Rock burst prediction method based on machine learning model

Abstract

The invention relates to the technical field of tunnel rockburst prediction, in particular to a rockburst prediction method based on a machine learning model, and the method comprises the following steps: S1, collecting rockburst case data, and constructing a database; s2, box diagram analysis is carried out on each type of data, abnormal values in the data are removed, and original rockburst case data are obtained; s3, a Borderline-SMOTE algorithm is adopted to carry out dynamic screening and synthesis on a minority class of samples in the original rockburst case data, and a more balanced data set is generated; s4, performing standardization processing on the rockburst data set, and dividing the rockburst data set into a training set and a test set through stratified sampling; s5, training a plurality of machine learning models by using the divided training set to obtain an optimal training model, and performing parameter tuning by using the divided test set to obtain an optimal hyper-parameter; and S6, based on the optimal hyper-parameter, establishing a rockburst grade prediction model, selecting an optimal model result, determining a rockburst grade, integrating prediction results of a plurality of rockburst grade prediction models, and outputting the rockburst grade.

Description

Technical Field

[0001] The present invention relates to the technical field of tunnel rockburst prediction, and particularly relates to a rockburst prediction method based on a machine learning model. Background Art

[0002] With the continuous expansion of the scale of underground engineering such as mine exploitation and tunnels, the problem of rockburst has become increasingly prominent. Especially in the engineering environment with a large burial depth, rockburst has become one of the major threats to the safety of on-site personnel and equipment. Rockburst prediction is one of the most effective means to reduce the harm of rockburst at present.

[0003] Based on summarizing the characteristics of a large number of rockburst cases, some prediction criteria have been proposed for the problem of rockburst grade classification prediction, the main conditions for the occurrence of rockburst have been summarized, and many methods and theories have been adopted, including: fuzzy comprehensive evaluation method, numerical analysis method, artificial neural network method, support vector machine method, etc. However, for the small sample situation of rockburst grade classification, the instability of a single classifier and the complexity of classifier parameter selection are challenges to the generalization ability of the model.

[0004] With the increase of burial depth, the high in-situ stress environment has brought some rock mechanics problems to underground engineering, and rockburst is one of them. The violent or sudden rockburst poses a huge threat to the safety of on-site personnel and equipment. Rockburst is generally divided into 4 grades according to its intensity (no rockburst, slight rockburst, moderate rockburst, and strong rockburst), and the prediction methods include empirical methods, numerical simulation, statistical and intelligent classification methods, etc.; with the increasing development of big data and artificial intelligence, machine learning intelligent algorithms such as ensemble learning are widely used in rockburst prediction.

[0005] However, the existing rockburst prediction models do not effectively handle outliers and class imbalance, directly use the original data for training, the model is sensitive to noise, and only relies on a single machine learning model, which is prone to overfitting in the small sample scenario, and the accuracy of the test set fluctuates greatly, resulting in poor model stability and low accuracy of rockburst prediction results. Summary of the Invention

[0006] The purpose of the present invention is to overcome the defects and deficiencies of the prior art, and provide a rockburst prediction method based on a machine learning model, which solves the problem of data imbalance through the Borderline - SMOTE algorithm, and then integrates with multiple machine learning models to improve the accuracy of rockburst prediction results.

[0007] The purpose of the present invention can be realized by the following technical solutions:

[0008] A rockburst prediction method based on a machine learning model, comprising the following steps:

[0009] S1. Collect rockburst case data and construct a database including burial depth, maximum shear stress, uniaxial compressive strength, uniaxial tensile strength, ratio of maximum shear stress to uniaxial compressive strength, brittleness coefficient, and elastic energy index;

[0010] S2. Perform box plot analysis on each type of data collected in step S1, remove outliers, and obtain original rock burst case data;

[0011] S3, using the Borderline-SMOTE algorithm to dynamically screen and synthesize the minority class samples in the original rockburst case data obtained in step S2 to generate a more balanced data set, thereby obtaining a processed rockburst data set;

[0012] S4, standardizing the rockburst dataset obtained in step S3 and dividing it into a training set and a test set by stratified sampling;

[0013] S5. Use the training set obtained in step S4 to train multiple machine learning models to obtain the optimal training model, and use the test set obtained in step S4 to perform parameter tuning to obtain the optimal hyperparameters;

[0014] S6. Establish a rockburst grade prediction model based on the optimal hyperparameters obtained in step S5, select the optimal model result by majority voting, determine the rockburst grade, integrate the prediction results of multiple rockburst grade prediction models, and output the rockburst grade;

[0015] Rockburst levels are divided into four levels: no rockburst, slight rockburst, moderate rockburst, and severe rockburst;

[0016] In step S3, each minority class sample is classified into noise data, dangerous data or safe data according to the proportion of other classes in its nearest neighbors, and then sample synthesis is performed based on the dangerous data to synthesize rockburst-free samples and slight rockburst samples.

[0017] As a preference, in said step S3, the following steps are included:

[0018] S31. Obtain a rockburst dataset T, where the minority class strong rockburst sample set is P = {p1, p2, ..., p a}, the majority class medium rockburst sample set is N = {n1, n2, ..., n b}, a is the number of strong rockburst samples, b is the number of moderate rockburst samples;

[0019] S32, for each strong rock burst sample p i , calculate its k nearest neighbors from the rockburst dataset T and count the number of medium rockburst samples m;

[0020] S33, according to the value of m, p iClassified as noise data, dangerous data, or safe data:

[0021] i) When m = k, determine p i as noise data;

[0022] ii) When k / 2 ≤ m < k, determine p i as dangerous data and add it to the dangerous data set;

[0023] iii) When 0 ≤ m < k / 2, determine p i as safe data and retain it;

[0024] S34. According to each sample q in the dangerous data set i synthesize strong rockburst samples. The formula for synthesizing strong rockburst samples is:

[0025] q new = q i + r j * × (q j - q i );

[0026] Among them, q new is a strong rockburst sample, p i is a strong rockburst sample, r j * is a random number between 0 and 1, and q j is a random neighboring sample of q i ;

[0027] S35. Repeat steps S31 - S34 to synthesize non - rockburst samples and minor rockburst samples to generate a balanced data set.

[0028] As an optimization, in step S33, eliminate noise data, where the noise data is a sample whose all nearest - neighbor samples are of other classes.

[0029] As an optimization, in step S2, eliminate outliers outside the upper and lower boundaries of the box plot. The lower boundary of the box plot is Q1 - 1.5×IQR, and the upper boundary is Q3 - 1.5×IQR, where Q1 is the value at 1 / 4 of the entire data set, Q3 is the value at 3 / 4 of the entire data set, and IQR = Q3 - Q1.

[0030] As an optimization, in step S4, the formula for normalizing the rockburst data set is:

[0031]

[0032] In the formula, μ is the mean of all sample feature indicators, σ is the standard deviation of all sample feature indicators, and y is a sample of the data set.

[0033] As an optimization, in the step S4, the ratio of the training set to the test set is 7:3.

[0034] As an optimization, in the step S5, 11 machine learning models in the machine learning library of Python are used to train the training set sorted in the step S4 to obtain 11 rock burst level prediction models; the machine learning models are SVM, KNN, DT, MLP, RF, XGBoost, GB, ET, AdaBoost, CatBoost, and TPOT respectively.

[0035] As an optimization, in the step S5, 5-fold grid search cross-validation is adopted to independently tune the parameters of each rock burst level prediction model to obtain the optimal hyperparameters when the algorithm accuracy rate is the highest.

[0036] As an optimization, in the step S6, by comparing the generalization performance of 11 rock burst level prediction models, three optimal rock burst level prediction models are determined and the tendency of rock burst is comprehensively judged, and the three rock burst level prediction models with the highest accuracy rate are selected. If the results of two of the rock burst level prediction models are the same, they are adopted; otherwise, the result with the most votes is selected.

[0037] As an optimization, in the step S6, 11 rock burst level prediction models are evaluated by precision, recall rate, and F1 value. The formulas for calculating precision, recall rate, and F1 value are as follows:

[0038] P = TP / (TP + FP);

[0039] R = TP / (TP + FN);

[0040] F1 = 2P·R / (P + R);

[0041] In the formula, P is the precision rate, TP is the number of positive samples predicted as positive by the model, TN is the number of negative samples predicted as negative by the model, FP is the number of negative samples predicted as positive by the model, FN is the number of positive samples predicted as negative by the model, R is the recall rate, and the F1 value is the harmonic mean of the precision rate and the recall rate.

[0042] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0043] 1. The present invention constructs a database through 7 key rockburst influence parameters such as burial depth, maximum shear stress, uniaxial compressive / tensile strength, etc., and constructs a comprehensive and quantifiable rockburst characteristic system by using multi-dimensional indexes, overcoming the prediction deviation caused by the traditional method relying on a single index; by adopting the Borderline-SMOTE algorithm to dynamically synthesize minority class samples, it significantly alleviates the problem of class imbalance in rockburst data (such as the scarcity of strong rockburst samples), makes the data balanced, enables more sufficient model training, and avoids ignoring small sample classes; by integrating the results of multiple machine learning models through the majority voting method, it solves the problem of poor generalization ability of a single model in small sample scenarios and improves the prediction stability.

[0044] 2. The traditional SMOTE algorithm generates all minority class samples without discrimination, which is easy to introduce noise at the classification boundary. The present invention determines noise data by the number of other class samples (m = k) among the k nearest neighbors and eliminates the noise data to avoid interference of wrong samples on model training and improve data quality; the present invention synthesizes only boundary dangerous samples (k / 2 ≤ m < k) to generate minority class samples close to the real distribution, avoiding the boundary blur problem caused by the blind generation of the traditional SMOTE.

[0045] 3. The present invention clearly eliminates noise data where all nearest neighbors are of other classes, reduces the risk of model overfitting, and improves the prediction reliability.

[0046] 4. The present invention objectively identifies outliers through the statistical rule of Q1 - 1.5×IQR to Q3 - 1.5×IQR, avoids subjective experience judgment errors, and ensures data reliability.

[0047] 5. The present invention adopts a standardization formula to eliminate the dimension difference, accelerates the model convergence speed, and improves the training efficiency.

[0048] 6. The present invention uses a 7:3 ratio of the training set to the test set, which balances the model training and verification requirements, avoids overfitting or underfitting, and is especially suitable for small sample scenarios.

[0049] 7. The present invention integrates 11 machine learning models, covering different types of models such as support vector machines, tree models, and neural networks, synthesizes the advantages of each algorithm, and improves the prediction robustness; through 5-fold grid search cross-validation, it ensures the optimal selection of hyperparameters, avoids local optimal solutions, and improves the model accuracy.

[0050] 8. The present invention uses the majority voting method ("adopt if two are consistent" rule) to reduce the misjudgment risk of a single model, and the result is more credible; by using multiple indicators such as precision (P), recall (R), and F1 value (harmonic mean) for evaluation, it comprehensively reflects the model's recognition ability for each category, especially paying attention to the performance of small sample categories (such as strong rockburst).

[0051] 9. The present invention improves the prediction accuracy of rockburst grades by balancing data, integrating models, and optimizing parameters; by using the Borderline-SMOTE algorithm and the stratified sampling strategy, the model can still maintain high stability when the sample size is ≤ 300; by adopting a standardized process (data cleaning → equalization → model training → voting decision) to adapt to the actual engineering requirements, complex on-site monitoring equipment is not required, and the application cost is reduced. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 is a flowchart of the rockburst prediction method based on a machine learning model of the present invention;

[0053] Figure 2 is a buried depth box plot of the rockburst prediction method based on a machine learning model of the present invention;

[0054] Figure 3 is a maximum shear stress box plot of the rockburst prediction method based on a machine learning model of the present invention;

[0055] Figure 4 is a uniaxial compressive strength box plot of the rockburst prediction method based on a machine learning model of the present invention;

[0056] Figure 5 is a uniaxial tensile strength box plot of the rockburst prediction method based on a machine learning model of the present invention;

[0057] Figure 6 is a ratio of maximum tangential stress to uniaxial compressive strength box plot of the rockburst prediction method based on a machine learning model of the present invention;

[0058] Figure 7 is a brittleness coefficient 1 box plot of the rockburst prediction method based on a machine learning model of the present invention;

[0059] Figure 8 is a brittleness coefficient 2 box plot of the rockburst prediction method based on a machine learning model of the present invention;

[0060] Figure 9 is an elastic energy index box plot of the rockburst prediction method based on a machine learning model of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0061] The present invention will be further described in detail below in conjunction with the embodiments and the drawings, but the embodiments of the present invention are not limited thereto.

[0062] A rockburst prediction method based on a machine learning model includes the following steps:

[0063] S1. Collect rockburst case data and construct a database including burial depth, maximum shear stress, uniaxial compressive strength, uniaxial tensile strength, ratio of maximum shear stress to uniaxial compressive strength, brittleness coefficient, and elastic energy index;

[0064] In step S1 of this embodiment, 222 sets of rockburst case data were collected. The characteristic parameters involved include: burial depth (D), maximum shear stress (MTS), uniaxial compressive strength (UCS), uniaxial tensile strength (UTS), ratio of maximum tangential stress to uniaxial compressive strength (SCF), brittleness coefficient (ratio of uniaxial compressive strength to uniaxial tensile strength B1 and ratio of the difference between the uniaxial compressive and tensile strengths to the sum of the two B2), and elastic energy index (Wet). The specific data values are shown in Table 1 below:

[0065]

[0066]

[0067]

[0068]

[0069]

[0070]

[0071]

[0072] Table 1

[0073] S2. Perform box plot analysis on each type of data collected in step S1, remove outliers, and obtain original rock burst case data;

[0074] In step S2 of this embodiment, the rockburst level is divided into four levels: no rockburst (Level I), slight rockburst (Level II), moderate rockburst (Level III), and severe rockburst (Level IV). The above data are classified according to the rockburst level, and a box plot is drawn for the collected data, where the value at 1 / 4 of the entire data set is Q1, and the value at 3 / 4 of the entire data set is Q3. Therefore, the height of the box plot is the interquartile range IQR = Q3-Q1, the lower boundary of the box plot is Q1-1.5×IQR, and the upper boundary is Q3-1.5×IQR. By eliminating outliers outside the upper and lower boundaries, the 222 data in Table 1 can be reduced to 201.

[0075] S3, using the Borderline-SMOTE algorithm to dynamically screen and synthesize the minority class samples in the original rockburst case data obtained in step S2 to generate a more balanced data set, thereby obtaining a processed rockburst data set;

[0076] Among them, when classifying according to the actual rockburst level, there are more rockburst level data in mild rockburst (Level II) and moderate rockburst (Level III), while there are fewer rockburst case data in no rockburst (Level I) and intense rockburst (Level IV). In order to improve the prediction performance of the prediction model as much as possible and obtain relevant data laws from the obtained rockburst data, the Borderline-SMOTE method is used to balance the rockburst case data in the training set, so that the number of rockburst cases in no rockburst (Level I) and intense rockburst (Level IV) can reach 80, respectively, to achieve the average distribution of various rockburst data.

[0077] In step S3, according to the proportion of other classes in the nearest neighbors of each minority class sample, it is classified as noise data, dangerous data or safe data, and then sample synthesis is performed according to the dangerous data to synthesize no rockburst samples and mild rockburst samples.

[0078] In step S3 of this embodiment, the following steps are included:

[0079] S31. Obtain the rockburst data set T, where the minority class intense rockburst sample set is P = {p1, p2,..., p a}, the majority class moderate rockburst sample set is N = {n1, n2,..., n b}, a is the number of intense rockburst samples, and b is the number of moderate rockburst samples;

[0080] S32. For each intense rockburst sample p i , calculate its k nearest neighbors from the rockburst data set T, and count the number m of moderate rockburst samples among them;

[0081] S33. Classify p i as noise data, dangerous data or safe data according to the m value:

[0082] i) When m = k, determine that p i is noise data;

[0083] ii) When k / 2 ≤ m < k, determine that p i is dangerous data and add it to the dangerous data set;

[0084] iii) When 0 ≤ m < k / 2, determine that p i is safe data and retain it;

[0085] As a preferred implementation manner, the noise data is removed in step S33, and the noise data is a sample in which all the nearest neighbor samples are of other classes.

[0086] S34. According to each sample q in the dangerous data set iSynthesize a strong rockburst sample, and the formula for synthesizing a strong rockburst sample is:

[0087] q new =q i +r j * ×(q j -q i );

[0088] Among them, q new is the strong rockburst sample, p i is the strong rockburst sample, r j * is a random number between 0 and 1, and q j is the random neighboring sample of q i ;

[0089] S35. Repeat steps S31 - S34 to synthesize non-rockburst samples and slightly rockburst samples to generate a balanced dataset.

[0090] S4. Standardize the rockburst dataset obtained in step S3 and divide it into a training set and a test set by stratified sampling;

[0091] In step S4 of this embodiment, the formula for standardizing the rockburst dataset is:

[0092]

[0093] In the formula, μ is the mean of all sample feature indicators, σ is the standard deviation of all sample feature indicators, and y is the dataset sample.

[0094] Use a Python program for stratified sampling to sample the standardized dataset.

[0095] Ensure that the ratio of the number of the training set to the number of the test set of the machine learning model is 7:3.

[0096] S5. Use the training set divided in step S4 to train multiple machine learning models to obtain an optimal training model, and use the test set divided in step S4 to optimize the parameters to obtain optimal hyperparameters;

[0097] In step S5 of this embodiment, use 11 machine learning models in the machine learning library of Python to train the training set divided in step S4. Specifically, use 5-fold grid search cross-validation to independently train each rockburst level prediction model to obtain 11 rockburst level prediction models. By comparing the accuracies under different parameter ranges, obtain the optimal hyperparameters of the 11 algorithms when the accuracy is relatively high, and obtain the optimal hyperparameters when the algorithm accuracy is the highest.

[0098] The 11 machine learning models are SVM, KNN, DT, MLP, RF, XGBoost, GB, ET, AdaBoost, CatBoost, and TPOT respectively.

[0099] S6. Establish a rockburst level prediction model based on the optimal hyperparameters obtained in step S5, select the optimal model result by the majority voting method to determine the rockburst level, integrate the prediction results of multiple rockburst level prediction models, and output the rockburst level;

[0100] In step S6 of this embodiment, by comparing the generalization performance of 11 rockburst level prediction models and the evaluation indexes of the model generalization ability of the 11 established rockburst level prediction models, namely accuracy, precision, recall, F1 value, macro-averaging, and micro-averaging indexes, three optimal classifiers are determined.

[0101] In step S6, 11 rockburst level prediction models are evaluated by precision, recall, and F1 value. The formulas for calculating precision, recall, and F1 value are as follows:

[0102] P = TP / (TP + FP);

[0103] R = TP / (TP + FN);

[0104] F1 = 2P·R / (P + R);

[0105] In the formula, P is precision, TP is the number of positive samples predicted as positive by the model, TN is the number of negative samples predicted as negative by the model, FP is the number of negative samples predicted as positive by the model, FN is the number of positive samples predicted as negative by the model, R is recall, and the F1 value is the harmonic mean of precision and recall.

[0106] Determine three optimal rockburst level prediction models and comprehensively judge the tendency of rockburst. Select the three rockburst level prediction models with the highest accuracy. If the results of two of the rockburst level prediction models are the same, adopt them; otherwise, select the result with the most votes. When two of the results of the three rockburst level prediction models with the best performance are the same, the classification result of rockburst can be determined. When the results of the three rockburst level prediction models are all different, compare the overall classification, and the result with the most votes is the rockburst prediction type.

[0107] A rockburst prediction method based on a machine learning model provided by the present invention effectively solves the problem that the instability of a single classifier and the complexity of hyperparameters in the case of small samples affect the accuracy of rockburst prediction. This method is applicable to known sample characteristic parameters of four or more, can directly learn and train the data related to the site, and directly realize rockburst prediction without investing a large number of devices for on-site monitoring. It is convenient to be popularized and applied in engineering, can effectively improve the rockburst prediction efficiency, and has strong practicability. It provides effective support for the accuracy of rockburst prediction and provides a certain reference basis for effectively preventing rockburst disasters in engineering.

[0108] Generally speaking, based on the characteristic data of rockburst cases, the present invention uses 11 machine learning methods to establish a rockburst prediction model. To achieve a higher-precision rockburst prediction result. It effectively solves the problem that the instability of a single classifier and the complexity of hyperparameters in the case of small samples affect the accuracy of rockburst prediction. This method is applicable to known sample characteristic parameters of four or more, such as buried depth (D), maximum shear stress (MTS), uniaxial compressive strength (UCS), uniaxial tensile strength (UTS), ratio of maximum tangential stress to uniaxial compressive strength (SCF), brittleness coefficient (ratio of uniaxial compressive strength to uniaxial tensile strength B1 and ratio of the difference between uniaxial compressive and tensile strengths to the sum of the two B2), and ignores the influence of the rock type in the sample. This method can directly learn and train the data related to the site and directly realize rockburst prediction without investing a large number of devices for on-site monitoring. It is convenient to be popularized and applied in engineering, can effectively improve the rockburst prediction efficiency, and has strong practicability. It provides effective support for the accuracy of rockburst prediction and provides a certain reference basis for effectively preventing rockburst disasters in engineering.

[0109] The above-described embodiments merely represent several implementation manners of the present invention, and their descriptions are relatively specific and detailed, but should not be construed as limiting the scope of the present invention patent. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present invention, several modifications and improvements can still be made, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the present invention patent shall be subject to the appended claims.

Claims

1. A rockburst prediction method based on a machine learning model, characterized in that, The following steps are involved: S1. Collect rockburst case data and construct a database including burial depth, maximum shear stress, uniaxial compressive strength, uniaxial tensile strength, ratio of maximum shear stress to uniaxial compressive strength, brittleness coefficient, and elastic energy index; S2. Perform box plot analysis on each type of data collected in step S1, remove outliers, and obtain original rock burst case data; S3, using the Borderline-SMOTE algorithm to dynamically screen and synthesize the minority class samples in the original rockburst case data obtained in step S2 to generate a more balanced data set, thereby obtaining a processed rockburst data set; S4, standardizing the rockburst dataset obtained in step S3 and dividing it into a training set and a test set by stratified sampling; S5. Use the training set obtained in step S4 to train multiple machine learning models to obtain the optimal training model, and use the test set obtained in step S4 to perform parameter tuning to obtain the optimal hyperparameters; S6. Establish a rockburst grade prediction model based on the optimal hyperparameters obtained in step S5, select the optimal model result by majority voting, determine the rockburst grade, integrate the prediction results of multiple rockburst grade prediction models, and output the rockburst grade; Rockburst levels are divided into four levels: no rockburst, slight rockburst, moderate rockburst, and severe rockburst; In step S3, each minority class sample is classified into noise data, dangerous data or safe data according to the proportion of other classes in its nearest neighbors, and then sample synthesis is performed based on the dangerous data to synthesize rockburst-free samples and slight rockburst samples.

2. The rock burst prediction method based on a machine learning model according to claim 1, wherein: In the step S3, the following steps are included: S31. Obtain the rockburst dataset T, where the minority-class strong rockburst sample set is P = {p1, p2, …, p a}, the majority-class moderate rockburst sample set is N = {n1, n2, …, n b}, a is the number of strong rockburst samples, and b is the number of moderate rockburst samples; S32. For each strong rockburst sample p i , calculate its k nearest neighbors from the rockburst dataset T, and count the number m of medium rockburst samples among them; S33. Classify p according to the value of m i into noise data, dangerous data or safe data: i) When m = k, determine that p i is noise data; ii) When k / 2 ≤ m < k, determine that p i is dangerous data and add it to the dangerous data set; iii) When 0 ≤ m < k / 2, determine that p i is secure data and retain it; S34. According to each sample q in the dangerous dataset i synthesize a strong rockburst sample, and the formula for synthesizing a strong rockburst sample is: q new = q i + r j * × (q j - q i ); Among them, q new is a strong rockburst sample, p i is a strong rockburst sample, r j * is a random number between 0 and 1, q j is the random nearest neighbor sample of q i ; S35 , repeating steps S31 - S34 to synthesize the samples without rockburst and the samples with slight rockburst to generate a balanced data set.

3. A rockburst prediction method based on a machine learning model according to claim 2, characterized in that: In step S33, noise data is removed, where all the nearest neighbor samples are samples of other categories.

4. A rockburst prediction method based on a machine learning model according to claim 1, characterized in that: In step S2, outliers outside the upper and lower boundaries of the box plot are removed. The lower boundary of the box plot is Q1-1.5×IQR, and the upper boundary is Q3-1.5×IQR, where Q1 is the value relative to 1 / 4 of the entire data set, Q3 is the value relative to 3 / 4 of the entire data set, and IQR=Q3-Q1.

5. The rockburst prediction method based on a machine learning model according to claim 1, characterized in that: In step S4, the formula for normalizing the rockburst dataset is: Where μ is the mean of all sample feature indicators, σ is the standard deviation of all sample feature indicators, and y is the sample of the data set.

6. A rockburst prediction method based on a machine learning model according to claim 1, characterized in that: In step S4, the ratio of the training set to the test set is 7:

3.

7. A rockburst prediction method based on a machine learning model according to claim 1, characterized in that: In step S5, the training set divided in step S4 is trained using 11 machine learning models in the Python machine learning library to obtain 11 rockburst grade prediction models; the machine learning models are SVM, KNN, DT, MLP, RF, XGBoost, GB, ET, AdaBoost, CatBoost and TPOT.

8. A rockburst prediction method based on a machine learning model according to claim 7, characterized in that: In step S5, 5-fold grid search cross validation is used to independently tune the parameters of each rockburst grade prediction model to obtain the optimal hyperparameters when the algorithm has the highest accuracy.

9. A rockburst prediction method based on a machine learning model according to claim 8, characterized in that: In step S6, by comparing the generalization performance of 11 rockburst level prediction models, three optimal rockburst level prediction models are determined and the tendency of rockburst is comprehensively judged. The three rockburst level prediction models with the highest accuracy are selected. If the results of two of the rockburst level prediction models are the same, they are adopted; otherwise, the result with the most votes is selected.

10. A rockburst prediction method based on a machine learning model according to claim 9, characterized in that: In step S6, 11 rockburst level prediction models are evaluated by precision, recall rate, and F1 value. The formulas for calculating precision, recall rate, and F1 value are as follows: P = TP / (TP + FP); R = TP / (TP + FN); F1 = 2P·R / (P + R); In the formula, P is the precision, TP is the number of positive samples predicted as positive by the model, TN is the number of negative samples predicted as negative by the model, FP is the number of negative samples predicted as positive by the model, FN is the number of positive samples predicted as negative by the model, R is the recall rate, and the F1 value is the harmonic mean of the precision and the recall rate.

Images