monitor the training process and record the loss, learning rate change and evaluation metrics for each epoch; if the validation set loss does not improve after 3 consecutive epochs, reduce the learning rate.

[0161] Preferably, step S331 of this embodiment, sample performance recording and weight allocation, involves evaluating the performance of validation set samples and recording the misclassification rate and cumulative loss of each sample. This allows the model to identify which samples contribute little to model training and which are boundary cases that are difficult to classify. Dynamic sample relabeling is performed based on the misclassification rate, focusing on those samples that trouble the model and increasing their importance in training. Significance: The weight allocation method guides the model to learn more effectively, significantly improving its performance when facing complex samples, thereby expanding the model's generalization ability; it allows the model to focus on imbalanced or boundary samples, improving the model's practicality in real-world scenarios and reducing the risk of overfitting. Step S332, adversarial example generation and re-evaluation, uses a strategy of generating adversarial examples to test the model's robustness to slight perturbations; during training, adversarial examples are used to further evaluate the model's sensitivity; the introduction of adversarial examples can reveal the model's potential weaknesses and shortcomings, making the model more resilient to real-world noise and attacks. Significance: By evaluating the model's performance on adversarial examples, its adaptability to uncertainty and anomalies can be enhanced, improving its stability and reliability in complex environments. It effectively strengthens the model's anti-interference capabilities, meaning it can better handle more challenging and complex data scenarios in practical applications. Step S333: Multi-task learning architecture design separates the main task and supplementary tasks, ensuring each task can be optimized independently while sharing underlying features at the hierarchical level. This allows for improved training efficiency through complementary learning between different tasks. The alternating training strategy allows shared layers to maintain consistency while providing timely updates for supplementary tasks, strengthening the synergistic effect between tasks. Significance: The multi-task learning strategy enhances the model's flexibility, especially when facing new tasks or datasets, enabling the model to quickly adapt and effectively transfer existing experience. During the same training process, sharing features avoids redundant computation, improving training speed and achieving better performance with limited resources. Step S334, Training Process Monitoring and Dynamic Adjustment, involves real-time monitoring of the training process, recording the loss, learning rate, and other evaluation metrics for each epoch. This helps to understand the model's learning status and monitor the training progress in a timely manner. Setting dynamic adjustment rules for the learning rate ensures that the model can adjust promptly under adverse training conditions (such as overfitting), thereby maintaining model performance. Significance: This ensures efficient training and avoids wasted computation time and resources. Dynamically adjusting the learning rate allows for adaptive adjustment based on actual training conditions, increasing the chance of successful convergence. The monitoring system provides a feedback mechanism for the model, making model training more controllable, improving the ability to identify potential problems during training, and enhancing the transparency of the training process.

[0162] In summary, this embodiment demonstrates how the organic integration of each step in the generalization testing process enhances the model's learning ability and application performance in complex environments. It not only emphasizes performance improvement but also strengthens its adaptability and stability in practical scenarios, ultimately improving the model's generalization ability and usability. This systematic and targeted approach lays a solid foundation for building more powerful machine learning models.

[0163] like Figure 7 As shown, this embodiment also provides an embodiment of a landslide surge height prediction system based on a gated loop unit. In this embodiment, the landslide surge height prediction system based on a gated loop unit is applied to the landslide surge height prediction method based on a gated loop unit as described in the above embodiment. The landslide surge height prediction system based on a gated loop unit includes a data preprocessing module 1, a model building module 2, and a model training module 3 that are electrically connected in sequence.

[0164] The data preprocessing module 1 collects relevant data from large-scale physical model tests in the landslide surge area. This data includes landslide volume, velocity, slope angle, and characteristic data of similar materials used to simulate real soil and rock materials. The preprocessed data undergoes data standardization and normalization, and is then used as input to the gated recurrent unit (GRU) model to obtain the surge height. The model building module 2 uses TensorFlow in a Python compilation environment to build a GRU network. Update and reset gates are introduced to control the flow of information. The update gate determines how much information from the previous state should be retained in the current state, while the reset gate controls how much information from previous states should be forgotten. The model training module 3 trains the GRU network model using historical landslide data to adjust model parameters. The trained GRU model is then deployed to the actual landslide surge prediction system for real-time monitoring.

[0165] Preferably, the data preprocessing module 1 in this embodiment collects important feature data related to landslides, such as landslide volume, landslide velocity, slope angle, and characteristics of similar materials; making the information input to the model as detailed and diverse as possible; standardizing and normalizing the feature data to eliminate the influence of different dimensions between features, making the data more suitable for training deep learning models. This step ensures the balance of data, making the training process more stable. Significance: Through effective data preprocessing, the sensitivity of the model to outliers or scale differences is reduced, thereby improving the model's learning ability and prediction accuracy; ensuring a clear and reliable data foundation for subsequent model construction, which helps to improve the model's prediction performance and generalization ability. The model building module 2 uses TensorFlow to build a GRU network, combined with update gate and reset gate design, which can effectively capture long-term and short-term dependencies in time series data; the update gate controls the retained information, while the reset gate controls the forgotten information. In this way, the model can efficiently process historical data and enhance information extraction capabilities. Significance: A well-designed GRU structure enhances the modeling capability for complex time-series data, meeting the needs of landslide surge height prediction; it improves the model's adaptability to different input data, resulting in more accurate predictions and reduced prediction errors. Model training module 3 uses historical landslide data for model training and employs backpropagation to adjust model parameters to minimize prediction errors; it can be continuously updated and trained with new landslide data, maintaining good predictive ability. Significance: Training with historical data ensures that the trained GRU model effectively reflects real landslide surge dynamics, possessing a certain degree of accuracy and reliability; deploying the trained model to a real-time landslide surge prediction system allows for timely responses between monitoring data and model output, providing decision support and improving early warning capabilities.

[0166] In summary, this embodiment of the landslide surge height prediction system based on gated cyclic units aims to achieve real-time monitoring and prediction of complex landslide activities through effective data preprocessing, sophisticated model construction, and efficient model training. Each module provides a scientific basis and strong technical support for landslide disaster prevention and mitigation, reducing disaster risks and providing necessary data support and decision-making assistance to related fields. This comprehensive approach not only improves the accuracy and reliability of the model but also enhances the system's practicality and response speed, enabling it to play a crucial role in real-world applications.

[0167] like Figure 8 As shown, this embodiment provides an embodiment of an electronic device 4, which includes a processor 41 and a memory 42 coupled to the processor 41.

[0168] The memory 42 stores program instructions for implementing the landslide surge height prediction method based on gated loop units in any of the above embodiments.

[0169] The processor 41 is used to execute program instructions stored in the memory 42 to perform landslide surge height prediction based on gated loop units.

[0170] The processor 41 can also be referred to as a CPU (Central Processing Unit). The processor 41 may be an integrated circuit chip with signal processing capabilities. The processor 41 can also be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. A general-purpose processor can be a microprocessor or any conventional processor.

[0171] Furthermore, Figure 9 This is a schematic diagram of the structure of a storage medium according to an embodiment of this application. The storage medium 5 of this embodiment stores program instructions 51 capable of implementing all the above methods. These program instructions 51 can be stored in the storage medium in the form of a software product, including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks, or terminal devices such as computers, servers, mobile phones, and tablets.

[0172] like Figure 10 As shown, the physical model test data is compared with the predicted values ​​of the gated cyclic unit method; Figure 11 The attached figure shows the results of K-fold cross-validation (K=5). As can be seen from the figure, the experimental data from the physical model show a high degree of overlap with the predicted values, proving that the model's predictions are highly accurate.

[0173] In the several embodiments provided by this invention, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be an indirect coupling or communication connection between apparatuses or units through some interfaces, and may be electrical, mechanical, or other forms.

[0174] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated units described above can be implemented in hardware or as software functional units. The above are merely embodiments of the present invention and do not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

[0175] The specific embodiments of the invention have been described in detail above, but these are merely examples, and the invention is not limited to the specific embodiments described above. For those skilled in the art, any equivalent modifications or substitutions to the invention are also within the scope of this invention. Therefore, all equivalent transformations, modifications, and improvements made without departing from the spirit and principles of this invention should be included within the scope of this invention.

Claims

1. A method for predicting landslide surge height based on gated cyclic units, characterized in that, The landslide surge height prediction method based on gated cyclic units includes: Data from large-scale physical model tests in the landslide surge area were collected, and the characteristic data were preprocessed by data standardization and normalization. The preprocessed characteristic data were used as input to the gated recurrent unit (GRU) model to obtain the surge height. In a Python compilation environment, a gated recurrent unit (GRU) network is built using tensor flow. Update gates and reset gates are introduced to control the flow of information. The update gate determines how much information from the previous state should be retained in the current state, while the reset gate controls how much information from the previous state should be forgotten. A gated recurrent unit (GRU) network model was trained using historical landslide data to adjust the model parameters. The trained GRU model was then deployed to an actual landslide surge prediction system for real-time monitoring.

2. The landslide surge height prediction method based on a gated cyclic unit according to claim 1, characterized in that, The relevant data includes landslide volume, landslide velocity, slope angle, and characteristic data of similar materials used to simulate the properties of real soil and rock materials.

3. The landslide surge height prediction method based on a gated cyclic unit according to claim 1, characterized in that, Constructing a gated recurrent unit (GRU) network includes: Design the input layer, GRU layer, fully connected layer, and output layer of a gated recurrent unit (GRU) network; Mean squared error is used as the loss function for the regression task, and the learning rate is dynamically adjusted according to gradient changes. When new data is input, the weights are updated through the backpropagation algorithm. When different environmental conditions change, the functions of resetting and updating the gates are also adjusted through feedback.

4. The landslide surge height prediction method based on a gated cyclic unit according to claim 3, characterized in that, The input layer is used to input feature vectors. The GRU layer captures long-term and short-term dependencies in time series data through a gating mechanism. The fully connected layer maps the output of the GRU layer to a new feature space and combines the time series information extracted from the GRU layer to generate a higher-level feature representation. The output layer maps the output of the fully connected layer to the specific prediction result, the height of the landslide surge.

5. The landslide surge height prediction method based on a gated cyclic unit according to claim 3, characterized in that, in, The input layer sets the input feature vector x t for: x t =[x t,1 ,x t,2 ,…,x t,n ] In the formula, n represents the number of features, including landslide-related parameters such as landslide volume, rate, and slope; environmental information such as rainfall, soil moisture, and temperature; and data on historical landslide events. The input data needs to be organized as a three-dimensional tensor: In the formula, m is the number of samples, T is the time step, and n is the number of features at each time step; GRU layer reset gate r t It is an important gating mechanism used to control how much historical information should be forgotten when calculating a new hidden state. Its calculation formula is: r t =σ(W r ·[h t-1 ,x t ]+b r ) In the formula, W r h represents the weight matrix of the reset gate. t-1 It represents the hidden state of the previous time step, σ(·) is the sigmoid activation function, and its output range is [0,1]. r This resets the door's bias. Update Gate Z t It is used to control the degree to which the previous hidden state retains the current state, and its calculation formula is as follows: z t =σ(W z ·[h t-1 ,x t ]+b z ) In the formula, W z It is the weight matrix of the updated gate, b z It updates the bias term of the gate; Candidate hidden state It is generated when the decision is made to update the current hidden state, and its calculation formula is: In the formula, ⊙ denotes element-wise multiplication, W is the weight matrix of the candidate state, and b is the bias term of the candidate state. Current hidden state h t The update formula is: By stacking multiple GRU layers, the GRU units in subsequent layers not only obtain information from the previous layer, but also combine the hidden states of the previous layers in their computation, forming a residual connection mechanism, i.e.: In the formula, This represents the current hidden state of the l-th layer, which is calculated based on the hidden states of the previous layer and the hidden states of the two preceding layers. This indicates the current hidden state of the (l-1)th layer, providing information about the previous layer. This indicates the current hidden state of layer l-2, providing information about earlier layers; The fully connected layer will process the output h of the GRU layer. t Input a fully connected layer and perform a nonlinear mapping using the ReLU activation function: h fc =ReLU(W fc h t +b fc ) In the formula, h fc W represents the output of the fully connected layer, specifically the activation value after the output from the GRU layer is passed to the fully connected layer. fc This represents the weight matrix of the fully connected layer, which determines the contribution of the input features to the output features. fc The bias term of the fully connected layer is used to adjust the output; the ReLU activation function maps negative values ​​to zero, i.e., ReLU(x) = max(0,x), introducing non-linear characteristics; The output layer maps features to predicted values ​​through a fully connected layer, using a linear activation function. y pred =W out ·h fc +b out In the formula, y pred This is the final predicted surge height; W out The weight matrix of the output layer is used to map the features of the fully connected layer to the output values, b out The bias term of the output layer.

6. The landslide surge height prediction method based on a gated cyclic unit according to claim 3, characterized in that, Loss function: In the formula, L is the mean squared error calculated by the loss function, which measures the difference between the wave height predicted by the model and the actual value, and N is the total number of samples used to calculate the average loss. The actual surge height of the i-th sample The model predicts the surge height for the i-th sample.

7. The landslide surge height prediction method based on a gated cyclic unit according to claim 1, characterized in that, The reset and update gates employ a dynamically coupled strategy, including: The interaction between the reset gate and the update gate is calculated. When calculating the reset gate at the current time step, the combined information of the previous hidden state and the current input is considered. The update gate from the previous time step is introduced, which simultaneously refers to the current input and the previous hidden state, and the amount of information retained at the current time step. Historical information is integrated during the generation of candidate hidden states. Through comprehensive analysis of historical information and new inputs, the candidate states reflect the characteristics of the current moment. The hidden state update process and its dynamic coupling effect: the update gate is responsible for determining how much information from the previous hidden state should be retained in the current hidden state.

8. The landslide surge height prediction method based on a gated cyclic unit according to claim 7, characterized in that, The generation process of candidate hidden states involves the interaction of reset gate and update gate. The weight of the reset gate affects the contribution of the previous hidden state in the candidate hidden state.

9. The landslide surge height prediction method based on a gated cyclic unit according to claim 7, characterized in that, Through a dynamic coupling mechanism, the reset gate and the update gate jointly determine the flow and storage of information; when updating the hidden state at the current moment, the reset gate determines the degree to which the previous state needs to be forgotten.

10. A landslide surge height prediction system based on a gated cyclic unit, applied to the landslide surge height prediction method based on a gated cyclic unit as described in any one of claims 1 to 9, characterized in that, The landslide surge height prediction system based on gated loop units includes: The data preprocessing module is used to collect relevant data from large-scale physical model tests in the landslide surge area, and to perform data standardization and normalization preprocessing on the feature data. The preprocessed feature data is used as input to the gated recurrent unit (GRU) model to obtain the surge height. The model building module is used to build a gated recurrent unit (GRU) network using tensor flow in a Python compilation environment. Update gates and reset gates are introduced to control the flow of information. The update gate determines how much information from the previous state should be retained in the current state, while the reset gate controls how much information from the previous state should be forgotten. The model training module is used to train the gated recurrent unit (GRU) network model. It uses historical landslide data to train the GRU network model and adjust the model parameters. The trained GRU model is then deployed to the actual landslide surge prediction system for real-time monitoring.

Images