A landslide risk prediction compensation method, system, device and medium

Abstract

The application relates to a landslide risk prediction compensation method, system, device and medium. The method comprises the following steps: acquiring rainfall monitoring data, soil body parameter data and historical slope stability evaluation data of a target area, wherein the soil body parameter data comprises soil body crack distribution data; performing time sequence segmentation feature extraction on the rainfall monitoring data to obtain rainfall time sequence segmentation feature data; performing infiltration response correlation analysis on the soil body parameter data and the rainfall time sequence segmentation feature data to obtain rainfall infiltration feature sequence data; inputting the rainfall infiltration feature sequence data into a dynamic lag compensation model for processing to obtain time sequence weighted response data; performing nonlinear mapping compensation based on the soil body crack distribution data to obtain dynamic pore pressure compensation data; and inputting the dynamic pore pressure compensation data and the historical slope stability evaluation data into a landslide risk prediction model to obtain landslide risk early warning data. The method can improve the accuracy and prediction period of landslide risk early warning and reduce false positives and false negatives.

Description

Technical Field

[0001] This invention belongs to the field of disaster monitoring, and in particular relates to a landslide risk forecasting and compensation method, system, equipment and medium. Background Technology

[0002] With the development of geological disaster monitoring and early warning technology, landslide risk prediction technology based on the relationship between rainfall and slope pore pressure response has emerged. This technology assesses the possibility of landslide instability by continuously monitoring rainfall data and establishing the correlation between rainfall and changes in pore water pressure inside the slope, and has become an important means of early warning for rainfall-induced landslides.

[0003] Traditional techniques typically employ rainfall threshold methods or linear statistical regression models to directly compare real-time rainfall monitoring data with preset empirical critical thresholds, or to establish a deterministic correspondence between rainfall input and pore pressure output through linear fitting to achieve landslide early warning and forecasting.

[0004] However, current forecasting methods have the following problems: the process of rainfall infiltrating into the deep part of the slope and causing the pore water pressure to rise has a complex time lag effect, and this lag effect is significantly affected by the non-uniform distribution of cracks in the soil, exhibiting strong spatiotemporal variability. The traditional methods of linear modeling and fixed threshold judgment cannot accurately characterize the dynamic lag relationship between rainfall and pore pressure response, nor can they reflect the spatial modulation effect of non-uniform crack distribution on pore pressure response, resulting in insufficient lead time for landslide warnings and easily causing delayed warning response or false alarms and omissions. Summary of the Invention

[0005] Therefore, it is necessary to provide a landslide risk prediction and compensation method, system, equipment, and medium to address the aforementioned technical problems.

[0006] Firstly, this application provides a landslide risk prediction and compensation method, including:

[0007] S1. Obtain rainfall monitoring data, soil parameter data, and historical slope stability assessment data for the target area; soil parameter data includes soil fissure distribution data.

[0008] S2. Extract time-series segmented features from rainfall monitoring data to obtain rainfall time-series segmented feature data;

[0009] S3. Perform infiltration response correlation analysis on soil parameter data and rainfall time-series segmented characteristic data to obtain rainfall infiltration characteristic sequence data;

[0010] S4. Input the rainfall infiltration characteristic sequence data into the dynamic lag compensation model to perform time-series lag dependency modeling and dynamic weight allocation to obtain time-series weighted response data.

[0011] S5. Based on soil fracture distribution data, nonlinear mapping compensation is performed on time-series weighted response data to obtain dynamic pore pressure compensation data; dynamic pore pressure compensation data is used to characterize the time-varying characteristics of pore pressure response under non-uniform fracture distribution conditions;

[0012] S6. Input the dynamic pore pressure compensation data and the historical slope stability assessment data into the landslide risk prediction model to predict and classify landslide risks, and obtain landslide risk early warning data.

[0013] Secondly, this application also provides a landslide risk prediction and compensation system, including:

[0014] The multidimensional data acquisition module is used to acquire rainfall monitoring data, soil parameter data, and historical slope stability assessment data for the target area; the soil parameter data includes soil fracture distribution data.

[0015] The rainfall time-series feature extraction module is used to extract time-series segmented features from rainfall monitoring data to obtain rainfall time-series segmented feature data;

[0016] The infiltration response correlation analysis module is used to perform infiltration response correlation analysis on soil parameter data and rainfall time-series segmented characteristic data to obtain rainfall infiltration characteristic sequence data.

[0017] The dynamic lag compensation module is used to input rainfall infiltration characteristic sequence data into the dynamic lag compensation model to perform time-series lag dependency modeling and dynamic weight allocation, and obtain time-series weighted response data.

[0018] The dynamic pore pressure compensation module is used to perform nonlinear mapping compensation on time-weighted response data based on soil fracture distribution data to obtain dynamic pore pressure compensation data; the dynamic pore pressure compensation data is used to characterize the time-varying features of pore pressure response under non-uniform fracture distribution conditions;

[0019] The risk classification and prediction early warning module is used to input dynamic pore pressure compensation data and historical slope stability assessment data into the landslide risk prediction model to predict and classify landslide risks, and obtain landslide risk early warning data.

[0020] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the method described in the first aspect.

[0021] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described in the first aspect.

[0022] The aforementioned landslide risk prediction and compensation method, system, equipment, and medium first acquire and preprocess rainfall monitoring data, soil parameter data, and historical slope stability assessment data for the target area to provide a reliable data source for subsequent steps. Then, time-series segmented feature extraction is performed on the rainfall monitoring data to capture local rainfall characteristics. Subsequently, infiltration response correlation analysis is conducted on the soil parameter data and the rainfall time-series segmented feature data to establish a nonlinear correlation between the two and capture the lag characteristics of rainfall infiltration. Finally, a dynamic lag compensation model is used to model the time-series lag dependency and dynamically allocate weights to obtain a time-weighted average. The system first analyzes the response data; then, based on the soil fissure distribution data, it performs nonlinear mapping compensation on the time-series weighted response data to accurately characterize the pore pressure response under the influence of fissures; finally, it inputs the dynamic pore pressure compensation data and historical slope stability assessment data into the landslide risk prediction model to achieve risk prediction and classification; it effectively captures the dynamic lag effect of rainfall infiltration and the spatial modulation effect of non-uniform distribution of soil fissures, improves the accuracy and lead time of landslide risk early warning, reduces prediction bias, reduces data redundancy, and can achieve accurate characterization of pore pressure response and accurate classification of landslide risk, providing reliable support for landslide early warning decision-making. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is a flowchart illustrating a landslide risk prediction and compensation method in one embodiment;

[0025] Figure 2 This is a schematic diagram of the structure of a landslide risk prediction and compensation system in one embodiment;

[0026] Figure 3 This is a schematic diagram of a computer device in one embodiment. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0028] refer to Figure 1 The application presents a flowchart illustrating a landslide risk prediction and compensation method, which includes the following steps:

[0029] S1. Obtain rainfall monitoring data, soil parameter data, and historical slope stability assessment data for the target area.

[0030] For example, rainfall monitoring data can be acquired through rain gauges or meteorological monitoring stations deployed in the target area. Rain gauges can employ tipping bucket or siphon metering principles, enabling real-time recording of the intensity and duration of precipitation events. The recording frequency of rainfall monitoring data typically corresponds to the sampling period, with each sampling moment corresponding to a rainfall intensity observation, measured in millimeters per hour. Raw rainfall monitoring data may contain missing data or outliers due to equipment malfunctions, communication interruptions, or environmental interference during acquisition; therefore, data quality control is necessary before proceeding to subsequent processing. Data quality control includes removing obvious equipment malfunction records, filling in missing time points using linear or spline interpolation methods, and eliminating random noise interference through sliding window smoothing. Furthermore, the timestamps of the rainfall monitoring data need to be uniformly calibrated to ensure consistency with soil parameter data and historical slope stability assessment data in the time dimension.

[0031] For example, soil parameter data may include soil fissure distribution data. Soil parameter data can be used to reflect the physical and mechanical properties and structural characteristics of the soil in a target area. Soil fissure distribution data within the soil parameter data is key information characterizing the internal structural features of the soil, and its acquisition can be achieved through various techniques such as field geological surveys, ground-penetrating radar detection, fiber optic sensing monitoring, and remote sensing image analysis. Soil fissure distribution data records the geometric morphological parameters of the fissures, including their length, width, depth, dip angle, orientation, and spatial distribution. In a geographic information system platform, these fissures can be represented as line segments or polyline elements, with each fissure element associated with corresponding attribute information. The presence of soil fissures has a significant impact on the rainwater infiltration process. Fissure channels provide preferential infiltration paths for rainwater, allowing it to quickly reach deeper parts of the soil, bypassing the infiltration resistance of normal soil. Therefore, accurately acquiring soil fissure distribution data is crucial for establishing the response relationship between rainfall infiltration and slope stability. In addition to soil fissure distribution data, soil parameter data may also include physical and mechanical parameters such as soil particle size distribution, porosity, permeability coefficient, water content, internal friction angle, and cohesion. These parameters together describe the engineering properties of the soil.

[0032] For example, historical slope stability assessment data can be derived from the accumulation of long-term monitoring and assessment of slope stability in the target area. Slope stability assessment can use the limit equilibrium method, finite element method, or other stability analysis methods to quantitatively evaluate the safety status of the slope. Historical assessment data is stored in time series form, with each time point corresponding to a stability status assessment result. The assessment results can be expressed in the form of safety factor values, stability level classifications, or risk probabilities. The safety factor is the core indicator for slope stability assessment, defined as the ratio of the slope's resisting moment to its sliding moment. When the safety factor is greater than 1, the slope is in a stable state; when the safety factor is equal to 1, the slope is in a limit equilibrium state; when the safety factor is less than 1, the slope is in an unstable state. The time span of the historical assessment data needs to cover a sufficiently long period to reflect the stability evolution of the slope under different meteorological conditions, groundwater level changes, and external loads.

[0033] S2. Extract time-series segmented features from rainfall monitoring data to obtain rainfall time-series segmented feature data.

[0034] For example, this step segments the rainfall monitoring data, which exhibits temporal fluctuations, to extract local rainfall characteristics for each time period. This addresses the problem of capturing local features by directly processing the overall time-series data, providing accurate support for subsequent infiltration response correlation analysis. Specifically, the segmented rainfall time-series characteristic data can be used for subsequent infiltration response correlation analysis, providing precise rainfall feature input. Rainfall monitoring data exhibits significant temporal fluctuations, with variations in intensity and duration across different rainfall periods. Directly processing the overall time-series data makes it difficult to capture local rainfall characteristics; therefore, segmenting the time series to extract local features clarifies the impact of rainfall at each time period on slope infiltration.

[0035] For example, to quantify the temporal volatility of rainfall monitoring data and support the rationality of segmentation, a rainfall intensity temporal volatility coefficient can be introduced, the calculation formula of which is: In the formula, The temporal fluctuation coefficient, representing the rainfall intensity over a specific period, can be used to characterize the degree of fluctuation in rainfall intensity during that period. This indicates the total number of sampling times within that period. Indicates the number of times within that time period Rainfall intensity values ​​at each sampling time, This represents the average rainfall intensity within a given period. The temporal fluctuation coefficient can be used to determine whether the rainfall intensity tends to stabilize within a certain period, while the gradient difference can be used to identify abrupt changes in rainfall intensity. Combining the two allows for precise segmentation of rainfall monitoring data, ensuring consistency in rainfall characteristics within each segment and providing a reliable foundation for subsequent feature extraction.

[0036] S3. Perform infiltration response correlation analysis on soil parameter data and rainfall time-series segmented characteristic data to obtain rainfall infiltration characteristic sequence data.

[0037] For example, the purpose of this step is to establish the correlation between soil parameter data and rainfall time-series segmented characteristic data, capture the dynamic lag characteristics of rainfall infiltration, and provide accurate input basis for subsequent dynamic lag compensation. Soil parameter data can be used to characterize the soil's permeability and water retention capacity, while rainfall time-series segmented characteristic data can be used to characterize the core characteristics of rainfall at different times. Together, they determine the rate, depth, and lag duration of rainfall infiltration. Through infiltration response correlation analysis, the influence of both can be effectively integrated to obtain rainfall infiltration characteristic sequence data that accurately reflects the dynamic characteristics of rainfall infiltration. This data can be used as input for subsequent dynamic lag compensation models, supporting accurate modeling of lag effects.

[0038] Optionally, to quantify the correlation between soil parameter data and rainfall time-series segmented characteristic data, an infiltration correlation degree calculation formula can be introduced, the expression of which is: In the formula, This represents the infiltration correlation between soil parameter data and rainfall time-series segmented characteristic data, with a value range of [-1, 1], and can be used to characterize the degree of linear correlation between the two. The covariance between soil parameter data and rainfall time-series segmented characteristic data can be used to measure the degree of co-variance between the two. This represents a soil parameter data matrix, which can be used to store the raw data of all soil-related parameters; This represents a feature data matrix for rainfall time series segments, which can be used to store feature data for all rainfall segments; The variance of soil parameter data can be used to characterize the dispersion of soil parameter data. The variance of the time-series feature data of rainfall can be used to characterize the dispersion of the time-series feature data of rainfall.

[0039] Optionally, to further clarify the influence weights of both factors on the rainfall infiltration process, an infiltration influence weight formula can be introduced. The formula for calculating the influence weight of soil parameter data on rainfall infiltration can be: The formula for calculating the weight of the impact of rainfall time-series segmented feature data on rainfall infiltration can be: In the formula, This indicates the weight of the influence of soil parameter data on rainfall infiltration. The weights represent the influence of rainfall time-series segmented feature data on rainfall infiltration. Both values ​​range from [0,1] and satisfy the following conditions: Infiltration correlation can be used to determine the degree of correlation between soil parameter data and rainfall time-series segmented characteristic data. Influence weight can be used to allocate the contribution ratio of the two to the infiltration process. Through infiltration correlation and influence weight, the relationship between the two can be accurately characterized, and finally, rainfall infiltration characteristic sequence data that integrates the influence of the two and the infiltration lag characteristics can be obtained.

[0040] S4. Input the rainfall infiltration characteristic sequence data into the dynamic lag compensation model to perform time-series lag dependency modeling and dynamic weight allocation to obtain time-series weighted response data.

[0041] For example, rainfall infiltration characteristic sequence data can be used to characterize the dynamic lag characteristics and infiltration fit relationship of rainfall infiltration; the dynamic lag compensation model achieves accurate modeling of time-series lag dependence through built-in forgetting gating structure, input gating structure and other components, and highlights the differences in the impact of lag characteristics on pore pressure response at different times through dynamic weight allocation; the time-series weighted response data can be used for subsequent nonlinear mapping compensation, providing support for the calculation of dynamic pore pressure compensation data.

[0042] For example, to achieve dynamic weight allocation, a dynamic weight calculation formula can be introduced, the expression of which can be: In the formula, Indicates the first The dynamic weights of the time-latency latent state data can be used to characterize the degree of influence of the time-latency latent state on the pore pressure response. This represents the total number of moments, used to define the time range of time-series data. This represents the current time sequence number, and its value is a positive integer, not greater than 1 / 2. ; This represents the weight decay coefficient, which can be used to control the decay rate of the weights. This represents the time index during the summation process, traversing from 1 to... Every moment.

[0043] For example, based on the calculated dynamic weights, the calculation formula for time-weighted response data can be obtained, and its expression can be: In the formula, This represents time-weighted response data, which can be used to integrate the effects of latent states at various time points; Indicates the first Dynamic weights at any given time; Indicates the first The lagged latent state data at a given time can be used to store the core information of the lagged characteristics at that time. Indicates the total number of moments. Represents the time sequence number, traversing from 1 to... The dynamic weighting calculation enables the allocation of weights to the lag latent states at different times, with greater weights for recent times and gradually decreasing weights for longer times. The time-series weighted response data formula integrates the lag characteristics of all times through weighted summation, resulting in time-series weighted response data that can accurately characterize the dynamic lag correlation between rainfall infiltration and pore pressure response.

[0044] S5. Based on the soil fracture distribution data, nonlinear mapping compensation is performed on the time-series weighted response data to obtain dynamic pore pressure compensation data.

[0045] Optionally, dynamic pore pressure compensation data can be used to characterize the time-varying features of pore pressure response under conditions of non-uniform fracture distribution.

[0046] For example, by incorporating the spatial modulation effect of non-uniform distribution of soil fissures, the pore pressure prediction bias caused by the failure of traditional methods to consider this factor is compensated. The time-series weighted response data (reflecting the dynamic lag correlation between rainfall infiltration and pore pressure response) and soil fissure distribution data (reflecting the spatial distribution, connectivity and development degree of fissures inside the soil) are fused together. Through nonlinear mapping compensation, dynamic pore pressure compensation data that can accurately characterize the time-varying characteristics of pore pressure under the influence of fissures is obtained, providing accurate pore pressure input for subsequent landslide risk prediction.

[0047] Optionally, soil fissure distribution data can be used to characterize the spatial distribution, connectivity, and development characteristics of fissures within the soil, while time-weighted response data can be used to characterize the dynamic hysteresis correlation between rainfall infiltration and pore pressure response. Together, they determine the response law of pore pressure under non-uniform fissure distribution conditions. Through nonlinear mapping compensation, the spatial modulation effect of fissures can be integrated into the pore pressure characterization, thereby improving the accuracy of pore pressure response characterization.

[0048] Optionally, to quantify the modulating effect of non-uniform fracture distribution on soil permeability, a formula for calculating the equivalent permeability coefficient of fractures can be introduced, the expression of which is: In the formula, Indicates the first The equivalent permeability coefficient at each fracture node location can be used to characterize the overall permeability at that location after considering the influence of fractures. This represents the original permeability coefficient of the soil, i.e., the soil's permeability capacity without considering the influence of cracks; Indicates the first The spatial attention weight of each fracture node can be used to characterize the degree to which the fracture node contributes to the permeability. Indicates the first The fracture permeability coefficient data of each fracture node can be used to characterize the extent to which the fracture enhances the permeability at that node.

[0049] For example, based on the calculated equivalent permeability coefficient of the fracture, a core formula for nonlinear mapping compensation can be established, and its calculation formula can be: In the formula, Indicates the first Dynamic pore pressure compensation data at time 1 can be used to characterize the first... Always consider the pore pressure response value after the non-uniform distribution of fractures; This represents the hyperbolic tangent activation function, which can be used to map the compensation result to the [-1,1] interval, facilitating subsequent scale adjustment; The weight matrix represents the nonlinear compensation and is used to perform linear transformation on the fused features; Indicates the first Time-weighted response data at each moment; Indicates the first Spatial attention weights for each fracture node; Indicates the first The equivalent permeability coefficient at each fracture node location; This represents the total number of fracture nodes; The bias term representing nonlinear compensation can be used to adjust the offset of the compensation result. The equivalent permeability coefficient can quantify the modulating effect of non-uniform fracture distribution on soil permeability. Nonlinear mapping compensation can fuse time-series weighted response data and fracture equivalent permeability coefficient characteristics to achieve spatiotemporal coupled compensation. It retains the dynamic hysteresis effect of rainfall infiltration and incorporates the spatial modulation effect of fractures, ultimately obtaining accurate dynamic pore pressure compensation data.

[0050] S6. Input the dynamic pore pressure compensation data and the historical slope stability assessment data into the landslide risk prediction model to predict and classify landslide risks, and obtain landslide risk early warning data.

[0051] For example, the purpose of this step is to integrate dynamic pore pressure compensation data (reflecting the time-varying characteristics of pore pressure response under non-uniform fracture distribution conditions) and historical slope stability assessment data (reflecting the past slope stability status and related influencing factors in the target area), and to achieve accurate probability prediction and classification of landslide risk through a landslide risk prediction model. This solves the problem of insufficient integration of historical stability information and current pore pressure characteristics in traditional prediction models, and provides direct support for landslide early warning decision-making.

[0052] Optionally, the landslide risk prediction model can achieve accurate prediction and classification of landslide risk by integrating pore pressure response characteristics and historical stability information, providing a basis for early warning decision-making. The landslide risk prediction model may include a feature encoding layer, a cross-attention fusion layer, a feature projection layer, a multi-scale dilated convolutional layer, a feature splicing layer, and a probabilistic regression prediction output layer.

[0053] Optionally, dynamic pore pressure compensation data can be used to characterize the time-varying characteristics of pore pressure response under non-uniform fracture distribution conditions. Pore pressure variation is a core factor affecting slope stability, and this data provides the core input for risk prediction of the current internal stress state of the slope. Historical slope stability assessment data can be used to provide information such as past slope stability status, pore pressure variation records, and rainfall conditions in the target area, providing historical reference for current risk prediction and helping the model to discover the correlation between current pore pressure characteristics and historical stability status.

[0054] For example, to effectively integrate current pore pressure characteristics and historical stability characteristics, a feature fusion calculation formula can be introduced, the expression of which can be: In the formula, Indicates the first The pore pressure-historical stability fusion feature data at any given time is used to integrate the current pore pressure features and historical stability features, providing a comprehensive input for subsequent risk probability prediction; The fusion weights of dynamic pore pressure compensation data can be used to characterize the contribution of dynamic pore pressure compensation data to the fusion features. Indicates the first Dynamic pore pressure compensation data at any given time; The fusion weights represent the fusion weights of historical slope stability assessment data, which can be used to characterize the contribution of historical slope stability assessment data to the fusion features. Indicates the first Historical stability feature data corresponding to the time; and satisfying This ensures the rationality of the fusion weights. Based on the fused feature data, a core formula for predicting landslide risk probability can be established, and its calculation formula is as follows: In the formula, Indicates the first The landslide risk probability data at time 1 can be used to characterize the landslide risk at time 2. The probability of a landslide occurring on the slope at time [0,1] is given. This represents the sigmoid activation function, which can be used to map the risk probability prediction results to the [0,1] interval to achieve probability normalization; This represents the transpose of the risk probability regression weight vector, which can be used to linearly weight fused feature data, highlighting features that have a significant impact on risk prediction. This represents the matrix transpose operation; Indicates the first Fusion feature data of pore pressure history stability at any given time; This represents the risk probability regression bias term, which can be used to adjust the offset of the risk probability prediction results; feature fusion can achieve weighted fusion of dynamic pore pressure compensation data and historical slope stability assessment data, taking into account both the current pore pressure status and historical stability information; risk probability prediction can transform the fused features into landslide risk probabilities, providing a quantitative basis for subsequent risk classification.

[0055] In the aforementioned landslide risk prediction and compensation method, the following steps are taken: First, rainfall monitoring data, soil parameter data, and historical slope stability assessment data of the target area are acquired and preprocessed to provide a reliable data source for subsequent steps. Then, time-series segmented feature extraction is performed on the rainfall monitoring data to capture local rainfall characteristics. Next, infiltration response correlation analysis is conducted on the soil parameter data and rainfall time-series segmented feature data to establish a nonlinear correlation between the two and capture the rainfall infiltration lag characteristics. Then, the rainfall infiltration feature sequence data is input into a dynamic lag compensation model to perform time-series lag dependency modeling and dynamic weight allocation, obtaining time-series weighted response data. Finally, based on the soil... Nonlinear mapping compensation is performed on the time-series weighted response data using volumetric fissure distribution data to obtain accurate dynamic pore pressure compensation data. Finally, the dynamic pore pressure compensation data and historical slope stability assessment data are input into the landslide risk prediction model to complete risk prediction and classification, resulting in landslide risk early warning data. This method can effectively capture the dynamic lag effect of rainfall infiltration and the spatial modulation effect of non-uniform distribution of soil fissures, improve the accuracy and lead time of landslide risk early warning, reduce prediction bias, reduce data redundancy and historical information interference, and achieve accurate compensation of pore pressure response and accurate classification of landslide risk, providing reliable support for landslide early warning decision-making.

[0056] In an optional embodiment, the rainfall monitoring data includes rainfall intensity values ​​at multiple sampling times, and S2 includes:

[0057] S21. Based on rainfall monitoring data, calculate the gradient difference of rainfall intensity between adjacent sampling times to obtain rainfall gradient sequence data.

[0058] Optionally, the rainfall gradient difference can reflect the rate of change of rainfall intensity between adjacent time points, thereby identifying abrupt changes in rainfall intensity and providing a basis for subsequent change point identification. The expression for the rainfall gradient difference can be: In the formula, Indicates the first Time and the The difference in rainfall gradient at time intervals, Indicates the first Rainfall intensity value at any given time Indicates the first Rainfall intensity value at any given time This represents the sampling time sequence number, and its value is a positive integer. The rainfall gradient difference at each time moment is calculated using this formula. The set of rainfall gradient differences at each time moment is the rainfall gradient sequence data. The rainfall gradient sequence data can clearly reflect the temporal trend of rainfall intensity. If the gradient difference is positive, it indicates that the rainfall intensity is increasing; if the gradient difference is negative, it indicates that the rainfall intensity is decreasing; the larger the absolute value of the gradient difference, the more drastic the change in rainfall intensity.

[0059] S22. Input the rainfall gradient sequence data into the change point recognition model to detect change points, obtain the change point probability value at each sampling time, and determine the sampling time when the change point probability value exceeds the preset probability threshold as the change point location, thereby obtaining the rainfall change point time data.

[0060] Optionally, the change point identification model can be used to identify moments in rainfall time series data where rainfall intensity changes significantly. These change points are key criteria for segmenting different rainfall feature segments, enabling the division of continuous rainfall time series data into multiple segments with similar characteristics. The change point identification model can include a one-dimensional convolutional feature extraction layer, a pooling dimensionality reduction layer, and a probability mapping layer, with each layer working collaboratively to achieve change point detection.

[0061] Optionally, the one-dimensional convolutional feature extraction layer can use multiple convolutional kernels of different sizes to extract local pattern features from the rainfall gradient sequence data, capture local variation features in the gradient sequence, and obtain local gradient feature data. The expression for the convolution operation can be: In the formula, Indicates the first Local gradient feature data at time step [time]. Indicates the kernel size. The convolution kernel is represented by the first... One weight parameter, Indicates the first The difference in rainfall gradient at time intervals, This represents the bias term for the convolution operation.

[0062] Optionally, the pooling dimensionality reduction layer uses max pooling to compress the local gradient feature data, reduce data redundancy, and retain key local features to obtain gradient dimensionality-reduced feature data. The expression for the pooling operation can be: In the formula, Indicates the first Gradient dimensionality reduction feature data at time step, Indicates the pooling window size. Indicates the first Local gradient feature data at time step;

[0063] Optionally, the probability mapping layer can employ a sigmoid activation function to perform variable-point probability mapping on the gradient-reduced feature data, mapping the feature data to variable-point probability values ​​in the interval [0,1]. The expression for the activation function can be: In the formula, Indicates the first The probability value of the change point at time t. This represents the sigmoid activation function. Indicates the first The gradient dimensionality reduction feature data at each time point. A preset probability threshold can be determined based on the geological conditions and rainfall characteristics of the target area, and can be used to distinguish between change points and non-change points. When the threshold is exceeded, that moment is determined to be a change point location, and all change point locations are aggregated to obtain the rainfall change point time data.

[0064] S23. Based on the location of the change point in the rainfall change point data, the rainfall monitoring data is segmented into time series to obtain the rainfall segment sequence data.

[0065] Optionally, the change point location can divide the complete time series of rainfall monitoring data into multiple consecutive segments, with consistent rainfall intensity variation patterns within each segment. Specifically, the segmentation is as follows: the first change point location is the end time of the first segment, and the distance from the start time of the rainfall monitoring data to this change point location constitutes the first segment; adjacent change point locations are used as the end time of the preceding segment and the start time of the following segment, respectively, to generate intermediate segments; the last change point location is the end time of the penultimate segment, and the distance from this change point location to the end time of the rainfall monitoring data constitutes the last segment. This segmentation method ensures that there are no significant abrupt changes in rainfall variation patterns within each segment, resulting in segmented rainfall sequence data.

[0066] S24. Perform multidimensional feature extraction on each segment of the rainfall segment sequence data to obtain rainfall segment feature vector data.

[0067] Optionally, the rainfall segment feature vector data may include the cumulative amount, peak intensity, and duration of rainfall within each segment. The purpose of multidimensional feature extraction is to transform the time-series data of each rainfall segment into a feature vector that can characterize the rainfall features of that segment, facilitating subsequent model processing and feature association.

[0068] Alternatively, the cumulative amount can be obtained by summing the rainfall intensity values ​​at all sampling times within the segment. The expression for the cumulative amount can be: In the formula, This represents the cumulative rainfall over a specific segment. This segment represents the set of all sampling times it contains. Indicates the first segment within this segment. The rainfall intensity value at any given time.

[0069] Optionally, the peak intensity is the maximum value of the rainfall intensity at all sampling times within this segment, and the expression for the peak intensity can be: In the formula, This indicates the peak rainfall intensity for that segment.

[0070] Optionally, the duration is the product of the number of sampling moments contained in the segment and the sampling interval. The expression for the duration can be: In the formula, This indicates the duration of rainfall in that segment. This indicates the number of sampling times contained in the segment. This indicates the sampling interval. The three features of cumulative amount, peak intensity, and duration for each segment are combined in sequence to obtain the rainfall segment feature vector for that segment. The feature vectors of all segments are then summed to obtain the rainfall segment feature vector data.

[0071] S25. Perform time-series decay position encoding on the segmented feature vector data of rainfall to obtain the time-series segmented feature data of rainfall.

[0072] Optionally, since different segments have different temporal positions, their impact on subsequent pore pressure response varies, and the impact of rainfall characteristics will show a decay trend over time. Therefore, it is necessary to use temporal decay position encoding to integrate the temporal position information and decay characteristics of the segments into the feature vector to improve the temporal correlation of the feature data.

[0073] Optionally, the expression for the timing decay position encoding can be: In the formula Indicates the first The first segment The temporal decay position encoding value of the dimensional feature. The time sequence number of the segment (arranged in chronological order, starting with the first segment). (increasing sequentially) The dimension of the feature vector ( (corresponding to cumulative amount, peak intensity, and duration, respectively). Represents the total dimension of the feature vector ( ), This represents the attenuation coefficient, used to control the attenuation rate of rainfall characteristics. The rainfall segment feature vector for each segment is added element-wise to the corresponding temporal attenuation location encoding value to obtain the rainfall temporal segment feature for that segment. The features of all segments are then aggregated to obtain the rainfall temporal segment feature data. This data contains both the core features of the rainfall segments and incorporates temporal location and attenuation characteristics, enabling a more accurate reflection of the temporal impact of rainfall on slope pore pressure.

[0074] In an optional embodiment, S3 includes:

[0075] S31. Extract infiltration characteristic parameters from the soil parameter data to obtain soil infiltration characteristic data.

[0076] Optionally, soil infiltration characteristic data can be used to characterize the soil's permeability and water retention capacity. Soil parameter data contains various infiltration-related parameters, and key infiltration characteristic parameters need to be selected through feature extraction. These mainly include saturated permeability coefficient, field capacity, and porosity. These parameters directly determine the soil's infiltration rate and water retention capacity for rainfall. The extraction process can employ Principal Component Analysis (PCA) to reduce the dimensionality of the original soil parameter data, retaining the principal components that can explain most of the data variance as soil infiltration characteristic data. The expression for principal component analysis can be: In the formula, This represents a data matrix representing soil infiltration characteristics. This represents the original soil parameter data matrix. The principal component loading matrix is ​​represented by the elements in the loading matrix, which represent the contribution of each original parameter to the principal component. This formula transforms high-dimensional original soil parameter data into low-dimensional soil infiltration characteristic data, thus preserving key infiltration information while reducing data redundancy.

[0077] S32. Calculate the infiltration capacity matching degree between the rainfall intensity and soil infiltration characteristic data of each segment in the rainfall time series segmented characteristic data to obtain segmented infiltration matching degree data.

[0078] Optionally, the infiltration capacity matching degree can be used to characterize the degree of fit between the rainfall intensity and the soil infiltration capacity of a certain segment. The higher the matching degree, the easier it is for the rainfall in that segment to infiltrate into the soil, and vice versa. The expression for calculating the infiltration capacity matching degree can be: In the formula, Indicates the first The infiltration capacity matching degree of each rainfall segment, with a value range of (0,1]. Indicates the first The average rainfall intensity of each rainfall segment (which can be obtained by dividing the cumulative rainfall of that segment by its duration) This represents the saturated permeability coefficient in soil infiltration characteristic data. This formula calculates the relative difference between rainfall intensity and saturated permeability coefficient, and uses an exponential function to map the difference to the (0,1) interval. The closer the rainfall intensity and saturated permeability coefficient are, the closer the matching degree is to 1; the greater the difference between the two, the closer the matching degree is to 0. This can accurately reflect the compatibility between rainfall and soil infiltration capacity.

[0079] S33. Based on the segmented infiltration matching degree data and soil infiltration characteristic data, the time delay of water transport after rainfall infiltration in each segment is calculated to obtain segmented infiltration time delay data.

[0080] Optionally, the change in pore pressure caused by rainfall infiltration into the deep slope has a time lag. The length of the time lag is closely related to the soil's infiltration capacity and the rainfall intensity. The stronger the infiltration capacity and the higher the match between rainfall intensity and infiltration capacity, the shorter the time lag; conversely, the longer the time lag. The expression for calculating the time lag can be: In the formula, Indicates the first Infiltration lag of each rainfall segment This indicates the slope monitoring depth (i.e., the distance from the pore pressure monitoring point to the ground surface). Represents the saturated permeability coefficient. Indicates the first The infiltration capacity matching degree of each rainfall segment. In the above formula, the slope monitoring depth... The saturated permeability coefficient is a fixed value. and matching degree Together, they determine the length of the time lag, enabling precise characterization of the infiltration lag characteristics corresponding to different rainfall segments.

[0081] S34. Based on the segmented infiltration time lag data, the rainfall time series segmented feature data are re-aligned on the time axis according to the infiltration time lag corresponding to each segment to obtain infiltration time lag aligned sequence data.

[0082] Optionally, infiltration delay aligned sequence data can be used to characterize the distribution of rainfall characteristics in each segment on the aligned time axis after infiltration delay.

[0083] Optionally, since the infiltration time lags differ across different rainfall segments, their impact on the pore pressure response also occurs at different times. Directly using the original time series for analysis would lead to a temporal misalignment between rainfall characteristics and the pore pressure response; therefore, a delay realignment is necessary. Specifically, the end time of each rainfall segment is used as a reference, and the time is delayed... The duration is used to determine the starting time at which the rainfall segment's characteristics affect the pore pressure response. All rainfall time series characteristics are then rearranged on the same time axis according to the adjusted influence time, resulting in infiltration delay aligned sequence data. The expression for the realignment process can be: In the formula Indicates the first The starting time of the impact after the adjustment of the rainfall segments Indicates the first The end time of each rainfall segment Indicates the first The infiltration time lag of each rainfall segment is determined. This realignment operation aligns the temporal relationship between rainfall characteristics and pore pressure response, eliminating the misalignment caused by time lag.

[0084] S35. Input the infiltration delay alignment sequence data and segmented infiltration matching degree data into the infiltration response association model for feature association fusion to obtain rainfall infiltration feature sequence data.

[0085] Optionally, the infiltration response association model includes a dual-path attention fusion layer and a sequence aggregation layer. The dual-path attention fusion layer can be used to calculate cross-attention weights on infiltration delay alignment sequence data and segmented infiltration matching degree data to obtain infiltration association fusion feature data. The sequence aggregation layer can be used to aggregate and splice the infiltration association fusion feature data in the time dimension to obtain rainfall infiltration feature sequence data.

[0086] Optionally, the core function of the dual-path attention fusion layer is to mine the correlation weights between the infiltration delay alignment sequence (rainfall characteristics) and the segmented infiltration matching degree (fitting relationship), so that the model focuses on rainfall characteristics with high fitting degree and large impact on pore pressure response. The expression for calculating its attention weights can be: In the formula, Indicates the first Attention weights for each segment, Indicates the first infiltration delayed aligned sequence data Feature vectors (query vectors) of each segment. Indicating the segmented infiltration matching degree data, the first... The matching degree value (key vector) of each segment. This indicates the total number of rainfall segments.

[0087] Optionally, after obtaining the attention weights, the infiltration delayed alignment sequence data is multiplied element-wise with the attention weights to obtain the infiltration association fusion feature data. The expression for the infiltration association fusion feature data can be: In the formula, Indicates the first Infiltration correlation fusion feature data of each segment, Indicates the first Attention weights for each segment, Indicates the first infiltration delayed aligned sequence data The feature vectors of each segment. The sequence aggregation layer uses a temporal concatenation method to concatenate the infiltration correlation fusion feature data of all segments in an adjusted temporal order to obtain the rainfall infiltration feature sequence data. The expression of the rainfall infiltration feature sequence data can be: In the formula, This represents the characteristic sequence data of rainfall infiltration. Indicates the first Infiltration correlation fusion feature data of each segment, This represents the total number of rainfall segments. This data fully preserves the correlation information of rainfall characteristics, infiltration time lag, and adaptation relationships, providing accurate input for subsequent dynamic lag compensation.

[0088] In an optional embodiment, the dynamic hysteresis compensation model includes a forgetting gating structure, an input gating structure, a cell state update structure, and an output gating structure, where S4 includes:

[0089] S41. Input the rainfall infiltration characteristic sequence data and the lag hidden state data of the preceding time step into the forgetting gating structure, calculate the historical information forgetting ratio of the rainfall infiltration characteristic sequence data and the lag hidden state data, and obtain the forgetting gating data.

[0090] Optionally, the forgetting gating structure determines the proportion of historical information in the lagging hidden states of previous time steps that needs to be retained or forgotten, thus avoiding interference from redundant historical information in the current computation. The expression for the forgetting gating data can be: In the formula, Indicates the first The forgetting gating data at each time step has a value range of [0,1]. This represents the sigmoid activation function. The weight matrix represents the forgetting gate structure. Indicates the first The lagged latent state data at time step 1 can be used to store the lagged characteristic information of previous time steps. Indicates the first Rainfall infiltration characteristic sequence data at different times, The bias term represents the forgetting gate structure. This indicates that preceding lagged latent state data is concatenated with current rainfall infiltration characteristic data. When When it is close to 1, it indicates that most of the historical information from the preceding time step is retained; when When the value is close to 0, it indicates that most of the historical information from previous moments has been forgotten, thus achieving selective retention of historical information.

[0091] S42. Input the rainfall infiltration characteristic sequence data and the hysteresis latent state data of the preceding time step into the input gating structure, calculate the retention ratio of the current input information for the rainfall infiltration characteristic sequence data and the hysteresis latent state data, and obtain the input gating data and candidate unit state data.

[0092] Optionally, the input gating structure determines the proportion of rainfall infiltration characteristic information to be retained at the current moment, and simultaneously generates candidate cell state data to update the cell state at the current moment. Its calculation expression includes two parts: the expression for the input gating data can be: The expression for the candidate cell state data can be: In the formula, Indicates the first The input gating data at time t, with values ​​ranging from [0,1], This represents the weight matrix of the input gating structure. This represents the bias term of the input gating structure; Indicates the first Candidate cell state data at time 10:00. This represents the hyperbolic tangent activation function, used to map candidate cell state data to the interval [-1, 1]. The weight matrix representing the state of the candidate unit. Bias terms representing the state of candidate cells. When the value is close to 1, it indicates that most of the input information at the current time is retained. This indicates that the preceding lagged latent state data is concatenated with the current rainfall infiltration characteristic data; When the value is close to 0, it indicates that the input information at the current moment is not retained, while the candidate unit state data stores the core features of the input information at the current moment.

[0093] S43. Input the forgotten gating data, input gating data, candidate cell state data and the lagging cell state data of the previous time step into the cell state update structure, selectively erase the historical information of the lagging cell state data of the previous time step and selectively write the current information of the candidate cell state data to obtain the lagging cell state data of the current time step.

[0094] Optionally, the expression for the state data of the hysteresis unit can be: In the formula, Indicates the first The state data of the lagging unit at a given time can be used to store the core information of the lagging characteristics at the current time. Indicates the first The state data of the lagging units at a given time step can be used to store the core information of the lagging characteristics of the preceding time steps. This represents an element-wise multiplication operation, which can be used to selectively fuse information. Indicates the first The forgetting gating data at each moment can be used to control the proportion of historical information erased from the previous unit's state. Indicates the first The input gating data at any given time can be used to control the proportion of current information written into the candidate cell state. Indicates the first The candidate unit state data at time step 1 can be used to provide new feature information for the current time step. The calculation process of this formula consists of two steps: the first step is through... To achieve selective erasure of historical information in the previous unit state, and to achieve forgetting-gated data. Element-wise multiplication of the previous unit state The first step is to retain necessary historical information and erase redundant information; the second step is to... To achieve selective writing of the current candidate cell status information, input gating data. Element-wise multiplication by candidate cell state The necessary current information is retained; finally, the results of the two steps are added together to obtain the state data of the lagging unit at the current moment. This enables dynamic updating of lagging feature information.

[0095] S44. Input the rainfall infiltration characteristic sequence data, the lag hidden state data of the previous time step and the lag unit state data of the current time step into the output gating structure, filter the output information of the lag unit state data of the current time step, and obtain the lag hidden state data of the current time step.

[0096] Optionally, the output gating structure filters out the lag feature information to be output from the lag unit state data at the current time step, updates the lag hidden state data, and provides input for the calculation of the next time step. The expression for the output gating data can be: The expression for the current lagged hidden state data can be: In the formula, Indicates the first The output gating data at each time point has a value range of [0,1]. This represents the weight matrix of the output gating structure. This represents the bias term of the output gating structure. This means concatenating the preceding lagged implicit state data, the current rainfall infiltration characteristic data, and the current lagged unit state data; Indicates the first Lag hidden state data at time tardiness. This indicates that the current lagging unit state data is mapped to the interval [-1, 1]. By multiplying each element by the mapping result, the output information can be filtered. When the value is close to 1, output more information about the current cell status; When the value is close to 0, less current cell state information is output, resulting in the final value. Used to store the hysteresis feature output information at the current moment.

[0097] S45. Based on the lagged hidden state data at each time point, dynamically assign weights to the lagged hidden state data at each time point to obtain time-series weighted response data.

[0098] Optionally, the influence of lagged latent state data on pore pressure response varies at different times, with closer-term lagged latent state data having a greater impact and the impact gradually decreasing in the longer term. Therefore, dynamic weight allocation is needed to highlight the influence at key moments. Weight allocation can be implemented using the Softmax function, and the expression for weight allocation can be: In the formula, Indicates the first Dynamic weights of time-lag hidden state data Indicates the total number of moments. Indicates the current time sequence number. This represents the weight decay coefficient, which can be used to control the decay rate of the weights.

[0099] Optionally, after obtaining the dynamic weights, the lagged latent state data at each time step are multiplied element-wise with the corresponding weights, and then the results at all times are summed to obtain the time-series weighted response data. The expression for the time-series weighted response data can be: In the formula, This represents time-weighted response data. Indicates the first Dynamic weights at time points, Indicates the first Lag hidden state data at time tardiness. This represents the total number of time points. This time-series weighted response data can accurately characterize the dynamic lag correlation between rainfall infiltration and pore pressure response, highlighting the lag effects at critical moments and providing a foundation for subsequent nonlinear mapping compensation.

[0100] In an optional embodiment, S5 includes:

[0101] S51. Extract fracture nodes from soil fracture distribution data to obtain fracture node feature vectors, and construct adjacency relationships from fracture node feature vectors to obtain fracture adjacency graph data; fracture adjacency graph data includes fracture node adjacency matrix.

[0102] Optionally, the crack adjacency graph data includes a crack node adjacency matrix. Soil crack distribution data contains spatial distribution information of cracks, which needs to be transformed into nodes in a graph structure through node extraction to facilitate subsequent graph convolution feature extraction: First, the soil crack distribution data is discretized, with the endpoints and turning points of each crack as crack nodes. The feature vector of each crack node can include information such as the node's spatial coordinates, crack aperture, and crack extension direction, forming a crack node feature vector; then, based on the spatial distance and connectivity between crack nodes, a crack adjacency matrix is ​​constructed. Represents the th node in the adjacency matrix of the fracture nodes. Line 1 Column elements, and These represent the indices of the fracture nodes, when the node... With nodes When there are connecting gaps between two nodes and the spatial distance is less than a preset threshold, it indicates that the two nodes are adjacent. ,otherwise The feature vectors of fracture nodes and the adjacency matrix of fracture nodes together constitute the fracture adjacency graph data, which can be used to characterize the spatial distribution and connectivity of fractures.

[0103] S52. Input the crack adjacency graph data into the crack spatial propagation model to perform multi-layer graph convolution feature propagation to obtain crack propagation feature data.

[0104] Optionally, the crack propagation model may include a multi-layer graph convolutional feature extraction layer and a node feature normalization layer. The multi-layer graph convolutional feature extraction layer can be used to aggregate neighborhood information from the crack adjacency graph data to obtain crack neighborhood aggregated feature data. The node feature normalization layer can be used to normalize the features of the crack neighborhood aggregated feature data to obtain crack propagation feature data. The expression for the crack neighborhood aggregated feature data can be: In the formula, Indicates the first The first layer of the graph convolutional layer The aggregated feature data of the fracture neighborhood of the fracture node is used to store the fracture neighborhood of the node after the fracture node. Neighborhood feature information aggregated after layer graph convolution; Represents the Corrected Linear Unit (ReLU) activation function, used to introduce nonlinear features. Its expression can be: This can effectively alleviate the gradient vanishing problem and improve model training performance; Indicates the first The weight matrix of the layered graph convolutional layer is used to perform a linear transformation on the features after neighborhood aggregation. This represents the convolutional layer number of the graph, and its value is a positive integer (the number of layers can be set according to actual needs). Indicates the first The set of neighboring nodes of the i-th fracture node, i.e., the set of neighboring nodes of the i-th fracture node. All nodes that are adjacent to each other; Indicates the first The degree of a crack node is the number of its neighboring nodes. Indicates the first The degree of each neighboring node; Indicates the first The first layer of the graph convolutional layer The feature vector of the crack node of each neighboring node, that is, the feature information of the neighboring node after the convolution of the previous layer graph; Indicates the first The bias vector of the layer graph convolutional layer is used to adjust the offset of the feature output; Represents the neighbor node index, traversal All neighboring nodes. The core function of this formula is to propagate the spatial features of the fracture by weighted aggregation of the features of the neighboring nodes of each fracture node. To avoid excessive influence of nodes with high degrees on aggregated features and ensure the rationality of feature aggregation, a multi-layer graph convolutional feature extraction layer performs multiple neighborhood aggregations to obtain fracture neighborhood aggregated feature data. This data is then normalized by a node feature normalization layer to eliminate feature scale differences, yielding fracture propagation feature data. The normalization expression can be: This represents the normalized fracture propagation characteristic data. This represents the mean of the aggregated feature data in the crack neighborhood. This represents the standard deviation of the aggregated feature data in the crack neighborhood.

[0105] S53. Perform nonlinear permeability coefficient mapping on the fracture propagation characteristic data to obtain fracture permeability coefficient data.

[0106] Optionally, fracture permeability coefficient data are used to characterize the extent to which the permeability of each fracture node location is enhanced under the influence of the fracture network.

[0107] Optionally, the presence of soil fissures significantly enhances the permeability of local areas. The more developed and interconnected the fissures, the greater the increase in permeability. Nonlinear permeability coefficient mapping can be achieved using a sigmoid activation function combined with a linear transformation, and its expression can be: In the formula, Indicates the first Data on fracture permeability coefficients at each fracture node location. This represents the sigmoid activation function, used to restrict the mapping result to the interval [0,1]. The weight matrix represents the permeability coefficient mapping. Indicates the first The output of the layer graph convolutional layer is the first Crack propagation characteristic data for each crack node. This indicates the total number of layers in the graph convolutional layer. This represents the bias term used in the permeability coefficient mapping. This represents the maximum value of the fracture permeability coefficient, used to scale the mapping result to a reasonable range of permeability coefficients. This formula can accurately map the increase in permeability at each fracture node location based on the spatial propagation characteristics of the fracture, and can be used to reflect the modulating effect of non-uniform fracture distribution on permeability.

[0108] S54. Calculate the spatial attention weights of the fracture permeability coefficient data to obtain the fracture spatial attention weight data.

[0109] Optionally, the fracture spatial attention weight data is used to characterize the relative contribution weight of each fracture node to the overall pore pressure response.

[0110] Optionally, different fracture nodes have different permeability coefficients, and their influence on the pore pressure response also varies. The larger the permeability coefficient, the greater its contribution weight to the pore pressure response. The expression for the spatial attention weight can be: In the formula, Indicates the first The spatial attention weights of each fracture node are in the range of [0,1], and the sum of the weights of all nodes is 1. Indicates the first Data on the fracture permeability coefficient of each fracture node; This represents the total number of fracture nodes; This represents the index of the fracture node, and the process iterates through all fracture nodes. The exponential function in the above formula amplifies the differences in permeability coefficients among different fracture nodes, achieving a precise allocation of contribution weights and highlighting the influence of fracture nodes with high permeability coefficients on pore pressure response.

[0111] S55. Based on the spatial attention weight data of the fracture, spatiotemporal coupling nonlinear compensation is performed on the time-series weighted response data to obtain dynamic pore pressure compensation data.

[0112] Optionally, time-weighted response data can be used to reflect the dynamic hysteresis effect of rainfall infiltration, and fracture spatial attention weight data can be used to reflect the spatial modulation effect of fracture non-uniform distribution. By integrating the effects of both through spatiotemporal coupling nonlinear compensation, dynamic pore pressure compensation data that can accurately characterize the time-varying characteristics of pore pressure response under fracture non-uniform distribution conditions is obtained. The compensation formula is as follows: In the formula, Indicates the first Dynamic pore pressure compensation data at any given time. This represents the hyperbolic tangent activation function, used to map the compensation result to the [-1,1] interval (which can be scaled to the actual pore pressure range according to actual needs). The weight matrix represents the nonlinear compensation. Indicates the first Time-weighted response data at each moment. Indicates the first Spatial attention weights for each fracture node Indicates the first Data on the fracture permeability coefficient of each fracture node. This represents the total number of fracture nodes. This represents the bias term for nonlinear compensation. The above formula allows for the fusion of time-series weighted response data with spatially attention-weighted permeability coefficient data from fractures, achieving spatiotemporal coupling compensation. This considers both the dynamic lag effect of rainfall infiltration and the spatial modulation effect of non-uniform distribution of soil fractures, effectively improving the accuracy of pore pressure response.

[0113] In an optional embodiment, the landslide risk prediction model includes a feature encoding layer, a cross-attention fusion layer, a feature projection layer, a multi-scale dilated convolutional layer, a feature splicing layer, and a probabilistic regression prediction output layer, S6 including:

[0114] S61. Input the historical slope stability assessment data into the feature coding layer, and perform stability state coding and temporal location coding on the historical slope stability assessment data to obtain historical stability coding sequence data.

[0115] Optionally, the historical slope stability assessment data contains slope stability state information from different periods, which needs to be converted into a feature sequence that the model can process through feature encoding. The stability state encoding can adopt the one-hot encoding method to convert different stability states (such as stable, basically stable, and unstable) into one-hot vectors. The temporal location encoding can be used to incorporate the temporal information of the historical data. In the formula, Indicates the first The first historical moment, the Temporal position encoding value of dimensional feature, Indicates the sequence number of a historical moment. This represents the dimension of the encoded vector. This represents the total dimension of the encoded vector. This represents the attenuation coefficient. The one-hot encoded vector at each historical moment is added element-wise to the corresponding temporal position encoded value to obtain the encoded feature for that historical moment. Arranging the encoded features of all historical moments in temporal order yields the historical stable encoded sequence data.

[0116] S62. Input the historical stability encoded sequence data and dynamic pore pressure compensation data into the cross-attention fusion layer, use the dynamic pore pressure compensation data as the query input and the historical stability encoded sequence data as the key input to calculate the attention weights and obtain the attention weighted feature data.

[0117] Optionally, the core function of the cross-attention fusion layer is to mine the correlation between dynamic pore pressure compensation data (current pore pressure characteristics) and historical stability encoding sequence data (historical stability characteristics), enabling the model to focus on historical stable states similar to the current pore pressure characteristics, thereby improving the accuracy of risk prediction. Its attention weight calculation expression can be: In the formula, Indicates the first Dynamic pore pressure compensation data at any time and the first Attention weights for stability encoding features at each historical moment. Indicates the first Real-time dynamic pore pressure compensation data (query vector). Indicates the first Stability encoding features (key vectors) at each historical moment. This represents the total number of historical moments. After obtaining the attention weights, the historical stability encoded sequence data is weighted and summed with the attention weights to obtain the attention-weighted feature data, expressed as: In the formula, Indicates the first Attention-weighted feature data at each moment Indicates the first Time and the The attention weight of each historical moment Indicates the first Stability coding features of each historical moment, This represents the total number of historical moments.

[0118] S63. Input the attention-weighted feature data into the feature projection layer for dimension mapping to obtain the pore pressure stability fusion feature data.

[0119] Optionally, the attention-weighted feature data and the dynamic pore pressure compensation data may have different dimensions, requiring dimension unification through a feature projection layer to achieve effective fusion. The feature projection layer can be implemented using a linear transformation, and its expression can be: In the formula, Indicates the first Fusion feature data of pore pressure stability at time step. This represents the weight matrix of the feature projection layer, used to map the attention-weighted feature data to the same dimension as the dynamic pore pressure compensation data. Indicates the first Attention-weighted feature data at each moment This represents the bias term of the feature projection layer. The mapped feature data is added element-wise with the dynamic pore pressure compensation data to obtain the final pore pressure stability fusion feature data, achieving deep fusion of current pore pressure features and historical stability features.

[0120] S64. Input the pore pressure stability fusion feature data into the multi-scale dilated convolutional layer, and use convolutional kernels with different dilation rates to perform parallel convolutional feature extraction to obtain multi-scale convolutional feature data.

[0121] Optionally, the multi-scale dilated convolutional layer is used to capture multi-scale features in the pore pressure stability fusion feature data. Convolutional kernels with different dilation rates can capture feature information of different ranges. The larger the dilation rate, the wider the range of features captured, and the more comprehensively the deep correlations in the feature data can be mined. Its convolution operation expression can be: In the formula, Indicates the first The time-inflation rate Convolutional feature data, Indicates the expansion rate (sets multiple different expansion rates, such as...) ), Indicates the kernel size. Indicates the expansion rate The first convolution kernel One weight parameter, Indicates the first Fusion feature data of pore pressure stability at time step. This represents the stride offset of the dilated convolution. Indicates the expansion rate The bias term of the convolution operation. By performing parallel operations with convolution kernels of different dilation rates, multiple sets of convolutional feature data at different scales are obtained, i.e., multi-scale convolutional feature data.

[0122] S65. Input the multi-scale convolutional feature data into the feature splicing layer to splice the channel dimensions and obtain multi-scale risk feature data.

[0123] Optionally, convolutional feature data at different scales contain different feature information. By concatenating them along the channel dimension, the multi-scale features are fused into a complete feature vector, retaining feature information at all scales and providing comprehensive feature support for subsequent risk prediction. The concatenation expression can be: In the formula, Indicates the first Multi-scale risk characteristic data at different times. Indicates the first The time-inflation rate Convolutional feature data, Indicates the first One expansion rate, The total number of dilation rates is represented by square brackets, which indicate the concatenation operation of channel dimensions, that is, combining feature vectors of different scales into a higher-dimensional feature vector according to the channel direction.

[0124] S66. Input the multi-scale risk feature data into the probability regression prediction output layer, perform risk probability regression prediction on the multi-scale risk feature data, and output landslide risk probability data.

[0125] The expression for the landslide risk probability data can be: In the formula, Indicates the first The landslide risk probability data at any given time ranges from [0,1]. The closer the value is to 1, the higher the landslide risk; the closer it is to 0, the lower the landslide risk. This represents the sigmoid activation function, used to map regression results to the [0,1] interval, thereby normalizing the risk probability; This represents the risk probability regression weight vector, used to weight multi-scale risk feature data and highlight features that have a significant impact on risk prediction. Indicates the first Multi-scale risk characteristic data at different times, including multi-scale pore pressure and stability characteristic information; This represents the risk probability regression bias term, used to adjust the offset of the regression results; This represents the transpose of the risk probability regression weight vector, ensuring the rationality of matrix operations. By combining linear regression with the sigmoid activation function using the above formula, multi-scale risk characteristic data is transformed into landslide risk probabilities, achieving accurate prediction of risk probabilities.

[0126] S67. Perform multi-level threshold classification on the landslide risk probability data to obtain landslide risk early warning data.

[0127] Optionally, to facilitate practical early warning applications, different risk levels need to be defined based on the probability of landslide risk. Each risk level corresponds to a different early warning level. The classification principle is determined based on factors such as the geological conditions, population distribution, and economic value of the target area. The risk level judgment rule is: when... At that time, the landslide risk warning level was low risk. At that time, the landslide risk warning level was medium risk. At that time, the landslide risk warning level was high. Among them, Indicates the first Landslide risk probability data at any given time. This indicates the threshold for distinguishing between low and medium risk. The threshold for distinguishing between medium and high risk ( Substituting the landslide risk probability data at each moment into the classification expression yields the corresponding risk warning level. The landslide risk warning data is obtained by summing up the warning levels at all moments. This data can be directly used for landslide warning decision-making, providing precise support for disaster prevention and control.

[0128] The aforementioned landslide risk prediction and compensation method first acquires and preprocesses rainfall monitoring data, soil parameter data, and historical slope stability assessment data for the target area. Then, it extracts time-series segmented features from the rainfall monitoring data to obtain rainfall time-series segmented feature data. Next, it performs infiltration response correlation analysis on the soil parameter data and rainfall time-series segmented feature data to obtain rainfall infiltration feature sequence data. This data is then input into a dynamic lag compensation model for time-series lag dependency modeling and dynamic weight allocation to obtain time-series weighted response data. Based on soil fissure distribution data, it performs nonlinear mapping compensation on this data to obtain dynamic pore pressure compensation data. Finally, it inputs the dynamic pore pressure compensation data and historical slope stability assessment data into a landslide risk prediction model for prediction and classification, ultimately obtaining landslide risk early warning data. This method effectively captures the dynamic lag effect of rainfall infiltration, incorporates the spatial modulation effect of non-uniform soil fissure distribution, improves the accuracy and lead time of landslide risk early warning, reduces prediction bias, minimizes data redundancy, and achieves accurate characterization of time-varying pore pressure response and accurate prediction and classification of landslide risk, reducing early warning errors and providing reliable support for landslide disaster prevention and control.

[0129] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0130] Based on the same inventive concept, this application also provides a landslide risk prediction and compensation system for implementing the landslide risk prediction and compensation method described above. The solution provided by this system is similar to the implementation scheme described in the above method; therefore, the specific limitations of one or more embodiments of the landslide risk prediction and compensation system provided below can be found in the limitations of the landslide risk prediction and compensation method described above, and will not be repeated here.

[0131] In one exemplary embodiment, such as Figure 2 As shown, a schematic diagram of a landslide risk prediction and compensation system 10 is provided, including:

[0132] The multidimensional data acquisition module 11 is used to acquire rainfall monitoring data, soil parameter data, and historical slope stability assessment data for the target area; the soil parameter data includes soil fissure distribution data.

[0133] The rainfall time-series feature extraction module 12 is used to extract time-series segmented features from rainfall monitoring data to obtain rainfall time-series segmented feature data.

[0134] The infiltration response correlation analysis module 13 is used to perform infiltration response correlation analysis on soil parameter data and rainfall time series segmented characteristic data to obtain rainfall infiltration characteristic sequence data.

[0135] The dynamic lag compensation module 14 is used to input rainfall infiltration characteristic sequence data into the dynamic lag compensation model to perform time-series lag dependency modeling and dynamic weight allocation to obtain time-series weighted response data.

[0136] The dynamic pore pressure compensation module 15 is used to perform nonlinear mapping compensation on time-weighted response data based on soil fracture distribution data to obtain dynamic pore pressure compensation data; the dynamic pore pressure compensation data is used to characterize the time-varying characteristics of pore pressure response under non-uniform fracture distribution conditions;

[0137] The risk classification and prediction early warning module 16 is used to input dynamic pore pressure compensation data and historical slope stability assessment data into the landslide risk prediction model to perform landslide risk prediction and risk classification, and obtain landslide risk early warning data.

[0138] In one embodiment, such as Figure 3 A computer device 300 is provided, comprising:

[0139] At least one processor 301, and at least one memory 302 communicatively connected to said processor 301; said memory stores application code executable by said processor, said application code being executed by said processor to enable said processor to perform the steps of a landslide risk prediction and compensation method as described above;

[0140] The computer device may also include: sensor 303;

[0141] The processor 301, memory 302, and sensor 303 can be connected via bus 304 or other means. The figure shows an example of connection via bus 304. Figure 3 The character is represented by a single thick line, but this does not mean that there is only one bus or a type of bus.

[0142] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0143] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0144] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A landslide risk prediction and compensation method, characterized in that, The method includes: S1. Acquire rainfall monitoring data, soil parameter data, and historical slope stability assessment data for the target area; the soil parameter data includes soil fissure distribution data. S2. Extract time-series segmented features from the rainfall monitoring data to obtain rainfall time-series segmented feature data; S3. Perform infiltration response correlation analysis on the soil parameter data and the rainfall time-series segmented characteristic data to obtain rainfall infiltration characteristic sequence data; S4. Input the rainfall infiltration characteristic sequence data into the dynamic lag compensation model to perform time-series lag dependency modeling and dynamic weight allocation to obtain time-series weighted response data. S5. Based on the soil fracture distribution data, nonlinear mapping compensation is performed on the time-series weighted response data to obtain dynamic pore pressure compensation data; the dynamic pore pressure compensation data is used to characterize the time-varying characteristics of pore pressure response under non-uniform fracture distribution conditions. S6. Input the dynamic pore pressure compensation data and the historical slope stability assessment data into the landslide risk prediction model to perform landslide risk prediction and risk classification, and obtain landslide risk early warning data.

2. The method according to claim 1, characterized in that, The rainfall monitoring data includes rainfall intensity values ​​at multiple sampling times, and S2 includes: S21. Based on the rainfall monitoring data, calculate the gradient difference of rainfall intensity between adjacent sampling times to obtain rainfall gradient sequence data; S22. The rainfall gradient sequence data is input into the change point recognition model for change point detection to obtain the change point probability value at each sampling time. The sampling time at which the change point probability value exceeds a preset probability threshold is determined as the change point position, thus obtaining rainfall change point time data. The change point recognition model includes a one-dimensional convolutional feature extraction layer, a pooling dimensionality reduction layer, and a probability mapping layer. The one-dimensional convolutional feature extraction layer is used to extract local pattern features from the rainfall gradient sequence data to obtain gradient local feature data. The pooling dimensionality reduction layer is used to reduce and compress the gradient local feature data to obtain gradient dimensionality reduction feature data. The probability mapping layer is used to map the gradient dimensionality reduction feature data to change point probability to obtain the change point probability value at each sampling time. S23. Based on the location of the change point in the rainfall change point time data, the rainfall monitoring data is segmented into time series segments to obtain rainfall segment sequence data; S24. Perform multidimensional feature extraction on each segment of the rainfall segment sequence data to obtain rainfall segment feature vector data; the rainfall segment feature vector data includes the cumulative amount of rainfall, peak intensity and duration features within each segment; S25. Perform time-series decay position encoding on the rainfall segment feature vector data to obtain the rainfall time-series segment feature data.

3. The method according to claim 2, characterized in that, S3 includes: S31. Extract infiltration characteristic parameters from the soil parameter data to obtain soil infiltration characteristic data; the soil infiltration characteristic data is used to characterize the soil's permeability and water retention capacity. S32. Calculate the infiltration capacity matching degree between the rainfall intensity of each segment in the rainfall time series segmented characteristic data and the soil infiltration characteristic data to obtain segmented infiltration matching degree data. S33. Based on the segmented infiltration matching degree data and the soil infiltration characteristic data, perform time delay calculation on the water transport delay after each segmented rainfall infiltration to obtain segmented infiltration time delay data. S34. Based on the segmented infiltration time lag data, the rainfall time series segmented feature data are re-aligned on the time axis according to the infiltration time lag corresponding to each segment to obtain infiltration delay alignment sequence data; the infiltration delay alignment sequence data is used to characterize the distribution of each segmented rainfall feature on the alignment time axis after infiltration delay. S35. Input the infiltration delay alignment sequence data and the segmented infiltration matching degree data into the infiltration response association model for feature association fusion to obtain the rainfall infiltration feature sequence data; wherein, the infiltration response association model includes a dual-path attention fusion layer and a sequence aggregation layer, the dual-path attention fusion layer is used to calculate cross-attention weights on the infiltration delay alignment sequence data and the segmented infiltration matching degree data to obtain infiltration association fusion feature data, and the sequence aggregation layer is used to aggregate and splice the infiltration association fusion feature data in the time dimension to obtain the rainfall infiltration feature sequence data.

4. The method according to claim 1, characterized in that, The dynamic hysteresis compensation model includes a forgetting gating structure, an input gating structure, a unit state update structure, and an output gating structure. S4 includes: S41. Input the rainfall infiltration characteristic sequence data and the lag hidden state data of the preceding time step into the forgetting gating structure, calculate the historical information forgetting ratio of the rainfall infiltration characteristic sequence data and the lag hidden state data, and obtain the forgetting gating data. S42. Input the rainfall infiltration feature sequence data and the hysteresis hidden state data of the preceding time step into the input gating structure, calculate the current input information retention ratio of the rainfall infiltration feature sequence data and the hysteresis hidden state data, and obtain the input gating data and candidate unit state data. S43. Input the forgotten gating data, the input gating data, the candidate unit state data, and the lagging unit state data from the previous time step into the unit state update structure. Perform selective erasure of historical information on the lagging unit state data from the previous time step and selective writing of current information on the candidate unit state data to obtain the lagging unit state data at the current time step; wherein, the expression of the lagging unit state data is: in, Indicates the first Lag unit state data at time t, Indicates the first Lag unit state data at time t, This indicates an element-wise multiplication operation. Indicates the first Forgetting gating data at any time Indicates the first Input gating data at any time, Indicates the first Candidate cell state data at any given time; S44. Input the rainfall infiltration characteristic sequence data, the hysteresis latent state data of the preceding time step and the hysteresis unit state data of the current time step into the output gating structure, filter the output information of the hysteresis unit state data of the current time step, and obtain the hysteresis latent state data of the current time step. S45. Based on the hysteresis latent state data at each time point, perform dynamic weight allocation on the hysteresis latent state data at each time point to obtain the time-series weighted response data.

5. The method according to claim 1, characterized in that, S5 includes: S51. Extract fracture nodes from the soil fracture distribution data to obtain fracture node feature vectors, and construct adjacency relationships from the fracture node feature vectors to obtain fracture adjacency graph data; the fracture adjacency graph data includes a fracture node adjacency matrix. S52. The crack adjacency graph data is input into the crack spatial propagation model for multi-layer graph convolutional feature propagation to obtain crack propagation feature data. The crack spatial propagation model includes a multi-layer graph convolutional feature extraction layer and a node feature normalization layer. The multi-layer graph convolutional feature extraction layer is used to aggregate neighborhood information of the crack adjacency graph data to obtain crack neighborhood aggregated feature data. The node feature normalization layer is used to normalize the features of the crack neighborhood aggregated feature data to obtain the crack propagation feature data. The expression for the crack neighborhood aggregated feature data is: in, Indicates the first The first layer of the graph convolutional layer Clustered feature data of the fracture neighborhood of each fracture node This indicates a modified linear unit activation function. Indicates the first The weight matrix of the layered graph convolutional layer, Indicates the first The set of neighboring nodes of a crack node. Indicates the first The degree of each fracture node. Indicates the first The degree of each neighboring node. Indicates the first The first layer of the graph convolutional layer The feature vector of the fracture node of each neighboring node. Indicates the first The bias vector of the layer graph convolutional layer, Indicates the sequence number of the graph convolutional layer. Indicates the neighbor node index; S53. Perform nonlinear permeability coefficient mapping on the fracture propagation characteristic data to obtain fracture permeability coefficient data; the fracture permeability coefficient data is used to characterize the enhancement of permeability at each fracture node location under the influence of the fracture network. S54. Calculate the spatial attention weights of the fracture permeability coefficient data to obtain fracture spatial attention weight data; the fracture spatial attention weight data is used to characterize the relative contribution weight of each fracture node to the overall pore pressure response. S55. Based on the fracture space attention weight data, perform spatiotemporal coupling nonlinear compensation on the time-series weighted response data to obtain the dynamic pore pressure compensation data.

6. The method according to claim 1, characterized in that, The landslide risk prediction model includes a feature encoding layer, a cross-attention fusion layer, a feature projection layer, a multi-scale dilated convolutional layer, a feature splicing layer, and a probabilistic regression prediction output layer. S6 includes: S61. Input the historical slope stability assessment data into the feature coding layer, and perform stability state coding and temporal location coding on the historical slope stability assessment data to obtain historical stability coding sequence data; S62. Input the historical stability encoding sequence data and the dynamic pore pressure compensation data into the cross-attention fusion layer, and use the dynamic pore pressure compensation data as the query input and the historical stability encoding sequence data as the key value input to calculate the attention weights and obtain attention weighted feature data. S63. Input the attention-weighted feature data into the feature projection layer for dimension mapping to obtain pore pressure stability fusion feature data; S64. Input the pore pressure stability fusion feature data into the multi-scale dilated convolutional layer, and use convolutional kernels with different dilation rates to perform parallel convolutional feature extraction to obtain multi-scale convolutional feature data. S65. Input the multi-scale convolutional feature data into the feature splicing layer to perform channel dimension splicing to obtain multi-scale risk feature data; S66. Input the multi-scale risk feature data into the probability regression prediction output layer, perform risk probability regression prediction on the multi-scale risk feature data, and output landslide risk probability data; wherein, the expression of the landslide risk probability data is: in, Indicates the first Landslide risk probability data at any given time. This represents the sigmoid activation function. This represents the risk probability regression weight vector. Indicates the first Multi-scale risk characteristic data at different times. This represents the risk probability regression bias term. Indicates the time sequence number; S67. Perform multi-level threshold classification on the landslide risk probability data to obtain the landslide risk early warning data.

7. A landslide risk prediction and compensation system, characterized in that, The system includes: The multidimensional data acquisition module is used to acquire rainfall monitoring data, soil parameter data, and historical slope stability assessment data for the target area; the soil parameter data includes soil fissure distribution data. The rainfall time-series feature extraction module is used to extract time-series segmented features from the rainfall monitoring data to obtain rainfall time-series segmented feature data; The infiltration response correlation analysis module is used to perform infiltration response correlation analysis on the soil parameter data and the rainfall time-series segmented characteristic data to obtain rainfall infiltration characteristic sequence data. The dynamic lag compensation module is used to input the rainfall infiltration characteristic sequence data into the dynamic lag compensation model to perform time-series lag dependency modeling and dynamic weight allocation to obtain time-series weighted response data. The dynamic pore pressure compensation module is used to perform nonlinear mapping compensation on the time-series weighted response data based on the soil fracture distribution data to obtain dynamic pore pressure compensation data; the dynamic pore pressure compensation data is used to characterize the time-varying features of pore pressure response under non-uniform fracture distribution conditions; The risk classification and prediction early warning module is used to input the dynamic pore pressure compensation data and the historical slope stability assessment data into the landslide risk prediction model to perform landslide risk prediction and risk classification, and obtain landslide risk early warning data.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 6.

Images