A geological disaster early warning method and system based on deep learning

Abstract

The application discloses a geological disaster early warning method and system based on deep learning, and relates to the technical field of disaster early warning.The system comprises a data mapping module, a feature extraction module, a probability value analysis module and a disaster early warning module.The application maps static geological attribute data into a geological permeation structure feature vector, and uses the geological permeation structure feature vector to generate a channel attention weight matrix for dynamic rainfall time series data, so as to perform weighted modulation on the rainfall data, extract effective hydrological response features constrained by geology, finally calculate the disaster occurrence probability and trigger the early warning, and achieve the accurate differentiation and dynamic early warning of disaster risks caused by the same rainfall in different geological environments, and solve the problem that the prior art cannot differentiate the risk differences of the same rainfall in different geological environments.

Description

Technical Field

[0001] This invention relates to the field of disaster early warning technology, and in particular to a geological disaster early warning method and system based on deep learning. Background Technology

[0002] With increasing demands for precision in geological disaster prevention and control, the industry urgently needs technologies that can integrate geological conditions and rainfall processes to achieve accurate early warnings. Current mainstream critical rainfall criterion methods rely on fixed rainfall thresholds for early warning. Their drawback lies in completely ignoring the crucial control of static geological attributes such as lithology and soil thickness on rainwater infiltration and runoff processes, leading to an inability to distinguish the risk differences of the same rainfall under different geological environments. Another approach uses machine learning models, simply concatenating geological parameters and rainfall data before inputting them into the model for prediction. However, in this method, continuous time-series rainfall data often dominates the model, overshadowing the effectiveness of static geological features. The model struggles to learn how geological conditions modulate the disaster-causing effects of rainfall, resulting in insufficient early warning accuracy. Summary of the Invention

[0003] This application provides a geological disaster early warning method and system based on deep learning, which solves the problem in the prior art that it is impossible to distinguish the risk differences of the same rainfall under different geological environments, and realizes accurate differentiation and dynamic early warning of disaster risks caused by the same rainfall under different geological environments.

[0004] This application provides a deep learning-based geological disaster early warning method, which is applied to a deep learning-based geological disaster early warning system, including:

[0005] Static geological attribute data and dynamic rainfall monitoring time series data of the target monitoring area are collected. The static geological attribute data is mapped into a geological permeability structure feature vector, and the dynamic rainfall monitoring time series data is mapped into a hydrological load input sequence.

[0006] The channel attention weight matrix for the hydrological load input sequence is generated by using the feature vector of geological permeability structure. The channel attention weight matrix is ​​then used to perform weighted modulation calculation on the hydrological load input sequence to extract effective hydrological response features constrained by geological conditions.

[0007] The effective hydrological response characteristics are input into a pre-set disaster evolution analysis network to obtain the probability value of disaster occurrence, which reflects the slope stability under the current geological conditions.

[0008] In response to a disaster occurrence probability value exceeding a preset dynamic safety threshold, a geological disaster early warning signal is generated through a signaling device and an alarm push operation is executed.

[0009] Furthermore, the extraction of effective hydrological response features constrained by geological conditions includes:

[0010] The geological permeability structure feature vector is input into a fully connected layer network for dimensional transformation, and the output is a geological gating adjustment vector that matches the time step dimension of the hydrological load input sequence.

[0011] By performing element-wise multiplication between the geological gating adjustment vector and the hydrological load input sequence, and by suppressing rainfall data during non-sensitive periods and enhancing rainfall data during sensitive periods, the effective hydrological response characteristics constrained by geological conditions are output.

[0012] Furthermore, the step of inputting the geological permeability structure feature vector into a fully connected layer network for dimensional transformation, and outputting a geological gating adjustment vector that matches the time step dimension of the hydrological load input sequence, includes:

[0013] Extract the first component representing the permeability coefficient of the rock and soil mass and the second component representing the soil layer thickness from the feature vector of the geological permeability structure.

[0014] Based on the first and second components, a time decay factor that can reflect the lag effect of rainfall infiltration by its numerical value is obtained by calculating the activation function;

[0015] Based on the time decay factor, the geological permeability structure feature vector is expanded temporally to construct a geological gating adjustment vector with time dimension attributes.

[0016] Furthermore, the extraction of the first component characterizing the permeability coefficient of the rock and soil mass and the second component characterizing the soil layer thickness from the geological permeability structure feature vector includes:

[0017] The static geological attribute data were subjected to unique thermal coding and normalization to construct an original geological data set containing lithology, soil porosity and slope structure.

[0018] The spatial neighborhood features in the original geological data set are aggregated by graph convolutional neural network to extract the potential hidden layer vectors that characterize the local micro-geomorphic and hydrological characteristics of the monitoring points, and the first and second components are extracted from the potential hidden layer vectors.

[0019] Furthermore, the output yields effective hydrological response characteristics constrained by geological conditions, including:

[0020] Identify the rainfall intensity and soil volumetric moisture content values ​​at each time step in the hydrological load input sequence;

[0021] The weight values ​​of the corresponding time steps in the geological gating adjustment vector are used as scaling factors to adjust the amplitude of the rainfall intensity values ​​and soil volumetric moisture content values.

[0022] Time series data with values ​​higher than the preset noise baseline after amplitude adjustment are retained, invalid data are removed, and effective hydrological response characteristics constrained by geological conditions are reconstructed.

[0023] Furthermore, the step of inputting effective hydrological response characteristics into a pre-set disaster evolution analysis network to obtain a disaster occurrence probability value reflecting the slope stability state under current geological conditions includes:

[0024] The effective hydrological response features are input into a long short-term memory network to extract the current hidden state vector that reflects the cumulative effect of the rainfall infiltration process.

[0025] The historical maximum effective rainfall carrying capacity of the target monitoring area is obtained as a benchmark comparison feature;

[0026] Calculate the Euclidean distance between the current hidden state vector and the baseline comparison feature, map the Euclidean distance to a value between 0 and 1, and determine it as the probability value of disaster occurrence.

[0027] Furthermore, obtaining the historical maximum effective rainfall carrying capacity of the target monitoring area as a benchmark comparison feature includes:

[0028] Retrieve disaster-causing rainfall event data from historical disaster records of the target monitoring area;

[0029] We use the geological permeability structure feature vector to perform weighted backtracking calculations on disaster-causing rainfall event data and reconstruct the critical hydrological feature vectors at historical disaster-causing moments;

[0030] Cluster centers were calculated for critical hydrological feature vectors at multiple historical disaster-causing moments, and the resulting cluster center vectors were marked as benchmark comparison features.

[0031] Furthermore, the calculation of the Euclidean distance between the current hidden state vector and the baseline comparison feature, mapping the Euclidean distance to a value between 0 and 1, and determining it as the probability value of disaster occurrence, includes:

[0032] Calculate the feature residual magnitude in the high-dimensional feature space of the feature comparison between the current hidden state vector and the benchmark.

[0033] The reciprocal of the feature residual magnitude is normalized using the Sigmoid activation function to obtain the disaster occurrence probability value. The smaller the feature residual magnitude, the higher the disaster occurrence probability value.

[0034] Furthermore, the step of generating a geological disaster early warning signal and performing an alarm push operation via a signaling device in response to a disaster occurrence probability value exceeding a preset dynamic safety threshold includes:

[0035] Read the threatened object level data of the current target monitoring area, and match the corresponding dynamic security threshold from the preset threshold lookup table based on the threatened object level data;

[0036] An alarm command is triggered when the probability of a disaster exceeds a dynamic safety threshold for a number of consecutive preset time periods.

[0037] The alarm command is converted into a digital data packet containing the probability value of the disaster and the handling suggestions, and then sent to the preset monitoring terminal.

[0038] This application provides a geological disaster early warning system based on deep learning, which is used to implement a geological disaster early warning method based on deep learning, including: a data mapping module, a feature extraction module, a probability value analysis module, and a disaster early warning module;

[0039] The data mapping module is used to collect static geological attribute data and dynamic rainfall monitoring time series data of the target monitoring area, map the static geological attribute data into a geological permeability structure feature vector, and map the dynamic rainfall monitoring time series data into a hydrological load input sequence.

[0040] The feature extraction module is used to generate a channel attention weight matrix for the hydrological load input sequence using the geological permeability structure feature vector, and to perform weighted modulation calculation on the hydrological load input sequence through the channel attention weight matrix to extract effective hydrological response features constrained by geological conditions.

[0041] The probability value analysis module is used to input effective hydrological response characteristics into a pre-set disaster evolution analysis network to obtain a disaster occurrence probability value that reflects the slope stability under the current geological conditions.

[0042] The disaster early warning module is used to generate a geological disaster early warning signal and perform an alarm push operation in response to the probability value of a disaster exceeding a preset dynamic safety threshold.

[0043] One or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages:

[0044] This application provides a deep learning-based geological disaster early warning method. It collects static geological attribute data and dynamic rainfall monitoring time-series data of the target monitoring area, maps the static geological attribute data to a geological permeability structure feature vector, and maps the dynamic rainfall monitoring time-series data to a hydrological load input sequence. Using the geological permeability structure feature vector, it generates a channel attention weight matrix for the hydrological load input sequence. This matrix is ​​then used to perform weighted modulation calculations on the hydrological load input sequence to extract effective hydrological response features constrained by geological conditions. These effective hydrological response features are input into a pre-set disaster evolution analysis network to obtain a disaster occurrence probability value reflecting the slope stability under current geological conditions. In response to the disaster occurrence probability value exceeding a preset dynamic safety threshold, a geological disaster early warning signal is generated and pushed out.

[0045] In this process, considering that geological conditions are the core factors controlling rainfall infiltration and runoff generation, the original rainfall time series data is weighted and modulated by mapping geological attribute data into feature vectors and generating channel attention weights. This enables dynamic enhancement of attention to rainfall data in sensitive periods and suppression of data in non-sensitive periods based on geophysical characteristics, thereby simulating the differences in rainfall response of different soil and rock masses and achieving targeted rainfall feature extraction.

[0046] Furthermore, when generating channel attention weights, the time decay factor is calculated by extracting the components representing the permeability coefficient and soil thickness from the geological permeability structure feature vector, and a geological gating adjustment vector is constructed. This directly transforms key geophysical parameters into the basis for attention allocation in the time dimension, forcing the learning of the correlation between geological attributes and the timing of rainfall effects, thus enhancing the transparency of the model's decision-making process.

[0047] Furthermore, when calculating the probability of disaster occurrence, the effective hydrological response characteristics after geological modulation are compared with the critical hydrological feature vector reconstructed based on historical disaster data. The distance between the two in the high-dimensional feature space is calculated and mapped to a probability. By shifting from absolute rainfall threshold warning to relative stability distance assessment, the warning results can be dynamically quantified to show how close the current geological and hydrological state is to its historical instability threshold. This improves the dynamic adaptability and accuracy of risk assessment, effectively reduces false alarms and missed alarms caused by ignoring geological differences in existing methods, and enhances the reliability and operational efficiency of geological disaster warnings. Attached Figure Description

[0048] Figure 1 A flowchart of a geological disaster early warning method based on deep learning is provided for embodiments of this application;

[0049] Figure 2This is a schematic diagram of the structure of a geological disaster early warning system based on deep learning, provided as an embodiment of this application. Detailed Implementation

[0050] This application provides a geological disaster early warning method and system based on deep learning, which solves the problem in the prior art that it is impossible to distinguish the risk differences of the same rainfall under different geological environments. By mapping static geological attribute data into geological permeability structure feature vectors and using them to generate channel attention weight matrices for dynamic rainfall time series data, the rainfall data is weighted and modulated to extract effective hydrological response features constrained by geology. Finally, the probability of disaster occurrence is calculated and an early warning is triggered, realizing accurate differentiation and dynamic early warning of disaster risks caused by the same rainfall conditions under different geological environments.

[0051] In related technologies, the mainstream critical rainfall criterion method relies on setting a fixed rainfall threshold. This method completely fails to couple the dynamic rainfall process with the static geological properties of the underlying surface, ignoring the crucial control role of lithology and soil thickness on rainwater infiltration. This results in an inability to distinguish the risk differences caused by the same rainfall conditions in different geological environments, leading to coarse warning results. Another approach is to use a simple machine learning model that simply concatenates geological parameters and rainfall data at the input layer. However, in such models, continuous time-series rainfall data often dominates the training process, causing the physical laws inherent in static geological features to be submerged. The model struggles to effectively learn how geological conditions modulate the disaster-causing effects of rainfall, resulting in limitations in both warning accuracy and mechanistic interpretability. Both using fixed rainfall thresholds or simple feature concatenation models for early warning will result in an inability to accurately characterize the disaster-causing process coupled with rainfall, leading to inaccurate warning results.

[0052] Based on the aforementioned technical problems, this application provides a geological disaster early warning method based on deep learning. By encoding geological attribute data into geological permeability structure feature vectors and using them as source data to generate channel attention weight matrices, the original rainfall monitoring time series data is weighted and modulated, making geological conditions a regulating valve for dynamically adjusting the rainfall data processing logic. By deeply embedding geological physical laws into the feedforward calculation process of the deep learning network, the method achieves deep fusion of static geological attributes and dynamic rainfall processes at the feature level, and finally outputs a dynamic disaster risk probability that can accurately reflect the pattern of different risks for the same rainfall.

[0053] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0054] like Figure 1As shown, this application provides a deep learning-based geological disaster early warning method, which is applied to a deep learning-based geological disaster early warning system, including:

[0055] Static geological attribute data and dynamic rainfall monitoring time series data of the target monitoring area are collected. The static geological attribute data is mapped into a geological permeability structure feature vector, and the dynamic rainfall monitoring time series data is mapped into a hydrological load input sequence.

[0056] Specifically, static geological attribute data is obtained by integrating geological survey reports, borehole data, remote sensing interpretation results, and digital elevation models of the target monitoring area, including lithology, soil layer thickness, permeability coefficient, porosity, slope structure, and fault zone distance. Dynamic rainfall monitoring time series data is automatically acquired through a network of rainfall monitoring stations deployed in and around the target area. It is a sequence of rainfall intensity recorded at fixed time intervals and can be integrated with the concurrent soil volumetric moisture content change sequence obtained from soil moisture sensors.

[0057] Mapping static geological attribute data to geological permeability structure feature vectors is specifically achieved through an encoder neural network, such as a multilayer perceptron. The static geological attribute data is input into the encoder neural network, which, through training, compresses and synthesizes these attributes into a low-dimensional, dense numerical vector. The geological permeability structure feature vector is a key feature abstracted from the raw geological data that comprehensively characterizes the permeability and water-holding capacity of soil and rock masses. Mapping dynamic rainfall monitoring time-series data to a hydrological load input sequence involves aligning the raw time-series data acquired from rain gauges according to time steps to form a regular matrix or tensor. The hydrological load input sequence is used to characterize the rainfall load applied to the geological body over time and the resulting changes in soil moisture content.

[0058] The channel attention weight matrix for the hydrological load input sequence is generated by using the feature vector of geological permeability structure. The channel attention weight matrix is ​​then used to perform weighted modulation calculation on the hydrological load input sequence to extract effective hydrological response features constrained by geological conditions.

[0059] The channel attention weight matrix is ​​a parametric matrix used to assign importance to different parts of the input data. It is a one-dimensional weight vector with a length equal to the time step dimension of the hydrological load input sequence. Each weight value corresponds to a time step, and its magnitude indicates the importance of the rainfall data at that moment to the rainfall-induced disaster process under the current geological conditions.

[0060] The channel attention weight matrix generation step includes: inputting the geological permeability structure feature vector into a small neural network, which includes fully connected layers with activation functions. The small neural network learns the mapping from static geological features to temporal attention preferences.

[0061] Weighted modulation calculation steps: The generated weight vector is multiplied with the hydrological load input sequence step by step to reflect the selective filtering and control of rainfall infiltration by geological conditions. Specifically, let the generated geological gating adjustment vector be... The hydrological load input sequence is ,in, Representing the Geological gating control elements at each time step Representing the A vector of rainfall data at each time step. This represents the total number of time steps. Weighted modulation calculation is essentially element-wise multiplication: ,in It is the first Features modulated at each time step. Performing this operation on all time steps yields an effective hydrological response feature sequence constrained by geological conditions. Geological gating adjustment vector Each element The value is used to directly reflect the model's response to the hydrological load input sequence. The level of attention given to data at each time step. Larger values... The corresponding time period is considered a sensitive period, and its data is augmented in subsequent calculations; smaller values... The corresponding time periods are considered non-sensitive, and their data is suppressed. This ability to distinguish is automatically learned through model training.

[0062] The effective hydrological response characteristics are input into a pre-set disaster evolution analysis network to obtain the probability value of disaster occurrence, which reflects the slope stability under the current geological conditions.

[0063] Specifically, a pre-built disaster evolution analysis network refers to a deep learning model that has been trained before the deployment of the early warning system and is used to simulate the disaster process.

[0064] In this embodiment, the Long Short-Term Memory (LSTM) network is a typical and preferred architecture choice for a pre-built disaster evolution analysis network. During the training phase, the LSM network uses historical data to learn the complex mapping relationship between effective hydrological response characteristics and the occurrence of disasters. After training, its structure and parameters are fixed, which is the pre-built disaster evolution analysis network.

[0065] The effective hydrological response features are input into the Long Short-Term Memory (LSTM) network. The LTM network processes the sequence step by step and updates its internal hidden state. After processing all time steps, the hidden state vector of the last time step is extracted. This is used to encode the cumulative information about the impact of the entire rainfall process on slope stability. Subsequently, this hidden state vector undergoes a nonlinear mapping through the output layer, which is a fully connected layer with a sigmoid activation function. The calculation process is as follows: ,in and The weight matrix and bias vector of the output layer, For the Sigmoid function, The output is the probability value of the disaster occurrence, and its range is... .

[0066] In response to a disaster occurrence probability value exceeding a preset dynamic safety threshold, a geological disaster early warning signal is generated through a signaling device and an alarm push operation is executed.

[0067] Specifically, the dynamic safety threshold is a pre-set probabilistic critical value based on the acceptable risk level and the importance of the protected object in early warning decisions. It is obtained through methods including: determination by domain experts in collaboration based on the probability distribution of historical disaster and non-disaster cases, combined with local disaster prevention and mitigation capabilities. This threshold may differ in areas with different protection levels.

[0068] Furthermore, the extraction of effective hydrological response features constrained by geological conditions includes:

[0069] The geological permeability structure feature vector is input into a fully connected layer network for dimensional transformation, and the output is a geological gating adjustment vector that matches the time step dimension of the hydrological load input sequence.

[0070] By performing element-wise multiplication between the geological gating adjustment vector and the hydrological load input sequence, and by suppressing rainfall data during non-sensitive periods and enhancing rainfall data during sensitive periods, the effective hydrological response characteristics constrained by geological conditions are output.

[0071] In this embodiment, a fully connected layer network refers to a basic feedforward neural network layer in which each input neuron is connected to each output neuron.

[0072] Furthermore, the step of inputting the geological permeability structure feature vector into a fully connected layer network for dimensional transformation, and outputting a geological gating adjustment vector that matches the time step dimension of the hydrological load input sequence, includes:

[0073] Extract the first component representing the permeability coefficient of the rock and soil mass and the second component representing the soil layer thickness from the feature vector of the geological permeability structure.

[0074] Based on the first and second components, a time decay factor that can reflect the lag effect of rainfall infiltration by its numerical value is obtained by calculating the activation function;

[0075] Based on the time decay factor, the geological permeability structure feature vector is expanded temporally to construct a geological gating adjustment vector with time dimension attributes.

[0076] In this embodiment, the first component and the second component are geological permeability structure feature vectors. Two specific dimension values ​​are denoted as follows: and .in, The numerical change can reflect the changing trend of the permeability coefficient of the rock and soil mass. The numerical change can reflect the trend of soil layer thickness change.

[0077] Calculate the time decay factor based on the first and second components. When this is the case, an activation function, such as the sigmoid function, is used. Specifically: ,in , and For learnable parameters, This is the Sigmoid function. The range of values ​​is Between these values, the rate at which geological influences decay over time is used to control the rate at which they decay.

[0078] Based on time decay factor Geological permeability structure characteristic vector Perform time-series unfolding to construct geological gating adjustment vectors This is used to reflect the timeliness of the impact of geological attributes on rainfall. Specifically, it involves generating decay weights for time series data. ,in It is a time step index. It is the total number of time steps. It is a scaling constant. Then all of them Combined into a geological gating adjustment vector .

[0079] Furthermore, the extraction of the first component characterizing the permeability coefficient of the rock and soil mass and the second component characterizing the soil layer thickness from the geological permeability structure feature vector includes:

[0080] The static geological attribute data were subjected to unique thermal coding and normalization to construct an original geological data set containing lithology, soil porosity and slope structure.

[0081] The spatial neighborhood features in the original geological data set are aggregated by graph convolutional neural network to extract the potential hidden layer vectors that characterize the local micro-geomorphic and hydrological characteristics of the monitoring points, and the first and second components are extracted from the potential hidden layer vectors.

[0082] In this embodiment, the static geological attribute data undergoes unique thermal encoding and normalization, converting it into a numerical format that the model can process. For example, lithology categories are converted into unique thermal vectors, and soil layer thicknesses are normalized to the [0,1] interval. The constructed raw geological data set contains the aforementioned processed attribute data for each monitoring point.

[0083] Since the geological attributes of adjacent monitoring points are often correlated, a graph structure is constructed to represent the monitoring points and their spatial adjacency relationships. Nodes represent monitoring points, node features represent processed geological attributes, and edges represent spatial adjacency relationships. A graph convolutional neural network is then used to aggregate information from neighboring nodes. The graph convolution operation extracts latent hidden layer vectors characterizing the local micro-geomorphological and hydrological properties of the monitoring points. .

[0084] Latent Hidden Layer Vector It is a comprehensive feature. Through the decoder network layer, it can be obtained from... The first component is mapped from the middle. Second component ,Right now ,in and For parameters.

[0085] Furthermore, the output yields effective hydrological response characteristics constrained by geological conditions, including:

[0086] Identify the rainfall intensity and soil volumetric moisture content values ​​at each time step in the hydrological load input sequence;

[0087] The weight values ​​of the corresponding time steps in the geological gating adjustment vector are used as scaling factors to adjust the amplitude of the rainfall intensity values ​​and soil volumetric moisture content values.

[0088] Time series data with values ​​higher than the preset noise baseline after amplitude adjustment are retained, invalid data are removed, and effective hydrological response characteristics constrained by geological conditions are reconstructed.

[0089] In this embodiment, the data at each time step of the hydrological load input sequence Includes rainfall intensity values and soil volumetric water content values ,Right now Rainfall intensity Soil volumetric moisture content, obtained from rainfall monitoring stations It is derived from soil moisture sensors or hydrological model inversion.

[0090] The weight value is the geological gating adjustment vector. Elements corresponding to the time step .

[0091] Will The amplitude is adjusted as a scaling factor, i.e., calculated. .

[0092] Preset noise baseline It is a very small positive value. For the adjusted data... Calculate its norm .like If the data at that time step is invalid, it will be discarded. All valid data will be retained. And reconstructed chronologically to form the final effective hydrological response characteristics constrained by geological conditions. Effective hydrological response characteristics are The matrix, This represents the effective time steps.

[0093] Furthermore, the step of inputting effective hydrological response characteristics into a pre-set disaster evolution analysis network to obtain a disaster occurrence probability value reflecting the slope stability state under current geological conditions includes:

[0094] The effective hydrological response features are input into a long short-term memory network to extract the current hidden state vector that reflects the cumulative effect of the rainfall infiltration process.

[0095] The historical maximum effective rainfall carrying capacity of the target monitoring area is obtained as a benchmark comparison feature;

[0096] Calculate the Euclidean distance between the current hidden state vector and the baseline comparison feature, map the Euclidean distance to a value between 0 and 1, and determine it as the probability value of disaster occurrence.

[0097] Furthermore, obtaining the historical maximum effective rainfall carrying capacity of the target monitoring area as a benchmark comparison feature includes:

[0098] Retrieve disaster-causing rainfall event data from historical disaster records of the target monitoring area;

[0099] We use the geological permeability structure feature vector to perform weighted backtracking calculations on disaster-causing rainfall event data and reconstruct the critical hydrological feature vectors at historical disaster-causing moments;

[0100] Cluster centers were calculated for critical hydrological feature vectors at multiple historical disaster-causing moments, and the resulting cluster center vectors were marked as benchmark comparison features.

[0101] In this embodiment, disaster-causing rainfall event data are retrieved from the historical disaster records of the target monitoring area to obtain several historical disaster events. Corresponding original rainfall sequence .

[0102] For each historical disaster event Calling the geological permeability structure feature vector And the weight generation step, for this historical rainfall sequence The same weighted modulation process is applied to obtain the critical hydrological feature vectors at historical disaster-causing moments. .

[0103] For the obtained set of all critical hydrological feature vectors Clustering algorithms are used to calculate cluster centers. The resulting cluster center vectors are then labeled as baseline contrast features. .

[0104] Furthermore, the calculation of the Euclidean distance between the current hidden state vector and the baseline comparison feature, mapping the Euclidean distance to a value between 0 and 1, and determining it as the probability value of disaster occurrence, includes:

[0105] Calculate the feature residual magnitude in the high-dimensional feature space of the feature comparison between the current hidden state vector and the benchmark.

[0106] The reciprocal of the feature residual magnitude is normalized using the Sigmoid activation function to obtain the disaster occurrence probability value. The smaller the feature residual magnitude, the higher the disaster occurrence probability value.

[0107] In this embodiment, the high-dimensional feature space refers to the space containing the hidden state vectors and the benchmark comparison features, and its dimension is equal to the dimension of the hidden state of the Long Short-Term Memory network. Feature residual magnitude. That is, the hidden state vector at the current time. Features compared with the benchmark The Euclidean distance between them is calculated using the following formula: ,in, It is the dimension of the hidden state.

[0108] Furthermore, the step of generating a geological disaster early warning signal and performing an alarm push operation via a signaling device in response to a disaster occurrence probability value exceeding a preset dynamic safety threshold includes:

[0109] Read the threatened object level data of the current target monitoring area, and match the corresponding dynamic security threshold from the preset threshold lookup table based on the threatened object level data;

[0110] An alarm command is triggered when the probability of a disaster exceeds a dynamic safety threshold for a number of consecutive preset time periods.

[0111] The alarm command is converted into a digital data packet containing the probability value of the disaster and the handling suggestions, and then sent to the preset monitoring terminal.

[0112] In this embodiment, the data on the level of threatened objects comes from the disaster-bearing body survey database and is divided according to factors such as the number of threatened people, the value of houses, and the importance of infrastructure, for example, into different levels such as Level 1 protection zone, Level 2 protection zone, and general zone. .

[0113] The pre-defined threshold lookup table stores different threat levels. Corresponding dynamic security threshold ,For example , .

[0114] An alarm command is triggered when the probability of a disaster occurring exceeds the dynamic safety threshold corresponding to the target monitoring area for K consecutive time periods. K is a preset integer used to avoid false alarms due to instantaneous fluctuations.

[0115] The response recommendations are text information generated according to preset rules, including the warning level, recommended measures, and responsible department. Alarm commands are converted into digital data packets containing the probability of disaster occurrence, location information, and response recommendations, and then transmitted wirelessly to the monitoring terminal.

[0116] like Figure 2 As shown in the figure, this application provides a geological disaster early warning system based on deep learning, which is used to implement the geological disaster early warning method based on deep learning, including: a data mapping module, a feature extraction module, a probability value analysis module, and a disaster early warning module;

[0117] The data mapping module is used to collect static geological attribute data and dynamic rainfall monitoring time series data of the target monitoring area, map the static geological attribute data into a geological permeability structure feature vector, and map the dynamic rainfall monitoring time series data into a hydrological load input sequence.

[0118] The feature extraction module is used to generate a channel attention weight matrix for the hydrological load input sequence using the geological permeability structure feature vector, and to perform weighted modulation calculation on the hydrological load input sequence through the channel attention weight matrix to extract effective hydrological response features constrained by geological conditions.

[0119] The probability value analysis module is used to input effective hydrological response characteristics into a pre-set disaster evolution analysis network to obtain a disaster occurrence probability value that reflects the slope stability under the current geological conditions.

[0120] The disaster early warning module is used to generate a geological disaster early warning signal and perform an alarm push operation in response to the probability value of a disaster exceeding a preset dynamic safety threshold.

[0121] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0122] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0123] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0124] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0125] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0126] In conclusion, 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, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A geological disaster early warning method based on deep learning, characterized in that, Includes the following steps: Static geological attribute data and dynamic rainfall monitoring time series data of the target monitoring area are collected. The static geological attribute data is mapped into a geological permeability structure feature vector, and the dynamic rainfall monitoring time series data is mapped into a hydrological load input sequence. The channel attention weight matrix for the hydrological load input sequence is generated by using the feature vector of geological permeability structure. The channel attention weight matrix is ​​then used to perform weighted modulation calculation on the hydrological load input sequence to extract the effective hydrological response features constrained by geological conditions. The extracted effective hydrological response features, constrained by geological conditions, include: The geological permeability structure feature vector is input into a fully connected layer network for dimensional transformation, and the output is a geological gating adjustment vector that matches the time step dimension of the hydrological load input sequence. By performing element-wise multiplication between the geological gating adjustment vector and the hydrological load input sequence, and by suppressing rainfall data in insensitive periods and enhancing rainfall data in sensitive periods, the effective hydrological response characteristics constrained by geological conditions are output. The process of inputting the geological permeability structure feature vector into a fully connected layer network for dimensional transformation, and outputting a geological gating adjustment vector that matches the time step dimension of the hydrological load input sequence, includes: Extract the first component representing the permeability coefficient of the rock and soil mass and the second component representing the soil layer thickness from the feature vector of the geological permeability structure. Based on the first and second components, a time decay factor that can reflect the lag effect of rainfall infiltration by its numerical value is obtained by calculating the activation function; Based on the time decay factor, the geological permeability structure feature vector is expanded in time series to construct a geological gating adjustment vector with time dimension attribute; The extraction of the first component characterizing the permeability coefficient of the rock and soil mass and the second component characterizing the soil layer thickness from the geological permeability structure feature vector include: The static geological attribute data were subjected to unique thermal coding and normalization to construct an original geological data set containing lithology, soil porosity and slope structure. The spatial neighborhood features in the original geological data set are aggregated by graph convolutional neural network, and the potential hidden layer vectors that characterize the local micro-geomorphic and hydrological characteristics of the monitoring points are extracted. The first component and the second component are then extracted from the potential hidden layer vectors. The effective hydrological response characteristics are input into a pre-set disaster evolution analysis network to obtain the probability value of disaster occurrence, which reflects the slope stability under the current geological conditions. In response to a disaster occurrence probability value exceeding a preset dynamic safety threshold, a geological disaster early warning signal is generated through a signaling device and an alarm push operation is executed.

2. The geological disaster early warning method based on deep learning as described in claim 1, characterized in that, The output yields effective hydrological response characteristics constrained by geological conditions, including: Identify the rainfall intensity and soil volumetric moisture content values ​​at each time step in the hydrological load input sequence; The weight values ​​of the corresponding time steps in the geological gating adjustment vector are used as scaling factors to adjust the amplitude of the rainfall intensity values ​​and soil volumetric moisture content values. Time series data with values ​​higher than the preset noise baseline after amplitude adjustment are retained, invalid data are removed, and effective hydrological response characteristics constrained by geological conditions are reconstructed.

3. The geological disaster early warning method based on deep learning as described in claim 1, characterized in that, The process of inputting effective hydrological response characteristics into a pre-set disaster evolution analysis network to obtain disaster occurrence probability values ​​reflecting the slope stability under current geological conditions includes: The effective hydrological response features are input into a long short-term memory network to extract the current hidden state vector that reflects the cumulative effect of the rainfall infiltration process. The historical maximum effective rainfall carrying capacity of the target monitoring area is obtained as a benchmark comparison feature; Calculate the Euclidean distance between the current hidden state vector and the baseline comparison feature, map the Euclidean distance to a value between 0 and 1, and determine it as the probability value of disaster occurrence.

4. The geological disaster early warning method based on deep learning as described in claim 3, characterized in that, The acquisition of the historical maximum effective rainfall carrying capacity of the target monitoring area as a benchmark comparison feature includes: Retrieve disaster-causing rainfall event data from historical disaster records of the target monitoring area; We use the geological permeability structure feature vector to perform weighted backtracking calculations on disaster-causing rainfall event data and reconstruct the critical hydrological feature vectors at historical disaster-causing moments; Cluster centers were calculated for critical hydrological feature vectors at multiple historical disaster-causing moments, and the resulting cluster center vectors were marked as benchmark comparison features.

5. The geological disaster early warning method based on deep learning as described in claim 3, characterized in that, The calculation of the Euclidean distance between the current hidden state vector and the baseline comparison feature, mapping the Euclidean distance to a value between 0 and 1, and determining it as the probability value of disaster occurrence includes: Calculate the feature residual magnitude in the high-dimensional feature space of the feature comparison between the current hidden state vector and the benchmark. The reciprocal of the feature residual magnitude is normalized using the Sigmoid activation function to obtain the disaster occurrence probability value. The smaller the feature residual magnitude, the higher the disaster occurrence probability value.

6. The geological disaster early warning method based on deep learning as described in claim 1, characterized in that, The response to a disaster occurrence probability exceeding a preset dynamic safety threshold, generating a geological disaster early warning signal via a signaling device and executing an alarm push operation, includes: Read the threatened object level data of the current target monitoring area, and match the corresponding dynamic security threshold from the preset threshold lookup table based on the threatened object level data; An alarm command is triggered when the probability of a disaster exceeds a dynamic safety threshold for a number of consecutive preset time periods. The alarm command is converted into a digital data packet containing the probability value of the disaster and the handling suggestions, and then sent to the preset monitoring terminal.

7. A deep learning-based geological disaster early warning system, used to implement the deep learning-based geological disaster early warning method according to any one of claims 1-5, characterized in that, include: Data mapping module, feature extraction module, probability value analysis module, disaster early warning module; The data mapping module is used to collect static geological attribute data and dynamic rainfall monitoring time series data of the target monitoring area, map the static geological attribute data into a geological permeability structure feature vector, and map the dynamic rainfall monitoring time series data into a hydrological load input sequence. The feature extraction module is used to generate a channel attention weight matrix for the hydrological load input sequence using the geological permeability structure feature vector, and to perform weighted modulation calculation on the hydrological load input sequence through the channel attention weight matrix to extract effective hydrological response features constrained by geological conditions. The extracted effective hydrological response features, constrained by geological conditions, include: The geological permeability structure feature vector is input into a fully connected layer network for dimensional transformation, and the output is a geological gating adjustment vector that matches the time step dimension of the hydrological load input sequence. By performing element-wise multiplication between the geological gating adjustment vector and the hydrological load input sequence, and by suppressing rainfall data in insensitive periods and enhancing rainfall data in sensitive periods, the effective hydrological response characteristics constrained by geological conditions are output. The process of inputting the geological permeability structure feature vector into a fully connected layer network for dimensional transformation, and outputting a geological gating adjustment vector that matches the time step dimension of the hydrological load input sequence, includes: Extract the first component representing the permeability coefficient of the rock and soil mass and the second component representing the soil layer thickness from the feature vector of the geological permeability structure. Based on the first and second components, a time decay factor that can reflect the lag effect of rainfall infiltration by its numerical value is obtained by calculating the activation function; Based on the time decay factor, the geological permeability structure feature vector is expanded in time series to construct a geological gating adjustment vector with time dimension attribute; The extraction of the first component characterizing the permeability coefficient of the rock and soil mass and the second component characterizing the soil layer thickness from the geological permeability structure feature vector include: The static geological attribute data were subjected to unique thermal coding and normalization to construct an original geological data set containing lithology, soil porosity and slope structure. The spatial neighborhood features in the original geological data set are aggregated by graph convolutional neural network, and the potential hidden layer vectors that characterize the local micro-geomorphic and hydrological characteristics of the monitoring points are extracted. The first component and the second component are then extracted from the potential hidden layer vectors. The probability value analysis module is used to input effective hydrological response characteristics into a pre-set disaster evolution analysis network to obtain a disaster occurrence probability value that reflects the slope stability under the current geological conditions. The disaster early warning module is used to generate a geological disaster early warning signal and perform an alarm push operation in response to the probability value of a disaster exceeding a preset dynamic safety threshold.

Images