Rock burst grade prediction method based on high-order feature decoupling and multi-scale integrated learning

Abstract

The invention discloses a rockburst grade prediction method based on high-order feature decoupling and multi-scale integrated learning, and belongs to the technical field of geotechnical engineering. The method comprises the following steps: acquiring original features, and performing multi-dimensional feature engineering on the original features to generate an extended feature set containing high-order information; a plurality of data scalers such as StandScaler, Robust Scaler and MinMaxScaler are adopted to carry out parallel processing on the extended feature set, and a plurality of scaled data sets with different scales are obtained; respectively training corresponding base learners by using the data sets; and after executing the same processing flow on a to-be-tested sample, inputting the to-be-tested sample into a plurality of trained base learners, and outputting a final prediction level by adopting a hard voting integration strategy. According to the method, through deep feature mining and multi-scale integrated learning, the accuracy and robustness of the prediction model are effectively improved, and meanwhile, the transparency and credibility of the model decision process are enhanced through quantitative interpretation.

Description

Technical Field

[0001] This invention relates to the field of geotechnical engineering technology, specifically to a rockburst severity prediction method based on high-order feature decoupling and multi-scale integrated learning. Background Technology

[0002] Rockbursts are a common geological hazard in deep underground engineering, posing a serious threat to the safety of personnel and equipment. Therefore, accurate and reliable prediction of rockburst levels is of great practical significance for disaster prevention and safe construction.

[0003] Currently, rockburst prediction has gradually evolved from traditional empirical formulas to intelligent methods based on machine learning. However, existing prediction methods still face challenges in practical applications. Rockbursts are the result of the nonlinear coupling of multiple geological and engineering factors. Directly using raw monitoring parameters as model inputs often fails to fully express the complex internal mechanisms influencing rockburst occurrence, thus limiting the predictive power of the models.

[0004] Furthermore, existing methods typically employ a single data preprocessing workflow and prediction model. The inherent noise and outliers in geological exploration data can lead to unstable performance of such single models. When faced with new and unknown engineering geological conditions, the model's generalization ability and robustness of prediction results are insufficient, making it difficult to meet the reliability requirements of engineering practice.

[0005] Meanwhile, many high-performing machine learning models have opaque internal decision-making logic, often considered "black boxes." While these models can provide predictions, they cannot clearly explain which key factors led to the predictions, nor can they quantify the degree of influence of each factor. This lack of interpretability severely hinders engineers' trust in and adoption of the predictions, and also prevents the models from being truly integrated into engineering decision-making and risk management processes. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a rockburst severity prediction method based on high-order feature decoupling and multi-scale ensemble learning. This method solves the problems of insufficient prediction accuracy, poor model robustness, and lack of interpretability in the prediction process when dealing with the high-dimensional, highly nonlinear, and uncertain characteristics of geological data.

[0007] To achieve the above objectives, the present invention provides the following technical solution: The first aspect of this invention provides a method for predicting rockburst severity based on high-order feature decoupling and multi-scale ensemble learning, the method comprising: First, obtain the original features containing multiple geotechnical engineering parameters.

[0008] Subsequently, multidimensional feature engineering is performed on the original features to generate an expanded feature set. This step aims to decouple and construct higher-order derived features that reflect underlying physical laws from the finite original features. This process may include one or more of the following transformations: Trigonometric function features are used to capture periodic or oscillatory patterns in data; The exponential reciprocal feature is constructed to characterize the nonlinear decay or growth effect of the parameter; Polynomial power transformation is used to increase the feature dimension to fit higher-order nonlinear relationships; Higher-order interactive feature construction is used to explicitly express the coupling effect between multiple original features; Ratio feature construction is used to generate dimensionless parameters with explicit physical meaning; Statistical feature construction is used to extract statistics on the distribution of a set of features.

[0009] The generated derived features are merged with the original features to form the extended feature set.

[0010] Next, multiple data scalers are used to process the expanded feature set in parallel, resulting in multiple scaled datasets at different scales. Different scalers, based on different statistical assumptions and transformation methods, can change the numerical range and data distribution of the features, thereby providing multi-perspective input for subsequent models. Specifically, the multiple data scalers may include: StandardScaler converts feature data into a standard normal distribution with a mean of 0 and a variance of 1, and is suitable for data that conforms to a Gaussian distribution. RobustScaler scales data using quantiles (such as interquartile range), providing good robustness to outliers in the data. MinMaxScaler linearly scales feature data to a given interval while preserving the original data distribution. This yields the datasets... Dataset and dataset .

[0011] Then, using the multiple scaled datasets at different scales and rockburst level labels, corresponding base learners are trained respectively, resulting in multiple trained base learners. For example, the XGBoost model can be selected as the base learner, and the dataset... Dataset and dataset Training is performed on the base learner to obtain the trained base learner. ), ( )and( ).

[0012] In the prediction phase, the original features of the sample to be tested are obtained, and the aforementioned multidimensional feature engineering and multiple data scaler processing are sequentially performed on the sample to be tested to obtain sample data at multiple scales corresponding to the training phase.

[0013] The processed test sample data at different scales are input into their respective trained base learners to obtain multiple independent prediction results.

[0014] Finally, a hard-vote integration strategy is used to decide the multiple independent prediction results, that is, the prediction level that appears most frequently is selected as the final prediction level and output.

[0015] Furthermore, to improve the interpretability of the model, this method also includes a step of performing SHAP (Shapley Additive ex Planations) analysis on the trained base learners: Applying the SHAP analysis framework, based on the trained base learners and the corresponding scaled dataset, the SHAP value for each feature of all samples in the dataset is calculated. This process, based on Shapley values ​​in cooperative game theory, precisely attributes the model's predicted output to each input feature. For the prediction of a single sample ii, the model can be interpreted as an additive feature attribution model: in, It is a base learner For the sample The original predicted score (e.g., logit value) for a specific rockburst severity category. Is this model on the reference dataset? The average predicted score (baseline value) on the [databaseline]. It is the first The contribution of each feature to the prediction.

[0016] Based on the calculated SHAP value, the following analysis can be performed: Calculating global feature importance: This quantifies the average impact of each feature on the overall model output by calculating the average absolute value of its SHAP values ​​across all samples. Features global importance The calculation is as follows: in: The total number of samples in the reference dataset; It is the absolute value of the SHAP value. The larger the value, the more significant the th . A feature (e.g., a certain ratio feature) or original features ) for the model The greater the impact on the overall forecast results.

[0017] Generate local interpretability analysis: For a single sample to be interpreted, extract and visualize its corresponding local SHAP value ( This is used to reveal how the prediction result of a sample is determined by the combined effect of its various features. Based on this, an example diagram of SHAP local interpretability can be output.

[0018] Analyze feature interactions: Calculate the SHAP value of the interaction between features ( ), and aggregate them to quantify features. and characteristics The synergistic or antagonistic effects between them can be analyzed. Based on this, an example diagram of SHAP feature interactions can be output.

[0019] A second aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of any of the aforementioned methods.

[0020] This invention provides a rockburst severity prediction method based on high-order feature decoupling and multi-scale ensemble learning. It has the following beneficial effects: 1. This invention utilizes multidimensional feature engineering to transform and combine original geotechnical engineering parameters into various derived features such as high-order interactive features and ratio features. This method effectively reveals the implicit nonlinear relationships and physical laws in the original data, providing more informative input for subsequent models, enhancing the expressive power of features, and thus improving the accuracy of rockburst level prediction.

[0021] 2. This invention employs multiple data scalers to process feature data in parallel, combined with an ensemble learning strategy. By training different base learners on datasets at different scales, complementary advantages are formed among the models. The final hard-voting ensemble method reduces the overall model's sensitivity to outliers or specific distribution patterns, resulting in more stable prediction results and improved robustness and generalization ability of the model.

[0022] 3. This invention introduces the SHAP analysis framework, enabling global and local quantitative interpretation of the prediction model. This method not only identifies key characteristic factors influencing rockburst severity but also clearly demonstrates the predictive attribution of individual samples, making the model's internal decision-making process transparent. This interpretability enhances the credibility of the technical solution in engineering applications, providing a reliable basis for on-site decision-making. Attached Figure Description

[0023] Figure 1 This is an overall flowchart of the method of the present invention; Figure 2 This is a data distribution and correlation analysis diagram for the present invention; Figure 3 This is the confusion matrix diagram of the present invention; Figure 4 This is a graph showing the total contribution of SHAP features in this invention. Figure 5 This is a diagram illustrating the importance of SHAP features in this invention. Figure 6 This is a detailed block diagram of the multidimensional feature engineering of the present invention; Figure 7 This is a schematic diagram of the training architecture for multi-scale ensemble learning according to an embodiment of the present invention. Figure 8 This is a flowchart illustrating the application of rockburst level prediction in an embodiment of the present invention. Detailed Implementation

[0024] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] Example: In step S101, the initial feature acquisition and label definition are performed first. This step prepares the basic dataset for subsequent feature engineering and model training.

[0026] In one embodiment, the step specifically includes: S101-1: Obtaining the Rockburst Case Dataset This dataset From multiple samples Composition, each sample is represented as a data pair ,in The total number of samples.

[0027] For the first The original feature vector of each sample is in the form of: This original feature vector contains four key indicators used to characterize the mechanical properties and stress state of the rock mass: Maximum shear stress (MTS) characterizes the maximum value of shear stress experienced by a rock mass at a specific point.

[0028] Uniaxial compressive strength (UCS) is the ultimate stress that a rock material can withstand in resisting uniaxial compressive failure.

[0029] Tensile strength (UTS) is the ultimate stress that a rock material can withstand to resist tensile failure.

[0030] The elastic energy index (WET) characterizes the ability of a rock to store and release energy during the elastic deformation stage.

[0031] For the first The rockburst level label corresponding to each sample is the prediction target of the model. Belongs to a predefined set of discrete categories ,in This represents the total number of rockburst severity categories. For example, this set... It can represent different levels of rockburst intensity: {no rockburst, slight rockburst, moderate rockburst, severe rockburst}.

[0032] By performing step S101-1, a structured dataset containing the original input features and corresponding target levels is obtained, providing a data foundation for subsequent data cleaning, feature construction, and model training.

[0033] Step S101 also includes outlier detection and cleaning of the raw feature data obtained in step S101-1. The purpose of this step is to identify and remove statistical outliers in the dataset that deviate from the population, so as to ensure the stability and reliability of subsequent model training.

[0034] In one embodiment, the outlier detection and cleaning employs the Z-score method, which is based on the assumption that the data follows a normal distribution and identifies outliers by calculating the standard deviation of data points from the mean. Specifically, it may include: S101-2: For the dataset obtained in step S101-1 (Include (samples), for One original feature (in this embodiment) Each feature in ) ( ), calculate its in all The mean of each sample and standard deviation .

[0035] Calculate the mean The formula is: ; Calculate the standard deviation The formula is: ; in: The total number of samples; For the first in the dataset The first sample The values ​​of each original feature.

[0036] S101-3: The mean calculated based on step S101-2 and standard deviation Traverse the dataset Each sample Each feature Calculate its corresponding Z-score value. : ; The value reflects the data point with its characteristic mean Distance (in standard deviation) (unit)

[0037] S101-4: Setting the anomaly detection threshold In a preferred embodiment, statistical methods are employed. The principle (or the Raida criterion) is to set a threshold. .

[0038] S101-5: For the first data in the dataset Sample Perform anomaly detection. Check all samples of this type of sample. The Z-score values ​​of each feature, if at least one feature... Meet the conditions (Right now ), then the sample Marked as an anomalous sample.

[0039] S101-6: From the original dataset Remove all samples that were marked as abnormal in steps S101-5.

[0040] After performing this step, you will obtain a cleaned dataset that does not contain statistical outliers. This dataset will serve as the input for feature engineering in step S102.

[0041] In step S102, the cleaned dataset output in step S101 is processed. Perform multi-dimensional feature engineering. This step is to achieve decoupling of higher-order features and to deeply explore the nonlinear relationship between the original physical quantities and the rockburst level.

[0042] Multidimensional feature engineering includes the construction of trigonometric function features. This process aims to enable the model to capture periodic patterns or nonlinear response modes that may exist in the original features through nonlinear periodic transformations. These modes are difficult to directly represent using the original linear features.

[0043] In one embodiment, the steps for constructing trigonometric function features may include: S102-1: For the cleaned dataset In One original feature ( In this embodiment ), that is, maximum tangential stress Uniaxial compressive strength ,tensile strength and elastic energy index .

[0044] S102-2: To Each of the original features Sine and cosine transforms are applied respectively to generate two new sets of derived features, namely sinusoidal features. Sum and cosine characteristics .

[0045] Sine characteristics The calculation method is as follows: ; Cosine characteristics The calculation method is as follows: ; in: For the first The values ​​of each original feature.

[0046] By executing step S102-2, for 1 original feature, co-generated These features are derived from trigonometric functions. They are incorporated into a feature set for processing in subsequent step S103.

[0047] The multidimensional feature engineering in step S102 also includes the construction of exponential and reciprocal features. This process aims to introduce exponential and inverse proportional transformations to capture the nonlinear contribution of the original features to the rockburst grade in different numerical ranges.

[0048] In one embodiment, the construction step may include: S102-3: Targeting One original feature That is, maximum tangential stress Uniaxial compressive strength ,tensile strength and elastic energy index Each feature Constructing exponential features and reciprocal features .

[0049] Exponential characteristics The calculation method is as follows: ; in: For the first The values ​​of the original features; This represents the natural exponential function.

[0050] Reciprocal characteristics The calculation method is as follows: ; in: For the first The values ​​of the original features; (epsilon) is a very small positive number (e.g., 10) pre-defined to ensure numerical stability. -6 ).

[0051] This setting is to avoid when When the observed value is equal to or close to zero, the calculation of the reciprocal can lead to a denominator of zero or numerical overflow.

[0052] By executing step S102-3, for 1 original feature, co-generated Each exponential and inverse derivative feature is incorporated into a feature set for processing in subsequent step S103.

[0053] The multidimensional feature engineering in step S102 also includes the construction of polynomial features and power transform features. This process aims to capture more complex nonlinear mapping relationships between the original physical quantities and rockburst levels, such as changes in the growth rate or responses approaching saturation, by introducing higher-order polynomials and nonlinear power functions (such as square roots and logarithms).

[0054] In one embodiment, the construction step may include: S102-4: Targeting One original feature That is, maximum tangential stress Uniaxial compressive strength ,tensile strength and elastic energy index Each feature Construct the characteristics of quadratic, cubic, square root, and logarithmic transformations.

[0055] Secondary features The calculation method is as follows: ; Tertiary features The calculation method is as follows: ; Square root characteristics The calculation method is as follows: in, This indicates taking the absolute value, used to ensure that for any The validity of square root calculation (including potentially negative stress values).

[0056] Logarithmic characteristics The calculation method is as follows: ; in: For the first The values ​​of the original features; Represents the natural logarithm function; Indicates taking the absolute value; (epsilon) is a very small positive number (e.g., 10) pre-defined to ensure numerical stability. -6 ), used to avoid when The problem of zero or negative input to the logarithmic function when the observed value is equal to or close to zero.

[0057] By executing step S102-4, for 1 original feature, co-generated These features are derived from polynomials and power transforms. These features are incorporated into a feature set for processing in subsequent step S103.

[0058] The multidimensional feature engineering in step S102 also includes the construction of high-order interactive features. This process aims to explicitly characterize the synergistic effects or composite relationships between different physical quantities by combining and operating on the original features, in order to decouple their comprehensive impact on rockburst severity.

[0059] In one embodiment, the construction step may include: S102-5: Targeting One original feature That is, maximum tangential stress Uniaxial compressive strength ,tensile strength and elastic energy index Select at least two different original features and combine them to construct the product term, the summation term, and the mean term.

[0060] In a preferred implementation, second-order interaction features are first constructed. For any two distinct original features... and ( Construct the following features respectively: Product interaction features : ; This feature is used to capture the multiplicative effect when two physical quantities act together.

[0061] Addition and interaction features : ; This feature is used to characterize the linear superposition effect when two physical quantities act together.

[0062] Mean interaction features : ; This feature is a linear transformation of the additive interaction feature, providing the average contribution of the two physical quantities.

[0063] in, and They represent the first The and the first The values ​​of each original feature.

[0064] To further support the construction of "higher-order" interactions, this process can be extended to combinations of ternary or higher-order features. For example, selecting three different original features. (in Constructing third-order product interaction features : ; By executing steps S102-5, the generated high-order interactive features are incorporated into the feature set for processing in subsequent step S103.

[0065] The multidimensional feature engineering in step S102 also includes the construction of ratio features. This process aims to introduce established rockburst criteria with clear physical meaning in the field of geotechnical engineering (such as stress-strength ratio, energy-strength ratio) as prior knowledge into the model in the form of feature ratios to explicitly characterize the relative relationships between different physical quantities.

[0066] In one embodiment, the construction step may include: S102-6: Targeting One original feature Select two original features and ( ), construct ratio features .

[0067] Ratio characteristics The general calculation method is as follows: ; in: For the first One original feature (as a molecule); For the first One original feature (as the denominator); (epsilon) is a very small positive number (e.g., 10) preset to ensure numerical stability. -6 ). This setting is to avoid when the denominator When the observed value is equal to or close to zero, the calculation of the ratio causes a numerical overflow problem.

[0068] In a preferred implementation, the focus is on constructing ratio characteristics based on classical rockburst criteria. For example: Structural stress-strength ratio characteristics: Selecting the maximum tangential stress (MTS) as a molecule, uniaxial compressive strength As the denominator, construct features : ; Construction of energy-strength ratio characteristics: selection of elastic energy index (WET) as a molecule, uniaxial compressive strength As the denominator, construct features : ; In addition, this step may also include constructing ratio combinations between other original features (e.g. / or / This allows for a comprehensive exploration of the relative relationships between features.

[0069] The multidimensional feature engineering in step S102 also includes the construction of statistical features. This process aims to characterize the overall numerical level, dispersion, or distribution range of the sample across all original dimensions by calculating aggregate statistics (such as mean, standard deviation, and extreme values) among multiple original features within the sample, providing the model with a macroscopic perspective that goes beyond a single feature.

[0070] Due to the rockburst case data in this embodiment ( The data is static cross-sectional data, not time series data, and the statistical features are specific to each sample. Across its The calculation is performed using the original features.

[0071] In one embodiment, the construction step may include: S102-7: For the cleaned dataset Each sample Based on its One original feature That is, maximum tangential stress Uniaxial compressive strength ,tensile strength and elastic energy index Calculate the following statistical characteristics of this sample: In-sample mean characteristics : ; This feature reflects the sample The average level of each original feature.

[0072] In-sample standard deviation characteristics : ; This feature reflects the sample The degree of dispersion among the original features.

[0073] In-sample maximum feature : ; In-sample minimum feature : ; in: For the current sample The values ​​of each original feature ; The total number of original features (in this embodiment) ); For this sample The arithmetic mean of the original features; and These are functions that take the maximum and minimum values, respectively.

[0074] By executing steps S102-7, the generated statistical features are incorporated into the feature set. At this point, step S102 (multidimensional feature engineering) is complete, forming an extended feature set containing the original features and all derived features (trigonometric functions, exponential reciprocals, polynomial power transformations, higher-order interactions, ratios, and statistics), for processing in subsequent step S103.

[0075] In step S103, the material constructed in step S102, containing... An extended feature set of features (including original features and all derived features) The data is normalized using at least three different data scalers to generate multi-scale training data. This step aims to provide inputs with different data distribution characteristics for the subsequent multi-scale ensemble learning in step S104.

[0076] In one embodiment, the first of the at least three different data scalers is a StandardScaler. This scaler performs zero-mean unit variance standardization, the purpose of which is to... Each feature in ( The linear transformation yields new features with a mean of 0 and a standard deviation of 1.

[0077] This standardization process may specifically include: S103-1: Targeting (This feature set contains) (number of training samples), calculate Each of the features exist The mean of the training samples and standard deviation .

[0078] mean The calculation method is as follows: ; Standard deviation The calculation method is as follows: ; in: The cleaned dataset output from step S101 The total number of samples in the sample; for The Middle The first sample The values ​​of each feature.

[0079] S103-2: The mean calculated based on step S103-1 and standard deviation traversal Each sample Each feature Calculate its standardized eigenvalues : ; in: These are the original eigenvalues; and These are the features The mean and standard deviation are calculated on the training set; (epsilon) is a minimal positive number preset to ensure numerical stability. This setting is to avoid when When the calculated value is equal to or close to zero, a calculation error occurs where the denominator is zero.

[0080] S103-3: All standardized features of all samples Combined, the first scaled training dataset is formed. .

[0081] S103-4: Save the results calculated in step S103-1 Each mean and one standard deviation These parameters are used as model parameters for StandardScaler. These parameters will be used in step S105 to perform the same standardization transformation on the test data to ensure consistency in data scale between the training and prediction phases.

[0082] The second of the at least three different data scalers described in step S103 is RobustScaler. This scaler uses statistics (i.e., median and interquartile range) that are insensitive to outliers in the dataset for data standardization. This method aims to provide a robust scaler... A data scale view with relatively small impact from potential extreme values.

[0083] This standardization process may specifically include: S103-5: Targeting (This feature set contains) (number of training samples), calculate Each of the features ( )exist The median of the training samples and interquartile range .

[0084] The median Features exist The 50th percentile of the sample observations.

[0085] The interquartile range The calculation method is as follows: ; in: Features exist The 25th percentile (i.e., the first quartile) of the sample observations; Features exist The 75th percentile (i.e., the third quartile) of the sample observations.

[0086] for , , For the calculation of percentiles, those skilled in the art can use the standard percentile calculation method, which will not be elaborated here.

[0087] S103-6: Median calculated based on step S103-5 and interquartile range traversal Each sample Each feature Calculate its robustly scaled eigenvalues : ; in: These are the original eigenvalues; and These are the features The median and interquartile range were calculated on the training set; (epsilon) is a minimal positive number preset to ensure numerical stability.

[0088] S103-7: All robustly scaled features of all samples Combined, a second scaled training dataset is formed. .

[0089] S103-8: Save the results calculated in step S103-5 Median and Interquartile range ( These parameters are used as model parameters for RobustScaler. These parameters will be used in step S105 to perform the same robust scaling transformation on the data to be tested.

[0090] The third of the at least three different data scalers mentioned in step S103 is MinMaxScaler (minimum-maximum scaler). This scaler performs linear interval scaling, the purpose of which is to... Each feature in ( This method linearly transforms the data to a predefined fixed interval, such as [0,1]. It preserves the original data's distribution pattern while unifying all features to the same numerical range.

[0091] This standardization process may specifically include: S103-9: Targeting (This feature set contains) (number of training samples), calculate Each of the features exist Minimum value on training samples and maximum value .

[0092] Minimum value The calculation method is as follows: ; Maximum value The calculation method is as follows: ; in: for The Middle The first sample The values ​​of each feature; The total number of training samples; and These are functions that take the minimum and maximum values, respectively.

[0093] S103-10: The minimum value calculated based on step S103-9 and maximum value traversal Each sample Each feature Scale it to the [0,1] interval and calculate its scaled eigenvalues. : ; in: These are the original eigenvalues; and These are the features The minimum and maximum values ​​are calculated on the training set; (epsilon) is a minimal positive number preset to ensure numerical stability.

[0094] S103-11: All min-max scaled features of all samples Combined, a third scaled training dataset is formed. .

[0095] S103-12: Save the results calculated in step S103-9 Minimum value { }and The maximum value {}( These parameters are used as model parameters for MinMaxScaler. These parameters will be used in step S105 to perform the same linear interval scaling transformation on the data to be tested.

[0096] At this point, step S103 (multi-scale data scaling) is complete, and a total of [data] has been generated. , and Three training datasets with different scale distribution characteristics provide input for the multi-scale ensemble learning in step S104.

[0097] In step S104, the base learner is trained in multi-scale ensemble learning. This step utilizes the three datasets at different scales generated in step S103. , and and the cleaned dataset The corresponding rockburst level label Multiple independent base learners are trained separately.

[0098] In one embodiment, the base learner is an XGBoost (Extreme Gradient Boosting) model. XGBoost is an algorithm based on Gradient Boosting Decision Tree (GBDT), which minimizes the loss function by iteratively adding decision trees and is suitable for handling high-dimensional data and complex nonlinear relationships.

[0099] The training process for this base learner specifically includes: S104-1: Using the first scaled training dataset (Includes features) ) and tags The first base learner is obtained through training. .

[0100] S104-2: Training dataset using the second scaled dataset (Includes features) ) and tags The second base learner is obtained through training. .

[0101] S104-3: Training dataset using the third scaled dataset (Includes features) ) and tags The third base learner is obtained through training. .

[0102] The training processes for S104-1, S104-2, and S104-3 are independent of each other. Each base learner (with...) is trained independently. For example, the XGBoost model optimizes a regularized objective function. To build A decision tree.

[0103] In the In each iteration, the objective function of the model It can be represented as: ; in: For sample index, This represents the total number of training samples in the cleaned dataset output from step S101. For the first The true rockburst level label for each sample. For the front tree to the first The cumulative prediction score for each sample. for The Middle The feature vector of each sample. The first to be learned in this iteration A decision tree. Let be the loss function. Since rockburst severity prediction is a multi-class classification task, here... Multi-class log-loss, also known as cross-entropy loss, is selected.

[0104] For the first The regularization term for each tree is used to control model complexity and prevent overfitting. It is defined as: ; in: For the first The number of leaf nodes in a tree. For the first The weight (or score) of each leaf node. and The regularization coefficients for the number of leaf nodes and the L2 norm of the leaf node weights are preset hyperparameters.

[0105] For the XGBoost model in the th How to optimize using Taylor second-order expansion in round iteration Determine Structure and The specific calculations, and how to use methods such as cross-validation to... , , Optimizing hyperparameters such as the number of trees and the maximum depth of the trees is a well-known technique in this field and will not be elaborated upon here.

[0106] After executing step S104, a total of , and Three pre-trained base learners based on different data scales. These base learners will be used as input to step S105.

[0107] Step S104 (Multi-scale Ensemble Learning) also includes defining an ensemble strategy for combining the predictions of multiple base learners. This strategy aims to comprehensively utilize the predictive capabilities of individual base learners trained at different data scales to improve the overall robustness and accuracy of rockburst level prediction.

[0108] In one embodiment, the integration strategy employs a voting mechanism. This strategy is applied in step S105 (rockburst level prediction) to aggregate data from... , and Independent judgment of the three base learners.

[0109] In a preferred embodiment, the voting mechanism is specifically hard voting, also known as majority voting.

[0110] When predicting a test sample, the application process of this strategy includes: S104-4 Obtain the features of the sample to be tested after StandardScaler transformation in steps S103-1 to S103-4. Input it into the first base learner The first predicted category label is obtained. .

[0111] S104-5 Obtain the features of the sample to be tested after RobustScaler transformation in steps S103-5 to S103-8. Input it into the second base learner The second predicted category label is obtained. .

[0112] S104-6 Obtain the features of the sample to be tested after the MinMaxScaler transformation in steps S103-9 to S103-12. Input it into the third base learner The third predicted category label is obtained. .

[0113] S104-7 applies a hard voting strategy, applying the three independent predicted category labels obtained in S104-4 to S104-6. A vote will be taken to produce the final integrated prediction results. .

[0114] The calculation method for the hard voting strategy can be expressed as follows: ; in: The final rockburst level prediction label output by the multi-scale integrated model; The mode function returns the input set. The element that appears most frequently (i.e., the category label with the most votes); They are respectively The predicted class label output by the base learner.

[0115] S104-8 defines the tie-breaking rules. Since this embodiment uses three base learners, a tie occurs when the predicted rank labels by the three base learners are all different (e.g., the prediction results are {Rank I, Rank II, Rank III}). In this case, a preset rule is needed to determine the unique tie. .

[0116] In one implementation, a pre-specified base learner can be selected (e.g., based on...). Training Prediction results As the final ruling. In another implementation, the lowest (or the one with the highest index) of the three predicted labels can be selected according to a preset order of rockburst levels (e.g., level I is better than level II, level II is better than level III, etc.) as the final ruling. .

[0117] By executing S104-1 to S104-8, the construction of the multi-scale ensemble learning model was completed, including the training of three independent basis learners and the definition of a hard-voting ensemble strategy for combining their predictions.

[0118] In step S105, the complete model (including feature engineering, multi-scale scaler parameters, and trained base learners) built in the training phases of steps S101 to S104 is used to perform rockburst level prediction on the new test data.

[0119] The prediction process may specifically include: S105-1: Receive one or more samples to be tested Each test sample contains One original feature ( In this embodiment, the maximum tangential stress is... Uniaxial compressive strength ,tensile strength and elastic energy index .

[0120] S105-2: Sample to be tested In The original features are used to perform multi-dimensional feature engineering as defined in step S102 (i.e., the trigonometric function, exponential reciprocal, polynomial power transform, higher-order interaction, ratio, and statistical feature construction processes described in S102-1 to S102-7). This process generates features for the test sample that include... Extended feature set of each feature .

[0121] S105-3: Extended feature set of the test sample The three sets of scaler model parameters saved in step S103 are loaded to generate three different scales of the test features.

[0122] S105-4: Load the StandardScaler parameters (mean) saved in step S103-4 and standard deviation Then, by applying the transformation formula in step S103-2, the first set of scaled features to be measured is calculated. .

[0123] S105-5: Load the RobustScaler parameters (median) saved in step S103-8 and interquartile range Then, by applying the transformation formula in step S103-6, the second set of scaled features to be measured is calculated. .

[0124] S105-6: Load the MinMaxScaler parameters (minimum value {) saved in step S103-12 and maximum value Then, by applying the transformation formula in steps S103-10, the third set of scaled features to be measured is calculated. .

[0125] S105-7: Will Input the base learner obtained from step S104-1 training Get the first predicted category label .

[0126] S105-8: Will Input the base learner obtained from step S104-2 Get the second predicted category label .

[0127] S105-9: Will Input the base learner obtained from step S104-3 Obtain the third predicted category label .

[0128] S105-10: Apply the hard-voting integration strategy defined in step S104 (S104-7) to the three independent prediction labels obtained in S105-7 to S105-9. A vote will be taken.

[0129] S105-11: Calculation The mode (i.e., the category label that appears most frequently) is used. If a tie occurs (i.e., all three predicted labels are different), the tie handling rules defined in step S104-8 are used to determine the unique final predicted rockburst level. If there are no tied votes, then the mode is... .

[0130] S105-12: Output the final predicted rockburst level .

[0131] To achieve the prediction of rockburst severity using the rockburst level model trained in step S104 (e.g., base learner) This method analyzes the internal decision-making logic of [the system / mechanism], and also includes an interpretability analysis step based on SHAP (SHapley Additive ex Planations). This step aims to use Shapley values ​​from game theory to quantitatively evaluate [the system / mechanism]. One extended feature (from) Each feature in ) Contribution to the model's predicted output.

[0132] In one embodiment, the analysis process may include: S106-1: Select one or more base learners trained in step S104 (e.g. () is used as the object of analysis.

[0133] S106-2: Select a reference dataset, such as the first scaled training dataset generated in step S103-3. (Includes features) ).

[0134] S106-3: Apply the SHAP analysis framework (specifically, an optimized interpreter suitable for XGBoost models, such as TreeExplainer), to compute the results for each sample in the reference dataset. Each feature SHAP value .

[0135] S106-4: Based on the calculated SHAP value A global interpretability analysis is performed. This analysis is conducted by calculating... Each of the features Average absolute SHAP value To assess its overall importance: ; in: The total number of samples in the reference dataset; It is the absolute value of the SHAP value. The larger the value, the more significant the th . A feature (e.g., a certain ratio feature) or original features ) for the model The greater the impact on the overall forecast results.

[0136] S106-5: Perform local interpretability analysis. This analysis is performed on a single test sample from step S105 (or in the reference dataset). Perform. Its SHAP value. Satisfying additive attribution: ; in: It is a base learner For the sample The original predicted score (e.g., logit value) for a specific rockburst severity category. Is this model on the reference dataset? The average predicted score (baseline value) on the [databaseline]. It is the first The contribution of each feature to the prediction.

[0137] S106-6: Through S106-5, the sample can be analyzed intuitively. Which specific feature values ​​(e.g.) Take a high value, (Taking low values) respectively affect the prediction results This produces a positive (driving the prediction to that level) or negative (hindering the prediction to that level) contribution, thus aiding in understanding the final prediction results in S105-12. This provides evidence for one of the reasons for its formation.

[0138] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A rockburst severity prediction method based on high-order feature decoupling and multi-scale ensemble learning, characterized in that, Includes the following steps: Obtain the original features; perform multidimensional feature engineering on the original features to obtain an extended feature set; The extended feature set is processed using multiple data scalers to obtain multiple scaled datasets at different scales; The corresponding base learners are trained using the multiple scaled datasets at different scales to obtain multiple trained base learners. Obtain the original features of the sample to be tested, and perform the multidimensional feature engineering and the various data scaler processing on the sample to be tested; The processed test sample is input into the multiple trained base learners to obtain multiple prediction results; The multiple prediction results are integrated to obtain the final prediction level.

2. The rockburst severity prediction method based on high-order feature decoupling and multi-scale ensemble learning according to claim 1, characterized in that, The multidimensional feature engineering includes: performing at least one of trigonometric function feature construction, exponential reciprocal feature construction, polynomial power transformation, higher-order interactive feature construction, ratio feature construction, and statistical feature construction to generate derived features and combine them with the original features to form the extended feature set.

3. The rockburst severity prediction method based on high-order feature decoupling and multi-scale ensemble learning according to claim 1, characterized in that, The various data scalers include: StandardScaler, RobustScaler, and MinMaxScaler; The multiple scaled datasets at different scales include: datasets Dataset and dataset .

4. The rockburst severity prediction method based on high-order feature decoupling and multi-scale ensemble learning according to claim 1, characterized in that, The base learner is the XGBoost model.

5. The rockburst severity prediction method based on high-order feature decoupling and multi-scale ensemble learning according to claim 1, characterized in that, The integration of the multiple prediction results adopts a hard-voting integration strategy.

6. The rockburst severity prediction method based on high-order feature decoupling and multi-scale ensemble learning according to claim 1, characterized in that, The method further includes: Based on the trained base learner and the scaled datasets at different scales, the SHAP analysis framework is applied to calculate the SHAP value of all samples.

7. The rockburst severity prediction method based on high-order feature decoupling and multi-scale ensemble learning according to claim 6, characterized in that, The method further includes: calculating the average absolute value of the SHAP values ​​to obtain the global feature importance, and outputting an example graph of the global feature importance of SHAP.

8. The rockburst severity prediction method based on high-order feature decoupling and multi-scale ensemble learning according to claim 6, characterized in that, The method further includes: for a single sample to be interpreted, extracting the local SHAP value corresponding to the sample, and outputting an example diagram of SHAP local interpretability.

9. The rockburst severity prediction method based on high-order feature decoupling and multi-scale ensemble learning according to claim 6, characterized in that, The method further includes: Based on the SHAP value and the scaled datasets at different scales, calculate the feature interaction value; The feature interaction values ​​are aggregated, and an example graph of SHAP feature interactions is output.

10. 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 as described in any one of claims 1 to 9.

Images