Deep learning-driven multi-level geological disaster precipitation early warning method

Abstract

This application provides a deep learning-driven multi-level geological disaster precipitation early warning method, including: dynamic permeability correction calculation of acquired real-time geological data based on static geological parameters to generate a dynamic permeability tensor; cumulative calculation of permeability lag perception of acquired precipitation sequence data based on the dynamic permeability tensor to generate multi-level cumulative rainfall; feature fusion of multi-level cumulative rainfall and dynamic permeability tensor through a cross-level attention mechanism constrained by geological compatibility to generate cross-level collaborative features; generation of a hierarchical disaster trigger probability matrix using a multi-layer perceptron based on the cross-level collaborative features and soil safety factors; and dynamic weighting of the hierarchical disaster trigger probability matrix based on static geological parameters to output geological disaster early warning data. This method can achieve a deep integration of multi-level early warning models with geological dynamic responses, improving the accuracy of geological disaster early warnings.

Description

Technical Field

[0001] The present invention belongs to the field of geological disaster early warning, and in particular relates to a deep learning-driven multi-level geological disaster precipitation early warning method. Background Art

[0002] With the rapid development of intelligent early warning technology for geological disasters, multi-level early warning methods driven by deep learning have gradually become a research hotspot in this field. Traditional geological disaster precipitation early warning technology primarily uses threshold judgment methods based on statistical analysis. By dividing the cumulative precipitation into fixed time windows (such as 24 hours and 72 hours) and combining it with regional geological classification to set graded warning rules, this method has initially achieved a time-series graded assessment of disaster risk by establishing an empirical relationship between precipitation intensity and historical disaster data. In recent years, with the introduction of deep learning technology, some improved solutions have begun to use models such as convolutional neural networks and long-short-term memory networks to automatically extract spatiotemporal features from meteorological data and improve the accuracy of disaster warnings.

[0003] However, existing technologies have not yet achieved deep integration of multi-level early warning mechanisms and geological dynamic response characteristics, resulting in a disconnect between early warning results and the actual disaster evolution process, affecting the accuracy of geological disaster early warnings. Summary of the Invention

[0004] Based on this, it is necessary to provide a deep learning-driven multi-level geological disaster precipitation warning method to address the above technical problems, which can achieve deep integration of multi-level warning models and geological dynamic responses, and improve the accuracy of geological disaster warnings.

[0005] In the first aspect, this application provides a deep learning-driven multi-level geological disaster precipitation early warning method, including:

[0006] Based on static geological parameters, dynamic permeability correction calculations are performed on the acquired real-time geological data to generate a dynamic permeability tensor. The real-time geological data includes soil moisture content and pore water pressure, and the static geological parameters include saturated permeability coefficient, residual moisture content, saturated moisture content, and rock and soil shear strength parameters.

[0007] Based on the dynamic permeability tensor, the accumulated rainfall data is calculated with infiltration lag perception to generate multi-level accumulated rainfall.

[0008] Through the cross-level attention mechanism constrained by geological compatibility, the multi-level accumulated rainfall and dynamic permeability tensors are fused to generate cross-level collaborative features.

[0009] Based on cross-level collaborative features and soil safety factors, a hierarchical disaster trigger probability matrix is ​​generated through a multi-layer perceptron. The soil safety factor is used to quantify the stability of the soil.

[0010] The hierarchical disaster trigger probability matrix is ​​dynamically weighted based on static geological parameters to output geological disaster warning data.

[0011] In one embodiment, based on static geological parameters, dynamic permeability correction calculation is performed on the acquired real-time geological data to generate a dynamic permeability tensor, including:

[0012] Obtain the geological unit coding table, match the geotechnical type code to each spatial grid based on the geological unit boundary vector in the static geological parameters, and generate the spatially distributed geotechnical pore structure parameters by looking up the table;

[0013] The dynamic permeability tensor is generated by the following formula;

[0014]

[0015] Among them, K eff (t) is the dynamic permeability tensor, K sat is the saturated permeability coefficient, θ res is the residual water content, θ(t) is the soil moisture content, k is the preset empirical correction coefficient, and λ is the rock and soil pore structure parameter.

[0016] In one embodiment, based on the dynamic permeability tensor, the accumulated rainfall data is calculated with infiltration hysteresis perception to generate multi-level accumulated rainfall, including:

[0017] Based on static geological parameters, a hierarchical time window is divided into short-term, medium-term, and long-term time windows. The short-term time window is used to characterize the time span of the permeability rapid response area, the medium-term time window is used to characterize the hysteresis response period of the geologically unstable area, and the long-term time window is used to characterize the long-term cumulative effect of the stable area.

[0018] The precipitation series data are intercepted by sliding windows based on short-term time windows, medium-term time windows and long-term time windows respectively to obtain hierarchical precipitation series;

[0019] The weighted attenuation accumulation calculation of the stratified precipitation sequence is performed based on the dynamic permeability tensor to obtain the multi-level accumulated rainfall.

[0020] In one embodiment, a cross-level attention mechanism constrained by geological compatibility is used to fuse the multi-level accumulated rainfall and dynamic permeability tensors to generate cross-level collaborative features, including:

[0021] Based on the geological unit boundary vectors, the target area is divided into transferable area boundaries to generate a geological compatibility matrix. The generation rules of the geological compatibility matrix include binary identification of the fault zone area and expansion of the transferable area boundaries through morphological dilation operations.

[0022] The multi-level accumulated rainfall and dynamic permeability tensors are input into the spatiotemporal convolution layer to extract the hierarchical feature vector;

[0023] Based on the geological compatibility matrix, the hierarchical feature vectors are fused through the attention weight allocation mechanism to generate cross-level collaborative features.

[0024] In one embodiment, based on the cross-level collaborative features and the soil safety factor, a hierarchical disaster trigger probability matrix is ​​generated by a multi-layer perceptron, including:

[0025] The soil safety factor is generated based on the effective cohesion and effective internal friction angle in the rock and soil shear strength parameters using the following formula:

[0026]

[0027] Among them, F s is the soil safety factor, c' is the effective cohesion, φ' is the effective internal friction angle, p(t) is the pore water pressure, σ is the normal stress, and τ is the shear stress;

[0028] The soil safety factor and cross-level collaborative features are input into a pre-trained multi-layer perceptron to obtain the predicted disaster probability, which is used to construct a hierarchical disaster triggering probability matrix.

[0029] In one embodiment, a method for training a multilayer perceptron includes:

[0030] A training set is constructed based on historical geological disaster data, where the training set includes an input feature set and a label set. The input feature set includes a dynamic permeability tensor, multi-level accumulated rainfall, and cross-level collaborative features, and the label set includes the true value of the disaster trigger probability.

[0031] Add geological disturbance noise to the input feature set to generate adversarial samples;

[0032] The physical constraint hybrid loss is calculated using the following formula:

[0033]

[0034] in, is the physical constraint mixed loss, α is the physical constraint weight, is the mean square error of the predicted safety factor and soil safety factor, β is the data-driven weight, is the cross entropy loss between the predicted disaster probability and the true label, γ is the regularization weight coefficient, and ||W||2 is the regularization term;

[0035] Based on the adversarial sample and physical constraint hybrid loss, the parameters of the multilayer perceptron are updated through the back propagation algorithm until the preset maximum number of iterations is reached to obtain the multilayer perceptron.

[0036] In one embodiment, the hierarchical disaster trigger probability matrix is ​​dynamically weighted based on static geological parameters to output geological disaster warning data, including:

[0037] Based on the geological unit boundary vectors and rock and soil type codes, the target area is divided into disaster sensitivity zones to obtain a disaster sensitivity mask. The disaster sensitivity mask is used to mark the disaster susceptibility level of different rock and soil type areas.

[0038] Generate a dynamic weighting coefficient matrix based on the hazard sensitivity mask and geotechnical shear strength parameters;

[0039] Perform element-by-element product operations on the short-term sub-matrix, medium-term sub-matrix, and long-term sub-matrix of the hierarchical disaster trigger probability matrix and the dynamic weight coefficient matrix to generate weighted hierarchical probability sub-matrix;

[0040] The hierarchical probability sub-matrix is ​​spatially superimposed to generate a warning probability distribution map, and the warning level is divided based on the preset probability threshold to output geological disaster warning data.

[0041] Secondly, this application also provides a deep learning-driven multi-level geological disaster precipitation early warning device, including:

[0042] The dynamic geological parameter acquisition module is used to perform dynamic permeability correction calculations on the acquired real-time geological data based on static geological parameters to generate a dynamic permeability tensor. The real-time geological data includes soil moisture content and pore water pressure, and the static geological parameters include saturated permeability coefficient, residual water content, saturated water content, and rock and soil shear strength parameters.

[0043] The effective accumulated rainfall calculation module is used to perform cumulative calculation of the acquired precipitation sequence data based on the dynamic permeability tensor and generate multi-level accumulated rainfall;

[0044] A cross-layer feature fusion module is used to fuse the multi-level accumulated rainfall and dynamic permeability tensors through a cross-level attention mechanism constrained by geological compatibility, generating cross-level collaborative features.

[0045] The physical constraint module is used to generate a hierarchical disaster trigger probability matrix through a multi-layer perceptron based on cross-level collaborative features and soil safety factors. The soil safety factor is used to quantify soil stability.

[0046] The geological disaster warning module is used to dynamically weight the hierarchical disaster trigger probability matrix based on static geological parameters and output geological disaster warning data.

[0047] In a third aspect, the present application also provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, it implements the above-mentioned deep learning-driven multi-level geological disaster precipitation warning method.

[0048] In a fourth aspect, the present application also provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, it implements the above-mentioned deep learning-driven multi-level geological disaster precipitation warning method.

[0049] The above-mentioned deep learning-driven multi-level geological disaster precipitation warning method performs dynamic infiltration correction calculation on real-time geological data based on static geological parameters to generate a dynamic permeability tensor, solving the problem of inaccurate geological dynamic response caused by the traditional warning model's reliance on fixed infiltration parameters; through the cumulative calculation of infiltration lag perception, the precipitation sequence data is converted into multi-level cumulative rainfall, the hysteresis effect of multi-level precipitation infiltration is modeled, and the temporal matching of precipitation action and geological response is achieved; further, a cross-level attention mechanism with geological compatibility constraints is used to fuse the features of multi-level cumulative rainfall and dynamic permeability tensor to generate cross-level collaborative features, forcing the geological response features of different time levels to maintain transmission consistency within geological units such as fault zones, eliminating logical conflicts in warnings between levels; the cross-level collaborative features and the soil safety factor are input into a multi-layer perceptron to generate a hierarchical disaster triggering probability matrix, the physical constraints are embedded in the deep learning framework, and the probability matrix is ​​dynamically weighted and fused based on static geological parameters, so that the warning results simultaneously reflect the synergistic influence of dynamic precipitation infiltration effects and static geological characteristics, thereby improving the accuracy of geological disaster warnings. The above technical methods can achieve multi-scale collaborative modeling of precipitation effects, rock and soil dynamic responses, and static geological characteristics through the integration of dynamic permeability correction and geological compatibility cross-level features, and embed soil mechanics laws into early warning decisions through a physically constrained deep learning framework, thereby improving the temporal and spatial adaptability and accuracy of geological disaster early warnings. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] In order to more clearly illustrate the technical solutions in the embodiments of the present application or related technologies, the following briefly introduces the drawings required for use in the embodiments or related technical descriptions. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0051] Figure 1A schematic diagram of a process flow of a deep learning-driven multi-level geological disaster precipitation early warning method provided by an embodiment of the present invention;

[0052] Figure 2 A schematic structural diagram of a deep learning-driven multi-level geological disaster precipitation early warning device provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0053] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0054] First, a brief introduction is given to the terms involved in the embodiments of this application.

[0055] The spatiotemporal convolutional layer is a key structure in deep learning for processing spatiotemporal sequence data. It extracts features from dynamically changing processes by jointly modeling the relationships between spatial and temporal dimensions (e.g., precipitation sequences and infiltration response cycles). While traditional convolutional neural networks (CNNs) primarily focus on local spatial features, the spatiotemporal convolutional layer, by introducing a time-axis convolution kernel, simultaneously captures spatial distribution patterns and temporal evolution within a sliding window, effectively mining long-range dependencies in spatiotemporal data. This technology has been widely used in fields such as video analysis, weather forecasting, and traffic flow modeling, and is particularly well-suited for modeling complex systems that require the integration of multi-scale spatiotemporal dynamic features.

[0056] The Multilayer Perceptron (MLP) is a basic feedforward artificial neural network structure consisting of an input layer, at least one hidden layer, and an output layer. It implements nonlinear mapping between multiple layers of neurons through nonlinear activation functions. Its core function is to extract high-order abstract features from input data and fit complex nonlinear relationships through layer-by-layer feature transformation and weight adjustment. In the field of deep learning, the MLP optimizes network parameters through a backpropagation algorithm and minimizes prediction errors using a gradient descent strategy. It is suitable for classification, regression, and feature fusion tasks. This model enhances representational capabilities by stacking hidden layers, effectively solving linear inseparability problems. It is one of the basic architectures for complex pattern recognition and prediction tasks and is commonly used in scenarios such as multidimensional data modeling, dynamic system simulation, and cross-modal feature interaction.

[0057] Based on the above explanation of terms, the implementation environment of the deep learning-driven multi-level geological disaster precipitation early warning method provided in the embodiment of this application is explained. Schematically, the implementation environment includes: multimodal sensors, terminals, storage devices, and processors. The terminals are connected to the multimodal sensors, storage devices, and processor signals via network devices; the multimodal sensors include but are not limited to soil parameter sensors, precipitation monitoring sensors, distributed geological stress monitors, geological radars, and meteorological parameter sensors; the storage devices can be distributed storage devices, centralized storage devices, or cloud storage spaces; and the processors can be central processing units, multi-core processors, or artificial intelligence chips, etc., which are not limited here.

[0058] In combination with the above explanations of terms and implementation environment, the application scenarios of the embodiments of this application are explained. The deep learning-driven multi-level geological disaster precipitation early warning method provided in the embodiments of this application can be applied to, but not limited to, the following scenarios:

[0059] In mountainous areas with frequent rainfall, traditional early warning models often generate false alarms or miss warnings due to their neglect of the hysteresis effect and dynamic response characteristics of rock and soil infiltration. This solution, through dynamic infiltration correction calculation and multi-level cumulative rainfall modeling, can accurately capture the rapid infiltration of short-term heavy rainfall into loose accumulation layers and the cumulative impact of long-term continuous rainfall on deep rock masses. Combined with cross-level feature fusion, it effectively identifies sudden changes in stability in fault zones and steep slopes, providing progressive early warnings for landslides and debris flows in mountainous areas, ranging from minutes to days, significantly improving the timeliness of emergency response.

[0060] To address the risk of soil instability caused by precipitation infiltration around projects like subway tunnels and underground pipeline corridors, traditional methods rely on manual inspections and static geological reports, making it difficult to respond to dynamic changes in a timely manner. This solution uses multimodal sensors to collect real-time pore water pressure, soil moisture content, and precipitation sequence data. Combined with cross-hierarchical feature analysis constrained by geocompatibility, it can dynamically assess the weakening effect of precipitation on the shear strength of the soil around the project, providing early warning of landslide risks around underground projects and providing data support for construction safety and operation and maintenance decisions.

[0061] Sustained heavy rainfall can cause seepage damage or piping in dams, but traditional empirical threshold methods are not adaptable to complex geological conditions. This solution, by integrating dam geotechnical shear strength parameters, real-time pore water pressure, and multi-level rainfall infiltration characteristics, can predict the stability degradation trends of different dam sections under different rainfall scenarios. This helps water conservancy departments develop tiered response strategies (such as reinforcement of key sections and flood discharge scheduling) and enhance the safety redundancy of flood control projects.

[0062] Illustratively, the deep learning-driven multi-level geological disaster precipitation warning method provided in the embodiment of the present application can also be applied to other application scenarios. It is only used as an example here and is not limited to the specific application scenario.

[0063] In an exemplary embodiment, Figure 1 As shown, a deep learning-driven multi-level geological disaster precipitation early warning method is provided. This embodiment uses the method as an example of a terminal in the aforementioned implementation environment. It is understandable that the method can also be applied to a server, and can also be applied to a system including a terminal and a server, and implemented through the interaction between the terminal and the server. In this embodiment, the method includes the following steps 101 to 105:

[0064] Step 101, based on static geological parameters, dynamic permeability correction calculation is performed on the acquired real-time geological data to generate a dynamic permeability tensor; wherein the real-time geological data includes soil moisture content and pore water pressure, and the static geological parameters include saturated permeability coefficient, residual moisture content, saturated moisture content and rock and soil shear strength parameters.

[0065] Specifically, soil moisture content and pore water pressure are collected in real time as real-time geological data through a geological sensor network. Static geological parameters such as saturated permeability coefficient, residual water content, saturated water content, and rock and soil shear strength parameters are extracted from geological survey reports. By combining static geological parameters with real-time geological data and applying permeability theory formulas such as Darcy's law, dynamic permeability correction calculations are performed on the real-time geological data, generating a dynamic permeability tensor that reflects the real-time geological permeability characteristics. For example, this calculation process can use finite element numerical simulation methods or existing empirical formula models. Through this step, a dynamic characterization of the permeability characteristics of the geological body is achieved, laying the foundation for subsequent analysis.

[0066] Step 102: Based on the dynamic permeability tensor, the acquired precipitation sequence data is cumulatively calculated based on the infiltration lag perception to generate multi-level cumulative rainfall.

[0067] Specifically, precipitation data processing involves correlating precipitation data series published by meteorological authorities with a dynamic permeability tensor. Based on the principle of permeability hysteresis, the time delay in precipitation infiltration within a geological body is considered, and the precipitation series data are cumulatively calculated to obtain multi-level accumulated rainfall at different time scales and depth levels. For example, a time series analysis algorithm can be used, or a specialized permeability-precipitation coupling model can be constructed to closely integrate precipitation data with the permeability characteristics of the geological body, thereby more realistically reflecting the impact of precipitation on the geological body.

[0068] In step 103, a cross-level attention mechanism constrained by geological compatibility is used to fuse the multi-level accumulated rainfall and dynamic permeability tensors to generate cross-level collaborative features.

[0069] Specifically, a cross-level attention mechanism based on geological compatibility constraints is used to fuse the features of multi-level accumulated rainfall and dynamic permeability tensors. The geological compatibility constraints are based on geological principles, for example, constraining the temporal and spatial distribution of accumulated rainfall and permeability to approximate Darcy's law or to conform to the physical laws governing soil structural stability. For example, the cross-level attention mechanism can employ the multi-head attention module within the Transformer architecture, customized to suit the characteristics of geological data. During the feature fusion process, the multi-level accumulated rainfall and dynamic permeability tensors are first encoded as feature vector sequences. Then, the attention mechanism is used to calculate the correlation weights between the two at different levels. The geological compatibility constraints provide supervision during this process, dynamically adjusting weights that do not conform to geophysical laws. After multiple rounds of iterative learning, cross-level collaborative features are ultimately generated. These features integrate bidirectional information about precipitation accumulation and geological permeability, calibrated by geological compatibility, providing more physically meaningful input variables for subsequent disaster probability assessments and enhancing the ability to capture geological hazard triggering mechanisms.

[0070] Step 104 : Based on the cross-level collaborative features and the soil safety factor, a hierarchical disaster triggering probability matrix is ​​generated by a multi-layer perceptron. The soil safety factor is used to quantify the stability of the soil.

[0071] Specifically, a multi-layer perceptron (MLP) is used to generate a hierarchical disaster trigger probability matrix based on cross-level collaborative features and soil safety factors. For example, the soil safety factor is calculated based on limit equilibrium theory, combining parameters such as slope and soil type to quantify soil stability. Its calculation formula can be determined using classical geomechanical methods such as the Bishop simplification method or the Janbu universal striping method. The MLP, as the classification component of a deep learning model, consists of an input layer, several hidden layers, and an output layer. The input layer receives data on cross-level collaborative features and soil safety factors, while the hidden layers use nonlinear activation functions to abstract and extract the input features layer by layer. During the training phase, a training set is constructed using a large amount of historical geological disaster data. The network weights are adjusted using a backpropagation algorithm to learn the mapping relationship between input features and disaster trigger probabilities. For example, using historical data from landslide-prone areas, the model learns that when cross-level collaborative features indicate high cumulative rainfall and the soil safety factor is below a critical threshold, the probability of disaster triggering increases significantly. After sufficient training, the multi-layer perceptron can output a hierarchical disaster trigger probability matrix based on new input data, reflecting the possibility of geological disasters in different regions and at different levels, providing a quantitative basis for early warning decisions and realizing the key transformation from feature fusion to disaster risk assessment.

[0072] Step 105 : Dynamically weighting the hierarchical disaster triggering probability matrix based on static geological parameters to output geological disaster warning data.

[0073] Specifically, static geological parameters are used as weighting vectors, with their weight coefficients determined by a pre-trained geological sensitivity model. This model assesses the vulnerability of each region to precipitation-induced disasters based on long-term, unchanging geological factors such as geological structure and lithologic distribution. For example, in areas with karst development, the geotechnical shear strength parameter, a static geological parameter, may have a greater influence on the weight; whereas in areas covered by loose soil layers, the saturated permeability coefficient is more dominant. During the dynamic weighting process, the static geological parameters are mapped into a weight matrix that matches the spatial dimensions of the hazard triggering probability matrix, and the original probability matrix is ​​adjusted using element-by-element multiplication. The weighted probability matrix can more accurately reflect the actual hazard risk differences between different geological units. The weighted probability matrix is ​​converted into visual warning data, such as an electronic map, with different warning levels of different regions indicated by color shades or numerical labels. This effectively transfers risk quantification to user-readable warning information, providing an intuitive and accurate scientific basis for geological disaster prevention and control decisions.

[0074] The above-mentioned deep learning-driven multi-level geological disaster precipitation warning method performs dynamic infiltration correction calculation on real-time geological data based on static geological parameters to generate a dynamic permeability tensor, solving the problem of inaccurate geological dynamic response caused by the traditional warning model's reliance on fixed infiltration parameters; through the cumulative calculation of infiltration lag perception, the precipitation sequence data is converted into multi-level cumulative rainfall, the hysteresis effect of multi-level precipitation infiltration is modeled, and the temporal matching of precipitation action and geological response is achieved; further, a cross-level attention mechanism with geological compatibility constraints is used to fuse the features of multi-level cumulative rainfall and dynamic permeability tensor to generate cross-level collaborative features, forcing the geological response features of different time levels to maintain transmission consistency within geological units such as fault zones, eliminating logical conflicts in warnings between levels; the cross-level collaborative features and the soil safety factor are input into a multi-layer perceptron to generate a hierarchical disaster triggering probability matrix, the physical constraints are embedded in the deep learning framework, and the probability matrix is ​​dynamically weighted and fused based on static geological parameters, so that the warning results simultaneously reflect the synergistic influence of dynamic precipitation infiltration effects and static geological characteristics, thereby improving the accuracy of geological disaster warnings. The above technical methods can achieve multi-scale collaborative modeling of precipitation effects, rock and soil dynamic responses, and static geological characteristics through the integration of dynamic permeability correction and geological compatibility cross-level features, and embed soil mechanics laws into early warning decisions through a physically constrained deep learning framework, thereby improving the temporal and spatial adaptability and accuracy of geological disaster early warnings.

[0075] In one embodiment, based on static geological parameters, dynamic permeability correction calculation is performed on the acquired real-time geological data to generate a dynamic permeability tensor, including:

[0076] Obtain the geological unit coding table, match the geotechnical type code to each spatial grid based on the geological unit boundary vector in the static geological parameters, and generate the spatially distributed geotechnical pore structure parameters by looking up the table.

[0077] Specifically, the boundary vector data of geological units is obtained through geological survey reports or remote sensing image interpretation, and the target area is divided into multiple geological units (such as fault zones, sedimentary layers, weathered rock areas, etc.). Based on the rock and soil type coding table, the corresponding rock and soil type code (such as clay coded as CL and sand coded as SM) is matched to each spatial grid, and the spatially distributed rock and soil pore structure parameters are generated by table lookup mapping. For example, the pore structure parameters of the clay area can be mapped to a low connectivity value, and the sand area can be mapped to a high connectivity value. The above method can solve the problem of insufficient accuracy caused by the dependence of rock and soil pore structure parameters on manual experience in traditional methods, realize the automated spatial matching of geological units and pore parameters, and provide high-resolution input for dynamic permeability calculation.

[0078] The dynamic permeability tensor is generated by the following formula;

[0079]

[0080] Among them, K eff (t) is the dynamic permeability tensor, K sat is the saturated permeability coefficient, θ res is the residual water content, θ(t) is the soil moisture content, k is the preset empirical correction coefficient, and λ is the rock and soil pore structure parameter.

[0081] Specifically, based on the real-time soil moisture data collected by the soil moisture sensor, combined with the saturated permeability coefficient, residual moisture content, and saturated moisture content in the static geological parameter library, the dynamic permeability tensor is calculated using the dynamic permeability correction formula. For example, the difference ratio between the real-time soil moisture content and the residual moisture content and saturated moisture content can be used as a correction factor, and the preset empirical correction coefficient and rock and soil pore structure parameters can be introduced to dynamically adjust the saturated permeability coefficient. For example, when the soil moisture content is close to saturation, the dynamic permeability approaches the saturated permeability coefficient; when the moisture content is low, the permeability decays exponentially with the pore structure parameters. The above method can break through the limitations of the traditional static permeability coefficient and accurately quantify the joint influence of soil moisture state and pore structure on permeability during precipitation infiltration.

[0082] In one embodiment, based on the dynamic permeability tensor, the accumulated rainfall data is calculated with infiltration hysteresis perception to generate multi-level accumulated rainfall, including:

[0083] Based on static geological parameters, hierarchical time windows are divided into short-term time windows, medium-term time windows and long-term time windows. Among them, the short-term time window is used to characterize the time span of the permeability rapid response area, the medium-term time window is used to characterize the lag response period of the geologically unstable area, and the long-term time window is used to characterize the long-term cumulative effect of the stable area.

[0084] Specifically, based on the static geological parameters in the geological survey report (such as rock and soil type and permeability classification), the target area is divided into a permeability rapid response zone, a geologically unstable zone, and a stable zone, and correspondingly divided into short-term, medium-term, and long-term time windows. For example, the loose accumulation layer area has high permeability, so a short-term time window is allocated to capture rapid response; the weathered rock layer has strong permeability hysteresis, so a medium-term time window is allocated; the deep bedrock area has a significant cumulative effect, so a long-term time window is allocated. The above method can solve the problem that the traditional single time window cannot adapt to the differences in permeability response speed in different geological regions. For example, in sandy soil areas, short-term heavy rainfall can quickly trigger infiltration, while clay areas need to rely on medium-term cumulative effects to assess risks. The hierarchical division significantly improves the adaptability of the time scale.

[0085] The precipitation series data are respectively intercepted by sliding windows based on short-term time windows, medium-term time windows and long-term time windows to obtain a hierarchical precipitation series.

[0086] Specifically, using precipitation spatial distribution data retrieved from weather radar and measured sequences from ground rain gauges, sliding window interceptions are performed according to short-term, medium-term, and long-term time windows. For example, for the short-term time window, peak precipitation intensity is intercepted at 10-minute intervals, for the medium-term time window, hourly sliding statistics are used to calculate cumulative values, and for the long-term time window, daily sliding statistics are used to calculate trend components, generating a hierarchical precipitation sequence. This method extracts precipitation characteristics at multiple time granularities, such as capturing the instantaneous impact of heavy rain at the short-term level, modeling the lag effect of infiltration at the medium-term level, and reflecting the gradual impact of continuous precipitation on deep rock masses at the long-term level, providing a data foundation for multi-scale early warning.

[0087] The weighted attenuation accumulation calculation of the stratified precipitation sequence is performed based on the dynamic permeability tensor to obtain the multi-level accumulated rainfall.

[0088] Specifically, a weighted attenuation accumulation calculation is performed on the hierarchical precipitation sequence based on the dynamic permeability tensor. The weighted attenuation accumulation calculation method here can be implemented in a variety of ways. For example, a weight can be assigned to each precipitation data according to the permeability change in the dynamic permeability tensor, and the weight decays as the precipitation time passes. For example, for precipitation data within a short-term time window, due to its fast infiltration response, a linear attenuation method can be used, that is, the most recent precipitation data has a higher weight, and the earlier precipitation data has a lower weight; for the medium-term time window, an exponential attenuation method can be used to make the weight decrease exponentially over time; for the long-term time window, a logarithmic attenuation method can be used to make the weight decrease slowly. Through this method, the accumulated rainfall at different levels can be obtained, which corresponds to the cumulative effect of rainfall under different geological conditions, thereby more accurately reflecting the impact of precipitation on geological disasters and improving the accuracy of early warning.

[0089] In one embodiment, a cross-level attention mechanism constrained by geological compatibility is used to fuse the multi-level accumulated rainfall and dynamic permeability tensors to generate cross-level collaborative features, including:

[0090] The target area is divided into transferable area boundaries based on the geological unit boundary vectors to generate a geological compatibility matrix. The generation rules of the geological compatibility matrix include binary identification of the fault zone area and expansion of the transferable area boundaries through morphological dilation operations.

[0091] Specifically, the geological unit boundary vector data can be obtained based on geological survey reports or remote sensing image interpretation, and the target area can be divided into transferable areas. For example, the fault zone area is binary identified (the transferable area is assigned a value of 1 and the non-transferable area is assigned a value of 0), and the transferable boundary is expanded through morphological dilation operations to eliminate grid alignment errors. For example, a 3×3 pixel rectangular kernel is used to dilate the fault zone boundary to generate a geological compatibility matrix. The above method can solve the problem of ignoring the geological unit boundary constraints in traditional feature fusion. For example, cross-level feature transfer is allowed in the upstream and downstream areas of the fault zone, while cross-level interaction is restricted in the stable bedrock area to avoid warning misjudgments caused by feature confusion.

[0092] The multi-level accumulated rainfall and dynamic permeability tensors are input into the spatiotemporal convolution layer to extract the hierarchical feature vector.

[0093] Specifically, the multi-level cumulative rainfall matrix and the dynamic permeability tensor are input into the spatiotemporal convolution layer, where a three-dimensional convolution kernel is used to simultaneously extract spatial distribution patterns and temporal evolution characteristics. For example, a small-scale convolution kernel is used to capture the spatial diffusion of precipitation peaks in the short-term layer, while a large-scale convolution kernel is used to model the cumulative trend in the long-term layer. The output layer feature vector effectively captures the lag effect and spatial heterogeneity of precipitation infiltration through joint spatiotemporal modeling. For example, in areas with loose accumulation layers, spatiotemporal convolution can simultaneously identify the rapid infiltration path of short-term precipitation and the delayed response of deep seepage.

[0094] Based on the geological compatibility matrix, the hierarchical feature vectors are fused through the attention weight allocation mechanism to generate cross-level collaborative features.

[0095] Specifically, attention weights are assigned to hierarchical feature vectors based on the geological compatibility matrix. For example, high weights are assigned to cross-level features in transferable areas (such as fault zones), allowing short-term and medium-term features to interact; for non-transferable areas (such as stable bedrock), only weighted fusion of same-level features is allowed to generate cross-level collaborative feature tensors. The above method achieves precise matching of multi-level feature fusion and geological dynamic response through the attention mechanism constrained by geological compatibility. For example, in the fault zone area, surface infiltration triggered by short-term heavy rainfall and mid-level seepage in the medium-term layer can interact through cross-level attention to collaboratively predict potential slip surfaces; while in stable areas, only same-level feature analysis is relied upon to avoid noise interference.

[0096] In one embodiment, based on the cross-level collaborative features and the soil safety factor, a hierarchical disaster trigger probability matrix is ​​generated by a multi-layer perceptron, including:

[0097] The soil safety factor is generated based on the effective cohesion and effective internal friction angle in the rock and soil shear strength parameters using the following formula:

[0098]

[0099] Among them, F s is the soil safety factor, c' is the effective cohesion, φ' is the effective internal friction angle, p(t) is the pore water pressure, σ is the normal stress, and τ is the shear stress.

[0100] The soil safety factor and cross-level collaborative features are input into a pre-trained multi-layer perceptron to obtain the predicted disaster probability, which is used to construct a hierarchical disaster triggering probability matrix.

[0101] Specifically, this technical approach embeds geotechnical laws into the early warning model, ensuring a strong correlation between the probability of disaster triggering and the actual physical state of the soil. For example, in clay regions, a low safety factor caused by high pore water pressure can trigger the model to output a high-probability early warning, avoiding the misjudgment caused by purely data-driven models that ignore physical mechanisms. By integrating physical laws with data-driven features, the credibility and interpretability of early warning results are improved. For example, in fault zones, when cross-level features indicate permeability anomalies and the safety factor is below a threshold, the model outputs a high-probability early warning. In stable bedrock regions, even if cross-level features are abnormal, the high safety factor will still suppress the early warning probability, effectively balancing data anomalies with physical rationality.

[0102] In one embodiment, a method for training a multilayer perceptron includes:

[0103] A training set is constructed based on historical geological disaster data, which includes an input feature set and a label set. The input feature set includes a dynamic permeability tensor, multi-level accumulated rainfall, and cross-level collaborative features, and the label set includes the true value of the disaster triggering probability.

[0104] Specifically, a training set is constructed by extracting spatiotemporal data from a database of historical geological disaster events (such as landslide and debris flow records). The input feature set includes a dynamic permeability tensor (derived from real-time sensor data), multi-level accumulated rainfall (derived from weather radar inversion), and cross-level collaborative features (generated through the fusion of geological compatibility constraints). The label set is the true value of the disaster trigger probability (labeled as a 0 / 1 binary label or probability distribution based on the location and time of historical disasters). By aligning the historical disaster data with multi-source features in spatiotemporal order, the training sample bias caused by data silos in traditional models is addressed. For example, in remote areas lacking disaster records, pseudo labels are generated by interpolating permeability and precipitation features to expand the diversity of training samples.

[0105] Add geological disturbance noise to the input feature set to generate adversarial samples.

[0106] Specifically, geological disturbance noise is added to the input feature set to simulate real-world interference such as sensor measurement errors and geological parameter fluctuations. For example, random Gaussian noise is applied to the dynamic permeability tensor to simulate the instantaneous drift of the soil moisture sensor; impulse noise is added to the multi-level accumulated rainfall to simulate abnormal echo interference from weather radar, generating an adversarial sample set.

[0107] The physical constraint hybrid loss is calculated using the following formula:

[0108]

[0109] in, is the physical constraint mixed loss, α is the physical constraint weight, is the mean square error of the predicted safety factor and soil safety factor, β is the data-driven weight, is the cross entropy loss between the predicted disaster probability and the true label, γ is the weight coefficient of the regularization term, and ||W||2 is the regularization term.

[0110] Specifically, by designing a hybrid loss function, the degree of physical law matching and data fit are jointly optimized. The mean square error loss is used to constrain the consistency of the predicted safety factor with the actual mechanical calculation value. The cross-entropy loss drives the predicted probability to approach the historical disaster label. At the same time, a regularization term is introduced to suppress model complexity. For example, the physical constraint weight is set higher than the data-driven weight, forcing the model to prioritize the mechanical equilibrium law of the soil. The above method can solve the problem that purely data-driven models may output physically unreliable results. For example, when the predicted safety factor exceeds the physical threshold, the hybrid loss function automatically penalizes the model output, ensuring that the warning probability is strictly positively correlated with the actual instability risk of the soil.

[0111] Based on the adversarial sample and physical constraint hybrid loss, the parameters of the multilayer perceptron are updated through the back propagation algorithm until the preset maximum number of iterations is reached to obtain the multilayer perceptron.

[0112] Specifically, the back-propagation algorithm iteratively optimizes the parameters of a multilayer perceptron (MLP) based on an adversarial sample set and a hybrid loss function. For example, an adaptive learning rate optimizer dynamically adjusts the parameter update step size, and an early stopping strategy is combined to terminate training when the validation set loss converges. Ultimately, a highly generalizable MLP early warning model is developed. Through the combined optimization of physical constraints and adversarial training, the model demonstrates strong generalization capabilities in unknown areas. For example, in newly surveyed fault zones, high-precision early warnings can be achieved with only a small amount of fine-tuning on labeled data, reducing model deployment costs.

[0113] In one embodiment, the hierarchical disaster trigger probability matrix is ​​dynamically weighted based on static geological parameters to output geological disaster warning data, including:

[0114] Based on the geological unit boundary vectors and rock and soil type codes, the target area is divided into disaster sensitivity zones to obtain a disaster sensitivity mask; among them, the disaster sensitivity mask is used to mark the disaster susceptibility level of areas with different rock and soil types.

[0115] Specifically, the target area is divided based on the geological unit boundary vectors and rock and soil type codes. Different rock and soil types are assigned different disaster susceptibility levels. For example, areas with loose sediments are classified as highly sensitive, while areas with bedrock are classified as less sensitive. This classification process can be implemented using Geographic Information System (GIS) software. Using its spatial analysis capabilities, regions are classified according to preset rules, ultimately generating a disaster sensitivity mask that intuitively demonstrates the differences in disaster susceptibility between different regions.

[0116] Based on the hazard sensitivity mask and geotechnical shear strength parameters, a dynamic weighting coefficient matrix is ​​generated.

[0117] Specifically, a dynamic weighting coefficient matrix is ​​generated by combining the disaster sensitivity mask with geotechnical shear strength parameters (such as effective cohesion and internal friction angle) through mapping rules. For example, higher weights are assigned to areas with high sensitivity and low shear strength, while lower weights are assigned to areas with low sensitivity and high shear strength, thus achieving a joint decision-making process of static geological characteristics and geotechnical parameters. By dynamically adjusting the weights based on shear strength, the impact of geotechnical stability on the warning results is accurately reflected. For example, due to the low shear strength of clay areas, even moderate accumulated rainfall will still trigger a high-weighted warning, while sandy areas require higher rainfall accumulation to trigger a warning due to their high shear strength.

[0118] The short-term sub-matrix, medium-term sub-matrix and long-term sub-matrix of the hierarchical disaster triggering probability matrix are respectively multiplied element-by-element with the dynamic weighting coefficient matrix to generate the weighted hierarchical probability sub-matrix.

[0119] Specifically, the short-term, medium-term, and long-term sub-matrices of the hierarchical disaster trigger probability matrix are each element-by-element multiplied with the dynamic weighting coefficient matrix. For example, a high weight is assigned to the short-term sub-matrix in sandy soil areas to quickly respond to the risk of short-term heavy rainfall; a high weight is assigned to the medium-term sub-matrix in clay regions to focus on monitoring the lag effect of seepage; and a high weight is assigned to the long-term sub-matrix in fault zones to quantify the cumulative risk of deep seepage. This generates a weighted hierarchical probability sub-matrix to achieve spatially differentiated integration of warning results at different time levels. For example, in areas with loose deposits, the short-term level dominates warning decisions; in areas with deep bedrock, the long-term level dominates trend prediction, improving the spatiotemporal adaptability of warnings.

[0120] The hierarchical probability sub-matrix is ​​spatially superimposed to generate a warning probability distribution map, and the warning level is divided based on the preset probability threshold to output geological disaster warning data.

[0121] Specifically, the weighted hierarchical probability submatrices are spatially superimposed to generate a comprehensive warning probability distribution map. Based on historical disaster statistics, probability thresholds are dynamically set. For example, areas with a probability greater than 0.7 are classified as red warnings (immediate risk avoidance), 0.4-0.7 as orange warnings (key attention), and less than 0.4 as blue warnings (normal monitoring). The geological disaster warning level data is then output. The above technical method effectively balances warning sensitivity and false alarm rate through dynamic weighted fusion driven by static geological parameters, thereby improving the accuracy of geological disaster warnings.

[0122] In summary, the deep learning-driven multi-level geological disaster precipitation warning method provided in the embodiment of the present application quantifies the dynamic changes of rock and soil pore structure during precipitation infiltration by performing dynamic permeability correction based on static geological parameters and real-time sensor data; constructs a multi-level cumulative rainfall matrix through hierarchical time window division (short-term, medium-term, and long-term) and cumulative calculation of infiltration lag perception, and simultaneously characterizes the three precipitation action modes of rapid response, lag effect, and long-term accumulation; further introduces a cross-level attention mechanism with geological compatibility constraints to achieve multi-level feature fusion in transferable areas such as fault zones to ensure that the interaction of spatiotemporal features conforms to the boundary laws of geological units; through a multi-layer perceptron, the cross-level collaborative features and the soil safety factor are jointly modeled to generate a hierarchical disaster triggering probability matrix, and the forced warning probability is strongly correlated with the mechanical stability of the soil; based on the dynamic weighted fusion of static geological parameters, the short-term, medium-term, and long-term warning results are enhanced in sensitivity in high-risk areas and false alarms in stable areas are suppressed. The above-mentioned technical solution overcomes the difficulties in traditional methods such as the separation of data-driven and physical laws, and the conflict of hierarchical warning logic through the deep integration of geological dynamic response mechanism and deep learning model, providing high-precision decision-making support for disaster prevention and control in complex geological environments.

[0123] It should be understood that, although the various steps in the flowcharts involved in the various embodiments described above are displayed in sequence according to the instructions of the arrows, these steps are not necessarily executed in sequence in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order restriction on the execution of these steps, and these steps can be executed in other orders. Moreover, at least a portion of the steps in the flowcharts involved in the various embodiments described above can include multiple steps or multiple stages, and these steps or stages are not necessarily executed and completed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a portion of steps or stages in other steps.

[0124] Based on the same inventive concept, the embodiment of the present application also provides a deep learning driven multi-level geological disaster precipitation warning device 10 for implementing the deep learning driven multi-level geological disaster precipitation warning method involved above. The implementation scheme for solving the problem provided by the device is similar to the implementation scheme recorded in the above method. Therefore, the specific limitations of the embodiments of one or more deep learning driven multi-level geological disaster precipitation warning devices 10 provided below can be referred to the limitations of the deep learning driven multi-level geological disaster precipitation warning method above, and will not be repeated here.

[0125] In an exemplary embodiment, Figure 2 As shown, a deep learning-driven multi-level geological disaster precipitation early warning device 10 is provided, comprising:

[0126] The dynamic geological parameter acquisition module 11 is used to perform dynamic permeability correction calculation on the acquired real-time geological data based on static geological parameters to generate a dynamic permeability tensor; wherein the real-time geological data includes soil moisture content and pore water pressure, and the static geological parameters include saturated permeability coefficient, residual water content, saturated water content and rock and soil shear strength parameters.

[0127] The effective accumulated rainfall calculation module 12 is used to perform cumulative calculation of infiltration lag perception on the acquired precipitation sequence data based on the dynamic permeability tensor to generate multi-level accumulated rainfall.

[0128] The cross-layer feature fusion module 13 is used to fuse the multi-level accumulated rainfall and dynamic permeability tensors through a cross-level attention mechanism constrained by geological compatibility, and generate cross-level collaborative features.

[0129] The physical constraint module 14 is used to generate a hierarchical disaster trigger probability matrix through a multi-layer perceptron based on cross-level collaborative features and soil safety factors; wherein the soil safety factor is used to quantify the stability of the soil.

[0130] The geological disaster warning module 15 is used to dynamically weight the hierarchical disaster trigger probability matrix based on static geological parameters and output geological disaster warning data.

[0131] In one embodiment, the dynamic geological parameter acquisition module 11 includes:

[0132] The geological unit code matching unit is used to obtain the geological unit code table, match the rock and soil type code to each spatial grid based on the geological unit boundary vector in the static geological parameters, and generate the spatially distributed rock and soil pore structure parameters by looking up the table.

[0133] The dynamic permeability calculation unit is used to generate a dynamic permeability tensor using the following formula;

[0134]

[0135] Among them, K eff (t) is the dynamic permeability tensor, K sat is the saturated permeability coefficient, θ res is the residual water content, θ(t) is the soil moisture content, k is the preset empirical correction coefficient, and λ is the rock and soil pore structure parameter.

[0136] In one embodiment, the effective accumulated rainfall calculation module 12 includes:

[0137] The hierarchical time window division unit is used to perform hierarchical time window division based on static geological parameters to obtain short-term time windows, medium-term time windows and long-term time windows; among them, the short-term time window is used to characterize the time span of the permeability rapid response area, the medium-term time window is used to characterize the lag response period of the geologically unstable area, and the long-term time window is used to characterize the long-term cumulative effect of the stable area.

[0138] The sliding window interception unit is used to perform sliding window interception on the precipitation sequence data based on the short-term time window, the medium-term time window and the long-term time window to obtain a hierarchical precipitation sequence.

[0139] The weighted attenuation accumulation unit is used to perform weighted attenuation accumulation calculation on the layered precipitation sequence based on the dynamic permeability tensor to obtain multi-level accumulated rainfall.

[0140] In one embodiment, the cross-layer feature fusion module 13 includes:

[0141] The geological compatibility matrix generation unit is used to divide the target area into transferable area boundaries based on the geological unit boundary vectors and generate a geological compatibility matrix; the generation rules of the geological compatibility matrix include binary identification of the fault zone area and expansion of the transferable area boundary through morphological expansion operation.

[0142] The spatiotemporal feature extraction unit is used to input the multi-level accumulated rainfall and dynamic permeability tensors into the spatiotemporal convolution layer to extract the hierarchical feature vector.

[0143] The attention weight allocation unit is used to fuse hierarchical feature vectors based on the geological compatibility matrix through the attention weight allocation mechanism to generate cross-level collaborative features.

[0144] In one embodiment, the physical constraint module 14 includes:

[0145] The soil safety factor calculation unit is used to generate the soil safety factor based on the effective cohesion and effective internal friction angle in the rock and soil shear strength parameters using the following formula:

[0146]

[0147] Among them, F s is the soil safety factor, c' is the effective cohesion, φ' is the effective internal friction angle, p(t) is the pore water pressure, σ is the normal stress, and τ is the shear stress.

[0148] The multi-layer perceptron prediction unit is used to input the soil safety factor and cross-level collaborative features into the pre-trained multi-layer perceptron to obtain the predicted disaster probability, which is used to construct a hierarchical disaster triggering probability matrix.

[0149] Furthermore, the training method of the multilayer perceptron includes the following steps:

[0150] A training set is constructed based on historical geological disaster data, which includes an input feature set and a label set. The input feature set includes a dynamic permeability tensor, multi-level accumulated rainfall, and cross-level collaborative features, and the label set includes the true value of the disaster triggering probability.

[0151] Add geological disturbance noise to the input feature set to generate adversarial samples.

[0152] The physical constraint hybrid loss is calculated using the following formula:

[0153]

[0154] in, is the physical constraint mixed loss, α is the physical constraint weight, is the mean square error of the predicted safety factor and soil safety factor, β is the data-driven weight, is the cross entropy loss between the predicted disaster probability and the true label, γ is the weight coefficient of the regularization term, and ||W||2 is the regularization term.

[0155] Based on the adversarial sample and physical constraint hybrid loss, the parameters of the multilayer perceptron are updated through the back propagation algorithm until the preset maximum number of iterations is reached to obtain the multilayer perceptron.

[0156] In one embodiment, the geological disaster early warning module 15 includes:

[0157] The disaster sensitivity zoning unit is used to divide the target area into disaster sensitivity zones based on the geological unit boundary vector and rock and soil type code to obtain a disaster sensitivity mask; wherein, the disaster sensitivity mask is used to mark the disaster susceptibility level of different rock and soil type areas.

[0158] The dynamic weighting unit is used to generate a dynamic weighting coefficient matrix based on the hazard sensitivity mask and the rock and soil shear strength parameters.

[0159] The hierarchical probability fusion unit is used to perform element-by-element product operations on the short-term sub-matrix, medium-term sub-matrix and long-term sub-matrix of the hierarchical disaster trigger probability matrix with the dynamic weighting coefficient matrix to generate a weighted hierarchical probability sub-matrix.

[0160] The warning data output unit is used to perform spatial superposition calculations on the hierarchical probability sub-matrices, generate a warning probability distribution map, divide the warning levels based on preset probability thresholds, and output geological disaster warning data.

[0161] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the steps of the deep learning-driven multi-level geological disaster precipitation warning method as described above are implemented.

[0162] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps in the above-mentioned method embodiments are implemented.

[0163] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to the partial description of the method embodiments. The device embodiments described above are merely illustrative, wherein the components described as separate parts may or may not be physically separated, and the parts displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the disclosed solution. A person of ordinary skill in the art can understand and implement it without expending creative work.

[0164] The above-described embodiments merely represent several implementation methods of the embodiments of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the patent application. It should be noted that a person skilled in the art may make various modifications and improvements without departing from the concept of the embodiments of the present application, and these modifications and improvements fall within the scope of protection of the embodiments of the present application.

Claims

1. A deep learning driven multi-level geological disaster precipitation early warning method, characterized in that: The method comprises: Based on static geological parameters, dynamic permeability correction calculation is performed on the acquired real-time geological data to generate a dynamic permeability tensor; wherein the real-time geological data includes soil moisture content and pore water pressure, and the static geological parameters include saturated permeability coefficient, residual water content, saturated water content, and rock and soil shear strength parameters; Based on the dynamic permeability tensor, the obtained precipitation sequence data is cumulatively calculated with infiltration hysteresis perception to generate multi-level cumulative rainfall; Through a cross-level attention mechanism constrained by geological compatibility, the multi-level accumulated rainfall and dynamic permeability tensors are fused to generate cross-level collaborative features; Based on the cross-level collaborative features and the soil safety factor, a hierarchical disaster trigger probability matrix is ​​generated through a multi-layer perceptron; wherein the soil safety factor is used to quantify the stability of the soil; Dynamically weighting the hierarchical disaster trigger probability matrix based on the static geological parameters to output geological disaster warning data; The method of performing dynamic permeability correction calculation on the acquired real-time geological data based on static geological parameters to generate a dynamic permeability tensor includes: Obtaining a geological unit coding table, matching the rock and soil type code to each spatial grid based on the geological unit boundary vector in the static geological parameters, and generating spatially distributed rock and soil pore structure parameters by looking up the table; The dynamic permeability tensor is generated by the following formula: ； in, is the dynamic permeability tensor, is the saturated permeability coefficient, is the saturated water content, is the residual water content, is the soil moisture content, is the preset empirical correction coefficient, is the rock and soil pore structure parameter; The step of dynamically weighting the hierarchical disaster trigger probability matrix based on the static geological parameters and outputting geological disaster warning data includes: Based on the geological unit boundary vector and the rock and soil type code, the target area is divided into disaster sensitivity zones to obtain a disaster sensitivity mask; wherein the disaster sensitivity mask is used to mark the disaster susceptibility level of different rock and soil type areas; generating a dynamic weighting coefficient matrix based on the disaster sensitivity mask and the rock and soil shear strength parameter; Performing element-by-element product operations on the short-term submatrix, the medium-term submatrix, and the long-term submatrix of the hierarchical disaster trigger probability matrix and the dynamic weighting coefficient matrix, respectively, to generate weighted hierarchical probability submatrices; The hierarchical probability submatrix is ​​spatially superimposed and calculated to generate a warning probability distribution map, and the warning level is divided based on the preset probability threshold, and the geological disaster warning data is output.

2. The method according to claim 1, characterized in that Based on the dynamic permeability tensor, the accumulated calculation of the infiltration hysteresis perception is performed on the obtained precipitation sequence data to generate multi-level accumulated rainfall, including: Based on the static geological parameters, hierarchical time window division is performed to obtain a short-term time window, a medium-term time window, and a long-term time window; wherein the short-term time window is used to characterize the time span of the permeability rapid response area, the medium-term time window is used to characterize the hysteresis response period of the geologically unstable area, and the long-term time window is used to characterize the long-term cumulative effect of the stable area; Performing sliding window interception on the precipitation sequence data based on the short-term time window, the medium-term time window, and the long-term time window, respectively, to obtain a hierarchical precipitation sequence; The multi-level accumulated rainfall is obtained by performing weighted attenuation accumulation calculation on the layered precipitation sequence based on the dynamic permeability tensor.

3. The method according to claim 1, characterized in that The cross-level attention mechanism constrained by geological compatibility performs feature fusion on the multi-level accumulated rainfall and dynamic permeability tensor to generate cross-level collaborative features, including: Based on the geological unit boundary vectors, the target area is divided into transferable area boundaries to generate a geological compatibility matrix; wherein the generation rule of the geological compatibility matrix includes binary identification of the fault zone area and expansion of the transferable area boundary through a morphological dilation operation; Inputting the multi-level accumulated rainfall and the dynamic permeability tensor into the spatiotemporal convolution layer to extract the hierarchical feature vector; Based on the geological compatibility matrix, the hierarchical feature vectors are fused through an attention weight allocation mechanism to generate the cross-hierarchical collaborative features.

4. The method according to claim 1, wherein The method of generating a hierarchical disaster triggering probability matrix based on the cross-level collaborative features and the soil safety factor through a multi-layer perceptron includes: The soil safety factor is generated according to the effective cohesion and effective internal friction angle in the rock and soil shear strength parameters using the following formula: ； in, is the soil safety factor, is the effective cohesion, is the effective internal friction angle, is the pore water pressure, is the normal stress, is the shear stress; The soil safety factor and the cross-level collaborative features are input into a pre-trained multi-layer perceptron to obtain a predicted disaster probability, which is used to construct the hierarchical disaster trigger probability matrix.

5. The method according to claim 4, characterized in that The training method of the multilayer perceptron includes: Constructing a training set based on historical geological disaster data, wherein the training set includes an input feature set and a label set, the input feature set includes the dynamic permeability tensor, the multi-level accumulated rainfall, and the cross-level collaborative feature, and the label set includes the true value of the disaster triggering probability; Adding geological disturbance noise to the input feature set to generate adversarial samples; The physical constraint hybrid loss is calculated using the following formula: ； in, is the physical constraint mixing loss, is the physical constraint weight, is the mean square error of the predicted safety factor and the soil safety factor, is the data driven weight, is the cross entropy loss between the predicted disaster probability and the true label, is the regularization term weight coefficient, is the regularization term; Based on the adversarial sample and the physical constraint mixed loss, the parameters of the multilayer perceptron are updated by a back propagation algorithm until a preset maximum number of iterations is reached to obtain the multilayer perceptron.

6. A deep learning driven multi-level geological disaster precipitation early warning device, characterized in that: The device comprises: A dynamic geological parameter acquisition module is used to perform dynamic permeability correction calculations on the acquired real-time geological data based on static geological parameters to generate a dynamic permeability tensor; wherein the real-time geological data includes soil moisture content and pore water pressure, and the static geological parameters include saturated permeability coefficient, residual water content, saturated water content, and rock and soil shear strength parameters; An effective accumulated rainfall calculation module is used to perform cumulative calculation of infiltration hysteresis perception on the acquired precipitation sequence data based on the dynamic permeability tensor to generate multi-level accumulated rainfall; A cross-layer feature fusion module is used to fuse the multi-level accumulated rainfall and dynamic permeability tensors through a cross-level attention mechanism constrained by geological compatibility to generate cross-level collaborative features; A physical constraint module is used to generate a hierarchical disaster trigger probability matrix through a multi-layer perceptron based on the cross-level collaborative features and the soil safety factor, wherein the soil safety factor is used to quantify the stability of the soil; A geological disaster warning module is used to dynamically weight the hierarchical disaster trigger probability matrix based on the static geological parameters and output geological disaster warning data; The dynamic geological parameter acquisition module includes: A geological unit code matching unit is used to obtain a geological unit code table, match the rock and soil type code to each spatial grid based on the geological unit boundary vector in the static geological parameters, and generate spatially distributed rock and soil pore structure parameters by looking up the table; A dynamic permeability calculation unit, configured to generate the dynamic permeability tensor using the following formula; ； in, is the dynamic permeability tensor, is the saturated permeability coefficient, is the saturated water content, is the residual water content, is the soil moisture content, is the preset empirical correction coefficient, is the rock and soil pore structure parameter; The geological disaster early warning module includes: a disaster sensitivity zoning unit, configured to divide the target area into disaster sensitivity zones based on the geological unit boundary vector and the rock and soil type code, to obtain a disaster sensitivity mask; wherein the disaster sensitivity mask is used to mark the disaster susceptibility level of different rock and soil type areas; A dynamic weighting unit, configured to generate a dynamic weighting coefficient matrix based on the disaster sensitivity mask and the rock and soil shear strength parameter; a hierarchical probability fusion unit, configured to perform element-by-element multiplication of the short-term submatrix, the medium-term submatrix, and the long-term submatrix of the hierarchical disaster trigger probability matrix with the dynamic weighting coefficient matrix, to generate a weighted hierarchical probability submatrix; The warning data output unit is used to perform spatial superposition calculation on the hierarchical probability submatrix to generate a warning probability distribution map, divide the warning level based on the preset probability threshold, and output geological disaster warning data.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the method according to any one of claims 1 to 5 is implemented.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 5 is implemented.

Images