A high-explainability method and system for predicting slope stability
By optimizing the hyperparameters of the EBM model and iteratively updating its main effect and interaction effect functions, the interpretability problem in slope stability prediction is solved, and high-precision and interpretable slope stability prediction is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- GUIZHOU UNIV
- Filing Date
- 2026-04-22
- Publication Date
- 2026-06-19
Smart Images
Figure CN122241154A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of geotechnical engineering technology, and in particular to a highly interpretable method and system for predicting slope stability. Background Technology
[0002] Slope stability prediction is a crucial topic in geotechnical engineering and geological disaster prevention. Traditional slope stability prediction methods mainly rely on physical model-based limit equilibrium methods and numerical simulations. These methods require detailed soil and rock parameters, are computationally complex, and are difficult to adapt to complex geological conditions. In recent years, with the development of artificial intelligence technology, machine learning models such as support vector machines, random forests, and neural networks have been widely applied to slope stability prediction. These data-driven methods can learn complex nonlinear relationships from historical data, improving the accuracy of predictions.
[0003] However, existing machine learning methods still have significant shortcomings in slope stability prediction. First, these models typically operate as "black boxes," with opaque decision-making processes that fail to provide engineers with predictive data they can understand and trust. In slope engineering, decisions often relate to the safety of life and property, making model interpretability crucial. Second, traditional machine learning models struggle to provide specific contributions of features to the prediction results, such as which geotechnical parameters are key factors influencing stability, how they affect stability, and the interactions between different features.
[0004] Interpretability is of particular importance in slope stability prediction. Engineers need to understand why a model makes a certain prediction in order to verify its rationality and formulate appropriate reinforcement or monitoring measures. For example, if the model can clearly indicate the specific impact of cohesion and internal friction angle on stability, engineers can design more effectively. Therefore, developing a slope stability prediction method that is both highly accurate and inherently interpretable is urgently needed. Summary of the Invention
[0005] To address the shortcomings of existing technologies, this invention provides a highly interpretable slope stability prediction method and system, which solves the problem that existing machine learning methods often operate in a "black box" manner and are difficult to provide interpretable prediction results.
[0006] According to an embodiment of the present invention, a highly interpretable slope stability prediction method includes: Obtain slope sample data, construct an EBM model, and optimize the hyperparameters of the EBM model using the BWO algorithm based on the slope sample data; The main effects and interaction effects of the EBM model are set to zero. Then, based on the slope sample data, the main effect function and interaction effect function of the EBM model are iteratively updated using the cyclic coordinate descent method. Obtain the current slope data and import it into the EBM model to obtain the current main effect and the current interaction effect. Then, calculate the current slope stability based on the current main effect and the current interaction effect.
[0007] Preferably, the method for optimizing the hyperparameters of the EBM model using the BWO algorithm based on slope sample data includes: S1: Initialize the population, reproduction rate, cannibalism rate and mutation rate. The population includes multiple parent individuals, and each parent individual is assigned random hyperparameters. S2: Calculate the fitness of each parent individual based on hyperparameters; S3: Based on the reproduction rate, crossbreed all individuals to obtain multiple offspring individuals and assign them corresponding fitness and hyperparameters. Then, sort all parent and offspring individuals in ascending order according to their fitness. S4: Calculate the number of individuals to be eliminated based on the cannibalism rate, and then remove the individuals ranked first that have the same number of individuals to be eliminated; S5: Determine whether each hyperparameter in each individual needs to be mutated based on the mutation rate, and change the value of each hyperparameter that needs to be mutated to another value; S6: Treat all individuals as parent individuals and repeat steps S2-S6 until the maximum number of iterations is reached. Then, take the hyperparameters corresponding to the individual with the highest fitness in the last iteration as the optimal hyperparameters and load the optimal hyperparameters into the EBM model.
[0008] Preferably, the method of crossbreeding all individuals according to the reproduction rate to obtain multiple offspring individuals and assigning corresponding fitness and hyperparameters includes: All parent individuals are sorted in descending order according to fitness. The number of breeding individuals N is calculated based on the number of parent individuals and the reproduction rate. The first N parent individuals are then arranged in pairs to obtain multiple breeding groups. For each breeding group, create a corresponding first and second offspring individuals, calculate the weighted hyperparameters of the two parent individuals, assign the weighted hyperparameters to the first offspring individual, and then calculate the corresponding fitness based on the weighted hyperparameters. Assign random selection parameters to the second offspring, and select one of the parent offspring based on the selection parameters. Assign the hyperparameters and fitness of the parent offspring to the second offspring.
[0009] Preferably, the slope sample data includes multiple feature vectors and corresponding sample labels. After setting the main effects and interaction effects of the EBM model to zero, the model bias is calculated based on the sample label corresponding to each sample feature.
[0010] Preferably, the method for iteratively updating the main effect function and interaction effect function of the EBM model using the cyclic coordinate descent method based on slope sample data includes: A1: For any feature vector in the slope sample data, select any feature in the feature vector, fix the main effect and interaction effect of other features, use the EBM model to calculate the first predicted value without the selected feature according to the model bias, and calculate the first pseudo residual based on the first predicted value. A2: Update all bins of the selected features based on the first pseudo residual, then generate new main effects based on all bins, then select other features and repeat A1-A2 until all features have been traversed; A3: Pair all features in the feature vector to obtain multiple interaction groups. Select one interaction group, fix the main effects and interaction effects of the other features, and use the EBM model to calculate the second predicted value without the two features in the interaction group according to the model bias. Calculate the second pseudo residual based on the second predicted value. A4: Update all two-dimensional bins between two features in the interaction group based on the second pseudo residual, then generate new interaction effects based on all two-dimensional bins, then select other interaction groups and repeat steps A3-A4 until all interaction groups have been traversed. A5: Select other feature vectors and repeat steps A1-A5 until the maximum number of iterations is reached.
[0011] Preferably, the calculation formula for the first pseudo-residual or the second pseudo-residual is as follows: in, Let the first or second pseudo-residual of the i-th eigenvector be the first pseudo-residual. Let i be the i-th eigenvector. Let be the sample label corresponding to the i-th feature vector.
[0012] Preferably, the formula for updating all bins of the selected feature is as follows: in, It's the learning rate. It fell into the box Number of samples within, It is the regularization parameter.
[0013] On the other hand, according to embodiments of the present invention, a highly interpretable slope stability prediction system is also provided, which uses the above-described highly interpretable slope stability prediction method, including: The data acquisition module is used to acquire slope sample data and current slope data; The parameter optimization module is used to construct the EBM model and optimize the hyperparameters of the EBM model using the BWO algorithm based on the slope sample data. The model optimization module is used to iteratively update the main effect function and interaction effect function of the EBM model based on the slope sample data and using the cyclic coordinate descent method. The stability prediction module is used to import the current slope data into the EBM model, obtain the current main effect and the current interaction effect, and then calculate the current slope stability based on the current main effect and the current interaction effect.
[0014] Compared with the prior art, the present invention has the following beneficial effects: This invention utilizes the Biometrics of the Model (BWO) to find the optimal hyperparameters of the EBM model. Then, using the EBM model with the optimal hyperparameters, in addition to outputting stability prediction results, it explains the influence between each feature and its contribution and impact on the stability prediction results by outputting the importance (main effect) of a single feature and the interaction effect between pairs of features, thus providing interpretable prediction results. Attached Figure Description
[0015] Figure 1 This is a diagram illustrating the slope stability prediction method according to an embodiment of the present invention.
[0016] Figure 2 This is a slope sample data diagram according to an embodiment of the present invention.
[0017] Figure 3 This is a graph showing the change in BWO fitness according to an embodiment of the present invention.
[0018] Figure 4 This is a confusion matrix diagram of the prediction results of the EBM model after loading the optimal hyperparameters according to an embodiment of the present invention.
[0019] Figure 5 This is a graph showing the prediction accuracy of the EBM model after loading the optimal hyperparameters according to an embodiment of the present invention.
[0020] Figure 6 This is a ranking diagram of the importance of features in the current slope data according to an embodiment of the present invention.
[0021] Figure 7 This is a graph showing the variation of the main effects of features in the current slope data of this invention.
[0022] Figure 8 This is a feature interaction effect diagram in the current slope data of this embodiment of the invention. Detailed Implementation
[0023] The technical solutions of the present invention will be further described below with reference to the accompanying drawings and embodiments.
[0024] like Figure 1 As shown in the figure, this invention proposes a highly interpretable slope stability prediction method, including: Obtain slope sample data, construct an EBM model, and optimize the hyperparameters of the EBM model using the BWO algorithm based on the slope sample data; Slope sample data were collected, including slope angle, slope height, cohesion, internal friction angle, unit weight, pore pressure ratio, and slope stability. The slope angle, slope height, cohesion, internal friction angle, unit weight, and pore pressure were combined into a feature vector, and the slope stability was defined as the sample label corresponding to the feature vector.
[0025] This invention collected a total of 462 samples, including 231 unstable samples and 231 stable samples, such as... Figure 2 As shown. Then, the above dataset is divided into training and test sets in a ratio of 8:2, and normalization is performed.
[0026] Then initialize the BWO algorithm parameters, such as the population size. Maximum number of iterations Reproduction rate cannibalism rate Variation rate and the EBM hyperparameter search space , interactions outer_bags learning_rate min_samples_leaf max_leaves early_stopping_rounds validation_size .
[0027] Population size This indicates that there are 15 parent individuals, and each parent individual contains a set of hyperparameters. , The values are: interactions, outer_bags, learning_rate, min_samples_leaf, max_leaves, early_stopping_rounds, and validation_size, where all values are random and fall within their respective search spaces.
[0028] Based on the initialization parameters mentioned above, and combined with the EBM model, the iterative loop begins, and the specific process is as follows: (1) Calculate the fitness of all parent individuals. The fitness function is calculated as follows: in, This refers to the hyperparameters of the i-th parent individual, i.e. , interaction, outer_bags, learning_rate, min_samples_leaf, max_leaves, early_stopping_rounds, validation_size; The number of cross-validations. It is the first The accuracy of the verification of the fold; Using hyperparameters The constructed EBM model.
[0029] (2) Reproduction stage First, select individuals for breeding: in, This indicates the order of fitness. In this invention, all parent individuals are sorted in descending order according to their fitness, and then the top-ranked parent individuals are selected. Each parent individual is used for reproduction.
[0030] The breeding operation includes two types: arithmetic crossover and discrete crossover. Since arithmetic crossover is essentially a weighted sum with different weights, the offspring of two parent individuals as "father and mother" and "mother and father" will not be the same. Therefore, it is necessary to arrange the parent individuals for the breeding operation in pairs to generate multiple breeding groups. For each breeding group, the following arithmetic crossover and discrete crossover operations are performed: Arithmetic cross: Discrete Crossover: In arithmetic cross calculation, This represents the value of the first offspring individual on the j-th parameter; This represents the value of the first parent individual on the j-th parameter; This represents the value of the second parent individual on the j-th parameter; It is a random number uniformly distributed in the range [0,1], that is... It determines the weight of the linear combination of the parameters of the two parent individuals. The hyperparameter of the first offspring individual is the weighted value of the hyperparameters of the two parent individuals. Then, based on its own hyperparameters, the fitness of the offspring individual is recalculated.
[0031] In discrete intersections, This represents the value of the second-generation individual on the j-th parameter; This represents the value of the first parent individual on the j-th parameter; This represents the value of the second parent individual on the j-th parameter. It is a random number uniformly distributed in the range [0,1], that is... ,according to Whether the value is less than 0.5 determines which parent's hyperparameters and corresponding fitness the second offspring inherits.
[0032] (3) Cannibalism The expression is as follows: in, This represents the total number of individuals after merging all parent and child individuals. The merged population The number of individuals that need to be eliminated is determined by the cannibalism rate. The result is obtained by multiplying by the combined population size and then rounding down. In this invention, all parent and offspring individuals are arranged in descending order of fitness, and then the top-ranked individuals are removed. individual.
[0033] The surviving population after elimination, that is, the worst-performing population removed from the merged population. Individual; To remove the previous After the individual, the i-th individual is sorted by fitness.
[0034] (4) The expression for the mutation stage is as follows: For each individual And for each parameter j, the probability of its mutation. After determining the parameters that need to be mutated, the mutated values are as follows: in, For the first The value of parameter j after mutation in each individual The search space range for hyperparameters.
[0035] Then, all parent and child individuals are treated as parent individuals, and steps (1)-(4) are repeated until the maximum number of iterations is reached. During the iteration process, the BWO iterative fitness curve is as follows: Figure 3 As shown, the hyperparameters corresponding to the individual with the highest fitness in the last iteration are taken as the optimal hyperparameters, and then the EBM model loads the optimal hyperparameters.
[0036] The EBM model with optimal hyperparameters was validated using a test set, and its prediction confusion matrix is shown below. Figure 4 As shown. The prediction accuracy, F1 score, balanced accuracy, AUC-ROC, and AUC-PR were 0.793, 0.793, 0.793, 0.885, and 0.905, respectively. Figure 5 As shown, the overall prediction accuracy of the model is considerable.
[0037] The main effects and interaction effects of the EBM model are set to zero. Then, based on the slope sample data, the main effect function and interaction effect function of the EBM model are iteratively updated using the cyclic coordinate descent method. Slope sample data ,in Let i be the i-th eigenvector. These are sample labels (in slope stability, 0 indicates instability and 1 indicates stability).
[0038] The original EBM model bins each continuous feature, discretizing it. The purpose of binning is to transform continuous features into discrete intervals to facilitate the learning of main effects. Let the first continuous feature be a bin. The number of bins for each feature is (Main effect binning) and (Interaction effect binning), binning methods typically use equal-frequency binning. For example, for features... Sort all values, determine quantile points, and divide the data into quantiles. Intervals, each interval contains approximately Each sample, after binning, has a feature value. Mapped to one of the bins , From the previous (Main features) (Interactive features) parameter settings.
[0039] The predicted value of EBM is expressed as the sum of main effects, including the bias term, main effects, and interaction effects, as shown in the following expression: in, It is a global model bias; It is a feature The main effect; It is a feature and The interaction effect; This represents the total number of input features.
[0040] In binary classification problems, the output value is the predicted probability, expressed as follows: in, It is the sigmoid activation function.
[0041] Main effect It is a piecewise constant function, composed of the score of each bin after binning. Let the characteristic be... The bin set is ,but: in, For boxing The score, For indicator functions, if If true, the value is 1; otherwise, the value is 0.
[0042] Similarly, interaction effect It is a two-dimensional piecewise constant function, consisting of the scores of binned combinations of two features: in, It is a combination of compartments. The score.
[0043] The EBM model was then trained using a cyclic coordinate descent method, updating feature by feature. The training process is as follows: (1) Initialization Initialize model bias , is the mean of the sample labels of all feature vectors (after logit transformation), i.e. , .
[0044] (2) Set the main effects and interaction effects of the EBM model to zero, i.e., all and It is 0.
[0045] (3) Update A1: For any feature vector in the slope sample data, select any feature in the feature vector, fix the main effect and interaction effect of other features, use the EBM model to calculate the first predicted value without the selected feature according to the model bias, and calculate the first pseudo residual based on the first predicted value. A2: Update all bins of the selected features based on the first pseudo residual, then generate new main effects based on all bins, then select other features and repeat A1-A2 until all features have been traversed; A3: Pair all features in the feature vector to obtain multiple interaction groups. Select one interaction group, fix the main effects and interaction effects of the other features, and use the EBM model to calculate the second predicted value without the two features in the interaction group according to the model bias. Calculate the second pseudo residual based on the second predicted value. A4: Update all 2D bins between the two features in the interaction group based on the second pseudo-residual, then generate new interaction effects based on all 2D bins, then select other interaction groups and repeat steps A3-A4 until all interaction groups have been traversed. A5: Select other feature vectors and repeat steps A1-A5 until the maximum number of iterations is reached.
[0046] The EBM model is updated in multiple iterations, with each main effect and interaction effect updated sequentially in each iteration. The logarithmic loss function is used during the iteration process, as shown in the following expression: in, .
[0047] During the iteration process, when updating a certain feature When determining the main effect, it is necessary to calculate the gradient of the loss function with respect to the main effect of that feature. In gradient boosting, pseudo-residuals (negative gradients) are generally used to fit the main effect. That is, for the feature vector... The loss function is related to The negative gradient is: Based on the gradients described above, the main effects and interaction effects of the EBM model can be updated.
[0048] Suppose we want to update the features in feature vector i. The main effects. At this point, temporarily fix the main effects and interaction effects of all other features, and then calculate the current EBM model excluding the main effects of the features. Under the main effect, the predicted value for each sample feature vector is expressed as follows: Specifically, the current model for samples Prediction (including features) The main effect is: Therefore, the first pseudo residual (negative gradient) can be obtained as follows: Then, the first pseudo residual can be used to update the features. The main effect. However, since the main effect is a piecewise constant, we only need to average the pseudo-residuals within each bin and then multiply by the learning rate to update the score of that bin. For features Boxes , as follows: in, It's the learning rate. It fell into the box Number of samples within, It is a regularization parameter to prevent overfitting.
[0049] Next, select other features and repeat the main effect update steps described above, using all features in feature vector i to update the main effects.
[0050] The update method for interaction effects is similar. For any two features in the feature vector, pair them up to generate multiple interaction groups. For one of these interaction groups... Fix all other items, and then calculate the predicted values excluding this interaction group: The current model prediction value is: Therefore, the second pseudo-residual (negative gradient) can be calculated using Equation 17, and then the interaction effect can be updated in each two-dimensional bin based on the second pseudo-residual. Score : in They fell into the boxes at the same time. and The number of samples.
[0051] Then select other interaction groups and repeat the interaction effect update steps above until all interaction groups have been traversed.
[0052] Then select other feature vectors and repeat the main effect and interaction effect update steps above until the maximum number of iterations is reached, or use an early stopping strategy, i.e., stop if the validation set loss no longer decreases.
[0053] Obtain the current slope data and import it into the EBM model to obtain the current main effect and the current interaction effect. Then, calculate the current slope stability based on the current main effect and the current interaction effect.
[0054] After the EBM model is trained, the current slope data is obtained, and for new samples... The possible classification (instability or stability) of the sample can be predicted using formulas 9 and 10.
[0055] In terms of interpretability analysis, the EBM model primarily utilizes main effects and interaction effects for internal interpretability analysis, mainly including feature importance, main effect plots, interaction effect plots, and local interpretability analysis for specific cases. Feature importance includes the importance of individual features (main effects) and the interaction effects between pairs of features, expressed as follows: Importance of main effects: Importance of interaction effects: Main effect plot: plotting each feature Main effect Follow The changes in can explain the specific contribution of this feature to the prediction.
[0056] Interaction effect heatmap: using heatmaps to illustrate With two features Analyze the changes in the values and their interaction effects.
[0057] Local interpretability analysis involves using importance ranking to interpret specific cases.
[0058] In terms of interpretability analysis, the feature importance of each input feature can be obtained based on the main effect, such as... Figure 6 As shown in the figure, cohesion, unit weight, and pore pressure ratio are highly important and have a significant impact on slope stability, while slope angle, slope height, and internal friction angle have a secondary impact. In addition to these features, the EBM model can also consider the interactive effects between features, such as the interaction between cohesion and slope height, and between cohesion and pore pressure ratio, which is a feature unmatched by other models. The main effect curves of each input feature are shown in the figure. Figure 7As shown, the contributions of cohesion, internal friction angle, and unit weight to the model increase with increasing values, indicating a higher probability of stable prediction. Conversely, pore pressure ratio, slope height, and slope angle show the opposite trend, consistent with the actual physical meaning of slope stability, demonstrating the high inherent interpretability of the EBM model. Furthermore, the model also satisfies the actual physical meaning of slope stability when considering characteristic interactions. Figure 8 As shown, when the cohesion and internal friction angle are sufficiently large, the slope remains stable even with a high slope angle, slope height, and pore pressure ratio.
[0059] On the other hand, embodiments of the present invention also provide a highly interpretable slope stability prediction system, which uses the above-described highly interpretable slope stability prediction method, including: The data acquisition module is used to acquire slope sample data and current slope data; The parameter optimization module is used to construct the EBM model and optimize the hyperparameters of the EBM model using the BWO algorithm based on the slope sample data. The model optimization module is used to iteratively update the main effect function and interaction effect function of the EBM model based on the slope sample data and using the cyclic coordinate descent method. The stability prediction module is used to import the current slope data into the EBM model, obtain the current main effect and the current interaction effect, and then calculate the current slope stability based on the current main effect and the current interaction effect.
[0060] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.
Claims
1. A highly interpretable slope stability prediction method, characterized in that: include: Obtain slope sample data, construct an EBM model, and optimize the hyperparameters of the EBM model using the BWO algorithm based on the slope sample data; The main effects and interaction effects of the EBM model are set to zero. Then, based on the slope sample data, the main effect function and interaction effect function of the EBM model are iteratively updated using the cyclic coordinate descent method. Obtain the current slope data and import it into the EBM model to obtain the current main effect and the current interaction effect. Then, calculate the current slope stability based on the current main effect and the current interaction effect.
2. The highly interpretable slope stability prediction method as described in claim 1, characterized in that: Based on slope sample data, methods for optimizing the hyperparameters of the EBM model using the BWO algorithm include: S1: Initialize the population, reproduction rate, cannibalism rate and mutation rate. The population includes multiple parent individuals, and each parent individual is assigned random hyperparameters. S2: Calculate the fitness of each parent individual based on hyperparameters; S3: Based on the reproduction rate, crossbreed all individuals to obtain multiple offspring individuals and assign them corresponding fitness and hyperparameters. Then, sort all parent and offspring individuals in ascending order according to their fitness. S4: Calculate the number of individuals to be eliminated based on the cannibalism rate, and then remove the individuals ranked first that have the same number of individuals to be eliminated; S5: Determine whether each hyperparameter in each individual needs to be mutated based on the mutation rate, and change the value of each hyperparameter that needs to be mutated to another value; S6: Treat all individuals as parent individuals and repeat steps S2-S6 until the maximum number of iterations is reached. Then, take the hyperparameters corresponding to the individual with the highest fitness in the last iteration as the optimal hyperparameters and load the optimal hyperparameters into the EBM model.
3. The highly interpretable slope stability prediction method as described in claim 2, characterized in that: Methods for crossbreeding all individuals based on reproduction rate to obtain multiple offspring and assigning corresponding fitness and hyperparameters include: All parent individuals are sorted in descending order according to fitness. The number of breeding individuals N is calculated based on the number of parent individuals and the reproduction rate. The first N parent individuals are then arranged in pairs to obtain multiple breeding groups. For each breeding group, create a corresponding first and second offspring individuals, calculate the weighted hyperparameters of the two parent individuals, assign the weighted hyperparameters to the first offspring individual, and then calculate the corresponding fitness based on the weighted hyperparameters. Assign random selection parameters to the second offspring, and select one of the parent offspring based on the selection parameters. Assign the hyperparameters and fitness of the parent offspring to the second offspring.
4. The highly interpretable slope stability prediction method as described in claim 1, characterized in that: The slope sample data includes multiple feature vectors and corresponding sample labels. After setting the main effects and interaction effects of the EBM model to zero, the model bias is calculated based on the sample label corresponding to each sample feature.
5. The highly interpretable slope stability prediction method as described in claim 4, characterized in that: Based on slope sample data, the methods for iteratively updating the main effect function and interaction effect function of the EBM model using the cyclic coordinate descent method include: A1: For any feature vector in the slope sample data, select any feature in the feature vector, fix the main effect and interaction effect of other features, use the EBM model to calculate the first predicted value without the selected feature according to the model bias, and calculate the first pseudo residual based on the first predicted value. A2: Update all bins of the selected features based on the first pseudo residual, then generate new main effects based on all bins, then select other features and repeat A1-A2 until all features have been traversed; A3: Pair all features in the feature vector to obtain multiple interaction groups. Select one interaction group, fix the main effects and interaction effects of the other features, and use the EBM model to calculate the second predicted value without the two features in the interaction group according to the model bias. Calculate the second pseudo residual based on the second predicted value. A4: Update all two-dimensional bins between two features in the interaction group based on the second pseudo residual, then generate new interaction effects based on all two-dimensional bins, then select other interaction groups and repeat steps A3-A4 until all interaction groups have been traversed. A5: Select other feature vectors and repeat steps A1-A5 until the maximum number of iterations is reached.
6. The highly interpretable slope stability prediction method as described in claim 5, characterized in that: The formulas for calculating the first or second pseudo-residual are as follows: in, For the first or second pseudo-residual of the i-th eigenvector, Let i be the i-th eigenvector. Let be the sample label corresponding to the i-th feature vector.
7. The highly interpretable slope stability prediction method as described in claim 5, characterized in that: The formula for updating all bins of the selected feature is as follows: in, It's the learning rate. It fell into the box Number of samples within, It is the regularization parameter.
8. A highly interpretable slope stability prediction system, characterized in that: The system uses a highly interpretable slope stability prediction method as described in any one of claims 1-7, comprising: The data acquisition module is used to acquire slope sample data and current slope data; The parameter optimization module is used to construct the EBM model and optimize the hyperparameters of the EBM model using the BWO algorithm based on the slope sample data. The model optimization module is used to iteratively update the main effect function and interaction effect function of the EBM model based on the slope sample data and using the cyclic coordinate descent method. The stability prediction module is used to import the current slope data into the EBM model, obtain the current main effect and the current interaction effect, and then calculate the current slope stability based on the current main effect and the current interaction effect.