Secondary landslide base friction coefficient prediction method and system based on deep neural network model

Abstract

The application discloses a secondary landslide base friction coefficient prediction method and system based on a deep neural network model, and the method comprises the following steps: obtaining a multi-factor data set of a real historical landslide; constructing a deep neural network model; training the deep neural network model by using the multi-factor data set to obtain a trained deep neural network model; obtaining a base friction coefficient prediction result; and obtaining a final base friction coefficient prediction result based on the base friction coefficient prediction result. The application automatically learns the nonlinear mapping relationship between factors through a deep learning model, realizes the change of landslide parameters from 'post-calibration' to 'pre-prediction', shortens the time consumption of parameter determination from 2 hours to less than 30 seconds, greatly improves the automation level and timeliness of parameter determination, realizes the dynamic deduction and rapid update of the secondary landslide disaster process, and provides scientific and timely technical support for on-site personnel evacuation and decision-making command.

Description

Technical Field

[0001] This invention belongs to the field of disaster prevention and dynamic simulation technology, and particularly relates to a method and system for predicting the base friction coefficient of secondary landslides based on a deep neural network model. Background Technology

[0002] During landslide disaster emergency rescue, the "secondary landslide" formed by the instability of the deformed body after the first landslide poses a significant threat to the safety of rescue personnel and equipment. Accurately predicting its sliding distance is the key to assessing the scope of the disaster and avoiding risks.

[0003] Currently, numerical simulations based on dynamic processes (such as Massflow) are the core method for predicting landslide movement patterns. However, the accuracy of the simulation results is highly dependent on the base friction coefficient. The possible values. Existing... The values ​​are mainly obtained by empirical assignment, that is, initial values ​​are given based on landslide volume and terrain slope, and then the actual sliding distance is fitted by repeated manual calculations and calibrations.

[0004] This "manual inversion" model has obvious drawbacks: first, the assignment of experience values ​​is highly subjective and uncertain; second, the repeated trial calculation process is time-consuming and cannot meet the urgent needs of emergency rescue for timeliness and dynamic simulation.

[0005] How to transform from experience-based inversion to intelligent prediction is a technical problem that urgently needs to be solved. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for predicting the friction coefficient of a secondary landslide base based on a deep neural network model, and to solve the technical problem of how to transform from empirical inversion to intelligent prediction.

[0007] This application provides a method for predicting the base friction coefficient of a secondary landslide based on a deep neural network model, comprising: S01. Obtain a multi-factor dataset of real historical landslides; S02, Construct a deep neural network model; S03. Train the deep neural network model using a multi-factor dataset to obtain a well-trained deep neural network model. S04. Input the multi-factor values ​​of the landslide whose base friction coefficient is to be measured, which correspond to the feature dimensions of the multi-factor dataset of real historical landslides, and obtain the base friction coefficient prediction results. S05. Based on the predicted base friction coefficient, perform a dynamic closed-loop iterative process of prediction-simulation-feedback to obtain the final predicted base friction coefficient.

[0008] Optionally, in some embodiments of this application, the multi-factor dataset includes landslide volume, slope angle, thickness, density, cohesion, internal friction angle, rainfall, pore water pressure, vertical drop height, and landslide movement distance; and / or The multi-factor dataset of real historical landslides was used to remove missing values ​​and outliers. Based on physical relationships, 24 landslide-specific derived features were constructed. These 24 landslide-specific derived features include: elevation difference-slip ratio, slope tangent, friction angle tangent, effective stress, stress ratio, shear strength proxy, thickness-slip ratio, volume per unit thickness, kinematic index, mass, potential energy, kinetic energy proxy, saturation index, water content proxy, slope-thickness coupling, cohesion-density coupling, friction angle-volume coupling, elevation difference-slip ratio squared, elevation difference-slip ratio cubed, slope cubed, stress ratio squared, volume logarithm, mass logarithm, and thickness square root.

[0009] Optionally, in some embodiments of this application, the deep neural network model includes an input projection layer, a residual block stacking layer, an attention mechanism layer, and an output projection layer; and / or A visualization interface is built based on the trained deep neural network model. The multi-factor values ​​of the landslide base friction coefficient, which correspond to the feature dimensions of the multi-factor dataset of real historical landslides, are input into the geometric parameters of the visualization interface to obtain the base friction coefficient prediction results.

[0010] Optionally, in some embodiments of this application, a multi-factor dataset of real historical landslides, 24 landslide-specific derived features, and a corrected base friction coefficient are used. The processed dataset is divided into training and testing sets in an 8:2 ratio to train and test the deep neural network model.

[0011] Optionally, in some embodiments of this application, ; In the formula, The noise term, which simulates random disturbances in a natural system, follows a normal distribution. , H This refers to the vertical drop of the landslide. L This is the gliding distance. For comprehensive correction factors; ; In the formula, This is a landslide volume correction factor. Pore ​​water pressure correction factor This is a correction factor for the cohesion of slip zone soil / rock. This is a correction factor for the internal friction angle of slip zone soil / rock. This is the correction factor for the average slope of the sliding surface. This is a correction factor for the average thickness of the landslide mass. This is the slide correction factor. This is a rainfall correction factor. ; In the formula, Let V be the cutoff function, and V be the landslide volume. ; In the formula, Pore ​​water pressure; ; In the formula, It refers to the cohesion or bond strength of the slip zone soil / rock; ; In the formula, The internal friction angle of the slip zone soil / rock; ; In the formula, α is the average slope of the sliding surface; ; In the formula, h is the average thickness of the landslide body; ; In the formula, ρ is the density of the sliding body; ; In the formula, R represents the rainfall.

[0012] Optionally, in some embodiments of this application, the multi-factor values ​​of the landslide with the base friction coefficient to be measured, corresponding to the feature dimensions of the multi-factor dataset of real historical landslides, are substituted into a trained deep neural network model to obtain an initial base friction coefficient prediction value μ1. S052. Input the multi-factor values ​​and μ1 of the landslide with the base friction coefficient to be measured, which correspond to the feature dimensions of the multi-factor dataset of real historical landslides, into the Massflow modeling system, set the simulation parameters, construct the landslide dynamics model, extract the intermediate group landslide movement distance L1 and vertical descent height H1 after simulation in ArcGIS, and substitute them into the trained base friction coefficient prediction model to obtain the intermediate prediction value μ2. S053. Update the intermediate groups L1 and H1 to the multi-factor values. Repeat S052 until the absolute value of the difference between adjacent intermediate predicted values ​​is less than or equal to 0.01. Then stop the iteration and output the final predicted value.

[0013] Optionally, in some embodiments of this application, the simulation parameters include inputting landslide density, pore water pressure coefficient, cohesion, and the time required for the landslide to complete its sliding. The formula for the pore water pressure coefficient is as follows: ru =u / (ρgH t / 1000); In the formula, u is the pore water pressure, ρ is the density of the sliding body, and g is the acceleration due to gravity, taken as 9.81 m / s². 2 , H t For the thickness of the sliding body; and / or The maximum number of iterations shall not exceed 5.

[0014] Accordingly, this application also provides a system for predicting the base friction coefficient of secondary landslides based on a deep neural network model, including: a multi-factor dataset acquisition module for obtaining multi-factor datasets of real historical landslides; The network model building module is used to build deep neural network models; The training module uses a multi-factor dataset to train a deep neural network model, resulting in a well-trained deep neural network model. The prediction module is used to input the multi-factor values ​​of the landslide base friction coefficient to be measured, which correspond to the feature dimensions of the multi-factor dataset of real historical landslides, and to obtain the base friction coefficient prediction result. The final substrate friction coefficient prediction module performs a dynamic closed-loop iterative process of prediction-simulation-feedback based on the substrate friction coefficient prediction results to obtain the final substrate friction coefficient prediction results.

[0015] Accordingly, embodiments of this application also provide a computer device, including a storage device and a processor, wherein the storage device stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the method described above.

[0016] Accordingly, embodiments of this application also provide a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the method described above.

[0017] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: The present invention provides a method for predicting the base friction coefficient of secondary landslides based on a deep neural network model. First, a multi-factor dataset of real historical landslides is collected, including landslide volume, slope angle, thickness, density, cohesion, internal friction angle, rainfall, pore water pressure, vertical drop height, and landslide movement distance. Then, a deep neural network model integrating residual neural networks and attention mechanisms is constructed. The collected data is cleaned, feature-engineered, enhanced, and normalized before model training to obtain optimal weights. Subsequently, the trained model is used to predict and verify landslide samples outside the training set. Finally, a visualization interface is built to provide real-time and intuitive display of multi-factor numerical inputs, base friction coefficient prediction results, and landslide characteristics. This invention utilizes a deep neural network model suitable for predicting the base friction coefficient of secondary landslides to address the landslide distance prediction problem, significantly improving the accuracy of landslide dynamics simulation compared to existing methods.

[0018] Specifically, the advantages and features of this invention are mainly reflected in the following aspects: First, it represents a technological leap from "manual inversion" to "intelligent prediction": Traditional methods rely on empirical assignment, typically requiring repeated calculations and calibrations of parameters after a landslide to fit the actual sliding distance. This process is highly subjective and time-consuming. This invention, through a deep learning model, automatically learns the nonlinear mapping relationships between factors, transforming landslide parameters from "post-accident calibration" to "pre-accident prediction." Parameter determination time has been reduced from 2 hours to less than 30 seconds, significantly improving the automation and timeliness of parameter determination.

[0019] Second, it possesses extremely high prediction accuracy and convergence stability: This invention introduces skip connections through a residual neural network, effectively solving the gradient vanishing problem in deep network training. The model validation set shows a stable R² above 0.90, RMSE < 0.008, MAPE < 3%, and prediction bias ≤ ±0.016 in multi-case validation, achieving an accuracy improvement of over 40% compared to the traditional empirical assignment method.

[0020] Third, it integrates deep learning with physical constraint logic: This invention introduces an attention mechanism that adaptively focuses on key physical factors, successfully identifying core features such as the H / L ratio and effective stress, with an average attention weight as high as 0.9. This "data-physics" dual-constraint modeling approach ensures that the prediction results conform to the dynamic laws of geological hazards, significantly improving the physical consistency of the simulation results.

[0021] Fourth, it provides real-time dynamic support for emergency rescue decision-making: By outputting the predicted μ value in real time through a visual interface, the Massflow modeling system and ArcGIS output the slip distance, velocity field and accumulation range in real time, realizing the dynamic simulation and rapid updating of the secondary landslide disaster process, and providing scientific and timely technical support for on-site personnel evacuation and decision-making. Attached Figure Description

[0022] Figure 1 This is a flowchart of the method for predicting the friction coefficient of a secondary landslide base based on a deep neural network model, as described in this invention. Figure 2 This is a graph showing the deep learning loss curve of this invention; Figure 3 This is the R² curve for deep learning in this invention; Figure 4 This is the attention distribution-important feature recognition map of the model in this invention; Figure 5 This is a comparison chart of the ranking predictions of the present invention; Figure 6 This is a map showing the accumulation range and thickness calculated using ArcGIS in this invention; Figure 6 In the equation, a represents the time when t=0 seconds, and b represents the time when t=50 seconds. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0024] The technical solution of this application is as follows: Firstly, please refer to Figure 1 This application provides a method for predicting the base friction coefficient of a secondary landslide based on a deep neural network model, including: S01. Obtain a multi-factor dataset of real historical landslides; S02, Construct a deep neural network model; S03. Train the deep neural network model using a multi-factor dataset to obtain a well-trained deep neural network model. S04. Input the multi-factor values ​​of the landslide whose base friction coefficient is to be measured, which correspond to the feature dimensions of the multi-factor dataset of real historical landslides, and obtain the base friction coefficient prediction results. S05. Based on the predicted base friction coefficient, perform a dynamic closed-loop iterative process of prediction-simulation-feedback to obtain the final predicted base friction coefficient.

[0025] In S01: In some embodiments, the multi-factor dataset includes landslide volume, slope angle, thickness, density, cohesion, internal friction angle, rainfall, pore water pressure, vertical drop height, and landslide movement distance.

[0026] For example, landslide volume: the total volume of the sliding body that has become unstable, which is calculated by generating the difference between the pre-disaster and post-disaster digital elevation models through UAV aerial survey or lidar. Slope angle: The average inclination angle of the sliding surface, obtained through high-precision DEM data and GIS spatial analysis; Thickness: The average normal thickness of the landslide body is determined by geophysical exploration methods such as on-site drilling, trench excavation, or ground-penetrating radar, combined with topographic profiles. Density: The natural physical density of the landslide soil and rock mass, obtained by taking undisturbed samples on site and conducting indoor geotechnical tests; Cohesion: The mutual attraction and cementation between adjacent particles within a rock or soil mass, determined by in-situ shear tests or indoor triaxial compression tests; Internal friction angle: reflects the frictional resistance characteristics between particles inside the rock and soil mass when sliding or rolling, and is obtained through shear test together with cohesion; Rainfall: The cumulative precipitation depth in the two months prior to the occurrence of the secondary landslide was obtained through monitoring data from on-site meteorological stations near the disaster site or from the regional meteorological station network; Pore ​​water pressure: The pressure exerted by groundwater on solid particles inside the pores of rock and soil, which is obtained through real-time monitoring by piezometers installed on site or estimated based on groundwater level monitoring data; Vertical drop height: The elevation difference between the highest point at the rear edge of the landslide source area and the lowest point at the front edge of the accumulation area, extracted from DEM data or on-site surveying; Landslide travel distance: The total horizontal projection of the landslide body sliding down from the unstable zone to the leading edge of the final accumulation zone, obtained through on-site investigation or post-disaster high-resolution UAV imagery.

[0027] In S02: In some embodiments, the deep neural network model includes an input projection layer, a residual block stacking layer, an attention mechanism layer, and an output projection layer.

[0028] Furthermore, the input projection layer includes Linear(34,512), BatchNorm1d(512), ReLU activation function, and Dropout(0.3).

[0029] Furthermore, the residual block stack layer includes three residual blocks connected in sequence, each residual block containing two Linear layers, two BatchNorm1d layers, and a Dropout layer.

[0030] It is understandable that skip connections are used to solve the gradient vanishing problem; the three residual blocks can capture the deep correlation of landslide features without losing physical meaning due to excessive network depth.

[0031] Alternatively, it can be understood that this application uses a strict dimensionality reduction sequence of 512-256-128-64 to forcibly remove environmental noise from the original landslide data layer by layer, mapping the 34-dimensional physical constraint feature vectors to a 512-dimensional high-dimensional feature space and then compressing them into the core mechanical features of 64 dimensions.

[0032] Furthermore, the attention mechanism layer includes Linear(64,16), ReLU activation function, Linear(16,64), and Sigmoid activation function.

[0033] It is understandable that ReLU can fit the nonlinear weight mapping of landslide physical characteristics very well.

[0034] Alternatively, it can be understood that by fixing the attention mechanism layer (through a 64-16-64 dimension mapping) in series with the deepest 64-dimensional residual output after three dimensionality reductions, the contribution of the "comprehensive mechanical state" to the friction coefficient can be accurately evaluated after all physical factors have completed deep nonlinear coupling, thereby locking in and amplifying the weight of the core physical factors (such as the H / L ratio) that determine landslide dynamics.

[0035] The upstream input (residual network) of this application: The residual network performs complex manifold compression on the original 34-dimensional features, outputting a "coarse" 64-dimensional latent space feature vector. Although the features are abstract at this point, they are still mixed with environmental noise.

[0036] This layer (feature self-attention mechanism module) processes: Through internal computation, it generates a set of 64-dimensional weight coefficient matrices (i.e., Figure 4 The height of the pillars in the diagram is used to perform element-wise multiplication (Hadamard product) with the original 64-dimensional features.

[0037] The highly deterministic 64-dimensional latent mechanical feature vector after adaptive physical saliency modulation (the feature clusters presented by the attention weight ranking are only a visual representation of the model's adaptive physical focusing ability, and do not remove other low-weight factors, because landslide is a complex nonlinear dynamic system, and low weights represent its existence as "background environment". The attention mechanism preserves the complete ecology of the data, only adaptively allocating it at each decision.) Downstream execution (regression prediction output layer): After receiving the features "purified and amplified" by the attention module, the spatial shrinkage is directly performed through the linear mapping layer (e.g., 64 → 32 → 1), and finally outputs a unique regression prediction scalar value - the basis friction coefficient.

[0038] In S03: In some embodiments, the multi-factor dataset of real historical landslides is subjected to missing value removal and outlier pruning, and 24 landslide-specific derived features are constructed through physical relationships. The 24 landslide-specific derived features include elevation difference-slip ratio, slope tangent, friction angle tangent, effective stress, stress ratio, shear strength proxy, thickness-slip ratio, volume per unit thickness, kinematic index, mass, potential energy, kinetic energy proxy, saturation index, water content proxy, slope-thickness coupling, cohesion-density coupling, friction angle-volume coupling, elevation difference-slip ratio square, elevation difference-slip ratio cube, slope cube, stress ratio square, volume logarithm, mass logarithm, and thickness square root.

[0039] Furthermore, the elevation difference-slip ratio is calculated using the formula: F1 = H / L, where H is the vertical drop in meters (m) and L is the slip distance in meters (m); the slope tangent is calculated using the formula: F2 = tan(θ), where θ is the average inclination angle of the sliding surface in degrees (°); the friction angle tangent is calculated using the formula: F3 = tan(φ), where φ is the internal friction angle in degrees (°); and the effective stress is calculated using the formula: F4 = c u, where c is cohesion (kPa) and u is pore water pressure (kPa); stress ratio, formula: F5=u / c, where c is cohesion (kPa) and u is pore water pressure (kPa); shear strength ratio, formula: F6=c+α·φ, where α is an empirical coefficient (0.5), c is cohesion (kPa), and φ is the internal friction angle (°); thickness-slip ratio, formula: F7=H t / L, where H t The thickness of the sliding body is in meters (m), and the sliding stroke is in meters (m). The volume per unit thickness is calculated using the formula: F8 = V / H t In the formula, V is the landslide volume in m³, and H... t The thickness of the sliding body is in meters (m); the kinematic performance index is calculated using the formula: F9 = L / H, where L is the sliding distance in meters (m) and H is the vertical drop in meters (m); the mass is calculated using the formula: F... 10 =ρ·V, where ρ is density (kg / m³), V is landslide volume (m³), and potential energy is expressed as F. 11 =ρ·V·g·H, where ρ is density (kg / m³), V is landslide volume (m³), g is gravitational acceleration (m / s²), and H is vertical drop (m); kinetic energy: F 12=ρ·V·H 2 In the formula, ρ is the density (kg / m³), V is the landslide volume (m³), H is the vertical drop (m), and the saturation index is F. 13 =R·u, where R is the rainfall (mm), u is the pore water pressure (kPa), and the water content is expressed as F. 14 =R·H t In the formula, R is the rainfall in mm, and H is the total rainfall. t The thickness of the sliding body is in meters (m); the slope-thickness coupling formula is: F 15 =tan(θ)·H t In the formula, θ is the average inclination angle of the sliding surface, in degrees, and H t The thickness of the sliding body is in meters; the cohesion-density coupling formula is: F 16 =c·ρ, where ρ is density (kg / m³) and c is cohesion (kPa); friction angle-volume coupling formula: F 17 =φ·log10(V+1), where φ is the internal friction angle in °, V is the landslide volume in m³, and F is the square of the elevation difference and the sliding distance. 18 =(H / L) 2 In the formula, H is the vertical drop in meters (m), L is the sliding distance in meters (m); the ratio of elevation difference to sliding distance is expressed as the cube of F. 19 =(H / L) 3 In the formula, H is the vertical drop (in meters), L is the sliding distance (in meters), and the slope is cubic, expressed as: F 20 =θ 2 In the formula, θ is the average inclination angle of the sliding surface, in degrees; the stress ratio squared is given by: F 21 = (u / c) 2 In the formula, u is the pore water pressure in kPa, c is the cohesion in kPa, and the volume logarithm is given by: F 22 =log10(V+1), where V is the landslide volume in m³; F is the logarithm of mass, calculated using the formula: 23 =log10(ρ·V+1), where ρ is the density (kg / m³), V is the landslide volume (m³), and the thickness is the square root of the thickness, calculated using the formula: F 24 =H t In the formula, H t The thickness of the sliding body is expressed in meters (m).

[0040] In some embodiments, a multi-factor dataset of real historical landslides, 24 landslide-specific derived features, and a corrected base friction coefficient are used. The processed dataset is divided into training and testing sets in an 8:2 ratio to train and test the deep neural network model.

[0041] Figure 2 The deep learning loss curves of this invention show the detailed changes in training and validation losses. Specifically, both curves exhibit a typical logarithmic decreasing trend, meaning that the model's loss value gradually and significantly decreases as training progresses. This trend not only reflects the model's gradual learning and optimization of its parameters during training but also indicates the effectiveness of the training data. Particularly after approximately 150 epochs of training, the training and validation loss curves clearly stabilize, no longer exhibiting significant fluctuations. This clearly demonstrates that the model has achieved good convergence, meaning that it effectively learns data features and maintains stable performance during training. Overall, this training process is efficient and stable, laying a solid foundation for the subsequent practical application of the model.

[0042] Figure 3 As shown in the R² curve diagram for deep learning, the R² value on the validation set exhibits a rapid upward trend in the early stages of model training. This rapid increase indicates that the model has already learned and captured the basic features of the data well in the initial phase. As the training process progresses, the R² value does not show significant fluctuations or decreases, but gradually stabilizes, eventually remaining at a high level above 0.90. This stable R² value not only reflects the model's very high goodness of fit on the validation set, but also means that the model can accurately identify and capture the inherent patterns and key features in the data. This significant result fully demonstrates that the model not only possesses excellent predictive ability but also exhibits good generalization performance when facing different data, further validating the model's reliability and effectiveness.

[0043] Figure 4This diagram illustrates the model attention distribution and key feature identification. It allows for the rapid and intuitive identification of core feature dimensions that significantly impact model performance. This discovery provides crucial reference for subsequent feature selection, feature engineering (including but not limited to feature enhancement and fusion), and model interpretability analysis. Specifically, for example, if the histogram heights for certain consecutive feature dimensions (such as dimensions 5, 9, 14, and 41) are significantly higher than other regions, this clearly indicates that these specific feature dimension clusters play a vital role in the model's information processing and decision-making processes, with an influence far exceeding that of other feature dimensions. This provides strong data support for optimizing model performance and improving model interpretability.

[0044] Based on the average attention weight, the top 10 important features identified by the model are: Features 35-36 (weight 0.9): H / L ratio and its squared terms (F1, F18); Features 12-13 (weight 0.9): Effective stress-related features (F4); Features 53-54 (weight 0.85): Energy-related features (F11, F12); Features 20-21 (weight 0.85): Geometric features (F7, F9); Features 60-61 (weight 0.85): Mechanical features (F6); Features 25-26 (weight 0.8): Energy / stress interaction terms (F16, F17); Features 37-38 (weight 0.8): Higher-order H / L terms (F19); Features 48-49 (weight 0.8): Stress ratio squared terms (F21); Other medium-weight features (weight 0.5-0.7): minor interaction terms; Low-weight features (weight < 0.3): Redundant or noisy features.

[0045] Model attention distribution - important feature identification: (1) Significantly improves the model's generalization performance and anti-overfitting ability: The causative factors of secondary landslides are extremely complex and exhibit high multicollinearity. The attention mechanism, by assigning extremely high weights to core physical features while significantly suppressing the weights of non-critical features (redundant background information), forcibly removes environmental noise from the underlying data. This ensures that the model can still output high-precision prediction results based on the dominant physical laws when facing completely new landslide geological conditions that have never been seen before.

[0046] (2) Enhanced the model's robustness against noise in extreme disaster environments: In real disaster emergency rescue sites, the basic data collected by front-end sensors or drones are often missing or have measurement errors. Since this model has anchored the decision-making focus on extremely stable macroscopic physical invariants (such as geometric shape and energy conversion characteristics) through the attention mechanism, even if some secondary underlying data fluctuate or are missing, the model's main control prediction logic remains stable, ensuring the anti-interference stability of the emergency rescue prediction system.

[0047] (3) It gives the deep learning “black box model” extremely high physical interpretability: the high-influence feature dimension identified by the model is verified with the core mechanism that determines the fluidization motion in the classic landslide dynamics, breaking the “black box” attribute of the traditional AI model, and at the same time providing solid data support for subsequent model lightweighting and feature engineering dimensionality reduction.

[0048] Figure 5 To compare the ranking predictions, after approximately 200 epochs of meticulous training, the resulting model was further validated on a test set containing 1000 landslide samples. To more accurately evaluate the model's performance, the samples in the test set were reordered for detailed analysis of their distribution. Through a series of rigorous tests and analyses, it is evident that the curve representing the base friction coefficient calculated by the model under the specific conditions provided by the test set fits well with the actual distribution of the test set data. This result fully demonstrates that the model not only performs excellently on this specific dataset but also possesses strong generalization ability, stably functioning in different scenarios and showcasing its broad application potential.

[0049] Understandably, the 24 landslide-specific derived features are dynamically generated through code scripts, without the need for manual calculation or collection.

[0050] It is understandable that a multi-factor dataset of real historical landslides undergoes missing value removal and outlier pruning. Twenty-four landslide-specific derived features are constructed based on physical relationships, and finally, a 34-dimensional dataset is input into a deep neural network model. This feature derivation method, with its "physically constrained boundaries," allows the neural network to acquire prior knowledge consistent with geological laws during the input stage.

[0051] For example, in the forward propagation: the processed dataset is input into the projection layer, and deep features are extracted through three residual blocks (data dimensions are 512→256→128→64 respectively), matching the hierarchical association of landslide features from "macro-geometry" to "micro-mechanics". Feature weights are generated through the attention layer (average attention weight ≥ 0.8) and weighted to ensure that the model focuses on key factors of landslide dynamics, achieving adaptive focusing on core physical parameters. Finally, the predicted value of the base friction coefficient is output through the output projection layer.

[0052] Backpropagation and optimization: Mean squared error (MSE) is used as the loss function, AdamW optimizer (initial learning rate 0.001, weight decay 1e-4, betas=(0.9,0.999)), CosineAnnealingWarmRestarts learning rate scheduler (T0=50, T_mult=2, eta_min=1e-6), gradient clipping threshold 1.0, mixed precision training (FP16 forward propagation, FP32 backpropagation), gradient accumulation steps 4, training epochs ≥150, early stopping patience value 50, training stops when the validation set R²≥0.90 and the loss converges.

[0053] Furthermore, ; In the formula, The noise term, which simulates random disturbances in a natural system, follows a normal distribution. , H This refers to the vertical drop of the landslide. L This is the gliding distance. For comprehensive correction factors; ; In the formula, This is a landslide volume correction factor. Pore ​​water pressure correction factor This is a correction factor for the cohesion of slip zone soil / rock. This is a correction factor for the internal friction angle of slip zone soil / rock. This is the correction factor for the average slope of the sliding surface. This is a correction factor for the average thickness of the landslide mass. This is the slide correction factor. This is a rainfall correction factor. ; In the formula, Let V be the cutoff function, and V be the landslide volume. ; In the formula, Pore ​​water pressure; ; In the formula, It refers to the cohesion or bond strength of the slip zone soil / rock; ; In the formula, The internal friction angle of the slip zone soil / rock; ; In the formula, α is the average slope of the sliding surface; ; In the formula, h is the average thickness of the landslide body; ; In the formula, ρ is the density of the sliding body; ; In the formula, R represents the rainfall.

[0054] It is understandable that the rainfall was the cumulative amount over the two months prior to the landslide.

[0055] In S04: In some embodiments, a visualization interface is built based on a trained deep neural network model. The multi-factor values ​​of the landslide with the base friction coefficient to be measured are input into the geometric parameters of the visualization interface, which correspond to the feature dimensions of the multi-factor dataset of real historical landslides, to obtain the base friction coefficient prediction result.

[0056] For example, the visual interface includes: Development framework: Built on Grado and Plotly, supporting web access; Geometric parameter input module: Provides two modes: slider input and precise numerical input. The parameters are consistent with the feature factors of S1, and the rationality is verified in real time after input. 3D Interactive Response Surface Module: Generates a 60×50 resolution mesh response surface with the H / L ratio as the X-axis, the comprehensive correction coefficient as the Y-axis, and the substrate friction coefficient as the Z-axis, and supports mouse click to select parameters; The prediction results display module includes the substrate friction coefficient value (accurate to 4 decimal places), dimensionality reduction parameters (H / L ratio, comprehensive correction coefficient), 3D coordinates (X=H / L ratio, Y=comprehensive correction coefficient, Z=predicted value), and physical interpretation (based on...). Value range classification: <0.15 indicates extremely low friction, 0.15≤ <0.25 indicates low friction, etc., and case references (the range and characteristics of μ values ​​matching real landslide cases).

[0057] In S05: In some embodiments, the method for obtaining the final predicted substrate friction coefficient includes: S051. Substitute the multi-factor values ​​of the landslide with the base friction coefficient to be measured, which correspond to the feature dimensions of the multi-factor dataset of real historical landslides, into the trained deep neural network model to obtain the initial base friction coefficient prediction value μ1. S052. Input the multi-factor values ​​and μ1 of the landslide with the base friction coefficient to be measured, which correspond to the feature dimensions of the multi-factor dataset of real historical landslides, into the Massflow modeling system, set the simulation parameters, construct the landslide dynamics model, extract the intermediate group landslide movement distance L1 and vertical descent height H1 after simulation in ArcGIS, and substitute them into the trained base friction coefficient prediction model to obtain the intermediate prediction value μ2. S053. Update the intermediate groups L1 and H1 to the multi-factor values. Repeat S052 until the absolute value of the difference between adjacent intermediate predicted values ​​is less than or equal to 0.01. Then stop the iteration and output the final predicted value.

[0058] In S052: In some embodiments, setting simulation parameters includes inputting landslide density, pore water pressure coefficient, cohesion, and the time required for the landslide to complete its sliding.

[0059] Furthermore, the formula for the pore water pressure coefficient is as follows: r u =u / (ρgH t / 1000); In the formula, u is the pore water pressure, ρ is the density of the sliding body, and g is the acceleration due to gravity, taken as 9.81 m / s². 2 , H t The thickness of the sliding body.

[0060] In S053: In some embodiments, the maximum number of iterations does not exceed 5.

[0061] Secondly, embodiments of this application provide a system for predicting the base friction coefficient of secondary landslides based on a deep neural network model, including: The multi-factor dataset acquisition module is used to obtain multi-factor datasets of real historical landslides; The network model building module is used to build deep neural network models; The training module uses a multi-factor dataset to train a deep neural network model, resulting in a well-trained deep neural network model. The prediction module is used to input the multi-factor values ​​of the landslide base friction coefficient to be measured, which correspond to the feature dimensions of the multi-factor dataset of real historical landslides, and to obtain the base friction coefficient prediction result. The final substrate friction coefficient prediction module performs a dynamic closed-loop iterative process of prediction-simulation-feedback based on the substrate friction coefficient prediction results to obtain the final substrate friction coefficient prediction results.

[0062] In the multi-factor dataset acquisition module: In some embodiments, the multi-factor dataset includes landslide volume, slope angle, thickness, density, cohesion, internal friction angle, rainfall, pore water pressure, vertical drop height, and landslide movement distance.

[0063] For example, landslide volume: the total volume of the sliding body that has become unstable, which is calculated by generating the difference between the pre-disaster and post-disaster digital elevation models through UAV aerial survey or lidar. Slope angle: The average inclination angle of the sliding surface, obtained through high-precision DEM data and GIS spatial analysis; Thickness: The average normal thickness of the landslide body is determined by geophysical exploration methods such as on-site drilling, trench excavation, or ground-penetrating radar, combined with topographic profiles. Density: The natural physical density of the landslide soil and rock mass, obtained by taking undisturbed samples on site and conducting indoor geotechnical tests; Cohesion: The mutual attraction and cementation between adjacent particles within a rock or soil mass, determined by in-situ shear tests or indoor triaxial compression tests; Internal friction angle: reflects the frictional resistance characteristics between particles inside the rock and soil mass when sliding or rolling, and is obtained through shear test together with cohesion; Rainfall: The cumulative precipitation depth in the two months prior to the occurrence of the secondary landslide was obtained through monitoring data from on-site meteorological stations near the disaster site or from the regional meteorological station network; Pore ​​water pressure: The pressure exerted by groundwater on solid particles inside the pores of rock and soil, which is obtained through real-time monitoring by piezometers installed on site or estimated based on groundwater level monitoring data; Vertical drop height: The elevation difference between the highest point at the rear edge of the landslide source area and the lowest point at the front edge of the accumulation area, extracted from DEM data or on-site surveying; Landslide travel distance: The total horizontal projection of the landslide body sliding down from the unstable zone to the leading edge of the final accumulation zone, obtained through on-site investigation or post-disaster high-resolution UAV imagery.

[0064] In the network model construction module: In some embodiments, the deep neural network model includes an input projection layer, a residual block stacking layer, an attention mechanism layer, and an output projection layer.

[0065] Furthermore, the input projection layer includes Linear(34,512), BatchNorm1d(512), ReLU activation function, and Dropout(0.3).

[0066] Furthermore, the residual block stack layer includes three residual blocks connected in sequence, each residual block containing two Linear layers, two BatchNorm1d layers, and a Dropout layer.

[0067] It is understandable that skip connections are used to solve the gradient vanishing problem; the three residual blocks can capture the deep correlation of landslide features without losing physical meaning due to excessive network depth.

[0068] Alternatively, it can be understood that this application uses a strict dimensionality reduction sequence of 512-256-128-64 to forcibly remove environmental noise from the original landslide data layer by layer, mapping the 34-dimensional physical constraint feature vectors to a 512-dimensional high-dimensional feature space and then compressing them into the core mechanical features of 64 dimensions.

[0069] Furthermore, the attention mechanism layer includes Linear(64,16), ReLU activation function, Linear(16,64), and Sigmoid activation function.

[0070] It is understandable that ReLU can fit the nonlinear weight mapping of landslide physical characteristics very well.

[0071] Alternatively, it can be understood that by fixing the attention mechanism layer (through a 64-16-64 dimension mapping) in series with the deepest 64-dimensional residual output after three dimensionality reductions, the contribution of the "comprehensive mechanical state" to the friction coefficient can be accurately evaluated after all physical factors have completed deep nonlinear coupling, thereby locking in and amplifying the weight of the core physical factors (such as the H / L ratio) that determine landslide dynamics.

[0072] In the training module: In some embodiments, the multi-factor dataset of real historical landslides is subjected to missing value removal and outlier pruning, and 24 landslide-specific derived features are constructed through physical relationships. The 24 landslide-specific derived features include elevation difference-slip ratio, slope tangent, friction angle tangent, effective stress, stress ratio, shear strength proxy, thickness-slip ratio, volume per unit thickness, kinematic index, mass, potential energy, kinetic energy proxy, saturation index, water content proxy, slope-thickness coupling, cohesion-density coupling, friction angle-volume coupling, elevation difference-slip ratio square, elevation difference-slip ratio cube, slope cube, stress ratio square, volume logarithm, mass logarithm, and thickness square root.

[0073] Furthermore, the elevation difference-slip ratio is calculated using the formula: F1 = H / L, where H is the vertical drop in meters (m) and L is the slip distance in meters (m); the slope tangent is calculated using the formula: F2 = tan(θ), where θ is the average inclination angle of the sliding surface in degrees (°); the friction angle tangent is calculated using the formula: F3 = tan(φ), where φ is the internal friction angle in degrees (°); and the effective stress is calculated using the formula: F4 = c u, where c is cohesion (kPa) and u is pore water pressure (kPa); stress ratio, formula: F5=u / c, where c is cohesion (kPa) and u is pore water pressure (kPa); shear strength ratio, formula: F6=c+α·φ, where α is an empirical coefficient (0.5), c is cohesion (kPa), and φ is the internal friction angle (°); thickness-slip ratio, formula: F7=H t / L, where H t The thickness of the sliding body is in meters (m), and the sliding stroke is in meters (m). The volume per unit thickness is calculated using the formula: F8 = V / H t In the formula, V is the landslide volume in m³, and H... t The thickness of the sliding body is in meters (m); the kinematic performance index is calculated using the formula: F9 = L / H, where L is the sliding distance in meters (m) and H is the vertical drop in meters (m); the mass is calculated using the formula: F... 10 =ρ·V, where ρ is density (kg / m³), V is landslide volume (m³), and potential energy is expressed as F. 11 =ρ·V·g·H, where ρ is density (kg / m³), V is landslide volume (m³), g is gravitational acceleration (m / s²), and H is vertical drop (m); kinetic energy: F 12 =ρ·V·H 2 In the formula, ρ is the density (kg / m³), V is the landslide volume (m³), H is the vertical drop (m), and the saturation index is F. 13 =R·u, where R is the rainfall (mm), u is the pore water pressure (kPa), and the water content is expressed as F. 14 =R·H t In the formula, R is the rainfall in mm, and H is the total rainfall. t The thickness of the sliding body is in meters (m); the slope-thickness coupling formula is: F 15 =tan(θ)·H t In the formula, θ is the average inclination angle of the sliding surface, in degrees, and H t The thickness of the sliding body is in meters; the cohesion-density coupling formula is: F 16 =c·ρ, where ρ is density (kg / m³) and c is cohesion (kPa); friction angle-volume coupling formula: F 17=φ·log10(V+1), where φ is the internal friction angle in °, V is the landslide volume in m³, and F is the square of the elevation difference and the sliding distance. 18 =(H / L) 2 In the formula, H is the vertical drop in meters (m), L is the sliding distance in meters (m); the ratio of elevation difference to sliding distance is expressed as the cube of F. 19 =(H / L) 3 In the formula, H is the vertical drop (in meters), L is the sliding distance (in meters), and the slope is cubic, expressed as: F 20 =θ 2 In the formula, θ is the average inclination angle of the sliding surface, in degrees; the stress ratio squared is given by: F 21 = (u / c) 2 In the formula, u is the pore water pressure in kPa, c is the cohesion in kPa, and the volume logarithm is given by: F 22 =log10(V+1), where V is the landslide volume in m³; F is the logarithm of mass, calculated using the formula: 23 =log10(ρ·V+1), where ρ is the density (kg / m³), V is the landslide volume (m³), and the thickness is the square root of the thickness, calculated using the formula: F 24 =H t In the formula, H t The thickness of the sliding body is expressed in meters (m).

[0074] In some embodiments, a multi-factor dataset of real historical landslides, 24 landslide-specific derived features, and a corrected base friction coefficient are used. The processed dataset is divided into training and testing sets in an 8:2 ratio to train and test the deep neural network model.

[0075] Understandably, the 24 landslide-specific derived features are dynamically generated through code scripts, without the need for manual calculation or collection.

[0076] It is understandable that a multi-factor dataset of real historical landslides undergoes missing value removal and outlier pruning. Twenty-four landslide-specific derived features are constructed based on physical relationships, and finally, a 34-dimensional dataset is input into a deep neural network model. This feature derivation method, with its "physically constrained boundaries," allows the neural network to acquire prior knowledge consistent with geological laws during the input stage.

[0077] For example, in the forward propagation: the processed dataset is input into the projection layer, and deep features are extracted through three residual blocks (data dimensions are 512→256→128→64 respectively), matching the hierarchical association of landslide features from "macro-geometry" to "micro-mechanics". Feature weights are generated through the attention layer (average attention weight ≥ 0.8) and weighted to ensure that the model focuses on key factors of landslide dynamics, achieving adaptive focusing on core physical parameters. Finally, the predicted value of the base friction coefficient is output through the output projection layer.

[0078] Backpropagation and optimization: Mean squared error (MSE) is used as the loss function, AdamW optimizer (initial learning rate 0.001, weight decay 1e-4, betas=(0.9,0.999)), CosineAnnealingWarmRestarts learning rate scheduler (T0=50, T_mult=2, eta_min=1e-6), gradient clipping threshold 1.0, mixed precision training (FP16 forward propagation, FP32 backpropagation), gradient accumulation steps 4, training epochs ≥150, early stopping patience value 50, training stops when the validation set R²≥0.90 and the loss converges.

[0079] Furthermore, ; In the formula, The noise term, which simulates random disturbances in a natural system, follows a normal distribution. , H This refers to the vertical drop of the landslide. L This is the gliding distance. For comprehensive correction factors; ; In the formula, This is a landslide volume correction factor. Pore ​​water pressure correction factor This is a correction factor for the cohesion of slip zone soil / rock. This is a correction factor for the internal friction angle of slip zone soil / rock. This is the correction factor for the average slope of the sliding surface. This is a correction factor for the average thickness of the landslide mass. This is the slide correction factor. This is a rainfall correction factor. ; In the formula, Let V be the cutoff function, and V be the landslide volume. ; In the formula, Pore ​​water pressure; ; In the formula, It refers to the cohesion or bond strength of the slip zone soil / rock; ; In the formula, The internal friction angle of the slip zone soil / rock; ; In the formula, α is the average slope of the sliding surface; ; In the formula, h is the average thickness of the landslide body; ; In the formula, ρ is the density of the sliding body; ; In the formula, R represents the rainfall.

[0080] It is understandable that the rainfall was the cumulative amount over the two months prior to the landslide.

[0081] In the prediction module: In some embodiments, a visualization interface is built based on a trained deep neural network model. The multi-factor values ​​of the landslide with the base friction coefficient to be measured are input into the geometric parameters of the visualization interface, which correspond to the feature dimensions of the multi-factor dataset of real historical landslides, to obtain the base friction coefficient prediction result.

[0082] For example, the visual interface includes: Development framework: Built on Grado and Plotly, supporting web access; Geometric parameter input module: Provides two modes: slider input and precise numerical input. The parameters are consistent with the feature factors of S1, and the rationality is verified in real time after input. 3D Interactive Response Surface Module: Generates a 60×50 resolution mesh response surface with the H / L ratio as the X-axis, the comprehensive correction coefficient as the Y-axis, and the substrate friction coefficient as the Z-axis, and supports mouse click to select parameters; The prediction results display module includes the substrate friction coefficient value (accurate to 4 decimal places), dimensionality reduction parameters (H / L ratio, comprehensive correction coefficient), 3D coordinates (X=H / L ratio, Y=comprehensive correction coefficient, Z=predicted value), and physical interpretation (based on...). Value range classification: <0.15 indicates extremely low friction, 0.15≤ <0.25 indicates low friction, etc., and case references (the range and characteristics of μ values ​​matching real landslide cases).

[0083] In the final substrate friction coefficient prediction module: In some embodiments, the method for obtaining the final predicted substrate friction coefficient includes: S051. Substitute the multi-factor values ​​of the landslide with the base friction coefficient to be measured, which correspond to the feature dimensions of the multi-factor dataset of real historical landslides, into the trained deep neural network model to obtain the initial base friction coefficient prediction value μ1. S052. Input the multi-factor values ​​and μ1 of the landslide with the base friction coefficient to be measured, which correspond to the feature dimensions of the multi-factor dataset of real historical landslides, into the Massflow modeling system, set the simulation parameters, construct the landslide dynamics model, extract the intermediate group landslide movement distance L1 and vertical descent height H1 after simulation in ArcGIS, and substitute them into the trained base friction coefficient prediction model to obtain the intermediate prediction value μ2. S053. Update the intermediate groups L1 and H1 to the multi-factor values. Repeat S052 until the absolute value of the difference between adjacent intermediate predicted values ​​is less than or equal to 0.01. Then stop the iteration and output the final predicted value.

[0084] In S052: In some embodiments, setting simulation parameters includes inputting landslide density, pore water pressure coefficient, cohesion, and the time required for the landslide to complete its sliding.

[0085] Furthermore, the formula for the pore water pressure coefficient is as follows: r u =u / (ρgH t / 1000); In the formula, u is the pore water pressure, ρ is the density of the sliding body, and g is the acceleration due to gravity, taken as 9.81 m / s². 2 , H t The thickness of the sliding body.

[0086] In S053: In some embodiments, the maximum number of iterations does not exceed 5.

[0087] Thirdly, this application provides a computer device including a storage device and a processor. The storage device stores a computer program, which, when executed by the processor, causes the processor to perform the steps of the above-described method for predicting the friction coefficient of a secondary landslide base based on a deep neural network model.

[0088] The computer device can be a desktop computer, laptop, handheld computer, or cloud server, etc. The computer device can interact with the user via a keyboard, mouse, remote control, touchpad, or voice control.

[0089] The memory includes at least one type of readable storage medium, including flash memory, hard disk, multimedia card, card-type memory (e.g., SD or D-interface display memory), random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, disk, optical disk, etc. In some embodiments, the memory may be an internal storage unit of the computer device, such as the hard disk or memory of the computer device. In other embodiments, the memory may also be an external storage device of the computer device, such as a plug-in hard disk, smart media card (SMC), secure digital card (SD), flash card, etc., equipped on the computer device. Of course, the memory may include both internal storage units and external storage devices of the computer device. In this embodiment, the memory is often used to store the operating system and various application software installed on the computer device, such as the program code of the method for predicting the friction coefficient of the secondary landslide base based on a deep neural network model. In addition, the memory can also be used to temporarily store various types of data that have been output or will be output.

[0090] In some embodiments, the processor may be a central processing unit (CPU), a controller, a microcontroller, a microprocessor, or other data processing chip. The processor is typically used to control the overall operation of the computer device. In this embodiment, the processor is used to run program code stored in the memory or process data, for example, to run program code for a method for predicting the base friction coefficient of a secondary landslide based on a deep neural network model.

[0091] Fourthly, this application provides a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the above-described method for predicting the friction coefficient of a secondary landslide base based on a deep neural network model.

[0092] The computer-readable storage medium stores an interface display program that can be executed by at least one processor to perform the steps of the above-described method for predicting the friction coefficient of a secondary landslide base based on a deep neural network model.

[0093] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk), and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the method for predicting the friction coefficient of secondary landslide base based on a deep neural network model as described in the embodiments of this application.

[0094] Application examples Taking the Junlian landslide that occurred on February 8, 2025, in Jinping Village, Mu'ai Town, Junlian County, Yibin City, Sichuan Province as an example, the dynamic constraint prediction method for the base friction coefficient of secondary landslides provided by this invention will be described in detail.

[0095] Step 1: Collect multi-factor dataset for the Junlian landslide. Based on field investigation, remote sensing monitoring, and literature, obtain the basic physical factor data of the landslide. Specific input parameters are as follows: landslide volume 3×10⁶ m³, thickness 20 m, slope angle 50°, density 2400 kg / m³, cohesion 35 kPa, internal friction angle 22°, rainfall 300 mm, pore water pressure 60 kPa, landslide movement distance 417 m, and vertical drop height 378 m.

[0096] It is understandable that by directly inputting the above 10 basic physical factor data into the visual interactive interface provided by the application system, the system will automatically perform multi-dimensional mapping and feature combination calculation based on the landslide dynamic mechanism and physical conservation relationship to calculate 24 landslide-specific derived features and construct a 34-dimensional physical constraint feature vector.

[0097] Steps 2-3: Construct and apply the pre-trained deep neural network model. A pre-constructed and trained deep neural network model is used. This model has been trained on a large-scale dataset of 11 real-world historical landslide cases, with a validation set R. 2 Stable above 0.90.

[0098] Step 4, Visualization Prediction. Input the physical factor values ​​obtained in step S1 into the visualization interface. The system maps the parameters to a high-dimensional space through the input projection layer, performs deep feature extraction using three residual blocks, and identifies key features such as the H / L ratio through the attention layer. Finally, it outputs a predicted value of 0.5230 for the initial group substrate friction coefficient.

[0099] Step 5: Execute a dynamic closed-loop iterative process of "prediction-simulation-feedback". To further constrain the accuracy of the predicted values, the following iterative steps are performed: S5-1, Initial Group Data: The multi-factor values ​​of the Junlian landslide input in S4 (landslide volume 3×10⁻⁶) 6 Substituting the following parameters into the trained model (m³, thickness 20m, slope angle 50°, density 2400kg / m³, cohesion 35kPa, internal friction angle 22°, rainfall 300mm, pore water pressure 60kPa, vertical drop height 378m, landslide movement distance 417m), we obtain the initial predicted base friction coefficient value μ1=0.5230. S5-2, Intermediate Group Data: Input the multi-factor values ​​of the initial group and μ1=0.5230 into the Massflow modeling system, set the simulation parameters with a time step of 0.1s and a calculation accuracy of 1e-5, construct the landslide dynamics model, extract the simulated landslide movement distance L1=447m and vertical descent height H1=363m in ArcGIS, calculate H1 / L1 as a physical constraint and substitute it into the model to obtain the intermediate predicted value μ2=0.5174; S5-3, Final Group Data: Update the intermediate group's L1=447m and H1=363m to the multi-factor values, keep other parameters unchanged, repeat the simulation process of S5-2, and obtain the simulated landslide movement distance L2=466m and vertical descent height H2=364m. Substitute them into the model to obtain the final predicted value μ3=0.5109. S5-4, Convergence Judgment: Calculate |μ3-μ2|=|0.5109-0.5174|=0.0065≤0.01, which satisfies the convergence condition. Stop the iteration and output μ3=0.5109 as the final result. S5-5, Result Verification: The final predicted value μ3=0.5109 is ≤±0.016 of the initial predicted base friction coefficient μ1=0.5230, with a deviation of 7.2%, which meets the verification standard of deviation ≤10%.

[0100] Figure 6 This is a map of the accumulation extent and thickness calculated using ArcGIS (final set). Figure 6 In the figure, (a) shows the three-dimensional morphology and spatial position of the landslide body at the initial moment of landslide initiation (t=0s); (b) shows the accumulation range, movement trajectory and thickness redistribution state of the landslide body after dynamic sliding driven by the predicted friction coefficient value and other simulation parameters, up to 50s.

[0101] This set of images confirms that the system can help fully reproduce the complex fluid dynamics process of landslides, from "instability initiation to high-speed sliding to cessation and evolution." It can clearly show emergency commanders the spatial spread trend of disasters over time, thus providing scientific and dynamic visual support for the planning of evacuation routes and the accurate prediction of the golden rescue window.

[0102] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting the base friction coefficient of a secondary landslide based on a deep neural network model, characterized in that, include: S01. Obtain a multi-factor dataset of real historical landslides; S02, Construct a deep neural network model; S03. Train the deep neural network model using a multi-factor dataset to obtain a well-trained deep neural network model. S04. Input the multi-factor values ​​of the landslide whose base friction coefficient is to be measured, which correspond to the feature dimensions of the multi-factor dataset of real historical landslides, and obtain the base friction coefficient prediction results. S05. Based on the predicted base friction coefficient, perform a dynamic closed-loop iterative process of prediction-simulation-feedback to obtain the final predicted base friction coefficient.

2. The method for predicting the base friction coefficient of a secondary landslide based on a deep neural network model according to claim 1, characterized in that, The multi-factor dataset includes landslide volume, slope angle, thickness, density, cohesion, internal friction angle, rainfall, pore water pressure, vertical drop height, and landslide movement distance; and / or The multi-factor dataset of real historical landslides was used to remove missing values ​​and outliers. Based on physical relationships, 24 landslide-specific derived features were constructed. These 24 landslide-specific derived features include: elevation difference-slip ratio, slope tangent, friction angle tangent, effective stress, stress ratio, shear strength proxy, thickness-slip ratio, volume per unit thickness, kinematic index, mass, potential energy, kinetic energy proxy, saturation index, water content proxy, slope-thickness coupling, cohesion-density coupling, friction angle-volume coupling, elevation difference-slip ratio squared, elevation difference-slip ratio cubed, slope cubed, stress ratio squared, volume logarithm, mass logarithm, and thickness square root.

3. The method for predicting the friction coefficient of a secondary landslide base based on a deep neural network model according to claim 1, characterized in that, Deep neural network models include input projection layers, residual block stacking layers, attention mechanism layers, and output projection layers; and / or A visualization interface is built based on the trained deep neural network model. The multi-factor values ​​of the landslide base friction coefficient, which correspond to the feature dimensions of the multi-factor dataset of real historical landslides, are input into the geometric parameters of the visualization interface to obtain the base friction coefficient prediction results.

4. The method for predicting the base friction coefficient of a secondary landslide based on a deep neural network model according to claim 2, characterized in that, The data includes a multi-factor dataset of real historical landslides, 24 landslide-specific derived features, and a corrected base friction coefficient. The processed dataset is divided into training and testing sets in an 8:2 ratio to train and test the deep neural network model.

5. The method for predicting the base friction coefficient of a secondary landslide based on a deep neural network model according to claim 4, characterized in that, ； In the formula, The noise term, which simulates random disturbances in a natural system, follows a normal distribution. , H This refers to the vertical drop of the landslide. L This is the gliding distance. For comprehensive correction factors; ； In the formula, This is a landslide volume correction factor. Pore ​​water pressure correction factor This is a correction factor for the cohesion of slip zone soil / rock. This is a correction factor for the internal friction angle of slip zone soil / rock. This is the correction factor for the average slope of the sliding surface. This is a correction factor for the average thickness of the landslide mass. This is the slide correction factor. This is a rainfall correction factor. ； In the formula, Let V be the cutoff function, and V be the landslide volume. ； In the formula, Pore ​​water pressure; ； In the formula, It refers to the cohesion or bond strength of the slip zone soil / rock; ； In the formula, The internal friction angle of the slip zone soil / rock; ； In the formula, α is the average slope of the sliding surface; ； In the formula, h is the average thickness of the landslide body; ； In the formula, ρ is the density of the sliding body; ； In the formula, R represents the rainfall.

6. The method for predicting the base friction coefficient of a secondary landslide based on a deep neural network model according to claim 1, characterized in that, Substitute the multi-factor values ​​of the landslide base friction coefficient, which correspond to the feature dimensions of the multi-factor dataset of real historical landslides, into the trained deep neural network model to obtain the initial base friction coefficient prediction value μ1. S052. Input the multi-factor values ​​and μ1 of the landslide with the base friction coefficient to be measured, which correspond to the feature dimensions of the multi-factor dataset of real historical landslides, into the Massflow modeling system, set the simulation parameters, construct the landslide dynamics model, extract the intermediate group landslide movement distance L1 and vertical descent height H1 after simulation in ArcGIS, and substitute them into the trained base friction coefficient prediction model to obtain the intermediate prediction value μ2. S053. Update the intermediate groups L1 and H1 to the multi-factor values. Repeat S052 until the absolute value of the difference between adjacent intermediate predicted values ​​is less than or equal to 0.

01. Then stop the iteration and output the final predicted value.

7. The method for predicting the base friction coefficient of a secondary landslide based on a deep neural network model according to claim 1, characterized in that, Setting simulation parameters includes inputting landslide density, pore water pressure coefficient, cohesion, and the time required for the landslide to complete its sliding. The formula for the pore water pressure coefficient is as follows: r u =u / (ρgH t / 1000); In the formula, u is the pore water pressure, ρ is the density of the sliding body, and g is the acceleration due to gravity, taken as 9.81 m / s². 2 , H t For the thickness of the sliding body; and / or The maximum number of iterations shall not exceed 5.

8. A system for predicting the friction coefficient of a secondary landslide base based on a deep neural network model, characterized in that, include: The multi-factor dataset acquisition module is used to obtain multi-factor datasets of real historical landslides; The network model building module is used to build deep neural network models; The training module uses a multi-factor dataset to train a deep neural network model, resulting in a well-trained deep neural network model. The prediction module is used to input the multi-factor values ​​of the landslide base friction coefficient to be measured, which correspond to the feature dimensions of the multi-factor dataset of real historical landslides, and to obtain the base friction coefficient prediction result. The final substrate friction coefficient prediction module performs a dynamic closed-loop iterative process of prediction-simulation-feedback based on the substrate friction coefficient prediction results to obtain the final substrate friction coefficient prediction results.

9. A computer device, characterized in that, It includes a storage device and a processor, the storage device storing a computer program that, when executed by the processor, causes the processor to perform the steps of the method as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The device stores a computer program that, when executed by a processor, causes the processor to perform the steps of the method as described in any one of claims 1-7.

Images