Rockburst prediction method based on few-sample metric and ensemble learning
By using a method based on the small sample measurement and integrated learning, combined with the decision tree and a random search algorithm, the rock burst prediction model is optimized, and the problem of low accuracy of rock burst prediction in the existing technology is solved, achieving higher prediction accuracy and model stability.
Patent Information
- Application Number
- CN202210381289.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-04-12
- Publication Date
- 2024-11-22
- Estimated Expiration
- 2042-04-12
AI Technical Summary
The existing rock burst prediction method based on artificial intelligence is simple due to insufficient sample size, making it difficult to accurately characterize the complex nonlinear relationship between rock burst grade and its influencing factors. The prediction accuracy is low, and the hyperparameters of the optimization model can only reach the sample. Optimization under conditions cannot improve the upper limit of model accuracy.
Using a method based on the small sample measurement and integrated learning, the rock burst hierarchical data set is constructed by analyzing quantitative predictive indicators and qualitative standards, data cross-reorganization and label optimization are carried out, and hyperparameter tuning is performed in combination with decision trees and random search algorithms to improve rock Generalization ability of explosive prediction model.
It improves the accuracy and reliability of rock burst prediction, enhances the scientificity and stability of the model, can effectively predict rock burst levels in actual projects, and improves the prediction accuracy.
Smart Images

Figure CN114662405B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of geotechnical engineering, and in particular relates to a rockburst prediction method based on few-sample metric and ensemble learning. Background Art
[0002] In order to meet the urgent needs of modernization and industrialization for space and resources, water conservancy, mining, tunnel transportation and other engineering construction and resource development are rapidly moving deep into the earth. In such deep-earth projects, the rock structure is subject to high ground stress and prominent energy accumulation. When disturbed by excavation, earthquakes, etc., rock burst disasters are very likely to occur, manifested as surrounding rock bursting, peeling, and high-speed ejection of fragments, posing a great threat to the safety of on-site workers and equipment. More seriously, the intensity and frequency of rock bursts further increase with the increase in the depth of the project, and it has become the main disaster threatening the safety of deep rock mass projects in my country. How to achieve accurate prediction of rock bursts is the key to disaster prevention and mitigation of deep-earth projects, and it is also a bottleneck problem that urgently needs to be broken through in the field of rock burst research.
[0003] In order to predict the occurrence of rockburst, people have proposed a number of rockburst criteria based on experiments and theories, which are mainly divided into two categories: stress-strength criteria (such as Hoek criterion, Russense criterion, Barton criterion, etc.) and energy criteria (elastic energy index criterion, impact energy index criterion, etc.). The above criteria provide theories and methods for scientifically predicting rockburst disasters, but due to the different factors and indicators considered by each criterion, the evaluation results of different methods are often contradictory, the prediction accuracy of rockburst disasters is not ideal, and the rockburst accident rate remains high.
[0004] As a branch of computer science, artificial intelligence technology can intelligently mine and learn hidden, effective, and understandable information in existing data, thereby solving unknown problems and providing new solutions to complex nonlinear problems in a big data environment. In recent years, a large number of scholars have applied various machine learning methods to the field of rockburst prediction, such as neural networks, support vector machines, linear discriminant analysis, decision trees, K nearest neighbor methods, ensemble learning, and other methods, and have achieved a large number of meaningful results. In addition, in order to improve the accuracy of prediction, many scholars have used intelligent optimization algorithms, such as particle swarm algorithms, genetic algorithms, ant colony algorithms, etc., to optimize the hyperparameters in existing machine learning methods, or use the advantages of each algorithm to combine multiple methods to predict rockbursts, and have achieved good results.
[0005] However, the current AI-based rockburst prediction methods still have the following problems: (1) In theory, machine learning requires a large number of samples to train in order to achieve the best results. However, a large number of projects lack detailed rockburst records, and the total amount of rockburst case data is only about 200 groups. Limited by the number of samples, the trained model is simple and it is difficult to accurately characterize the complex nonlinear relationship between the rockburst level and its influencing factors, resulting in low prediction accuracy; (2) In order to improve the prediction accuracy, many scholars have optimized the model hyperparameter adjustment process and improved the generalization ability of the model, but this method is only an optimization within the model, so that the model reaches the optimal value under the sample conditions, and cannot fundamentally improve the upper limit of the model accuracy. Summary of the invention
[0006] The object of the present invention is to provide a rockburst prediction method based on few-sample metric and ensemble learning which is highly reliable, accurate, scientific and stable.
[0007] The rockburst prediction method based on few-sample metric and ensemble learning provided by the present invention comprises the following steps:
[0008] S1. Analyze and select quantitative prediction indicators that affect rockburst levels and qualitative standards for rockburst level classification;
[0009] S2. Collect and analyze existing rockburst engineering cases, obtain quantitative prediction indicators and rockburst grades for each case, and construct a rockburst classification data set;
[0010] S3. The rockburst classification data set obtained in step S2 is randomly divided into a training set and a test set according to the proportion, while ensuring that the proportion of cases of each rockburst grade is consistent;
[0011] S4. Cross-recombining each group of data in the training set obtained in step S3 with all the data in the training set in turn and re-calibrating the data labels to form a model training set;
[0012] S5. Using an ensemble learning method, the model training set obtained in step S4 is trained and hyperparameters are optimized to obtain an initial model for rockburst prediction;
[0013] S6. Cross-recombining each set of data in the test set obtained in step S3 with all the data in the training set in pairs and re-calibrating the data labels to form a model test set;
[0014] S7. Bring the model test set obtained in step S6 into the initial rockburst prediction model obtained in step S5 for testing, and optimize it according to the test results to obtain a rockburst prediction model;
[0015] S8. Use the rockburst prediction model obtained in step S7 to predict rockburst in actual projects.
[0016] The quantitative prediction index affecting the rockburst level described in step S1 specifically includes the following steps:
[0017] Select surrounding rock tangential stress σ θ , rock uniaxial compressive strengthσ c , rock uniaxial tensile strength σ t , the ratio of surrounding rock tangential stress to rock uniaxial compressive strength Ratio of uniaxial compressive strength to uniaxial tensile strength and the elastic deformation energy index W of rock et as a quantitative predictor.
[0018] The qualitative standard for rockburst grade classification described in step S1 specifically includes the following steps:
[0019] According to the existing national standards, rock burst levels are divided into four levels: high, medium, low and none.
[0020] The step S2 of obtaining the quantitative prediction index and rockburst grade of each case and constructing a rockburst classification data set specifically includes the following steps:
[0021] Get the quantitative prediction index of the ith case: surrounding rock tangential stress σ θi Using the letter a i Indicates that the uniaxial compressive strength of rock σ ci Using the letter b i Indicates that the uniaxial tensile strength of rock σ ti Using the letter c i It means that the ratio of surrounding rock tangential stress to rock uniaxial compressive strength is Using the letter d i It indicates the ratio of uniaxial compressive strength to uniaxial tensile strength. Using the letter e i Indicates that the elastic deformation energy index W of rock eti Using the letter f i Indicates that the rockburst grade is represented by the letter g i Indicates that the i-th case is then represented in the rockburst classification dataset as the i-th group of data [a i ,b i ,c i ,d i ,e i ,f i ,g i ]; where the rockburst grade value g i The values are 3, 2, 1 and 0, which correspond to high, medium, low and no rockburst levels respectively; i is a positive integer from 1 to n;
[0022] The quantitative prediction indicators and rockburst grades of all n cases are obtained, and finally a rockburst classification dataset is constructed.
[0023] Step S4 of cross-combining each set of data in the training set obtained in step S3 with all the data in the training set in turn and re-calibrating the data labels specifically includes the following steps:
[0024] The i-th group of data [a i ,b i ,c i ,d i ,e i ,f i ,g i ], and then with the n sets of data [a 1 ,b 1 ,c 1 ,d 1 ,e 1 ,f 1 ,g 1 ]~[a n ,b n ,c n ,d n ,e n ,f n ,g n ] to perform crossover recombination, and the n groups of data obtained after crossover recombination are:
[0025] [a i ,b i ,c i ,d i ,e i ,f i ,a 1 ,b 1 ,c 1 ,d 1 ,e 1 ,f 1 ,G i1 ]~[a i ,b i ,c i ,d i ,e i ,f i ,a n ,b n ,c n ,d n ,e n ,f n ,G in ]; the jth group of data is [a i ,b i ,c i ,d i ,ei ,f i ,a j ,b j ,c j ,d j ,e j ,f j ,G ij ], where G ij is the training data label after cross-combination and The value of j is a positive integer from 1 to n;
[0026] The n groups of data in the training set are cross-combined with all the data in the training set in pairs according to the above steps, and the data labels are recalibrated to finally construct the model training set.
[0027] The integrated learning method described in step S5 is used to train and optimize the hyperparameters of the model training set obtained in step S4 to obtain an initial model for rockburst prediction, which specifically includes the following steps:
[0028] Use decision tree as base learner;
[0029] The following hyperparameters are tuned using a random search algorithm:
[0030] Hyperparameters: data integration method; the search scope includes Bagging integration, AdaBoost integration, RUSBoost integration, GentleBoost integration and LogitBoost integration;
[0031] Hyperparameters: number of base learners; search range is [10,500];
[0032] Hyperparameters: learning rate; search range is [0.001, 1];
[0033] Hyperparameters: maximum number of splits; search range is [1, N]; where N is the amount of data in the model training set;
[0034] The K-fold cross-validation method is used to verify the model training process.
[0035] Step S6, in which each group of data in the test set obtained in step S3 is cross-combined with all the data in the training set in pairs and the data labels are recalibrated to form a model test set, specifically includes the following steps:
[0036] The qth group of data in the test set [a' q ,b' q ,c' q ,d' q ,e' q ,f′ q ,g' q], and then with the n sets of data [a 1 ,b 1 ,c 1 ,d 1 ,e 1 ,f 1 ,g 1 ]~[a n ,b n ,c n ,d n ,e n ,f n ,g n ] to perform crossover recombination, and the n groups of data obtained after crossover recombination are:
[0037] [a' q ,b' q ,c' q ,d' q ,e' q ,f′ q ,a 1 ,b 1 ,c 1 ,d 1 ,e 1 ,f 1 ,G' q1 ]~[a' q ,b' q ,c' q ,d' q ,e' q ,f′ q ,a n ,b n ,c n ,d n ,e n ,f n ,G' qn ]; the rth group of data is [a' q ,b' q ,c' q ,d' q ,e' q ,f′ q ,a r ,b r ,c r ,d r ,e r ,f r ,G' qr ], where G' qr is the test data label after cross-combination and The value of q is a positive integer from 1 to m; the value of r is a positive integer from 1 to n;
[0038] The m groups of data in the test set are cross-combined with all the data in the training set in pairs according to the above steps, and the data labels are recalibrated to finally construct the model test set.
[0039] Step S7 of bringing the model test set obtained in step S6 into the initial rockburst prediction model obtained in step S5 for testing specifically includes the following steps:
[0040] Substitute the model test set of step S6 into the initial rockburst prediction model obtained in step S5 for testing, and restore the "0-1" label to the rockburst grade result in a maximum probability attribution manner;
[0041] During the test, the primary rockburst prediction model outputs the following results for the input model test set: 1 ,b′ 1 ,c′ 1 ,d′ 1 ,e′ 1 ,f′ 1 ,a 1 ,b 1 ,c 1 ,d 1 ,e 1 ,f 1 ,G″′ 11 ]~[a' m ,b' m ,c' m ,d' m ,e' m ,f′ m ,a n ,b n ,c n ,d n ,e n ,f n ,G″′ mn ]; Rockburst prediction results G″′ for the primary model of rockburst prediction xy , x is a positive integer from 1 to m, and y is a positive integer from 1 to n;
[0042] Separate statistics g y = 0, g y =1, g y =2 and g y =3:G″′ xy = 0, and the statistics g y = 0, g y =1, g y =2 and g y =3:G″′ xy = the number of 1s;
[0043] The probability of belonging to each level is calculated using the following formula:
[0044]
[0045] Where I(G″′ xy =1|g y =i) means when g y =i when G″′ xy = the number of 1s; I(G″′ xy =0|g y =i) means when g y =i when G″′ xy = the number of 0; i takes values of 0, 1, 2 and 3; N is the number of data in the model training set; P i is the probability that the test set data belongs to each level; is the probability that the test set data does not belong to each level; R represents P i The rockburst level corresponding to the maximum value; express The rockburst level corresponding to the maximum value is used for reverse verification;
[0046] The level corresponding to the calculated maximum probability value is used as the rockburst prediction level output by the primary rockburst prediction model.
[0047] Step S8 uses the rockburst prediction model obtained in step S7 to predict rockburst in actual projects, specifically obtaining corresponding data in the actual project, processing it in the manner of step S6, and then entering the rockburst prediction model obtained in step S7 for prediction, and restoring the prediction result to obtain the final rockburst grade result, thereby completing the final rockburst prediction.
[0048] The rockburst prediction method based on few-sample metric and ensemble learning provided by the present invention performs cross-recombination data enhancement processing on the original data set of a small number of samples. The processing process does not add new information nor delete any information in the original data. The amount of data after enhancement is the square of the amount of original data. The increase in the amount of data improves the effectiveness of the machine learning training process. At the same time, the present invention adopts a data metric method and focuses on finding the similarities between data of the same rockburst level and the differences between data of different rockburst levels through data training. The types of training labels are small and the training process is simple. In addition, the method of the present invention has high reliability, high accuracy, and is scientific and stable. BRIEF DESCRIPTION OF THE DRAWINGS
[0049] Figure 1 It is a schematic diagram of the method flow of the present invention.
[0050] Figure 2 Schematic diagram of the minimum classification error curve of the new training set iterative process in an embodiment of the present invention.
[0051] Figure 3 Schematic diagram of the minimum classification error curve of the original training set iteration process of an embodiment of the present invention.
[0052] Figure 4 Schematic diagram of the new test set classification confusion matrix of an embodiment of the present invention.
[0053] Figure 5 Schematic diagram of the original test set classification confusion matrix of an embodiment of the present invention.
[0054] Figure 6 It is a schematic diagram of the accuracy obtained by two models of an embodiment of the present invention.
[0055] Figure 7 Schematic diagram of the confusion matrix of the final classification result of the test set of an embodiment of the present invention. DETAILED DESCRIPTION
[0056] like Figure 1 The figure shows a schematic flow chart of the method of the present invention: the rockburst prediction method based on few-sample metric and ensemble learning provided by the present invention comprises the following steps:
[0057] S1. Analyze and select quantitative prediction indicators that affect rockburst levels and qualitative standards for rockburst level classification;
[0058] In specific implementation, the quantitative prediction indicators affecting the rockburst level include:
[0059] Select surrounding rock tangential stress σ θ , rock uniaxial compressive strengthσ c , rock uniaxial tensile strength σ t , the ratio of surrounding rock tangential stress to rock uniaxial compressive strength Ratio of uniaxial compressive strength to uniaxial tensile strength and the elastic deformation energy index W of rock et As a quantitative predictor;
[0060] The qualitative standards for rockburst classification include:
[0061] According to the existing national standards, rockburst levels are divided into four levels: high, medium, low and none; the specific classification standards are shown in Table 1:
[0062] Table 1 Schematic diagram of qualitative classification of rockburst grades
[0063]
[0064] S2. Collect and analyze existing rockburst engineering cases, obtain quantitative prediction indicators and rockburst grades for each case, and construct a rockburst classification data set; specifically, the following steps are included:
[0065] Get the quantitative prediction index of the ith case: surrounding rock tangential stress σ θi Using the letter a i Indicates that the uniaxial compressive strength of rock σ ci Using the letter b i Indicates that the uniaxial tensile strength of rock σ ti Using the letter c i It means that the ratio of surrounding rock tangential stress to rock uniaxial compressive strength is Using the letter d i It indicates the ratio of uniaxial compressive strength to uniaxial tensile strength. Using the letter e i Indicates that the elastic deformation energy index W of rock eti Using the letter f i Indicates that the rockburst grade is represented by the letter g i Indicates that the i-th case is then represented in the rockburst classification dataset as the i-th group of data [a i ,b i ,c i ,d i ,e i ,f i ,g i ]; where the rockburst grade value g i The values are 3, 2, 1 and 0, which correspond to high, medium, low and no rockburst levels respectively; i is a positive integer from 1 to n;
[0066] Obtain the quantitative prediction indicators and rockburst grades of all n cases, and finally construct a rockburst classification dataset;
[0067] S3. The rockburst classification data set obtained in step S2 is randomly divided into a training set and a test set according to the proportion, while ensuring that the proportion of cases of each rockburst grade is consistent;
[0068] S4. Cross-recombining each set of data in the training set obtained in step S3 with all the data in the training set in turn and re-calibrating the data labels to form a model training set; specifically comprising the following steps:
[0069] The i-th group of data [a i ,b i ,c i ,d i ,e i ,f i ,g i ], and then with the n sets of data [a 1 ,b 1 ,c 1 ,d 1 ,e 1 ,f 1 ,g1 ]~[a n ,b n ,c n ,d n ,e n ,f n ,g n ] to perform crossover recombination, and the n groups of data obtained after crossover recombination are:
[0070] [a i ,b i ,c i ,d i ,e i ,f i ,a 1 ,b 1 ,c 1 ,d 1 ,e 1 ,f 1 ,G i1 ]~[a i ,b i ,c i ,d i ,e i ,f i ,a n ,b n ,c n ,d n ,e n ,f n ,G in ]; the jth group of data is [a i ,b i ,c i ,d i ,e i ,f i ,a j ,b j ,c j ,d j ,e j ,f j ,G ij ], where G ij is the training data label after cross-combination and The value of j is a positive integer from 1 to n;
[0071] The n groups of data in the training set are cross-combined with all the data in the training set in pairs according to the above steps and the data labels are recalibrated to finally construct the model training set;
[0072] After two-by-two crossover recombination, the number of features in the data set becomes twice as much as before, and the amount of data in the original training set is n, while the amount of data in the new training set becomes n. 2, the data labels are changed from rockburst level to “0” and “1”;
[0073] S5. Using an ensemble learning method, the model training set obtained in step S4 is trained and hyperparameters are optimized to obtain an initial model for rockburst prediction; specifically, the steps include:
[0074] Decision tree is used as the base learner because the decision tree algorithm has fast calculation speed, high accuracy, can process data sets containing various types, and does not require any domain knowledge and parameter assumptions;
[0075] The random search method is more efficient when tuning multiple hyperparameters at the same time and the amount of data is large. Therefore, the random search algorithm is used to tune the following hyperparameters:
[0076] Hyperparameter: data integration method; the search range includes Bagging integration, AdaBoost integration, RUSBoost integration, GentleBoost integration and LogitBoost integration; different integration methods have a great impact on the accuracy of the model, so the integration method is used as a hyperparameter for subsequent adjustment;
[0077] Hyperparameters: number of base learners; search range is [10,500];
[0078] Hyperparameters: learning rate; search range is [0.001, 1];
[0079] Hyperparameters: maximum number of splits; search range is [1, N]; where N is the amount of data in the model training set;
[0080] The specific hyperparameters are shown in Table 2:
[0081] Table 2 Hyperparameters
[0082]
[0083] The K-fold cross-validation method (preferably the 10-fold cross-validation method) is used to validate the model training process; that is, the training set is randomly divided into ten parts of substantially equal size. The first time, the first nine parts are used as training subsets, and the last part is used as a validation set. The training subset is used to train the model, and the validation set is used to evaluate the model. The second time, the ninth part is used as a validation set, and the remaining parts are used as training subsets. This process is repeated ten times until each part is selected as a validation set once.
[0084] S6. Cross-recombining each group of data in the test set obtained in step S3 with all the data in the training set in turn and re-calibrating the data labels to form a model test set; specifically comprising the following steps:
[0085] The qth group of data in the test set [a'q ,b' q ,c' q ,d' q ,e' q ,f q ',g' q ], and then with the n sets of data [a 1 ,b 1 ,c 1 ,d 1 ,e 1 ,f 1 ,g 1 ]~[a n ,b n ,c n ,d n ,e n ,f n ,g n ] to perform crossover recombination, and the n groups of data obtained after crossover recombination are:
[0086] [a' q ,b' q ,c' q ,d' q ,e' q ,f′ q ,a 1 ,b 1 ,c 1 ,d 1 ,e 1 ,f 1 ,G' q1 ]~[a' q ,b' q ,c' q ,d' q ,e' q ,f′ q ,a n ,b n ,c n ,d n ,e n ,f n ,G' qn ]; the rth group of data is [a' q ,b' q ,c' q ,d' q ,e' q ,f q ',a r ,b r ,c r ,d r ,e r ,f r ,G' qr], where G' qr is the test data label after cross-combination and The value of q is a positive integer from 1 to m; the value of r is a positive integer from 1 to n;
[0087] The m groups of data in the test set are cross-combined with all the data in the training set in pairs according to the above steps and the data labels are recalibrated to finally construct the model test set;
[0088] S7. Bring the model test set obtained in step S6 into the initial rockburst prediction model obtained in step S5 for testing, and optimize it according to the test results to obtain a rockburst prediction model;
[0089] Substitute the model test set of step S6 into the initial rockburst prediction model obtained in step S5 for testing, and restore the "0-1" label to the rockburst grade result in a maximum probability attribution manner;
[0090] During the test, the primary rockburst prediction model outputs the following results for the input model test set: 1 ,b′ 1 ,c′ 1 ,d′ 1 ,e′ 1 ,f′ 1 ,a 1 ,b 1 ,c 1 ,d 1 ,e 1 ,f 1 ,G″′ 11 ]~[a' m ,b' m ,c' m ,d' m ,e' m ,f′ m ,a n ,b n ,c n ,d n ,e n ,f n ,G″′ mn ]; Rockburst prediction results G″′ for the primary model of rockburst prediction xy , x is a positive integer from 1 to m, and y is a positive integer from 1 to n;
[0091] Separate statistics g y = 0, g y =1, g y =2 and g y =3:G″′ xy = 0, and the statistics g y = 0, gy =1, g y =2 and g y =3:G″′ xy = the number of 1s;
[0092] The probability of belonging to each level is calculated using the following formula:
[0093]
[0094] Where I(G″′ xy =1|g y =i) means when g y =i when G″′ xy = the number of 1s; I(G″′ xy =0|g y =i) means when g y =i when G″′ xy = the number of 0; i takes values of 0, 1, 2 and 3; N is the number of data in the model training set; P i is the probability that the test set data belongs to each level; is the probability that the test set data does not belong to each level; R represents P i The rockburst level corresponding to the maximum value; express The rockburst level corresponding to the maximum value is used for reverse verification;
[0095] The level corresponding to the calculated maximum probability value is used as the rockburst prediction level output by the primary rockburst prediction model;
[0096] Optimize based on the test results, specifically:
[0097] According to the test results, if the accuracy of the tested initial rockburst prediction model reaches the set requirements, then the initial rockburst prediction model is used as the final rockburst prediction model;
[0098] If the accuracy of the tested initial rockburst prediction model does not meet the set requirements, it means that the initial rockburst prediction model trained in step S5 does not meet the requirements. At this time, it is necessary to return to step S5 and adjust the training parameters to obtain a new initial rockburst prediction model and test it again until the obtained initial rockburst prediction model meets the set requirements, thereby obtaining the final rockburst prediction model.
[0099] S8. Use the rockburst prediction model obtained in step S7 to predict rockburst in actual projects; in specific implementation, obtain corresponding data in the actual project, process it in the manner of step S6, and then enter the rockburst prediction model obtained in step S7 for prediction, and restore the prediction result to obtain the final rockburst grade result to complete the final rockburst prediction
[0100] The method of the present invention is further described below in conjunction with an embodiment:
[0101] S1. Selected surrounding rock tangential stress σ θ , rock uniaxial compressive strengthσ c , rock uniaxial tensile strength σ t , the ratio of surrounding rock tangential stress to rock uniaxial compressive strength Ratio of uniaxial compressive strength to uniaxial tensile strength and the elastic deformation energy index W of rock et As a quantitative predictor;
[0102] The rockburst levels are divided into four levels: high, medium, low and none, represented by 3, 2, 1 and 0 respectively:
[0103] S2. Collect and analyze existing rockburst engineering cases, obtain quantitative prediction indicators and rockburst levels for each case, and build a database containing 212 sets of data; the distribution of each indicator in the database is shown in Table 3:
[0104] Table 3 Schematic diagram of rockburst database information
[0105] <![CDATA[σ θ ]]> <![CDATA[σ c ]]> <![CDATA[σ t ]]> <![CDATA[σ θ / s t ]]> <![CDATA[σ c / s t ]]> <![CDATA[W et ]]> Minimum 2.60 20.00 0.40 0.05 0.15 0.81 Maximum 297.80 304.20 22.60 3.45 76.67 30.00 average value 54.77 119.70 7.16 0.50 21.15 4.86 Standard Deviation 46.43 49.95 4.37 0.48 13.10 3.69
[0106] S3. The rockburst classification data set obtained in step S2 is divided into a training set and a test set in a ratio of 3:1, while ensuring that the proportion of cases of each rockburst grade is consistent; the training set contains 156 groups of data, and the test set contains 56 groups of data;
[0107] S4. Cross-recombining each set of data in the training set obtained in step S3 with all the data in the training set in turn and re-calibrating the data labels to form a model training set containing 24,336 sets of data;
[0108] S5. Construct an integrated learning model, and use the model training set obtained in step S4 to train the integrated learning model to obtain a primary model for rockburst prediction;
[0109] An ensemble learning algorithm is used, and the base learner is a decision tree. The hyperparameters that need to be adjusted are the ensemble method, the maximum number of splits, the number of learners, and the learning rate. The random search method is used for parameter tuning, and the number of iterations is 30. During the training process, a 10-fold cross-validation method is used for verification to obtain a model trained with a new training set. At the same time, another model is trained with the original training set in the same way as above for comparison; Figure 2 and Figure 3 They are the minimum error curves during the training of the new training set and the original training set, as well as the optimal hyperparameter results. Figure 2Train the model for the new training set, point A in the figure, the ensemble method is LogitBoost, the number of learners is 212, the learning rate is 0.4353, and the maximum number of splits is 6311; Figure 3 The model trained with the original training set is point B in the figure. The ensemble method is AdaBoost, the number of learners is 87, the learning rate is 0.0108, and the maximum number of splits is 7. Finally, the validation set accuracy of the model trained with the new training set reached 97.7%, while the accuracy of the model trained with the original training set on the validation set was only 56.4%.
[0110] S6. Cross-recombining each set of data in the test set obtained in step S3 with all the data in the training set in pairs and re-calibrating the data labels to form a model test set;
[0111] S7. Use the model test set obtained in step S6 to test the primary rockburst prediction model obtained in step S5, and optimize it according to the test results to obtain a rockburst prediction model; input the model test set into the model to test the generalization ability of the model. The classification confusion matrices of the two models on the test set are as follows: Figure 4 (corresponding to the model test set confusion matrix) and Figure 5 (corresponding to the confusion matrix of the original test set) is shown; the accuracy of the two models on the validation set and test set is shown in Figure 6 As shown, it can be seen that the model trained by the method of the present invention has higher reliability;
[0112] Then, the classification results of the model on the new test are analyzed, the number of labels "0" and "1" in each group of data is counted, the probability of belonging to each level is calculated, and the level with the largest probability is used as the predicted final rockburst level. The two labels are restored to the rockburst level results. The actual rockburst results and predicted classification results of the 56 groups of data in the test set are shown in Table 4, and the classification confusion matrix is shown in the attached figure. Figure 7 shown.
[0113] Table 4 Schematic diagram of classification results of model test set
[0114]
[0115]
[0116]
[0117]
[0118] Among the 56 groups of data in the test set, the method of the present invention has 14 groups of data classified incorrectly, and its classification accuracy rate reaches 75.00%; the original method has 26 groups of classification errors, and the accuracy rate is only 53.57%; the accuracy rate of the method of the present invention has been improved by more than 20 percentage points, indicating that the model trained by the method of the present invention has good generalization performance and has a high prediction accuracy rate for unknown data, and can be used for actual rock burst prediction;
[0119] S8. Use the rockburst prediction model obtained in step S7 to predict rockburst in actual projects.
Claims
1. A rockburst prediction method based on few-sample metric and ensemble learning, comprising the following steps: S1. Analyze and select quantitative prediction indicators that affect rockburst levels and qualitative standards for rockburst level classification; S2. Collect and analyze existing rockburst engineering cases, obtain quantitative prediction indicators and rockburst grades for each case, and construct a rockburst classification data set; S3. The rockburst classification data set obtained in step S2 is randomly divided into a training set and a test set according to the proportion, while ensuring that the proportion of cases of each rockburst grade is consistent; S4. Cross-recombining each group of data in the training set obtained in step S3 with all the data in the training set in turn and re-calibrating the data labels to form a model training set; S5. Using an ensemble learning method, the model training set obtained in step S4 is trained and hyperparameters are optimized to obtain an initial model for rockburst prediction; S6. Cross-recombining each set of data in the test set obtained in step S3 with all the data in the training set in pairs and re-calibrating the data labels to form a model test set; S7. Bring the model test set obtained in step S6 into the initial rockburst prediction model obtained in step S5 for testing, and optimize it according to the test results to obtain a rockburst prediction model; S8. Use the rockburst prediction model obtained in step S7 to predict rockburst in actual projects.
2. The rockburst prediction method based on few-sample metric and ensemble learning according to claim 1 is characterized in that The quantitative prediction index affecting the rockburst level described in step S1 specifically includes the following steps: Select surrounding rock tangential stress σ θ , rock uniaxial compressive strength σ c , rock uniaxial tensile strength σ t , the ratio of surrounding rock tangential stress to rock uniaxial compressive strength Ratio of uniaxial compressive strength to uniaxial tensile strength and the rock elastic deformation energy index W et as a quantitative predictor.
3. The rockburst prediction method based on few-sample metric and ensemble learning according to claim 2 is characterized in that The qualitative standard for rockburst grade classification described in step S1 specifically includes the following steps: According to the existing national standards, rock burst levels are divided into four levels: high, medium, low and none.
4. The rockburst prediction method based on few-sample metric and ensemble learning according to claim 3 is characterized in that The step S2 of obtaining the quantitative prediction index and rockburst grade of each case and constructing a rockburst classification data set specifically includes the following steps: Get the quantitative prediction index of the ith case: surrounding rock tangential stress σ θi Using the letter a i Indicates that the uniaxial compressive strength of rock σ ci Using the letter b i Indicates that the uniaxial tensile strength of rock σ ti Using the letter c i It means that the ratio of surrounding rock tangential stress to rock uniaxial compressive strength is Using the letter d i It indicates the ratio of uniaxial compressive strength to uniaxial tensile strength. Using the letter e i Indicates that the elastic deformation energy index W of rock eti Using the letter f i Indicates that the rockburst grade is represented by the letter g i Indicates that the i-th case is then represented in the rockburst classification dataset as the i-th group of data [a i ,b i ,c i ,d i ,e i ,f i ,g i ];in Rockburst grade value g i The values are 3, 2, 1 and 0, which correspond to high, medium, low and no rockburst levels respectively; i is a positive integer from 1 to n; The quantitative prediction indicators and rockburst grades of all n cases are obtained, and finally a rockburst classification dataset is constructed.
5. The rockburst prediction method based on few-sample metric and ensemble learning according to claim 4 is characterized in that Step S4, in which each set of data in the training set obtained in step S3 is cross-combined with all the data in the training set in pairs and the data labels are recalibrated, specifically includes the following steps: The i-th group of data [a i ,b i ,c i ,d i ,e i ,f i ,g i ], and then compare them with the n sets of data [a1,b1,c1,d1,e1,f1,g1]~[a n ,b n ,c n ,d n ,e n ,f n ,g n ] to perform crossover recombination, and the n groups of data obtained after crossover recombination are: [a i ,b i ,c i ,d i ,e i ,f i ,a1,b1,c1,d1,e1,f1,G i1 ]~[a i ,b i ,c i ,d i ,e i ,f i ,a n ,b n ,c n ,d n ,e n ,f n ,G in ]; the jth group of data is [a i ,b i ,c i ,d i ,e i ,f i ,a j ,b j ,c j ,d j ,e j ,f j ,G ij ], where G ij is the training data label after cross-combination and The value of j is a positive integer from 1 to n; The n groups of data in the training set are cross-combined with all the data in the training set in pairs according to the above steps, and the data labels are recalibrated to finally construct the model training set.
6. The rockburst prediction method based on few-sample metric and ensemble learning according to claim 5 is characterized in that The integrated learning method described in step S5 is used to train and optimize the hyperparameters of the model training set obtained in step S4 to obtain an initial model for rockburst prediction, which specifically includes the following steps: Use decision tree as base learner; The following hyperparameters are tuned using a random search algorithm: Hyperparameters: data integration method; the search scope includes Bagging integration, AdaBoost integration, RUSBoost integration, GentleBoost integration and LogitBoost integration; Hyperparameters: number of base learners; search range is [10,500]; Hyperparameters: learning rate; search range is [0.001, 1]; Hyperparameters: maximum number of splits; search range is [1, N]; where N is the amount of data in the model training set; The K-fold cross-validation method is used to verify the model training process.
7. The rockburst prediction method based on few-sample metric and ensemble learning according to claim 6 is characterized in that Step S6, in which each group of data in the test set obtained in step S3 is cross-combined with all the data in the training set in pairs and the data labels are recalibrated to form a model test set, specifically includes the following steps: The qth group of data in the test set [a' q ,b' q ,c' q ,d' q ,e' q ,f' q ,g' q ], and then compare them with the n sets of data [a1,b1,c1,d1,e1,f1,g1]~[a n ,b n ,c n ,d n ,e n ,f n ,g n ] to perform crossover recombination, and the n groups of data obtained after crossover recombination are: [a' q ,b' q ,c' q ,d' q ,e' q ,f' q ,a1,b1,c1,d1,e1,f1,G' q1 ]~[a' q ,b' q ,c' q ,d' q ,e' q ,f' q ,a n ,b n ,c n ,d n ,e n ,f n ,G' qn ]; the rth group of data is [a' q ,b' q ,c' q ,d' q ,e' q ,f' q ,a r ,b r ,c r ,d r ,e r ,f r ,G' qr ], where G' qr is the test data label after cross-combination and The value of q is a positive integer from 1 to m; the value of r is a positive integer from 1 to n; The m groups of data in the test set are cross-combined with all the data in the training set in pairs according to the above steps, and the data labels are recalibrated to finally construct the model test set.
8. The rockburst prediction method based on few-sample metric and ensemble learning according to claim 7 is characterized in that Step S7 of bringing the model test set obtained in step S6 into the initial rockburst prediction model obtained in step S5 for testing specifically includes the following steps: Substitute the model test set of step S6 into the initial rockburst prediction model obtained in step S5 for testing, and restore the "0-1" label to the rockburst grade result in a maximum probability attribution manner; During the test, the output of the primary rockburst prediction model for the input model test set is [a′1,b′1,c′1,d′1,e′1,f1′,a1,b1,c1,d1,e1,f1,G″′ 11 ]~[a' m ,b' m ,c' m ,d' m ,e' m ,f' m ,a n ,b n ,c n ,d n ,e n ,f n ,G″′ mn ]; Rockburst prediction results G″′ for the primary model of rockburst prediction xy , x is a positive integer from 1 to m, and y is a positive integer from 1 to n; Separate statistics g y = 0, g y =1, g y =2 and g y =3:G″′ xy = 0, and the statistics g y = 0, g y =1, g y =2 and g y =3:G″′ xy = the number of 1s; The probability of belonging to each level is calculated using the following formula: Where I(G″′ xy =1|g y =i) means when g y =i when G″′ xy = the number of 1s; I(G″′ xy =0|g y =i) means when g y =i when G″′ xy = the number of 0; i takes values of 0, 1, 2 and 3; N is the number of data in the model training set; P i is the probability that the test set data belongs to each level; is the probability that the test set data does not belong to each level; R represents P i The rockburst level corresponding to the maximum value; express The rockburst level corresponding to the maximum value is used for reverse verification; The level corresponding to the calculated maximum probability value is used as the rockburst prediction level output by the primary rockburst prediction model.
9. The rockburst prediction method based on few-sample metric and ensemble learning according to claim 8 is characterized in that Step S8 uses the rockburst prediction model obtained in step S7 to predict rockburst in actual projects, specifically obtaining corresponding data in the actual project, processing it in the manner of step S6, and then entering the rockburst prediction model obtained in step S7 for prediction, and restoring the prediction result to obtain the final rockburst grade result, thereby completing the final rockburst prediction.
Citation Information
Patent Citations
SVM-LDA rockburst machine learning prediction model method based on data analysis principle
CN113076700A
Rock burst early warning method based on deep learning micro-seismic monitoring data
CN113469342A
