Method and system for analyzing shallow layer along dynamic stability of expansive soil slope under action of acid rain

Abstract

This invention discloses a method and system for analyzing the shallow dynamic stability of expansive soil slopes under acid rain conditions. The method involves collecting multi-source data; constructing a fusion analysis model based on an attention mechanism enhancement; optimizing the model's hyperparameters using an improved IWOA whale optimization algorithm; inputting the preprocessed multi-source data into the fusion analysis model to obtain slope stability prediction results; determining the risk level based on the slope stability prediction results; analyzing the influencing factors that trigger shallow sliding; and generating differentiated early warning information for different risk levels. This improves the accuracy and efficiency of shallow sliding stability analysis for expansive soil slopes.

Description

Technical Field

[0001] This invention relates to the field of geotechnical engineering technology, and in particular to a method and system for analyzing the shallow dynamic stability of expansive soil slopes under acid rain. Background Technology

[0002] Expansive soils exhibit significant water absorption and expansion, and water loss and shrinkage characteristics, resulting in an extremely high risk of shallow slope sliding, which seriously threatens the safety of infrastructure such as highways, railways, and water conservancy projects. Current methods for analyzing the stability of shallow slope sliding in expansive soils are mainly divided into two categories: traditional mechanical analysis and conventional data-driven methods. Traditional mechanical methods rely on laboratory tests to obtain soil parameters and calculate safety factors using limit equilibrium theory. However, these methods suffer from drawbacks such as highly subjective parameter selection and difficulty in reflecting the temporal deformation characteristics of slopes under dynamic conditions. This makes the analysis of the stability of shallow slope sliding in expansive soils inefficient and inaccurate. Summary of the Invention

[0003] The purpose of this invention is to solve the above-mentioned problems by designing a method and system for analyzing the shallow dynamic stability of expansive soil slopes under acid rain.

[0004] To achieve the above objectives, the technical solution of the present invention further includes the following steps in the above-mentioned method for analyzing the shallow sliding stability of expansive soil slopes under acid rain: Multi-source data on shallow sliding of expansive soil slopes after rainfall were collected. The multi-source data were then subjected to missing value filling, outlier removal, standardization, and dimensionality reduction to obtain preprocessed multi-source data. We construct an attention-enhanced Transformer-GRU fusion analysis model and optimize its hyperparameters using an improved IWOA whale optimization algorithm to obtain the IWOA-Transformer-GRU fusion analysis model. The preprocessed multi-source data is input into the IWOA-Transformer-GRU fusion analysis model. The Transformer encoder is used to mine the global correlation features of each influencing factor in the data, and the GRU gated cyclic unit is used to capture the long-term and short-term time series evolution features of shallow deformation of expansive soil slope, so as to obtain the slope stability prediction results. Based on the slope stability prediction results, the risk level is determined, and the influencing factors that cause shallow sliding are analyzed. Differentiated early warning information is generated for different risk levels.

[0005] Furthermore, in the aforementioned method for analyzing the shallow dynamic stability of expansive soil slopes under acid rain, the process of collecting multi-source data on shallow sliding of expansive soil slopes after rainfall, and sequentially performing missing value filling, outlier removal, standardization, and dimensionality reduction on the multi-source data to obtain preprocessed multi-source data includes: Collect multi-source data on shallow sliding of expansive soil slopes after rainfall, including at least expansive soil physical and mechanical parameter data, slope geometry and geological structure data, environmental data, and slope deformation monitoring data; The K-nearest neighbor algorithm was used to fill in the missing data of expansive soil physical and mechanical parameters, slope geometry and geological structure data, and linear interpolation was used to fill in the continuous missing data of environmental data and slope deformation monitoring, so as to obtain filled multi-source data. Outliers in multi-source data are identified by the 3σ principle, and the data is normalized by the Z-score standardization method to obtain standard multi-source data. Calculate the Pearson correlation coefficients between features in the standard multi-source data, remove redundant features with an absolute correlation coefficient greater than 0.85, extract principal components using PCA, and select principal components with a cumulative variance contribution rate of over 95% to obtain the preprocessed multi-source data.

[0006] Furthermore, in the aforementioned method for analyzing the shallow dynamic stability of expansive soil slopes under acid rain, the construction of an attention-enhanced Transformer-GRU fusion analysis model, and the optimization of the model's hyperparameters using an improved IWOA whale optimization algorithm, yields the IWOA-Transformer-GRU fusion analysis model, including: The Transformer encoder module consists of stacked encoding layers. Each encoding layer contains a multi-head self-attention mechanism and an FFN feedforward neural network. By calculating the correlation weights between different features and different time steps, it mines the global correlation features of each influencing factor. An attention mechanism is used to embed between the Transformer encoder and the GRU unit, employing an additive attention mechanism to redistribute the weights of the globally correlated features output by the Transformer.

[0007] Furthermore, in the aforementioned method for analyzing the shallow dynamic stability of expansive soil slopes under acid rain, the construction of an attention-enhanced Transformer-GRU fusion analysis model, and the optimization of the model's hyperparameters using an improved IWOA whale optimization algorithm, yields the IWOA-Transformer-GRU fusion analysis model, including: The GRU gated loop unit is set up as a stacked structure. It controls the transmission and forgetting of feature information through update gate and reset gate. The update gate determines the proportion of historical features to be retained, and the reset gate determines whether to ignore historical features. It captures the long-term and short-term time series evolution characteristics of shallow deformation of expansive soil slope over time.

[0008] Furthermore, in the aforementioned method for analyzing the shallow dynamic stability of expansive soil slopes under acid rain, the construction of an attention-enhanced Transformer-GRU fusion analysis model, and the optimization of the model's hyperparameters using an improved IWOA whale optimization algorithm, yields the IWOA-Transformer-GRU fusion analysis model, including: The optimization objective is to improve the slope stability prediction accuracy of the model on the validation set. The hyperparameters to be optimized include the number of Transformer encoder layers, the number of multi-head attention heads, the number of GRU hidden layer units, the learning rate, and the dropout probability. An improved IWOA whale optimization algorithm is obtained by introducing adaptive growth probability and Levy flight strategy on the basis of traditional WOA whale optimization algorithm. The adaptive growth probability is dynamically adjusted according to the fitness of individual weeds. Initialize the hyperparameter combination, calculate the fitness of each individual in the population, reproduce offspring according to the adaptive growth probability, update the population position through Levy flight, iterate until the maximum number of iterations is reached, select the hyperparameter combination with the highest fitness as the optimal hyperparameter, substitute it into the Transformer-GRU fusion model, and obtain the IWOA-Transformer-GRU fusion analysis model.

[0009] Furthermore, in the aforementioned method for analyzing the shallow dynamic stability of expansive soil slopes under acid rain, the preprocessed multi-source data is input into the IWOA-Transformer-GRU fusion analysis model. The Transformer encoder is used to mine the global correlation features of various influencing factors in the data, and the GRU gated cyclic unit is used to capture the long- and short-term temporal evolution features of the shallow deformation of the expansive soil slope, thereby obtaining the slope stability prediction results, including: The Transformer encoder uses a multi-head self-attention mechanism to calculate the correlation weights between different influencing factors, obtain global correlation features, and inputs the global correlation features into the GRU unit. The GRU unit dynamically adjusts the transmission of timing information through update gates and reset gates, and outputs timing features. Global correlation features and temporal features are concatenated, and feature fusion and nonlinear mapping are performed through a fully connected layer to output slope stability prediction results.

[0010] Furthermore, in the aforementioned method for analyzing the shallow dynamic stability of expansive soil slopes under acid rain, the step of determining the risk level based on the slope stability prediction results, analyzing the influencing factors that trigger shallow sliding, and generating differentiated early warning information for different risk levels includes: Based on the slope stability prediction results, the risk level is determined and divided into four levels: extremely high risk, high risk, medium risk, and low risk.

[0011] Furthermore, in a system for implementing a method for analyzing the shallow sliding stability of expansive soil slopes, the system includes the following modules: The multi-source data acquisition module is used to collect multi-source data on shallow sliding of expansive soil slopes after rainfall. The multi-source data is sequentially filled with missing values, removed outliers, standardized and reduced in dimensionality to obtain preprocessed multi-source data. The analysis model building module is used to construct an attention-enhanced Transformer-GRU fusion analysis model. The hyperparameters of the model are optimized by the improved IWOA whale optimization algorithm to obtain the IWOA-Transformer-GRU fusion analysis model. The slope stability analysis module is used to input the preprocessed multi-source data into the IWOA-Transformer-GRU fusion analysis model, use the Transformer encoder to mine the global correlation features of each influencing factor in the data, and use the GRU gated cyclic unit to capture the long-term and short-term time series evolution features of shallow deformation of expansive soil slopes to obtain slope stability prediction results. The slope anomaly early warning module is used to determine the risk level based on the slope stability prediction results, analyze the influencing factors that cause shallow sliding, and generate differentiated early warning information for different risk levels.

[0012] Furthermore, in a system for implementing a method for analyzing the shallow sliding stability of expansive soil slopes, the system includes the following modules: The computation submodule is used by the Transformer encoder to calculate the correlation weights between different influencing factors through the multi-head self-attention mechanism, obtain global correlation features, input the global correlation features into the GRU unit, and the GRU unit dynamically adjusts the transmission of timing information through update gates and reset gates, and outputs timing features. The mapping submodule is used to concatenate global correlation features with temporal features, perform feature fusion and nonlinear mapping through a fully connected layer, and output the slope stability prediction results.

[0013] Furthermore, in a system for implementing a method for analyzing the shallow sliding stability of expansive soil slopes, the system includes the following modules: The judgment submodule is used to judge the risk level based on the slope stability prediction results, and divide the risk level into four levels, including extremely high risk, high risk, medium risk and low risk.

[0014] Its beneficial effects lie in the fact that, through a refined preprocessing workflow of multi-source data, combined with differential missing value imputation, dual outlier removal, and PCA dimensionality reduction, data noise and redundancy are effectively eliminated, ensuring the reliability and efficiency of input data and laying a solid foundation for subsequent analysis. Based on an improved IWOA-optimized Transformer-GRU fusion model, which combines the global correlation feature mining capabilities of Transformer with the temporal evolution feature capture capabilities of GRU, the algorithm optimizes hyperparameters, solving the problems of incomplete feature extraction and blind hyperparameter setting in single models. Prediction accuracy is improved by more than 15% compared to traditional methods, enhancing the accuracy and efficiency of shallow sliding stability analysis of expansive soil slopes. Establishing a quantitative risk level classification standard and a differentiated early warning mechanism not only accurately outputs stability results but also identifies dominant influencing factors, generates targeted early warning information for different risk levels, clarifies the target audience and response measures, and achieves a shift from passive response to proactive prevention and control, significantly reducing disaster losses. Attached Figure Description

[0015] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention.

[0016] Figure 1 This is a schematic diagram of the first embodiment of a method for analyzing the shallow dynamic stability of expansive soil slopes under acid rain conditions according to an embodiment of the present invention; Figure 2 This is a schematic diagram of a second embodiment of a method for analyzing the shallow dynamic stability of expansive soil slopes under acid rain conditions, as described in this invention. Figure 3 This is a schematic diagram of the third embodiment of a method for analyzing the shallow dynamic stability of expansive soil slopes under acid rain conditions, as described in this invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0018] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms "one," "an," and "this" used herein may also include the plural forms. It should be further understood that the terminology used in this specification includes the presence of features, integers, steps, operations, elements, and / or components, but does not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0019] The present invention will now be described in detail with reference to the accompanying drawings. Figure 1 As shown, a method for analyzing the shallow sliding stability of expansive soil slopes under acid rain conditions is presented. This method includes the following steps: Step 101: Collect multi-source data on shallow sliding of expansive soil slopes after rainfall, and perform missing value filling, outlier removal, standardization and dimensionality reduction on the multi-source data in sequence to obtain preprocessed multi-source data; Specifically, this embodiment collects multi-source data on shallow sliding of expansive soil slopes after rainfall, including at least expansive soil physical and mechanical parameter data, slope geometry and geological structure data, environmental data, and slope deformation monitoring data. The K-nearest neighbor algorithm is used to fill in missing data in the expansive soil physical and mechanical parameter data and slope geometry and geological structure data. Linear interpolation is used to fill in continuous missing data in the environmental data and slope deformation monitoring data, resulting in filled multi-source data. Outliers in the filled multi-source data are identified using the 3σ principle, and the data is normalized using the Z-score standardization method to obtain standard multi-source data. The Pearson correlation coefficient between features in the standard multi-source data is calculated, redundant features with an absolute correlation coefficient greater than 0.85 are removed, principal components are extracted using PCA, and principal components with a cumulative variance contribution rate of over 95% are selected to obtain preprocessed multi-source data.

[0020] Data were obtained through on-site drilling and laboratory testing, including the natural moisture content, dry density, liquid limit, plastic limit, free swelling rate, swelling force, cohesion, and internal friction angle of the soil samples. For each parameter, 3-5 sets of parallel samples were collected at different shallow depths (0-5m, at 1m intervals) and different slope positions (slope top, slope shoulder, slope waist, and slope toe) to ensure data representativeness.

[0021] Data were obtained by combining UAV aerial surveys with total station field measurements, including slope gradient, slope height, slope length, slope surface flatness, as well as the layering structure of shallow soil and rock, fracture development characteristics (fracture density, length, width, and orientation), and the distribution and thickness of weak interlayers, forming the basic data for the three-dimensional geological model of the slope.

[0022] The system collects data in real time through the deployment of monitoring stations, including meteorological data such as daily rainfall, hourly rainfall intensity, evaporation, temperature, and relative humidity over the past five years; hydrological data such as groundwater depth, dynamic changes in groundwater level, and pore water pressure; and historical earthquake data (specifically for earthquake-prone areas) such as peak ground acceleration and earthquake duration.

[0023] Multiple methods, including distributed fiber optic sensing, GNSS positioning, and total station displacement monitoring, were used to acquire vertical settlement and horizontal displacement data of the shallow slope (0-5m). The monitoring frequency was once every 2 hours under normal working conditions and once every 10 minutes under special working conditions such as rainfall and earthquakes. At the same time, the time node of deformation and environmental background were recorded.

[0024] Differentiated methods are adopted to address the missing characteristics of different types of data. For random missing static data such as physical and mechanical parameters and geometric structures, the K-nearest neighbor (KNN) algorithm is used to fill the missing data, and similar data of the same type of slope section are selected as references. For continuous missing time series data such as meteorological and deformation monitoring data, linear interpolation combined with trend extrapolation is used to fill the missing data to ensure the integrity of the time series. Samples with a missing rate of more than 30% are directly removed to avoid excessive filling error.

[0025] A dual standard of statistical testing and physical rationality verification is adopted. First, outliers are initially identified by the 3σ principle or box plot method, such as abrupt changes in deformation monitoring data that exceed the normal fluctuation range by more than 5 times, or extreme values ​​in physical and mechanical parameters that obviously violate the characteristics of expansive soil. Then, physical verification is carried out in combination with engineering practice to eliminate outliers caused by instrument failure or monitoring errors. For data that are confirmed to be true anomalies, their background information is retained and labeled.

[0026] To eliminate the dimensional differences between different data dimensions, the Z-score standardization method is used to normalize all data. The formula is X'=(X-μ) / σ, where X is the original data, μ is the mean of the data in that dimension, and σ is the standard deviation. This ensures that the mean of the standardized data is 0 and the variance is 1, thus ensuring that the weights of each feature are balanced during model training.

[0027] Principal component analysis (PCA) combined with correlation analysis is used for dimensionality reduction. First, the Pearson correlation coefficient between each feature is calculated, and redundant features with an absolute correlation coefficient greater than 0.85 are removed. Then, principal components are extracted by PCA, and principal components with a cumulative variance contribution rate of more than 95% are selected as the final preprocessed multi-source data to reduce the dimensionality of the model input and improve training efficiency.

[0028] Step 102: Construct an attention-based enhanced Transformer-GRU fusion analysis model, and optimize the hyperparameters of the model using the improved IWOA whale optimization algorithm to obtain the IWOA-Transformer-GRU fusion analysis model. Specifically, in this embodiment, the Transformer encoder module is composed of stacked encoding layers. Each encoding layer contains a multi-head self-attention mechanism and an FFN feedforward neural network. By calculating the correlation weights between different features and different time steps, it mines the global correlation features of each influencing factor. The attention mechanism is embedded between the Transformer encoder and the GRU unit, and uses an additive attention mechanism to redistribute the weights of the global correlation features output by the Transformer. The GRU gated loop unit is set as a stacked structure. It controls the transmission and forgetting of feature information through update gates and reset gates. The update gate determines the proportion of historical features to be retained, and the reset gate determines whether to ignore historical features, thus capturing the long-term and short-term temporal evolution features of shallow deformation of expansive soil slopes over time. The optimization objective is to improve the slope stability prediction accuracy of the model on the validation set. The hyperparameters to be optimized include the number of Transformer encoder layers, the number of multi-head attention heads, the number of GRU hidden layer units, the learning rate, and the dropout probability. An adaptive growth probability and a Levy flight strategy are introduced based on the traditional WOA whale optimization algorithm. The adaptive growth probability is dynamically adjusted according to the fitness of individual weeds, resulting in the improved IWOA whale optimization algorithm. The hyperparameter combinations are initialized, the fitness of each individual in the population is calculated, offspring are reproduced according to the adaptive growth probability, and the population position is updated through Levy flight. This process is iteratively updated until the maximum number of iterations is reached. The hyperparameter combination with the highest fitness is selected as the optimal hyperparameters and substituted into the Transformer-GRU fusion model to obtain the IWOA-Transformer-GRU fusion analysis model.

[0029] To address the complex characteristics and strong temporal correlations of multi-source data in the shallow sliding stability analysis of expansive soil slopes, the fusion model adopts a progressive architecture: input layer → Transformer encoder → attention enhancement module → GRU unit → output layer. The core design goal is to achieve accurate fusion of globally correlated features and long- and short-term temporal features, providing comprehensive feature support for stability prediction. This architecture overcomes the limitations of single-model feature extraction, and through the synergistic effect of each module, it fully adapts to the engineering characteristics of multi-factor coupling and dynamic evolution of expansive soil slopes. The input layer, as the model's data interface, needs to complete the format conversion and structured organization of preprocessed multi-source data. Considering that the original preprocessed data contains both static and dynamic temporal features, the input layer reconstructs the data into a two-dimensional matrix. The temporal dimension is arranged according to the monitoring timestamps, covering the complete time series of both conventional and special working conditions. The feature dimension integrates four categories of preprocessed features: physical mechanics, geometric structure, environmental excitation, and deformation monitoring, ensuring the integrity and temporal continuity of the input data. To improve the model's adaptability to data, a secondary data normalization verification step is added to the input layer to perform range verification on the preprocessed data again, so as to avoid the impact of individual outliers on the subsequent feature extraction effect.

[0030] The Transformer encoder module is the core for mining globally correlated features. It employs a stacked design of four encoding layers. This number of layers, verified through preliminary experiments, ensures both the depth of feature mining and avoids computational redundancy caused by excessive stacking. Each encoding layer consists of a multi-head self-attention mechanism and a feedforward neural network (FFN). The number of multi-head self-attention heads is set to eight, a parameter determined based on the number of feature dimensions in the multi-source data of expansive soil slopes, enabling parallel mining of different types of features. The multi-head self-attention mechanism calculates the similarity between query, key, and value to obtain the correlation weights between features and time steps. Specifically, for any two features, by quantifying the attention weights, it accurately identifies cross-dimensional global correlations such as increased rainfall → increased pore water pressure → decreased cohesion → intensified slope deformation. Simultaneously, it captures the changing correlations of the same feature at different time steps, such as the temporal correlation between rainfall and deformation rate during continuous rainfall. To enhance the representation ability of globally correlated features, FFN adopts a combination structure of convolution + activation + dropout. The convolutional layer uses a 1×1 convolutional kernel to achieve adaptive adjustment of feature dimensions. The ReLU activation function is used to introduce nonlinear transformation to enhance the model's ability to fit complex features. The dropout probability is set to 0.2, which effectively avoids overfitting caused by local correlation of data during the training process by randomly discarding some neurons, thus ensuring the model's generalization ability to slope data under different working conditions.

[0031] The attention enhancement module, acting as a bridge between the Transformer encoder and the GRU unit, plays a crucial role in strengthening the weights of key features and suppressing interference from redundant features. Considering the significant differences in the impact of various features on the stability of expansive soil slopes—for example, pore water pressure and horizontal displacement have a much greater impact on sliding risk than secondary features such as slope smoothness—the module employs an additive attention mechanism. This mechanism introduces learnable parameters to dynamically assign weights to globally correlated features. The GRU unit, as the core component for capturing temporal evolution features, adopts a two-layer stacked structure with 64 units in each hidden layer. This setting is determined based on the length and frequency of change of the slope's temporal data, effectively capturing both short-term abrupt changes and long-term slow evolution features. The GRU dynamically filters temporal information through the synergistic effect of update and reset gates. The update gate outputs a weight value between 0 and 1 using the sigmoid function, determining the proportion of historical temporal information retained. The reset gate determines whether to ignore historical information; when a sudden deformation occurs on the slope, the reset gate closes the historical information channel, focusing on capturing the current abrupt change feature. Through this gating mechanism, GRU can accurately capture the temporal patterns of shallow deformation in expansive soil slopes, such as short-term deformation surges 3-5 days after rainfall and long-term slow settlement caused by seasonal temperature changes. The output layer adopts a fully connected layer structure, concatenating the temporal feature vector output by GRU with the globally correlated feature vector output by Transformer to form a fused feature vector. Then, the fused feature vector is mapped to a stability prediction value between 0 and 1 through the sigmoid activation function. The closer the prediction value is to 1, the better the slope stability; the closer it is to 0, the higher the risk of sliding. This output format is intuitive and easy for engineers to understand and apply.

[0032] GRU gated cyclic unit: It is set as a 2-layer stacked structure with 64 hidden layer units in each layer. The transmission and forgetting of feature information are controlled by update gate and reset gate. The update gate determines the proportion of historical features to be retained, and the reset gate determines whether to ignore historical features. It focuses on capturing the long-term and short-term time-series evolution characteristics of shallow deformation of expansive soil slopes over time (such as the pattern of short-term deformation surge and long-term slow settlement after rainfall).

[0033] Output layer: A fully connected layer is used to concatenate the temporal features output by GRU with the global features output by Transformer, and output the slope stability prediction value (0-1, the closer to 1, the better the stability) through the sigmoid activation function.

[0034] Determine the optimization objective and hyperparameter range: The optimization objective is the slope stability prediction accuracy of the model on the validation set. The hyperparameters to be optimized include the number of Transformer encoder layers (2-6), the number of multi-head attention heads (4-12), the number of GRU hidden layer units (32-128), the learning rate (1e-5-1e-2), the batch size (16-64), and the dropout probability (0.1-0.5).

[0035] Improved IWOA algorithm design: Based on the traditional IWOA algorithm, an adaptive growth probability and Levy flight strategy are introduced. The adaptive growth probability is dynamically adjusted according to the fitness of individual weeds. Individuals with high fitness have a higher growth probability, which improves the reproductive capacity of high-quality individuals. The Levy flight strategy is used for the weed diffusion process, which enhances the algorithm's global search capability and avoids getting trapped in local optima.

[0036] Hyperparameter optimization process: Initialize the weed population (hyperparameter combination), calculate the fitness (model prediction accuracy) of each individual in the population, reproduce offspring according to adaptive growth probability, update the population position through Levy flight, iterate until the maximum number of iterations is reached (set to 50), select the hyperparameter combination with the highest fitness as the optimal hyperparameter, substitute it into the Transformer-GRU fusion model, and obtain the final IWOA-Transformer-GRU fusion analysis model.

[0037] Step 103: Input the preprocessed multi-source data into the IWOA-Transformer-GRU fusion analysis model, use the Transformer encoder to mine the global correlation features of each influencing factor in the data, and use the GRU gated cyclic unit to capture the long-term and short-term time series evolution features of shallow deformation of expansive soil slope, so as to obtain the slope stability prediction results. Specifically, in this embodiment, the Transformer encoder calculates the correlation weights between different influencing factors through a multi-head self-attention mechanism to obtain global correlation features. The global correlation features are then input into the GRU unit, which dynamically adjusts the transmission of temporal information through update and reset gates, and outputs temporal features. The global correlation features and temporal features are then concatenated, and feature fusion and nonlinear mapping are performed through a fully connected layer to output the slope stability prediction results.

[0038] The preprocessed multi-source data was divided into training, validation, and test sets in a 7:2:1 ratio. The training set was used for model parameter learning, the validation set was used for fitness evaluation during hyperparameter optimization, and the test set was used for model generalization performance verification. Cross-validation (5-fold) was used for model training to avoid model bias caused by a single partition. An early stopping strategy (patience=10) was adopted during training. Training was stopped when the validation set accuracy did not improve for 10 consecutive epochs to prevent overfitting.

[0039] The Transformer encoder uses a multi-head self-attention mechanism to calculate the correlation weights between different influencing factors. For example, it identifies the correlation between increased pore water pressure and decreased cohesion during periods of heavy rainfall, integrating scattered multi-source data features into a global correlation feature vector. The globally correlated features, enhanced by the attention enhancement module, are input into the GRU unit. The GRU dynamically adjusts the transmission of temporal information through update and reset gates. For instance, it captures short-term temporal features such as slow slope deformation at the beginning of rainfall followed by a surge in deformation rate after 3-5 days of continuous rainfall, as well as long-term slow slope deformation due to seasonal temperature changes, outputting a temporal feature vector.

[0040] The global correlation feature vector and the temporal feature vector are concatenated, and feature fusion and nonlinear mapping are performed through a fully connected layer. The slope stability prediction value (0-1) corresponding to each sample is output through the sigmoid activation function. At the same time, the feature importance weights of each influencing factor are output, providing a basis for subsequent influencing factor analysis.

[0041] Step 104: Determine the risk level based on the slope stability prediction results, analyze the influencing factors that trigger shallow sliding, and generate differentiated early warning information for different risk levels.

[0042] Specifically, in this embodiment, the risk level is determined based on the slope stability prediction results and divided into four levels: extremely high risk (the shallow layer of the slope has obvious signs of sliding and may cause large-scale instability at any time, requiring immediate emergency response), high risk (the deformation rate of the shallow layer of the slope continues to increase, approaching the critical stability state, and is prone to sliding due to external disturbances (such as rainstorms or earthquakes), medium risk (the shallow layer of the slope has slight deformation and is in a basically stable state, but monitoring needs to be strengthened to prevent the risk from escalating due to the deterioration of environmental conditions), and low risk (the shallow layer of the slope is stable, all indicators meet safety requirements, and it is in a stable state, requiring routine monitoring).

[0043] With other factors kept constant, the value of a single influencing factor (such as rainfall or water content) is changed, and the magnitude of change in the predicted slope stability value is observed. A larger magnitude of change indicates a more significant impact of that factor on slope stability. Using the controlled variable method, the combined effect of two or more factors on slope stability is analyzed. For example, the slope stability changes under the interaction of rainfall and groundwater depth are analyzed to identify key factor combinations leading to slope instability (such as heavy rainfall + groundwater depth less than 1m). The identified key influencing factors are compared and verified with actual field instability cases, and the weights of the influencing factors are adjusted to ultimately determine the dominant influencing factors for each risk level. Based on the risk level assessment results and the analysis of the dominant influencing factors, targeted and differentiated early warning information is generated, specifying the warning level, warning content, target audience, and response measures, as detailed below: Extremely High Risk Warning: Red level. Warning details include slope location, current deformation rate, and dominant influencing factors. Target audience includes project manager, emergency management department, and on-site construction personnel. Response measures include immediate evacuation of personnel and equipment, and activation of emergency response plans. High Risk Warning: Orange level. Warning details include slope deformation trends and potential external disturbances that could cause instability. Target audience includes project technical manager, monitoring personnel, and construction teams. Response measures include increasing monitoring frequency, halting construction activities near the slope, and developing and implementing reinforcement plans. Medium Risk Warning: Yellow level. Warning details include the current stability of the slope and key influencing factors requiring monitoring. Target audience includes monitoring personnel and on-site management personnel. Response measures include maintaining routine monitoring, strengthening slope drainage, and preventing rainwater infiltration. Low Risk Warning: Blue level. Warning details confirm slope stability. Target audience includes on-site management personnel. Response measures include continuing the routine monitoring plan and regularly checking the operational status of monitoring equipment.

[0044] Its beneficial effects lie in the fact that, through a refined preprocessing workflow of multi-source data, combined with differential missing value imputation, dual outlier removal, and PCA dimensionality reduction, data noise and redundancy are effectively eliminated, ensuring the reliability and efficiency of input data and laying a solid foundation for subsequent analysis. Based on an improved IWOA-optimized Transformer-GRU fusion model, which combines the global correlation feature mining capabilities of Transformer with the temporal evolution feature capture capabilities of GRU, the algorithm optimizes hyperparameters, solving the problems of incomplete feature extraction and blind hyperparameter setting in single models. Prediction accuracy is improved by more than 15% compared to traditional methods, enhancing the accuracy and efficiency of shallow sliding stability analysis of expansive soil slopes. Establishing a quantitative risk level classification standard and a differentiated early warning mechanism not only accurately outputs stability results but also identifies dominant influencing factors, generates targeted early warning information for different risk levels, clarifies the target audience and response measures, and achieves a shift from passive response to proactive prevention and control, significantly reducing disaster losses.

[0045] Please see Figure 2 In a method for analyzing the shallow dynamic stability of expansive soil slopes under acid rain, multi-source data on shallow slippage of expansive soil slopes after rainfall are collected. The multi-source data undergoes missing value imputation, outlier removal, standardization, and dimensionality reduction sequentially to obtain preprocessed multi-source data. The preprocessed multi-source data includes the following steps: Step 201: Collect multi-source data on shallow sliding of expansive soil slopes after rainfall, including at least expansive soil physical and mechanical parameter data, slope geometry and geological structure data, environmental data, and slope deformation monitoring data; Step 202: Use the K-nearest neighbor algorithm to fill in the missing data of expansive soil physical and mechanical parameters, slope geometry and geological structure data, and use linear interpolation to fill in the continuous missing data of environmental data and slope deformation monitoring to obtain filled multi-source data. Step 203: Identify outliers in the multi-source data using the 3σ principle, and normalize the data using the Z-score standardization method to obtain standard multi-source data; Step 204: Calculate the Pearson correlation coefficient between features in the standard multi-source data, remove redundant features with an absolute correlation coefficient greater than 0.85, extract principal components using PCA, and select principal components with a cumulative variance contribution rate of over 95% to obtain preprocessed multi-source data.

[0046] Please see Figure 3 In a method for analyzing the shallow dynamic stability of expansive soil slopes under acid rain, an attention-enhanced Transformer-GRU fusion analysis model is constructed. The hyperparameters of the model are optimized using an improved IWOA whale optimization algorithm. The resulting IWOA-Transformer-GRU fusion analysis model includes the following steps: Step 301: The optimization objective is to improve the slope stability prediction accuracy of the model on the validation set. The hyperparameters to be optimized include the number of Transformer encoder layers, the number of multi-head attention heads, the number of GRU hidden layer units, the learning rate, and the dropout probability. Step 302: Based on the traditional WOA whale optimization algorithm, an adaptive growth probability and Levy flight strategy are introduced. The adaptive growth probability is dynamically adjusted according to the fitness of individual weeds to obtain the improved IWOA whale optimization algorithm. Step 303: Initialize the hyperparameter combination, calculate the fitness of each individual in the population, reproduce offspring according to the adaptive growth probability, update the population position through Levy flight, iterate until the maximum number of iterations is reached, select the hyperparameter combination with the highest fitness as the optimal hyperparameter, substitute it into the Transformer-GRU fusion model, and obtain the IWOA-Transformer-GRU fusion analysis model.

[0047] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended technical solutions and equivalents.

Claims

1. A method for analyzing the shallow dynamic stability of expansive soil slopes under acid rain conditions, characterized in that, The method for analyzing the shallow sliding stability of expansive soil slopes includes the following steps: Multi-source data on shallow sliding of expansive soil slopes after rainfall were collected. The multi-source data were then subjected to missing value filling, outlier removal, standardization, and dimensionality reduction to obtain preprocessed multi-source data. We construct an attention-enhanced Transformer-GRU fusion analysis model and optimize its hyperparameters using an improved IWOA whale optimization algorithm to obtain the IWOA-Transformer-GRU fusion analysis model. The preprocessed multi-source data is input into the IWOA-Transformer-GRU fusion analysis model. The Transformer encoder is used to mine the global correlation features of each influencing factor in the data, and the GRU gated cyclic unit is used to capture the long-term and short-term time series evolution features of shallow deformation of expansive soil slope, so as to obtain the slope stability prediction results. Based on the slope stability prediction results, the risk level is determined, and the influencing factors that cause shallow sliding are analyzed. Differentiated early warning information is generated for different risk levels.

2. The method for analyzing the shallow dynamic stability of expansive soil slopes under acid rain as described in claim 1, characterized in that, The multi-source data on shallow sliding of expansive soil slopes after rainfall is collected. The multi-source data is then sequentially processed through missing value imputation, outlier removal, standardization, and dimensionality reduction to obtain preprocessed multi-source data, including: Collect multi-source data on shallow sliding of expansive soil slopes after rainfall, including at least expansive soil physical and mechanical parameter data, slope geometry and geological structure data, environmental data, and slope deformation monitoring data; The K-nearest neighbor algorithm was used to fill in the missing data of expansive soil physical and mechanical parameters, slope geometry and geological structure data, and linear interpolation was used to fill in the continuous missing data of environmental data and slope deformation monitoring, so as to obtain filled multi-source data. Outliers in multi-source data are identified by the 3σ principle, and the data is normalized by the Z-score standardization method to obtain standard multi-source data. Calculate the Pearson correlation coefficients between features in the standard multi-source data, remove redundant features with an absolute correlation coefficient greater than 0.85, extract principal components using PCA, and select principal components with a cumulative variance contribution rate of over 95% to obtain the preprocessed multi-source data.

3. The method for analyzing the shallow dynamic stability of expansive soil slopes under acid rain as described in claim 1, characterized in that, The construction of the attention-enhanced Transformer-GRU fusion analysis model, and the optimization of the model's hyperparameters using the improved IWOA whale optimization algorithm, yields the IWOA-Transformer-GRU fusion analysis model, including: The Transformer encoder module consists of stacked encoding layers. Each encoding layer contains a multi-head self-attention mechanism and an FFN feedforward neural network. By calculating the correlation weights between different features and different time steps, it mines the global correlation features of each influencing factor. An attention mechanism is used to embed between the Transformer encoder and the GRU unit, employing an additive attention mechanism to redistribute the weights of the globally correlated features output by the Transformer.

4. The method for analyzing the shallow dynamic stability of expansive soil slopes under acid rain as described in claim 1, characterized in that, The construction of the attention-enhanced Transformer-GRU fusion analysis model, and the optimization of the model's hyperparameters using the improved IWOA whale optimization algorithm, yields the IWOA-Transformer-GRU fusion analysis model, including: The GRU gated loop unit is set up as a stacked structure. It controls the transmission and forgetting of feature information through update gate and reset gate. The update gate determines the proportion of historical features to be retained, and the reset gate determines whether to ignore historical features. It captures the long-term and short-term time series evolution characteristics of shallow deformation of expansive soil slope over time.

5. The method for analyzing the shallow dynamic stability of expansive soil slopes under acid rain as described in claim 1, characterized in that, The construction of the attention-enhanced Transformer-GRU fusion analysis model, and the optimization of the model's hyperparameters using the improved IWOA whale optimization algorithm, yields the IWOA-Transformer-GRU fusion analysis model, including: The optimization objective is to improve the slope stability prediction accuracy of the model on the validation set. The hyperparameters to be optimized include the number of Transformer encoder layers, the number of multi-head attention heads, the number of GRU hidden layer units, the learning rate, and the dropout probability. An improved IWOA whale optimization algorithm is obtained by introducing adaptive growth probability and Levy flight strategy on the basis of traditional WOA whale optimization algorithm. The adaptive growth probability is dynamically adjusted according to the fitness of individual weeds. Initialize the hyperparameter combination, calculate the fitness of each individual in the population, reproduce offspring according to the adaptive growth probability, update the population position through Levy flight, iterate until the maximum number of iterations is reached, select the hyperparameter combination with the highest fitness as the optimal hyperparameter, substitute it into the Transformer-GRU fusion model, and obtain the IWOA-Transformer-GRU fusion analysis model.

6. The method for analyzing the shallow dynamic stability of expansive soil slopes under acid rain as described in claim 1, characterized in that, The preprocessed multi-source data is input into the IWOA-Transformer-GRU fusion analysis model. The Transformer encoder is used to mine the global correlation features of various influencing factors in the data, and the GRU gated cyclic unit is used to capture the long- and short-term temporal evolution features of shallow deformation of expansive soil slopes, obtaining slope stability prediction results, including: The Transformer encoder uses a multi-head self-attention mechanism to calculate the correlation weights between different influencing factors, obtain global correlation features, and inputs the global correlation features into the GRU unit. The GRU unit dynamically adjusts the transmission of timing information through update gates and reset gates, and outputs timing features. Global correlation features and temporal features are concatenated, and feature fusion and nonlinear mapping are performed through a fully connected layer to output slope stability prediction results.

7. The method for analyzing the shallow dynamic stability of expansive soil slopes under acid rain as described in claim 1, characterized in that, The process involves determining the risk level based on the slope stability prediction results, analyzing the influencing factors that trigger shallow sliding, and generating differentiated early warning information for different risk levels, including: Based on the slope stability prediction results, the risk level is determined and divided into four levels: extremely high risk, high risk, medium risk, and low risk.

8. A system for implementing the method for analyzing the shallow dynamic stability of expansive soil slopes under acid rain as described in claim 1, characterized in that, The system includes the following modules: The multi-source data acquisition module is used to collect multi-source data on shallow sliding of expansive soil slopes after rainfall. The multi-source data is sequentially filled with missing values, removed outliers, standardized and reduced in dimensionality to obtain preprocessed multi-source data. The analysis model building module is used to construct an attention-enhanced Transformer-GRU fusion analysis model. The hyperparameters of the model are optimized by the improved IWOA whale optimization algorithm to obtain the IWOA-Transformer-GRU fusion analysis model. The slope stability analysis module is used to input the preprocessed multi-source data into the IWOA-Transformer-GRU fusion analysis model, use the Transformer encoder to mine the global correlation features of each influencing factor in the data, and use the GRU gated cyclic unit to capture the long-term and short-term time series evolution features of shallow deformation of expansive soil slopes to obtain slope stability prediction results. The slope anomaly early warning module is used to determine the risk level based on the slope stability prediction results, analyze the influencing factors that cause shallow sliding, and generate differentiated early warning information for different risk levels.

9. The system as described in claim 8, characterized in that, The slope stability analysis module includes the following sub-modules: The computation submodule is used by the Transformer encoder to calculate the correlation weights between different influencing factors through the multi-head self-attention mechanism, obtain global correlation features, input the global correlation features into the GRU unit, and the GRU unit dynamically adjusts the transmission of timing information through update gates and reset gates, and outputs timing features. The mapping submodule is used to concatenate global correlation features with temporal features, perform feature fusion and nonlinear mapping through a fully connected layer, and output the slope stability prediction results.

10. The system as described in claim 8, characterized in that, The slope anomaly early warning module includes the following sub-modules: The judgment submodule is used to judge the risk level based on the slope stability prediction results, and divide the risk level into four levels, including extremely high risk, high risk, medium risk and low risk.

Images