A Machine Learning-Based Method and System for Early Warning of Geological Disaster Meteorological Risks

Abstract

This application relates to the field of data processing technology and discloses a method and system for geological disaster meteorological risk early warning based on machine learning. The method includes: collecting rainfall intensity, soil saturation, groundwater level changes, and slope runoff coefficients from multiple sensors to construct a geological disaster meteorological dataset; assigning sensitivity weights to meteorological factors according to geological conditions to obtain a weight matrix; weighting and fusing the weight matrix with meteorological time-series data, and extracting features through a geologically constrained long short-term memory network to obtain a risk probability vector; dynamically adjusting the early warning threshold based on the rate of change of the safety factor; and performing Bayesian fusion of the risk probability vector and the adaptive early warning threshold to obtain a graded early warning result. This application solves the technical problem of the lack of intelligent weight allocation for multiple meteorological factors and adaptive threshold adjustment mechanisms for geological conditions in existing geological disaster early warning technologies.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a method and system for early warning of geological disaster meteorological risks based on machine learning. Background Technology

[0002] Existing geological disaster early warning technologies are mainly based on traditional statistical models and simple threshold judgment methods, using rainfall intensity-duration (ID) curves built from single rainfall monitoring data to provide early warnings. These methods typically employ fixed warning thresholds, combined with historical disaster statistics and expert experience, to conduct risk assessments of geological disasters such as landslides and debris flows in specific areas. Some technologies also incorporate simple machine learning algorithms, such as support vector machines and random forests, to analyze and process multidimensional meteorological data.

[0003] However, existing technologies have significant shortcomings: First, traditional methods mainly rely on a single rainfall data source, ignoring the complex coupling relationships between multiple meteorological factors such as soil saturation, groundwater level changes, and slope runoff, resulting in low accuracy of early warnings; second, existing machine learning methods use general algorithms and lack specific designs for geological disaster characteristics, failing to fully consider the sensitivity differences of various meteorological factors under different geological conditions; third, traditional early warning systems use fixed early warning thresholds and cannot adaptively adjust according to the dynamic changes in slope stability, reducing the timeliness and accuracy of early warnings.

[0004] Based on the above analysis, the core problem with existing technologies lies in the lack of an intelligent fusion mechanism for multiple meteorological factors and an adaptive weight allocation strategy for geological conditions. Because the contribution of various meteorological factors to disaster triggering differs significantly under different geological environments, traditional unified weight allocation methods cannot accurately reflect this difference, thus affecting the accuracy of early warning models. Simultaneously, fixed-threshold early warning mechanisms ignore the dynamic evolution of slope stability; when the safety factor continuously decreases, the early warning system cannot adjust its sensitivity in a timely manner, leading to delayed warnings. Therefore, there is an urgent need for a machine learning-based early warning method that can achieve intelligent weight allocation of multiple meteorological factors, integrate geological and physical constraints, and possess the ability to adaptively adjust thresholds based on changes in slope stability. Summary of the Invention

[0005] This application provides a geological disaster meteorological risk early warning method and system based on machine learning, which is used to solve the technical problem that existing geological disaster early warning technologies lack intelligent weight allocation of multiple meteorological factors and adaptive threshold adjustment mechanism for geological conditions.

[0006] Firstly, this application provides a machine learning-based method for early warning of meteorological risks of geological disasters. The method includes: real-time data collection and processing of rainfall intensity, soil saturation, groundwater level changes, and slope runoff coefficient in a target area using multi-source sensors to obtain a meteorological dataset for geological disasters; sensitivity weight allocation processing of each meteorological factor in the meteorological dataset according to geological conditions to obtain a geological disaster sensitivity weight matrix; weighted fusion processing of the geological disaster sensitivity weight matrix with meteorological time-series data, and nonlinear feature extraction of the fused data using a geologically constrained long short-term memory network to obtain a geological disaster risk probability vector; dynamic adjustment of the warning threshold based on the rate of change of the safety factor, increasing the threshold sensitivity when the rate of change of the safety factor continuously decreases to obtain an adaptive warning threshold parameter; and Bayesian fusion processing of the geological disaster risk probability vector and the adaptive warning threshold parameter to obtain a graded geological disaster warning result.

[0007] Secondly, this application provides a machine learning-based geological disaster meteorological risk early warning system, which includes:

[0008] The data acquisition module is used to collect and process data on rainfall intensity, soil saturation, groundwater level changes, and slope runoff coefficient in the target area in real time through multi-source sensors, so as to obtain a geological disaster meteorological dataset.

[0009] The allocation module is used to perform sensitivity weight allocation processing on each meteorological factor in the geological disaster meteorological dataset according to geological conditions, so as to obtain a geological disaster sensitivity weight matrix.

[0010] The weighting module is used to perform weighted fusion processing on the geological hazard sensitivity weight matrix and meteorological time series data, and to extract nonlinear features from the fused data through a geologically constrained long short-term memory network to obtain a geological hazard risk probability vector.

[0011] The adjustment module is used to dynamically adjust the warning threshold according to the rate of change of the safety factor. When the rate of change of the safety factor continues to decrease, the threshold sensitivity is increased to obtain an adaptive warning threshold parameter.

[0012] The fusion module is used to perform Bayesian fusion processing on the geological disaster risk probability vector and the adaptive early warning threshold parameter to obtain the graded geological disaster early warning result.

[0013] Thirdly, a machine learning-based geological disaster meteorological risk early warning device is provided, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor invokes the instructions in the memory to cause the machine learning-based geological disaster meteorological risk early warning device to execute the aforementioned machine learning-based geological disaster meteorological risk early warning method.

[0014] Fourthly, a computer-readable storage medium is provided, wherein instructions are stored therein, which, when executed on a computer, cause the computer to perform the aforementioned machine learning-based geological disaster meteorological risk early warning method.

[0015] The technical solution provided in this application utilizes multi-source sensors to collect real-time data on diverse meteorological factors such as rainfall intensity, soil saturation, groundwater level changes, and slope runoff coefficients. This overcomes the limitations of traditional geological disaster early warning technologies that rely solely on single rainfall data, establishing a comprehensive geological disaster meteorological dataset and significantly improving the completeness and reliability of the data source. By classifying meteorological factors according to geological conditions and assigning sensitivity weights, this application innovatively constructs a geological disaster sensitivity weight matrix. This matrix can accurately quantify the different contributions of each meteorological factor to disaster triggering under different geological environments, solving the problem of low early warning accuracy caused by fixed weight allocation in traditional methods. The core algorithm uses a geologically constrained long short-term memory network (LSTM), incorporating geological physical constraints to enable the neural network to learn feature patterns consistent with geological engineering principles. Compared to general time-series prediction algorithms, this network can more accurately capture the induction patterns of geological disasters, and the extracted geological disaster risk probability vector has stronger physical interpretability and predictive reliability.

[0016] The dynamic adjustment mechanism of the safety factor change rate in this application achieves adaptive optimization of the early warning threshold. When the safety factor change rate continuously decreases, the threshold sensitivity is automatically increased, effectively solving the technical problem that traditional fixed-threshold early warning systems cannot respond to dynamic changes in slope stability, significantly improving the timeliness and accuracy of early warnings. Bayesian fusion processing probabilistically fuses machine learning predictions with physical model calculations, combining the learning capabilities of data-driven methods with the physical constraints of mechanistic models. The resulting graded geological disaster early warning results have higher credibility and practicality. Especially in the specific application field of geological disaster meteorological risk early warning, the gating modulation mechanism of the geologically constrained long short-term memory network can automatically adjust the information processing strategy according to different geological conditions, enabling the features learned by the network to better reflect the disaster risk characteristics of different geological environments such as granite weathered soil and shale fracture zones. Compared with traditional machine learning methods, the early warning accuracy is significantly improved under complex geological-meteorological coupling conditions. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a schematic diagram of one embodiment of the geological disaster meteorological risk early warning method based on machine learning in this application.

[0019] Figure 2 This is a schematic diagram of one embodiment of the geological disaster meteorological risk early warning system based on machine learning in this application.

[0020] Figure 3 This is a schematic block diagram of the geological disaster meteorological risk early warning device based on machine learning in an embodiment of the present invention. Detailed Implementation

[0021] This application provides a method and system for early warning of meteorological risks of geological disasters based on machine learning. The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data used can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or device that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0022] For ease of understanding, the specific process of the embodiments of this application is described below. Please refer to [link / reference]. Figure 1 One embodiment of the geological disaster meteorological risk early warning method based on machine learning in this application includes:

[0023] Step S101: Real-time data collection and processing of rainfall intensity, soil saturation, groundwater level changes, and slope runoff coefficient in the target area using multi-source sensors to obtain a geological disaster meteorological dataset;

[0024] Step S102: Based on the classification of geological conditions, perform sensitivity weight allocation on each meteorological factor in the geological disaster meteorological dataset to obtain the geological disaster sensitivity weight matrix;

[0025] Step S103: The geological hazard sensitivity weight matrix and meteorological time series data are weighted and fused. The fused data are then subjected to nonlinear feature extraction through a geologically constrained long short-term memory network to obtain a geological hazard risk probability vector.

[0026] Step S104: Dynamically adjust the warning threshold according to the rate of change of the safety factor. When the rate of change of the safety factor continues to decrease, increase the threshold sensitivity to obtain the adaptive warning threshold parameter.

[0027] Step S105: Perform Bayesian fusion processing on the geological disaster risk probability vector and the adaptive early warning threshold parameter to obtain the graded geological disaster early warning results.

[0028] It is understood that the executing entity of this application can be a machine learning-based geological disaster meteorological risk early warning system, or it can be a terminal or a server; the specific implementation is not limited here. This application's embodiments use a server as an example for illustration.

[0029] Specifically, multi-source sensor data acquisition forms the data foundation of the entire early warning system. When automatic weather stations collect rainfall intensity data, sensors record rainfall hourly. This data is compared and filtered against a preset effective rainfall intensity range, eliminating outliers to form an effective rainfall intensity sequence. Soil moisture sensors monitor soil saturation in real time. The raw saturation values ​​are subject to boundary constraints according to a physically reasonable range to ensure the data remains within a reasonable interval. Groundwater level monitors continuously collect groundwater level change data. By calculating the difference in water level between adjacent time points divided by the time interval, a groundwater level change rate sequence is obtained. Slope flow monitoring equipment collects runoff data. The measured runoff is divided by the corresponding rainfall to calculate the runoff coefficient, forming a slope runoff coefficient sequence.

[0030] A geological hazard sensitivity weight matrix is ​​constructed, which is one of the core innovations of this invention. Based on geological exploration data of the target area, slope, lithology, and soil type are classified and identified to obtain geological condition classification labels. Historical geological hazard cases are categorized and statistically processed based on these geological condition classification labels, and the correlation coefficients between different meteorological factors and hazard occurrence under various geological conditions are calculated. The gradient descent algorithm, an optimization algorithm, iteratively updates the weight parameters until convergence by calculating the gradient of the loss function. This algorithm iteratively optimizes the weight parameters to obtain optimized weight parameters. Sensitivity weight allocation processing, based on the optimized weight parameters, assigns weights to rainfall intensity, soil saturation, groundwater level change, and slope runoff coefficient, constructing a matrix structure with geological conditions as rows and meteorological factors as columns.

[0031] This paper implements feature extraction functionality for a geologically constrained Long Short-Term Memory (LSTM) network. LSTM is a special type of recurrent neural network capable of remembering long-term dependencies, controlling information flow through forget gates, input gates, and output gates. The geologically constrained LSTM network incorporates geophysical constraints on top of the standard LSTM. Geological condition embedding vectors undergo vector encoding processing through embedding layers, concatenating and combining slope classification, lithology, and soil type codes to construct a comprehensive geological condition encoding sequence. Geophysical constraint modulation processing modulates the network's three gates based on the geological condition embedding vectors, enabling the network to learn feature patterns that conform to geophysical laws.

[0032] The safety factor is dynamically calculated using the Bishop circular sliding method. The Bishop circular sliding method is a commonly used slope stability analysis method in geological engineering. This method assumes the sliding surface is circular and calculates the safety factor based on the principle of moment balance. The rate of change of the safety factor is obtained by performing differential operations on the real-time safety factor over a continuous time period, calculating the ratio of the difference in safety factor between adjacent time points to the time interval. When the rate of change of the safety factor is continuously negative and its absolute value exceeds a preset decrease threshold, a sensitivity adjustment mechanism is triggered, and the baseline warning threshold is increased by a percentage. The soil moisture content correction factor is calculated by using the ratio of the current saturation to the field capacity saturation as a correction factor to correct for the impact of moisture on the adjusted warning threshold.

[0033] Bayesian fusion processing probabilistically fuses machine learning predictions with physical model calculations. The physical model risk probability is calculated using a normal distribution based on the real-time safety factor, with the difference between the safety factor and the benchmark value divided by the standard deviation used as the input parameter. Monte Carlo sampling, a numerical calculation method based on random sampling, calculates the variance of the prediction results by repeatedly sampling parameters such as effective cohesion and internal friction angle, yielding the uncertainty quantification probability. Weighted linear combination processing probabilistically fuses the geological hazard risk probability vector, the physical model risk probability, and the uncertainty quantification probability according to preset weighting coefficients.

[0034] In one specific embodiment, the process of performing step S101 may specifically include the following steps:

[0035] Rainfall intensity data is collected and processed hourly by automatic weather stations. The collected data is then compared and filtered with the set effective rainfall intensity range to obtain an effective rainfall intensity sequence.

[0036] Soil saturation data is monitored and processed in real time by a soil moisture sensor. The collected raw saturation values ​​are constrained according to a preset physical reasonable range to obtain a standard soil saturation sequence.

[0037] By continuously collecting and processing groundwater level change data through a groundwater level monitoring instrument, calculating the water level difference between adjacent time points and dividing by the time interval, a groundwater level change rate sequence is obtained.

[0038] Based on the slope flow monitoring equipment, runoff data is collected and processed, and the measured runoff is divided by the corresponding rainfall to calculate the runoff coefficient value, thus obtaining the slope runoff coefficient sequence.

[0039] The effective rainfall intensity sequence, standard soil saturation sequence, groundwater level change rate sequence, and slope runoff coefficient sequence were time-synchronized and aligned to obtain a geological disaster meteorological dataset.

[0040] Specifically, when automatic weather stations collect and process rainfall intensity data hourly, the rainfall intensity sensor records the rainfall value once per hour. The collected raw data needs to be compared and filtered against a set effective rainfall intensity range. This comparison and filtering process is implemented through a numerical comparison algorithm, comparing each collected data point with preset upper and lower thresholds one by one. Abnormal data points that exceed the range are marked and removed, and the retained data points are arranged in chronological order to form an effective rainfall intensity sequence. When soil moisture sensors monitor and process soil saturation data in real time, the sensor probe is inserted into the soil to measure the soil moisture content. The raw saturation value is converted into a digital signal through changes in the sensor's resistance or frequency. Boundary constraint processing refers to comparing the collected raw saturation value with a preset physical reasonable range. The physical reasonable range is determined based on soil type and geological conditions. Values ​​that exceed the reasonable range are adjusted to the boundary value, and the adjusted data are arranged in chronological order to form a standard soil saturation sequence.

[0041] The groundwater level monitoring instrument continuously collects and processes groundwater level change data, including two data processing processes: water level difference calculation and change rate calculation. The water level difference calculation is obtained by subtracting the water level readings at adjacent time points. The change rate calculation divides the water level difference by the time interval, which is usually one hour. The result of the division operation represents the rate of change of the groundwater level per unit time. The continuous change rate values ​​form a groundwater level change rate sequence in chronological order.

[0042] When slope flow monitoring equipment collects and processes runoff data, flow sensors measure the water flow through a specific cross-section. The runoff coefficient is calculated by dividing the measured runoff by the rainfall within the corresponding time period. The physical meaning of this division operation is the proportion of rainfall converted into surface runoff. The runoff coefficient reflects the slope's permeability and runoff generation capacity. The calculated runoff coefficient values ​​are arranged chronologically to form a slope runoff coefficient sequence. Time synchronization and alignment processing aligns four data sequences from different sources according to a unified time reference. The alignment process first establishes a unified time axis, then matches each data sequence according to its corresponding timestamp. When a certain data point is missing, linear interpolation is used to fill the missing value. Linear interpolation calculates the missing value by weighted averaging of data values ​​from adjacent time points. The aligned data forms a unified data matrix structure, where rows represent time points and columns represent different meteorological factors. This data matrix is ​​the geological disaster meteorological dataset.

[0043] In one specific embodiment, the process of performing step S102 may specifically include the following steps:

[0044] Based on geological exploration data of the target area, slope, lithology, and soil type are classified and identified to obtain geological condition classification labels;

[0045] Based on geological condition classification labels, historical geological disaster cases are classified and statistically processed. The correlation coefficients between different meteorological factors and disaster occurrence under various geological conditions are calculated to obtain the factor correlation coefficient matrix.

[0046] The factor correlation coefficient matrix is ​​normalized according to the geological condition dimension, and the weight parameters are iteratively optimized using the gradient descent algorithm to obtain the optimized weight parameters.

[0047] Based on the optimized weight parameters, sensitivity weights were assigned to rainfall intensity, soil saturation, groundwater level change, and slope runoff coefficient. A matrix structure with geological conditions as rows and meteorological factors as columns was constructed to obtain the initial sensitivity weight matrix.

[0048] Based on geological engineering experience, the initial sensitivity weight matrix is ​​verified by expert knowledge to obtain the geological hazard sensitivity weight matrix.

[0049] Specifically, the geological condition classification and identification process is based on geological exploration data of the target area. Slope classification and identification involves measuring the surface tilt angle and discretizing it according to preset intervals, converting continuous slope values ​​into classification labels. Lithology classification and identification determines rock types through mineral composition analysis and physical property testing of rock samples. Soil type classification and identification determines soil classification through soil particle analysis, plasticity index determination, bearing capacity testing, and other methods. The three types of geological condition information are combined to form a geological condition classification label, which serves as an index for subsequent data processing.

[0050] Historical geological disaster cases are categorized and statistically processed based on geological condition classification labels. Past landslides, debris flows, collapses, and other geological disasters are grouped according to the geological conditions of their locations. Each geological condition group includes information such as the time of occurrence, meteorological conditions, and disaster scale. Correlation coefficient calculation is a statistical method for measuring the strength of the linear relationship between two variables. The calculation process involves calculating the covariance between each meteorological factor and the number of disaster occurrences, then dividing by the product of the standard deviations of the two variables. The correlation coefficient ranges from negative to positive, with positive values ​​indicating positive correlation and negative values ​​indicating negative correlation; the larger the absolute value, the stronger the correlation. The correlation coefficients between different meteorological factors and disaster occurrences under various geological conditions are arranged in matrix form, forming a factor correlation coefficient matrix.

[0051] Normalization is a crucial step in data preprocessing. The factor correlation coefficient matrix is ​​normalized according to the geological condition dimension. Normalization is calculated by subtracting the minimum correlation coefficient under that geological condition from each correlation coefficient, and then dividing by the difference between the maximum and minimum correlation coefficients. The normalized values ​​range from zero to one, eliminating the influence of differences in numerical magnitudes under different geological conditions. Gradient descent is a commonly used optimization algorithm in machine learning. It determines the gradient direction by calculating the partial derivative of the loss function with respect to the weight parameters. The weight parameters are updated in the opposite direction of the gradient, with the update step size controlled by the learning rate. The iterative process continues until the loss function converges or the preset number of iterations is reached. The optimized weight parameters are obtained after iterative optimization.

[0052] Sensitivity weight allocation is based on optimized weight parameters. It quantifies and assigns sensitivity weights to four meteorological factors—rainfall intensity, soil saturation, groundwater level change, and slope runoff coefficient—under different geological conditions. The principle of weight allocation is that meteorological factors with stronger correlations receive higher weight values. Weight values ​​are calculated by multiplying the optimized weight parameters by the correlation coefficients. The matrix structure is constructed with geological conditions as rows and meteorological factors as columns. Each matrix element represents the sensitivity weight of a specific meteorological factor under specific geological conditions. This matrix structure facilitates subsequent matrix operations and data indexing. The completed matrix is ​​called the initial sensitivity weight matrix.

[0053] Expert knowledge verification, based on professional experience and theoretical knowledge in the field of geological engineering, verifies and adjusts the rationality of the weight allocation in the initial sensitivity weight matrix. The verification process includes checking whether the weight allocation conforms to the mechanism of geological disaster occurrence; for example, the weight of groundwater level change should be higher under high-permeability soil conditions, and the weight of rainfall intensity should be higher under low-permeability soil conditions. The verification process corrects unreasonable weight values ​​through expert experience, including adjusting, lowering, or redistributing weight values. After verification, the geological disaster sensitivity weight matrix is ​​obtained.

[0054] In one specific embodiment, the process of executing step S103 may specifically include the following steps:

[0055] The geological disaster meteorological dataset is processed by sequential segmentation according to time windows to construct meteorological time-series data with continuous time steps;

[0056] Based on the geological hazard sensitivity weight matrix, each meteorological factor in the meteorological time series data is processed by element-wise weighted multiplication to obtain a weighted meteorological feature sequence.

[0057] Geological condition information such as slope, lithology, and soil type is vector-encoded through an embedding layer to obtain geological condition embedding vectors.

[0058] Based on the geological condition embedding vector, the forget gate, input gate and output gate of the geological constraint long short-term memory network are subjected to geophysical constraint modulation processing. The weighted meteorological feature sequence is input into the modulated network for temporal feature learning to obtain the geological constraint feature vector.

[0059] The geological constraint feature vector is processed by calculating the probability distribution of the geological hazard risk vector through a fully connected layer and a softmax activation function.

[0060] Specifically, the time window sequence segmentation process segments the geological disaster meteorological dataset according to a preset time window length. The time window refers to the length of a continuous time series used to train the neural network. The segmentation process is implemented using the sliding window technique. The sliding window starts from the beginning time point of the dataset and moves forward at a fixed step size. At each window position, a corresponding length of continuous data is extracted to form independent time series samples. These samples constitute meteorological time series data with a continuous time step. The structure of the time series data is a three-dimensional tensor. The first dimension represents the number of samples, the second dimension represents the time step, and the third dimension represents the number of meteorological factor features. The element-wise weighted product processing adjusts the weights of each meteorological factor in the meteorological time series data based on the geological hazard sensitivity weight matrix. The weighted product refers to the operation of multiplying the weight values ​​in the weight matrix with the corresponding meteorological factor values ​​one by one. The multiplication process needs to consider the matching relationship of geological conditions. Each time series sample corresponds to specific geological conditions. The weight values ​​of the row corresponding to the geological conditions in the weight matrix are multiplied element-wise with the meteorological factor values ​​in that sample. The result of the product operation is the adjusted meteorological factor value, which reflects the importance of different meteorological factors under specific geological conditions. The weighted data is reorganized according to the original time series structure to form a weighted meteorological feature sequence.

[0061] The embedding layer vector encoding process transforms discrete geological condition information into continuous numerical vectors. Embedding layers are a commonly used feature representation method in deep learning, mapping discrete identifiers to dense vectors through a trainable weight matrix. The slope information encoding process first discretizes continuous slope values ​​according to preset intervals, with each slope level corresponding to a unique integer identifier. These identifiers are then converted into fixed-dimensional vector representations through the embedding layer. The lithology information encoding process converts lithology categories such as granite, shale, and sandstone into one-hot encoded forms. One-hot encoding refers to a binary vector with only one position set to 1 and the rest to zero. These one-hot encodings are further converted into dense vectors through the embedding layer. The soil type information encoding process assigns numerical labels to categories such as clay, sand, and silt. These labels are also converted into vector representations through the embedding layer. The vector representations of the three geological conditions are combined through a concatenation operation to form the geological condition embedding vector.

[0062] Geophysical constraint modulation processing modulates the gating mechanism of a Long Short-Term Memory (LSTM) network based on geological condition embedding vectors. The LTM network comprises three gating structures: a forget gate, an input gate, and an output gate. Each gating structure controls the flow of information through a sigmoid activation function. The forget gate determines which information is discarded from the cell state. Its calculation involves a linear combination of the current input, the previous hidden state, and the geological condition embedding vector. This linear combination is achieved through matrix multiplication of the weight matrix and the input vector. The product is then processed by the sigmoid function to obtain a gating value between zero and one. The input gate determines which new information is stored in the cell state. Its calculation is similar to the forget gate, but uses different weight matrices and bias parameters. The output gate controls which parts of the cell state are output. Its calculation also involves the geological condition embedding vector. Modulation processing modifies the weight parameters of each gating mechanism using the geological condition embedding vector, enabling the network to adjust its information processing strategy according to different geological conditions. A weighted meteorological feature sequence is input into the modulated network for forward propagation. The network learns temporal patterns through multi-level feature transformations to obtain geological constraint feature vectors. The probability distribution calculation process transforms the geological constraint feature vector into a risk probability output. The fully connected layer, a linear transformation layer in a neural network, calculates the linear output through the matrix multiplication of the weight matrix and the input vector. This linear output, plus a bias term, forms the output of the fully connected layer. The softmax activation function, commonly used in multi-class classification problems, transforms any real-valued vector into a probability distribution. The transformation process first calculates an exponential function for each element of the input vector, then normalizes all exponential values ​​by dividing each exponential value by the sum of all exponential values. The output of the softmax function is a probability vector, where each element represents the probability value of the corresponding class, and the sum of all probability values ​​equals one. The geological constraint feature vector undergoes a linear transformation through the fully connected layer. The transformed vector is then used to calculate the geological hazard risk probability vector through the softmax activation function, containing probability values ​​for multiple risk levels.

[0063] In one specific embodiment, the process of performing vector encoding of geological condition information such as slope, lithology, and soil type through an embedding layer can specifically include the following steps:

[0064] The slope values ​​are discretized in a hierarchical manner, and the continuous slope values ​​are classified and coded according to the preset slope intervals to obtain the slope classification code.

[0065] Based on lithological type, rock types are numerically mapped to convert granite, shale and sandstone lithological categories into unique thermal coding forms to obtain lithological coding vectors.

[0066] Based on the soil classification standard, soil types are labeled and coded, and clay, sandy soil, and silt are numerically identified to obtain soil type codes.

[0067] The slope classification code, lithology code vector and soil type code are spliced ​​and combined to construct a comprehensive geological condition code sequence, and the geological condition input vector is obtained.

[0068] The geological condition input vector is subjected to linear transformation and nonlinear activation processing through a trainable weight matrix to obtain the geological condition embedding vector.

[0069] Specifically, the generation process of the geological condition embedding vector includes five core data encoding and transformation steps, each involving specific data processing algorithms and numerical calculations. Slope classification discretization converts continuous slope measurements into classification identifiers. Classification discretization involves dividing the continuous numerical space into several non-overlapping intervals, each corresponding to a discrete category. Slope values ​​are obtained through topographic surveying equipment, with the measurement results representing the angle between the ground surface and the horizontal plane. Pre-defined slope intervals are divided according to geological engineering standards, including categories such as gentle slopes, medium slopes, steep slopes, and dangerous slopes. The classification encoding process is implemented through a conditional judgment algorithm, which compares the slope value with the boundary values ​​of each interval, assigning the slope value to the corresponding interval. Each interval corresponds to a unique integer code, and the encoding result forms the slope classification code. Lithology numerical mapping converts qualitative rock type information into quantitative numerical representations. Lithology types are determined through geological exploration and rock sample analysis, including attributes such as mineral composition, structural characteristics, and formation environment. One-hot encoding is a commonly used categorical variable encoding method in machine learning. The encoding process creates a binary vector for each lithology category, with the vector length equal to the total number of lithology categories. Positions corresponding to the lithology category are set to 1, and the remaining positions are set to zero. Granite, shale, and sandstone each have different one-hot encoding patterns. The encoding process is achieved by creating a zero vector and setting the corresponding position to 1. The encoding results maintain the independence between lithology categories, avoiding artificial numerical relationships, and the lithology encoding vector contains lithology type classification information.

[0070] Soil labeling and coding is based on soil classification standards, which convert soil types into numerical identifiers. These standards are determined according to physical properties such as soil particle composition, plasticity, and bearing capacity. Categories like clay, sand, and silt possess different engineering geological characteristics. Labeling and coding involves assigning a unique integer identifier to each soil category. This assignment follows predetermined coding rules, and the coding process is implemented using lookup tables or mapping dictionaries. This maps soil category names to corresponding numerical identifiers, resulting in a soil type code. The coding maintains the discreteness and uniqueness of soil classification.

[0071] The splicing and combination process merges the coding results of the three geological conditions into a unified data structure. The splicing operation refers to connecting multiple vectors or values ​​into a longer vector in a specific order. The slope classification code, as a scalar value, is first converted into vector form. The lithology code vector retains its original vector structure, and the soil type code is also converted into vector form. The three vectors are spliced ​​end-to-end, and the length of the spliced ​​vector is equal to the sum of the lengths of the individual component vectors. The splicing order follows a fixed sequence of slope, lithology, and soil type. The splicing result constructs a comprehensive geological condition coding sequence, which contains all the classification information of the geological conditions, forming the geological condition input vector.

[0072] Linear transformation and nonlinear activation processing convert the geological condition input vector into a dense feature representation. Linear transformation is achieved through matrix multiplication of a trainable weight matrix and the input vector. The dimension of the weight matrix is ​​determined by the length of the input vector and the dimension of the target embedding. The matrix multiplication process involves row-by-row and column-by-column multiplication and addition. The result, plus the bias vector, forms the output of the linear transformation. Nonlinear activation processing applies a nonlinear mapping to the linear transformation result using activation functions, including ReLU, tanh, and sigmoid. The choice of activation function affects the numerical distribution and expressive power of the embedded vector. The vector after activation processing becomes the geological condition embedding vector, which has fixed dimensions and continuous numerical features, capable of representing the complex relationships between geological conditions.

[0073] For example, geological exploration results at a monitoring point in a mountainous area show a slope of 35 degrees, granite as the lithology, and weathered soil as the soil. Slope classification discretization classifies 35 degrees into the medium slope category, with a corresponding classification code of integer value two. In the lithology numerical mapping process, the unique thermal codes for granite, shale, and sandstone are three in length. The code vector for granite is a vector with the first position set to one and the remaining positions set to zero. Shale and sandstone each correspond to vectors with different positions set to one. Soil labeling and coding classifies weathered soil into a specific category, with a corresponding numerical identifier of integer value one. The splicing and combination process converts the slope code into a single-element vector, which is then connected end-to-end with the lithology and soil code vectors. The resulting geological condition input vector contains all the coded information for slope, lithology, and soil. During the linear transformation, the input vector is multiplied by the weight matrix. The parameters of the weight matrix are learned through machine learning training. The transformation result is processed by the ReLU activation function, which sets negative values ​​to zero and keeps positive values ​​unchanged. The processed vector becomes the geological condition embedding vector of the monitoring point. The value of the embedding vector reflects the geological characteristics of the moderately sloped granite weathered soil slope.

[0074] In one specific embodiment, the process of executing step S104 may specifically include the following steps:

[0075] The slope stability is calculated in real time based on the Bishop circular sliding method. The current safety factor is calculated based on the effective cohesion, internal friction angle, soil weight, sliding body thickness, and slope angle parameters to obtain the real-time safety factor.

[0076] The real-time safety factor within a continuous time period is processed by differential operation, and the ratio of the difference in safety factor between adjacent time points to the time interval is calculated to obtain the rate of change of safety factor.

[0077] The rate of change of the safety factor is compared with a preset decrease threshold for judgment and processing. When the rate of change is continuously negative and the absolute value exceeds the threshold, the sensitivity adjustment mechanism is triggered to obtain the threshold adjustment trigger signal.

[0078] Based on the threshold adjustment trigger signal, the baseline warning threshold is adjusted upwards by a percentage, and the three warning thresholds of danger, warning and attention are adjusted upwards simultaneously to obtain the adjusted warning threshold.

[0079] Based on the soil moisture content correction coefficient, the adjusted early warning threshold is corrected for water influence. The ratio of the current saturation to the field capacity saturation is calculated as the correction factor to obtain the adaptive early warning threshold parameter.

[0080] Specifically, the Bishop circular arc sliding method is based on slope stability theory. This method assumes the sliding surface is circular and calculates the slope safety factor through the principle of moment balance. The safety factor calculation requires key parameters such as effective cohesion, internal friction angle, soil weight, sliding mass thickness, and slope angle. Effective cohesion is determined through triaxial compression or direct shear tests, reflecting the bond strength between soil particles. The internal friction angle is determined through soil mechanics tests, representing the magnitude of internal frictional resistance in the soil. Soil weight is obtained through field sampling and indoor density testing. The sliding mass thickness is determined through geological exploration and borehole data, and the slope angle is obtained through topographic surveys. The Bishop circular arc sliding method divides the sliding mass into several vertical blocks. The weight, bottom inclination angle, and effective stress of each block are calculated separately. The safety factor is equal to the ratio of the resisting moment to the sliding moment. The calculation process uses an iterative method because the safety factor itself appears in the denominator of the calculation formula. The iterative calculation starts from the assumed initial value of the safety factor and gradually corrects it until convergence, yielding the real-time safety factor. Differential operations are used to perform numerical analysis of real-time safety factors over continuous time periods. Differential operations are a fundamental method for calculating the rate of change of a function in numerical analysis. The rate of change of the safety factor is calculated by subtracting the safety factor values ​​at adjacent time points to obtain the difference. This difference reflects the magnitude of change in the safety factor per unit time. The time interval is typically set to one hour. The rate of change is calculated by dividing the difference by the time interval. The result of the division operation represents the rate of change of the safety factor per hour; a positive value indicates an increase in the safety factor, and a negative value indicates a decrease. The absolute value of the rate of change reflects the drasticness of the change. The rates of change at multiple consecutive time points form a rate of change sequence, which reflects the dynamic evolution trend of slope stability.

[0081] The comparative judgment process compares the rate of change of the safety factor with a preset decline threshold. This comparison is implemented using a conditional judgment algorithm. The decline threshold is determined based on geological engineering experience and historical disaster statistics. The threshold setting needs to balance the sensitivity and accuracy of the early warning; a threshold that is too small will lead to frequent false alarms, while a threshold that is too large will reduce the timeliness of the early warning. The comparative judgment involves a logical AND operation of two conditions. The first condition checks whether the rate of change is negative, indicating a declining trend in the safety factor. The second condition checks whether the absolute value of the rate of change exceeds the preset threshold; exceeding the threshold indicates an excessively rapid decline. When both conditions are met simultaneously, a sensitivity adjustment mechanism is triggered. This triggering process is implemented using Boolean logic operations. A true value is output when the conditions are met, and a false value is output when the conditions are not met. The true value signal serves as the threshold adjustment trigger signal.

[0082] The percentage increase processing adjusts the baseline warning threshold based on the threshold adjustment trigger signal. Percentage increase refers to increasing the threshold value by a fixed proportion. The baseline warning threshold includes three levels: Danger, Warning, and Caution. Each level corresponds to a different risk probability range, with Danger having the lowest threshold and Caution having the highest. The increase processing applies the thresholds for all three levels simultaneously. The increase percentage is determined based on the severity of the decrease in the safety factor. The increase calculation is obtained by multiplying the original threshold by an increase coefficient, which is usually set to a value greater than one. The increased threshold improves the sensitivity of the warning, enabling the warning system to issue alerts earlier. The adjusted warning threshold reflects the current deterioration of slope stability.

[0083] The moisture impact correction process further adjusts the pre-warning threshold based on a soil moisture content correction coefficient, which reflects the influence of soil moisture conditions on slope stability. The current saturation is obtained in real-time through a soil moisture sensor. Field capacity saturation, the maximum water content that soil can maintain under gravity, is determined through soil physical experiments. The correction factor is calculated by dividing the current saturation by the field capacity saturation. A correction factor greater than one indicates that the soil moisture content exceeds the normal level, while a correction factor less than one indicates that the soil is relatively dry. The correction is obtained by dividing the adjusted pre-warning threshold by the correction factor. This division operation takes into account the influence of soil moisture on slope stability; the threshold decreases at high moisture content and increases at low moisture content. The corrected value becomes the adaptive pre-warning threshold parameter.

[0084] In one specific embodiment, the process of executing step S105 may specifically include the following steps:

[0085] The risk probability of the physical model is calculated using a normal distribution based on the real-time safety factor. The difference between the safety factor and the benchmark value is divided by the standard deviation as the input parameter to obtain the risk probability of the physical model.

[0086] The geological parameters were randomly perturbed by Monte Carlo sampling, and the variance of the prediction results was calculated by multiple random samplings of the effective cohesion and internal friction angle parameters to obtain the uncertainty quantification probability.

[0087] The geological hazard risk probability vector, physical model risk probability, and uncertainty quantification probability are weighted and linearly combined, and then probability fusion calculation is performed according to preset weight coefficients to obtain the fused risk probability.

[0088] The fusion risk probability is graded and judged according to the adaptive early warning threshold parameter. The fusion probability is compared with the four levels of thresholds: danger, warning, attention, and safety, to obtain the early warning level judgment result.

[0089] Based on the early warning level determination results, the early warning information is formatted and output to generate an early warning report containing risk level, probability value, and duration prediction, thus obtaining graded geological disaster early warning results.

[0090] Specifically, the normal distribution calculation quantifies the risk probability of the physical model based on the real-time safety factor. The normal distribution is a commonly used continuous probability distribution in probability theory, characterized by a bell-shaped curve. Its distribution parameters include the mean and standard deviation. The benchmark value is usually set as the critical value of the safety factor. When the safety factor equals the benchmark value, it indicates that the slope is in a critical stable state. The difference between the safety factor and the benchmark value reflects the degree to which the current stability deviates from the critical state; a positive difference indicates good stability, and a negative difference indicates insufficient stability. The standard deviation parameter is determined through statistical analysis of historical safety factor data, reflecting the dispersion of the safety factor. A larger standard deviation indicates more drastic fluctuations in the safety factor. The process of dividing the difference by the standard deviation is a standardization process. The standardized value serves as the input parameter for the normal distribution. The normal distribution calculation obtains the corresponding cumulative probability through table lookup or numerical integration methods. This probability represents the likelihood that the safety factor will be lower than the current value, and the probability value reflects the risk level predicted by the physical model. The Monte Carlo sampling method randomly perturbs the geological parameters. The Monte Carlo method is a numerical calculation technique based on random sampling, obtaining numerical solutions through a large number of random experiments. Effective cohesion and internal friction angle are key geological parameters affecting slope stability. These parameters exhibit uncertainty in practical engineering, stemming from factors such as measurement errors, spatial variability, and differences in experimental conditions. Random disturbance processing is achieved by adding random noise to the original parameter values. This random noise typically follows a normal or uniform distribution, and its amplitude is determined based on the uncertainty level of the parameters. Multiple random sampling processes generate numerous parameter combinations, each corresponding to a safety factor calculation. The results of these multiple calculations form a probability distribution of the safety factor. The variance of the predicted result is obtained by summing the squares of the differences between all results and the mean, then dividing by the sample size. Variance reflects the degree of uncertainty in the prediction; a larger variance indicates lower prediction reliability. The quantified probability of uncertainty is obtained by comparing the variance with a preset threshold.

[0091] Weighted linear combination (WLC) integrates probabilistic information from three different sources, serving as a fundamental method for multi-source information fusion. The geological hazard risk probability vector originates from machine learning predictions using a geologically constrained long short-term memory network (LSTM). The physical model risk probability is derived from mechanical calculations using Bishop's circular sliding method, and the uncertainty quantification probability comes from statistical analysis using Monte Carlo sampling. These three probabilistic information sources possess different physical meanings and reliability levels. Preset weight coefficients reflect the importance and credibility of each information source. Determining these weight coefficients requires comprehensive consideration of factors such as the accuracy, stability, and applicability of each method. The sum of the weight coefficients equals one to ensure the normalization of the probability. Probability fusion is achieved by summing the product of each probability and its corresponding weight coefficient. The fused probability integrates the data-driven capabilities of machine learning, the mechanistic explanatory power of the physical model, and the uncertainty quantification capabilities of statistical methods. The fused risk probability reflects the comprehensive judgment result of multi-source information.

[0092] The tiered judgment process classifies the probability of fusion risk into levels based on adaptive warning threshold parameters. This tiered judgment is a crucial step in risk assessment. The adaptive warning threshold parameters include four levels: Danger, Warning, Caution, and Safety. Each threshold corresponds to a different risk probability range, and the thresholds are determined based on historical disaster statistics and engineering experience. The step-by-step comparison process starts with the highest risk level. First, it checks whether the fusion probability exceeds the Danger level threshold. If it does, it is classified as Dangerous; otherwise, it checks whether it exceeds the Warning level threshold, and so on, until a specific risk level is determined. The comparison process is implemented using a conditional judgment algorithm with an if-else logic structure. Each condition corresponds to a risk level judgment, and the judgment result is expressed as a warning level in numerical or textual form. The warning level determination result lays the foundation for subsequent warning information generation.

[0093] The formatted output processing generates standardized early warning information based on the early warning level determination results. Formatted output is the standardization process of information expression. The early warning information includes key elements such as risk level, probability value, and duration prediction. The risk level is directly obtained from the early warning level determination results, the probability value comes from the calculation results of fused risk probabilities, and the duration prediction is based on meteorological forecast data and statistical analysis of historical disaster durations. The early warning report generation process includes steps such as information collection, formatting, and content verification. Information collection summarizes the results of each processing step, formatting organizes the information content according to a predetermined template structure, and content verification checks the completeness and consistency of the information. The graded geological disaster early warning results are output in the form of structured data or documents. The output content includes a clear identifier of the early warning level, a quantitative value of the risk probability, and a specific range of the effective warning time. The output format facilitates rapid understanding and response by emergency management departments.

[0094] The above describes the machine learning-based geological disaster meteorological risk early warning method in the embodiments of this application. The following describes the machine learning-based geological disaster meteorological risk early warning system in the embodiments of this application. Please refer to [link / reference]. Figure 2 One embodiment of the geological disaster meteorological risk early warning system based on machine learning in this application includes:

[0095] The data acquisition module is used to collect and process data on rainfall intensity, soil saturation, groundwater level changes, and slope runoff coefficient in the target area in real time through multi-source sensors, so as to obtain a geological disaster meteorological dataset.

[0096] The allocation module is used to perform sensitivity weight allocation processing on each meteorological factor in the geological disaster meteorological dataset according to geological conditions, so as to obtain a geological disaster sensitivity weight matrix.

[0097] The weighting module is used to perform weighted fusion processing on the geological hazard sensitivity weight matrix and meteorological time series data, and to extract nonlinear features from the fused data through a geologically constrained long short-term memory network to obtain a geological hazard risk probability vector.

[0098] The adjustment module is used to dynamically adjust the warning threshold according to the rate of change of the safety factor. When the rate of change of the safety factor continues to decrease, the threshold sensitivity is increased to obtain an adaptive warning threshold parameter.

[0099] The fusion module is used to perform Bayesian fusion processing on the geological disaster risk probability vector and the adaptive early warning threshold parameter to obtain the graded geological disaster early warning result.

[0100] above Figure 2 The geological disaster meteorological risk early warning system based on machine learning in this embodiment of the invention will be described in detail from the perspective of modular functional entities. The geological disaster meteorological risk early warning device based on machine learning in this embodiment of the invention will be described in detail from the perspective of hardware processing.

[0101] Reference Figure 3 This invention also provides a machine learning-based geological disaster meteorological risk early warning device, which can be a server, and its internal structure can be as follows: Figure 3As shown, the machine learning-based geological disaster meteorological risk early warning device includes a processor, memory, display screen, input device, network interface, and database connected via a system bus. The processor provides computational and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database stores the data corresponding to this embodiment. The network interface allows communication with external terminals via a network connection. The computer program is executed by the processor to implement the aforementioned method.

[0102] Those skilled in the art will understand that Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present invention and does not constitute a limitation on the machine learning-based geological disaster meteorological risk early warning device to which the present invention is applied.

[0103] The present invention also provides a computer-readable storage medium, which can be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium, wherein the computer-readable storage medium stores instructions that, when the instructions are executed on a computer, cause the computer to perform the steps of the machine learning-based geological disaster meteorological risk early warning method.

[0104] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0105] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a machine learning-based geological disaster meteorological risk early warning device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0106] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for early warning of meteorological risks of geological disasters based on machine learning, characterized in that, The method includes: By using multi-source sensors to collect and process real-time data on rainfall intensity, soil saturation, groundwater level changes, and slope runoff coefficient in the target area, a geological disaster meteorological dataset is obtained. Based on the geological condition classification, sensitivity weights are assigned to each meteorological factor in the geological disaster meteorological dataset to obtain a geological disaster sensitivity weight matrix. The geological hazard sensitivity weight matrix is ​​weighted and fused with meteorological time series data. The fused data is then subjected to nonlinear feature extraction through a geologically constrained long short-term memory network to obtain a geological hazard risk probability vector. The warning threshold is dynamically adjusted based on the rate of change of the safety factor. When the rate of change of the safety factor continues to decrease, the threshold sensitivity is increased to obtain an adaptive warning threshold parameter. The geological disaster risk probability vector and the adaptive early warning threshold parameter are fused using Bayesian processing to obtain the graded geological disaster early warning results.

2. The geological disaster meteorological risk early warning method based on machine learning according to claim 1, characterized in that, The process involves real-time data collection and processing of rainfall intensity, soil saturation, groundwater level changes, and slope runoff coefficient in the target area using multi-source sensors to obtain a geological disaster meteorological dataset, including: Rainfall intensity data is collected and processed hourly by automatic weather stations. The collected data is then compared and filtered with the set effective rainfall intensity range to obtain an effective rainfall intensity sequence. Soil saturation data is monitored and processed in real time by a soil moisture sensor. The collected raw saturation values ​​are constrained according to a preset physical reasonable range to obtain a standard soil saturation sequence. By continuously collecting and processing groundwater level change data through a groundwater level monitoring instrument, calculating the water level difference between adjacent time points and dividing by the time interval, a groundwater level change rate sequence is obtained. Based on the slope flow monitoring equipment, runoff data is collected and processed, and the measured runoff is divided by the corresponding rainfall to calculate the runoff coefficient value, thus obtaining the slope runoff coefficient sequence. The effective rainfall intensity sequence, standard soil saturation sequence, groundwater level change rate sequence, and slope runoff coefficient sequence are time-synchronized and aligned to obtain a geological disaster meteorological dataset.

3. The geological disaster meteorological risk early warning method based on machine learning according to claim 1, characterized in that, The sensitivity weight allocation process for each meteorological factor in the geological disaster meteorological dataset, based on geological condition classification, yields a geological disaster sensitivity weight matrix, including: Based on geological exploration data of the target area, slope, lithology, and soil type are classified and identified to obtain geological condition classification labels; Based on the geological condition classification labels, historical geological disaster cases are classified and statistically processed, and the correlation coefficients between different meteorological factors and disaster occurrence under various geological conditions are calculated to obtain the factor correlation coefficient matrix. The factor correlation coefficient matrix is ​​normalized according to the geological condition dimension, and the weight parameters are iteratively optimized using the gradient descent algorithm to obtain the optimized weight parameters. Based on the optimized weight parameters, sensitivity weights are assigned to rainfall intensity, soil saturation, groundwater level change, and slope runoff coefficient. A matrix structure with geological conditions as rows and meteorological factors as columns is constructed to obtain the initial sensitivity weight matrix. Based on geological engineering experience, the initial sensitivity weight matrix is ​​subjected to expert knowledge verification to obtain the geological hazard sensitivity weight matrix.

4. The geological disaster meteorological risk early warning method based on machine learning according to claim 1, characterized in that, The process of weighting and fusing the geological hazard sensitivity weight matrix with meteorological time-series data, and then extracting nonlinear features from the fused data using a geologically constrained long short-term memory network to obtain a geological hazard risk probability vector includes: The geological disaster meteorological dataset is processed by sequential segmentation according to time windows to construct meteorological time-series data with continuous time steps; Based on the geological hazard sensitivity weight matrix, each meteorological factor in the meteorological time series data is processed by element-wise weighted multiplication to obtain a weighted meteorological feature sequence. Geological condition information such as slope, lithology, and soil type is vector-encoded through an embedding layer to obtain geological condition embedding vectors. Based on the geological condition embedding vector, the forget gate, input gate, and output gate of the geological constraint long short-term memory network are subjected to geophysical constraint modulation processing. The weighted meteorological feature sequence is then input into the modulated network for temporal feature learning to obtain the geological constraint feature vector. The geological constraint feature vector is processed by probability distribution calculation using a fully connected layer and a softmax activation function to obtain a geological disaster risk probability vector.

5. The geological disaster meteorological risk early warning method based on machine learning according to claim 4, characterized in that, The geological condition information, including slope, lithology, and soil type, is vector-encoded through an embedding layer to obtain a geological condition embedding vector, including: The slope values ​​are discretized in a hierarchical manner, and the continuous slope values ​​are classified and coded according to the preset slope intervals to obtain the slope classification code. Based on lithological type, rock types are numerically mapped to convert granite, shale and sandstone lithological categories into unique thermal coding forms to obtain lithological coding vectors. Based on the soil classification standard, soil types are labeled and coded, and clay, sandy soil, and silt are numerically identified to obtain soil type codes. The slope classification code, lithology code vector, and soil type code are spliced ​​and combined to construct a comprehensive geological condition code sequence, thus obtaining the geological condition input vector. The geological condition input vector is subjected to linear transformation and nonlinear activation processing through a trainable weight matrix to obtain the geological condition embedding vector.

6. The geological disaster meteorological risk early warning method based on machine learning according to claim 1, characterized in that, The process of dynamically adjusting the warning threshold based on the rate of change of the safety factor, and increasing the threshold sensitivity when the rate of change of the safety factor continues to decrease, yields adaptive warning threshold parameters, including: The slope stability is calculated in real time based on the Bishop circular sliding method. The current safety factor is calculated based on the effective cohesion, internal friction angle, soil weight, sliding body thickness, and slope angle parameters to obtain the real-time safety factor. The real-time safety factor within a continuous time period is processed by differential operation, and the ratio of the difference in safety factor between adjacent time points to the time interval is calculated to obtain the rate of change of safety factor. The change rate of the safety factor is compared with a preset decrease threshold for judgment and processing. When the change rate is continuously negative and the absolute value exceeds the threshold, the sensitivity adjustment mechanism is triggered to obtain the threshold adjustment trigger signal. Based on the threshold adjustment trigger signal, the baseline warning threshold is adjusted upward by a percentage, and the three warning thresholds of danger, warning and attention are adjusted upward simultaneously to obtain the adjusted warning threshold. Based on the soil moisture content correction coefficient, the adjusted early warning threshold is corrected for water influence. The ratio of the current saturation to the field capacity saturation is calculated as a correction factor to obtain the adaptive early warning threshold parameter.

7. The geological disaster meteorological risk early warning method based on machine learning according to claim 6, characterized in that, The step of performing Bayesian fusion processing on the geological disaster risk probability vector and the adaptive early warning threshold parameter to obtain graded geological disaster early warning results includes: Based on the real-time safety factor, the risk probability of the physical model is calculated using a normal distribution. The difference between the safety factor and the benchmark value is divided by the standard deviation as the input parameter to obtain the risk probability of the physical model. The geological parameters were randomly perturbed by Monte Carlo sampling, and the variance of the prediction results was calculated by multiple random samplings of the effective cohesion and internal friction angle parameters to obtain the uncertainty quantification probability. The geological hazard risk probability vector, physical model risk probability, and uncertainty quantification probability are weighted and linearly combined, and probability fusion calculation is performed according to preset weight coefficients to obtain the fused risk probability. The fusion risk probability is graded and judged according to the adaptive warning threshold parameter. The fusion probability is compared with the four level thresholds of danger, warning, attention and safety level to obtain the warning level judgment result. Based on the warning level determination results, the warning information is formatted and output to generate a warning report containing risk level, probability value, and duration prediction, thus obtaining graded geological disaster warning results.

8. A geological disaster meteorological risk early warning system based on machine learning, characterized in that, For implementing the machine learning-based geological disaster meteorological risk early warning method as described in any one of claims 1-7, the machine learning-based geological disaster meteorological risk early warning system comprises: The data acquisition module is used to collect and process data on rainfall intensity, soil saturation, groundwater level changes, and slope runoff coefficient in the target area in real time through multi-source sensors, so as to obtain a geological disaster meteorological dataset. The allocation module is used to perform sensitivity weight allocation processing on each meteorological factor in the geological disaster meteorological dataset according to geological conditions, so as to obtain a geological disaster sensitivity weight matrix. The weighting module is used to perform weighted fusion processing on the geological hazard sensitivity weight matrix and meteorological time series data, and to extract nonlinear features from the fused data through a geologically constrained long short-term memory network to obtain a geological hazard risk probability vector. The adjustment module is used to dynamically adjust the warning threshold according to the rate of change of the safety factor. When the rate of change of the safety factor continues to decrease, the threshold sensitivity is increased to obtain an adaptive warning threshold parameter. The fusion module is used to perform Bayesian fusion processing on the geological disaster risk probability vector and the adaptive early warning threshold parameter to obtain the graded geological disaster early warning result.

9. A geological disaster meteorological risk early warning device based on machine learning, characterized in that, The system includes a memory and a processor, wherein the memory stores a computer program that can run on the processor, and the processor executes the computer program to implement the machine learning-based geological disaster meteorological risk early warning method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is run by the processor, it causes the processor to execute the machine learning-based geological disaster meteorological risk early warning method as described in any one of claims 1 to 7.

Images