Rockburst intensity classification prediction method based on normalization and balanced integrated Voting
Through normalization and balanced processing and multi-model integration Voting methods, the outliers and imbalance problems in the rock burst data set are solved, the accuracy of rock burst intensity prediction is improved, and it is suitable for disaster prevention in global engineering construction.
Patent Information
- Application Number
- CN202210011696.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-01-06
- Publication Date
- 2024-11-29
- Estimated Expiration
- 2042-01-06
AI Technical Summary
The existing rock burst intensity prediction methods have poor prediction results in large-scale rock burst data sets, weak generalization ability, and ignore the problems of unevenness and data imbalance of rock burst phenomena, resulting in a decline in prediction performance.
Using a method based on normalization and balanced integration of Voting, data is processed through Yeo-Johnson transformation and K-means SMOTE, multiple machine learning models are constructed, and the comprehensive weight is calculated using macro-average accuracy and Pearson correlation coefficients, and the comprehensive weight is filtered out. The optimal Voting model is used to predict.
It significantly improves the accuracy of rock burst intensity prediction, reduces the impact of outliers and data imbalance, integrates a variety of excellent machine learning algorithms to obtain optimal prediction performance, and is suitable for rock burst disaster prevention and control in global engineering construction.
Smart Images

Figure CN114547965B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of rockburst intensity classification prediction, and in particular to a rockburst intensity classification prediction method based on normalization and balanced integrated Voting. Background Art
[0002] Rockburst refers to a complex dynamic geological disaster in which the elastic strain energy stored in the rock mass is suddenly released under the influence of engineering excavation disturbance under high ground stress environment, causing violent peeling, ejection, and projection of rock blocks accompanied by loud noises and vibrations. In recent years, rockburst disasters caused by global engineering construction have become more frequent with the increase of engineering burial depth or the increase of ground stress level. Rockburst not only affects the progress of engineering construction, but also threatens the lives and property safety of workers. If the occurrence of rockburst can be accurately predicted and protective measures can be taken in advance to reduce the occurrence of rockburst, the casualties and property losses of workers can be greatly reduced. Therefore, the accurate prediction of rockburst is particularly important for the construction progress of underground engineering and the safety of workers.
[0003] In order to accurately predict the occurrence of rockburst disasters, technicians have conducted in-depth research and technology development: the Chinese invention patent with announcement number CN106407493B discloses "a rockburst grade evaluation method based on a multidimensional Gaussian cloud model", and the Chinese invention patent with announcement number CN112699553B discloses "a rockburst intensity grade intelligent prediction system method". Both patented technologies use a method combining a multidimensional normal cloud model with indicator weights to determine the intensity grade of rockburst samples. The parameter expectation, entropy and super entropy values in the multidimensional normal cloud model are determined according to the rockburst prediction index classification standard in Table 1. However, the rockburst prediction index grade division is too rigid, and the distinction between different rockburst grades is too clear. The rockburst grade is determined by exceeding a certain value or being less than a certain value. As a result, the model has poor prediction effect on large rockburst data sets consisting of more than one hundred samples, and the generalization ability of the model is not strong. The maximum tangential stress of the surrounding rock (σ θ ), uniaxial compressive strength of rock (σ c ), uniaxial tensile strength of rock (σ t ), rock elastic strain energy index (w et ), rock stress coefficient (σ θ / σ c ) and rock brittleness coefficient (σ c / σ t ) is a rockburst prediction index.
[0004] Table 1: Rockburst prediction index classification standard
[0005]
[0006] The Chinese patent application with publication number CN110889440A discloses "Rockburst grade prediction method and system based on principal component analysis and BP neural network". This technical solution ignores the uneven occurrence of rockburst, that is, rockburst disasters of level I and IV occur less frequently, while rockburst disasters of level II and III occur more frequently, and the rockburst data of each level are unbalanced. In addition, compared with the integrated method, the prediction performance of a single machine learning model for rockburst grade prediction is far inferior to the integrated learning prediction performance that can integrate multiple machine learning models. Summary of the invention
[0007] In order to solve the problems existing in the prior art, the present invention provides a rockburst intensity classification prediction method based on normalization and balanced integrated Voting with high prediction accuracy.
[0008] In order to solve the above technical problems, the rockburst intensity classification prediction method based on normalization and balanced integrated voting of the present invention comprises the following steps:
[0009] Step S1: Collect, organize and summarize the rockburst intensity classification data at home and abroad, determine the maximum tangential stress of surrounding rock, rock uniaxial compressive strength, rock uniaxial tensile strength, rock elastic strain energy index, rock stress coefficient and rock brittleness coefficient as the influencing factors of rockburst, and the rockburst level is divided into: level I is no rockburst, level II is slight rockburst, level III is moderate rockburst, and level IV is strong rockburst, so as to construct a rockburst data set;
[0010] Step S2: For the rockburst dataset constructed in step S1, the Yeo-Johnson transformation and K-meansSMOTE oversampling are used to normalize and balance the data;
[0011] Step S3: using the normalized and balanced rockburst dataset in step S2 as input, constructing multiple machine learning rockburst intensity classification prediction models;
[0012] Step S4: For the constructed model, the macro-average precision is used to characterize the prediction performance of the model, the Pearson correlation coefficient is used to characterize the difference in prediction results between the models, and the comprehensive weight of the prediction performance of each model and the difference in prediction results is calculated;
[0013] Step S5: All machine learning models are put into voting for training and prediction, and the models with small weights are eliminated in turn by exhaustive selection, thereby screening out the optimal voting rockburst intensity classification prediction model. The majority principle is adopted for the prediction results of the base model, and the final prediction results are determined by voting.
[0014] Furthermore, the number of rockburst samples collected in step S1 is not less than 100, and all rockburst samples have no duplicate values or missing values, ensuring that the collected rockburst samples are effective and reliable.
[0015] Furthermore, the Yeo-Johnson transformation formula in step S2 is:
[0016]
[0017] Among them, y is the rockburst characteristic data, and λ is the parameter estimated by the maximum likelihood method.
[0018] Furthermore, the K-means SMOTE algorithm in step S2 is divided into three steps:
[0019] Clustering: The rockburst data is divided into k clusters using the K-means algorithm;
[0020] Filtering: Filter out clusters with a high proportion of minority class samples;
[0021] Oversampling: Perform SMOTE oversampling on clusters with a high proportion of minority class samples. The sparser the minority class samples in the cluster, the more minority class samples are added.
[0022] Furthermore, in step S3, the machine learning model selects an algorithm with high prediction accuracy, such as support vector machine, random forest, multi-layer perceptron, XgBoost, Lightgbm, etc. No less than 8 machine learning rockburst intensity classification prediction models are constructed, and grid search with cross-validation is used to adjust parameters during model training to obtain the optimal parameters when the model accuracy is the highest.
[0023] Furthermore, the machine learning rockburst intensity classification prediction model is selected from AdaBoost, XgBoost, Lightgbm, GradientBoosting, CatBoost, Bagging KNN, Bagging MLP, Bagging GP, Bagging NBM, Random Forest, Multi-layer Perception, Bayes Model, Decision Tree, K-NearestNeighbor, Gaussian Process, among which Bagging KNN, Bagging MLP, Bagging GP, Bagging NBM, and Random Forest are integrated bagging algorithms with KNN, MLP, GP, NBM, and DT as base models respectively.
[0024] Furthermore, the macro-average accuracy calculation formula in step S4 is:
[0025]
[0026] Where P i is the rockburst prediction accuracy of the i-th level.
[0027] Furthermore, the Pearson correlation coefficient calculation formula in step S4 is:
[0028]
[0029] Where X and Y are the prediction results of the two models respectively.
[0030] Furthermore, in step S4, the performance of each model is determined by using the size of the comprehensive weight, which is composed of the size of the difference weight of the model prediction results and the size of the model macro-average accuracy weight.
[0031] Furthermore, the calculation formula for the difference weight of the model prediction results is:
[0032]
[0033] Among them, b is the average value of the correlation between the model and the other models;
[0034] The calculation formula for the precision weight of the model macro average is:
[0035]
[0036] Where a is the macro-average accuracy of the model;
[0037] The comprehensive weight of the model is: i =λ i +w i .
[0038] Furthermore, all steps from step S1 to step S5 are implemented by Python programming.
[0039] Compared with the prior art, the rockburst intensity classification prediction method based on normalization and balanced integrated voting provided by the present invention has the following beneficial effects:
[0040] (1) The present invention is based on the actual situation of underground engineering at home and abroad, and comprehensively considers the internal and external factors that lead to rock burst disasters. The maximum tangential stress of the surrounding rock (σ θ ), uniaxial compressive strength of rock (σ c ), uniaxial tensile strength of rock (σ t ), rock elastic strain energy index (w et ), rock stress coefficient (σ θ / σc ) and rock brittleness coefficient (σ c / σ t ) is the rockburst prediction index. The machine learning algorithm that does not rely on the rockburst prediction index classification standard to determine the model parameters is used for rockburst prediction, which overcomes the problem that the rockburst prediction index classification standard is too rigid, resulting in a decrease in prediction performance.
[0041] (2) The present invention uses Yeo-Johnson transformation and K-means SMOTE oversampling to normalize and balance the data, reducing the poor separability of rockburst data caused by the presence of many outliers in the rockburst data and the uneven occurrence of rockburst phenomena. The number of outliers is reduced at the data level, the number of rockburst samples of each level is balanced, and the rockburst data structure is significantly improved.
[0042] (3) The present invention uses multiple machine learning algorithms to predict rockbursts, and selects models with excellent prediction performance as integrated voting models. The advantages of multiple excellent machine learning algorithms are integrated at the algorithm level to obtain a voting model with the best prediction performance. BRIEF DESCRIPTION OF THE DRAWINGS
[0043] The specific implementation of the present invention will be further explained below in conjunction with the accompanying drawings.
[0044] Figure 1 It is a flow chart of the rockburst intensity classification prediction method based on normalization and balanced integrated Voting of the present invention;
[0045] Figure 2 The macro-average precision of 15 models in the embodiment and the ranking diagram of precision are shown in FIG.
[0046] Figure 3 This is a diagram of the Voting-based model screening results in the embodiment. DETAILED DESCRIPTION
[0047] Combination Figure 1 The rockburst intensity classification prediction method based on normalization and balanced integrated Voting in this embodiment includes the following steps:
[0048] Step S1: Collect, organize and summarize rockburst intensity classification data at home and abroad, determine the maximum tangential stress of surrounding rock, rock uniaxial compressive strength, rock uniaxial tensile strength, rock elastic strain energy index, rock stress coefficient and rock brittleness coefficient as the influencing factors of rockburst, and the rockburst level is divided into: level I is no rockburst, level II is slight rockburst, level III is moderate rockburst, and level IV is strong rockburst, so as to construct a rockburst data set;
[0049] Preferably, in this embodiment, the number of rockburst samples collected in step S1 is not less than 100, and all rockburst samples have no duplicate values or missing values, thereby ensuring that the collected rockburst samples are effective and reliable.
[0050] Step S2: For the rockburst dataset constructed in step S1, the Yeo-Johnson transformation and K-meansSMOTE oversampling are used to normalize and balance the data;
[0051] In this embodiment, the Yeo-Johnson transformation formula in step S2 is preferably:
[0052]
[0053] Among them, y is the rockburst characteristic data, and λ is the parameter estimated by the maximum likelihood method.
[0054] In this embodiment, the K-means SMOTE algorithm in step S2 is preferably divided into three steps:
[0055] Clustering: The rockburst data is divided into k clusters using the K-means algorithm;
[0056] Filtering: Filter out clusters with a high proportion of minority class samples;
[0057] Oversampling: Perform SMOTE oversampling on clusters with a high proportion of minority class samples. The sparser the minority class samples in the cluster, the more minority class samples are added.
[0058] Step S3: using the normalized and balanced rockburst dataset in step S2 as input, constructing multiple machine learning rockburst intensity classification prediction models;
[0059] In this embodiment, the machine learning model in step S3 preferably selects an algorithm with high prediction accuracy, such as support vector machine, random forest, multi-layer perceptron, XgBoost, Lightgbm, etc. No less than 8 machine learning rockburst intensity classification prediction models are constructed, and grid search with cross-validation is used to adjust parameters during model training to obtain the optimal parameters when the model accuracy is the highest.
[0060] In this embodiment, 15 machine learning models are constructed, and the 15 models are: AdaBoost, XgBoost, Lightgbm, GradientBoosting, CatBoost, BagKNN (Bagging KNN), BagMLP (Bagging MLP), BagGP (Bagging GP), BagNBM (Bagging NBM), RF (Random Forest), MLP (Multi-layerPerception), NBM ( Bayes Model), DT (Decision Tree), KNN (K-Nearest Neighbor), GP (Gaussian Process). BagKNN, BagMLP, BagGP, BagNBM, RF are integrated bagging algorithms based on KNN, MLP, GP, NBM, and DT as the base models. The macro-average accuracy of the 15 models and the ranking of the accuracy are shown in the figure below. Figure 2 shown.
[0061] Step S4: For the constructed model, the macro-average precision is used to characterize the prediction performance of the model, the Pearson correlation coefficient is used to characterize the difference in prediction results between the models, and the comprehensive weight of the prediction performance of each model and the difference in prediction results is calculated;
[0062] In this embodiment, the macro-average accuracy calculation formula in step S4 is preferably:
[0063]
[0064] Where P i is the rockburst prediction accuracy of the i-th level.
[0065] In this embodiment, the Pearson correlation coefficient calculation formula in step S4 is preferably:
[0066]
[0067] Where X and Y are the prediction results of the two models respectively.
[0068] Preferably, in step S4, the performance of each model is determined by using a comprehensive weight, and the comprehensive weight is composed of a difference weight of the model prediction results and a macro-average precision weight of the model.
[0069] In this embodiment, the calculation formula for the difference weight of the model prediction results is preferably:
[0070]
[0071] Among them, b is the average value of the correlation between the model and the other models;
[0072] The calculation formula for the precision weight of the model macro average is:
[0073]
[0074] Where a is the macro-average accuracy of the model;
[0075] The comprehensive weight of the model is: i=λ i +w i .
[0076] In this embodiment, the prediction performance weights, prediction result difference weights and comprehensive weights of various models are shown in Table 2.
[0077] Table 2: Weights of 15 machine learning models
[0078]
[0079]
[0080] Step S5: Put the 15 models in this embodiment into voting for training and prediction, and use the exhaustive method to eliminate the models with small weights in turn, thereby screening out the optimal voting rockburst intensity classification prediction model, and use the principle of minority obeys majority for the prediction results of the base model, and vote to determine the final prediction results.
[0081] In this embodiment, the Voting-based model screening results are as follows: Figure 3 shown. Figure 3 The middle horizontal axis represents the number of models with high comprehensive weights retained.
[0082] Preferably, in this embodiment, all steps from step S1 to step S5 are implemented by Python programming.
[0083] The rockburst intensity classification prediction method based on normalized and balanced integrated Voting in this embodiment reduces the influence of outliers and data imbalance in data optimization, integrates multiple excellent models in algorithm optimization, effectively improves the rockburst prediction performance, and has higher prediction accuracy. The implementation of the present invention is of great significance to improving the accuracy of rockburst prediction, and has important guiding value for the prevention and control of rockburst disasters in global engineering construction processes.
[0084] Many specific details are described in the above description to facilitate a full understanding of the present invention. However, the above description is only a preferred embodiment of the present invention. The present invention can be implemented in many other ways different from those described herein, so the present invention is not limited to the specific implementation disclosed above. At the same time, any person familiar with the art can make many possible changes and modifications to the technical solution of the present invention using the methods and technical contents disclosed above without departing from the scope of the technical solution of the present invention, or modify it into an equivalent embodiment of equivalent changes. Any simple modification, equivalent change and modification made to the above embodiments according to the technical essence of the present invention without departing from the content of the technical solution of the present invention still falls within the scope of protection of the technical solution of the present invention.
Claims
1. A rockburst intensity classification prediction method based on normalization and balanced integrated voting, characterized by: The steps include: Step S1: Collect, organize and summarize the rockburst intensity classification data at home and abroad, determine the maximum tangential stress of surrounding rock, rock uniaxial compressive strength, rock uniaxial tensile strength, rock elastic strain energy index, rock stress coefficient and rock brittleness coefficient as the influencing factors of rockburst, and the rockburst level is divided into: level I is no rockburst, level II is slight rockburst, level III is moderate rockburst, and level IV is strong rockburst, so as to construct a rockburst data set; Step S2: For the rockburst dataset constructed in step S1, the Yeo-Johnson transformation and K-means SMOTE oversampling are used to normalize and balance the data; Step S3: using the normalized and balanced rockburst dataset in step S2 as input, constructing multiple machine learning rockburst intensity classification prediction models; Step S4: For the constructed model, the macro-average precision is used to characterize the prediction performance of the model, the Pearson correlation coefficient is used to characterize the difference in prediction results between the models, and the comprehensive weight of the prediction performance of each model and the difference in prediction results is calculated; Step S5: All machine learning models are put into voting for training and prediction, and the models with small weights are eliminated in turn by exhaustive selection, thereby screening out the optimal voting rockburst intensity classification prediction model, and the majority principle is adopted for the prediction results of the base model, and the final prediction results are determined by voting; In step S4, the performance of each model is determined by using the size of the comprehensive weight, which is composed of the size of the difference weight of the model prediction results and the size of the model macro-average precision weight; The calculation formula for the difference weight of the model prediction results is: Among them, b is the average value of the correlation between the model and the other models; The calculation formula for the precision weight of the model macro average is: Where a is the macro-average accuracy of the model; The comprehensive weight of the model is: i =λ i +w i .
2. The rockburst intensity classification prediction method based on normalization and balanced integrated voting according to claim 1 is characterized by: The number of rockburst samples collected in step S1 is not less than 100, and all rockburst samples have no duplicate values or missing values, ensuring that the collected rockburst samples are valid and reliable.
3. The rockburst intensity classification prediction method based on normalization and balanced integrated voting according to claim 1 is characterized by: The Yeo-Johnson transformation formula in step S2 is: Among them, y is the rockburst characteristic data, and λ is the parameter estimated by the maximum likelihood method.
4. The rockburst intensity classification prediction method based on normalization and balanced integrated voting according to claim 1 is characterized by: The K-means SMOTE algorithm in step S2 is divided into three steps: Clustering: The rockburst data is divided into k clusters using the K-means algorithm; Filtering: Filter out clusters with a high proportion of minority class samples; Oversampling: Perform SMOTE oversampling on clusters with a high proportion of minority class samples. The sparser the minority class samples in the cluster, the more minority class samples are added.
5. The rockburst intensity classification prediction method based on normalization and balanced integrated voting according to claim 1 is characterized by: There are no less than 8 machine learning rockburst intensity classification prediction models constructed in step S3. During the model training process, grid search with cross-validation is used to adjust parameters to obtain the optimal parameters when the model accuracy is the highest.
6. The rockburst intensity classification prediction method based on normalization and balanced integrated voting according to claim 5 is characterized by: The machine learning rockburst intensity classification prediction model is selected from AdaBoost, XgBoost, Lightgbm, GradientBoosting, CatBoost, Bagging KNN, Bagging MLP, Bagging GP, Bagging NBM, Random Forest, Multi-layer Perception, Bayes Model, Decision Tree, K-NearestNeighbor, Gaussian Process, among which Bagging KNN, Bagging MLP, Bagging GP, Bagging NBM, and Random Forest are integrated bagging algorithms with KNN, MLP, GP, NBM, and DT as base models respectively.
7. The rockburst intensity classification prediction method based on normalization and balanced integrated voting according to claim 1 is characterized by: The calculation formula of the macro average precision in step S4 is: Where P i is the rockburst prediction accuracy of the i-th level.
8. The rockburst intensity classification prediction method based on normalization and balanced integrated voting according to claim 1 is characterized by: The Pearson correlation coefficient calculation formula in step S4 is: Where X and Y are the prediction results of the two models respectively.
Citation Information
Patent Citations
A Rockburst Severity Evaluation Method Based on a Multidimensional Gaussian Cloud Model
CN106407493B
Rockburst grade prediction method and system based on principal component analysis and BP neural network
CN110889440A
A method for intelligent prediction system of rockburst intensity level
CN112699553B
Multiple target comprehensive decision evaluation method based on scene
CN103065042A
Vocabulary entry classification method and audit information extraction method
CN109635289A
