An artificial intelligence-based multi-disaster type hierarchical prediction method

Abstract

The application discloses a kind of multi-disaster grading prediction methods based on artificial intelligence, comprising the following steps: mine disaster data is collected, and mine disaster database is formed;Mine disaster prediction index system is constructed;Mine multi-disaster prediction database is constructed;According to the index selection evaluation characteristics in mine disaster prediction index system, according to each algorithm in the preset mine disaster prediction algorithm bank respectively to the data in mine multi-disaster prediction database is graded prediction, under the different evaluation indexes of preset, the prediction results of each algorithm graded prediction are compared, and the algorithm with the optimal comprehensive evaluation index score of prediction result is obtained, and the algorithm is used as the algorithm of mine disaster prediction optimal model;Mine disaster data is input into mine disaster prediction optimal model, and multi-disaster grading prediction is carried out.The application significantly improves the imbalance of mine disaster data set, realizes data dimension reduction, and improves the accuracy of prediction.

Description

Technical Field

[0001] The present invention belongs to the technical field of mine disaster classification prediction, and specifically relates to a multi-disaster classification prediction method based on artificial intelligence. Background Art

[0002] The geological conditions of many mines in my country are complex. During the mining process, they are inevitably threatened by disasters such as water disasters, fire, dust disasters, roof disasters, and slope instability. Deep mining areas are also plagued by various disasters such as rock bursts, pillar instability, and goaf instability. These disasters not only delay construction schedules and cause huge economic losses, but also threaten the lives and safety of construction workers, bringing severe challenges to the operation and safety of underground projects.

[0003] Mine disaster prediction is a difficult problem that needs to be solved urgently. However, the mechanisms of many mine disasters are complex and have not yet been unified. Its prediction is a complex nonlinear problem. It is the result of the combined action of multiple factors and has randomness, fuzziness, and uncertainty. No mathematical or mechanical method can fully and accurately describe it. Therefore, it is necessary to explore how to improve existing prediction methods and increase the accuracy of mine disaster prediction to provide a scientific basis for safe protection and reasonable construction in my country's mines. In addition, due to the particularity of mine disasters and the relatively harsh data collection environment, they lead to problems such as low data quality and difficulty in selecting evaluation indicators. Therefore, it is meaningful to study how to improve the quality of mine disaster datasets and the optimization of indicators.

[0004] There is not just one single disaster in a mine. If one disaster corresponds to a set of algorithms, it will lead to the redundancy of the smart mining system, reduce its response sensitivity, greatly increase the cost of software and hardware, cause waste of resources, and make use and management difficult. Summary of the Invention

[0005] The purpose of the present invention is to provide an artificial intelligence-based multi-hazard classification prediction method, which significantly improves the imbalance of mine disaster data sets, achieves data dimensionality reduction, and improves the accuracy of prediction.

[0006] To solve the above technical problems, the technical solution of the present invention is: a multi-disaster graded prediction method based on artificial intelligence, comprising the following steps:

[0007] S1. Collect mining hazard data to form a mining hazard database; the mining hazard data at least includes hanging wall stability, slope stability, rock burst, pillar stability and goaf stability;

[0008] S2. Constructing a mine disaster prediction index system: Screening the data in the mine disaster database, eliminating indicators with high feature overlap and low feature importance, and forming a mine disaster prediction index system after dimensionality reduction based on the screened data;

[0009] S3. Constructing a mine multi-disaster prediction database: reconstructing the data set in the mine disaster database, processing the outliers and missing values, obtaining a balanced data set and using it as the data in the mine multi-disaster prediction database;

[0010] S4. Select evaluation features based on the indicators in the mine disaster prediction index system, perform hierarchical prediction on the data in the mine multi-disaster prediction database according to each algorithm in the preset mine disaster prediction algorithm library, summarize and compare the prediction results of the hierarchical prediction of each algorithm under different preset evaluation indicators, obtain the algorithm with the best comprehensive evaluation index score of the prediction result, and use this algorithm as the algorithm of the optimal model for mine disaster prediction;

[0011] S5. Input the mine disaster data into the optimal mine disaster prediction model to conduct multi-disaster graded prediction.

[0012] In S2, data indicators are screened based on the Pearson correlation coefficient and feature importance. The specific steps are as follows:

[0013] Calculate the feature importance of any disaster type in the mine disaster data, eliminate the features with low feature importance, and calculate the Pearson correlation coefficient of the remaining features;

[0014] The remaining features are weighted according to their absolute values. The larger the absolute value of the Pearson correlation coefficient between the features, the greater the weight. The Pearson correlation coefficient is used to indicate the strength of the correlation between features. The larger the absolute value of the Pearson correlation coefficient between two features, the stronger the correlation.

[0015] According to the preset different feature importance algorithms, the features corresponding to the absolute values ​​of the weighted Pearson correlation coefficients are ranked by importance, and the filtered data is obtained based on the importance ranking results and the correlation between the features.

[0016] The Pearson correlation coefficient is a statistical indicator used to measure the strength and direction of the linear relationship between two continuous variables. Its value range is between -1 and 1. The larger the absolute value, the stronger the correlation. If the value is positive, it is a positive correlation, and if the value is negative, it is a negative correlation. Its calculation formula is expressed as:

[0017]

[0018] Among them, ρ X,Y is the Pearson correlation coefficient, cov(X,Y) represents the covariance between the two variables X and Y, σ X , σ Y are the standard deviations of variables X and Y, respectively.

[0019] The feature importance is output by the random forest algorithm and measured by the Gini coefficient. The principle is as follows: suppose there are J features X1, X2, X3, ..., XJ, a total of I decision trees, and C categories; for each feature Xj, calculate its contribution to each tree in the random forest, and then take the average to compare the importance of the features;

[0020] When calculating the Gini index score, the impurity change of each node of each tree is considered, which is expressed as the Gini coefficient of the node; among them, the Gini coefficient of node q is The calculation formula is expressed as:

[0021]

[0022] Where C represents the number of categories, represents the proportion of category c in node q;

[0023] The importance of feature Xj in the i-th tree node q, that is, the change in the Gini coefficient before and after the node q branches for:

[0024]

[0025] in, and Respectively represent the Gini coefficients of the two new nodes after branching.

[0026] The specific steps for reconstructing the data set in the mine disaster database in S3 are:

[0027] The Smote algorithm is improved by Tomek Link algorithm to reduce the outliers generated in the data set reconstruction;

[0028] The steps of reconstructing the data set using the improved SmoteTomek algorithm are as follows:

[0029] (1) The mine disaster database is used as a sample set, and the data in it is used as samples. After determining the minority class and the majority class based on the sample set, the Euclidean distance between each minority class sample and all minority class samples is calculated to obtain its k-nearest neighbor samples; then, the same operation is performed on other minority class samples in sequence;

[0030] (2) According to the sample ratio of the majority class samples to the minority class samples, the number of samples to be generated is determined. For each minority class sample k, a corresponding number of samples are randomly selected from its nearest neighbor set. Assume that the selected nearest neighbor is x;

[0031] (3) For each selected neighbor sample x, construct a new sample x according to the following formula new :

[0032]

[0033] (4) Construct Tomek Link pairs with the new samples and their original samples, delete the paired synthetic samples, and clean up each minority class sample in turn;

[0034] (5) Repeat steps (1) to (4) until the sample categories of the reconstructed data are balanced.

[0035] The specific method for processing data outliers in the reconstructed dataset in S3 is:

[0036] The Mahalanobis distance discriminant method is used to screen the data, and its calculation method is expressed as:

[0037]

[0038] The above formula is the Mahalanobis distance in n-dimensional space for a multivariate vector with mean μ and covariance matrix Σ;

[0039] Among them, D M (x) is the Mahalanobis distance between each individual sample Y and the average sample μ; T represents the transpose of the matrix.

[0040] The method also includes the following steps: performing algorithm analysis on the mine multi-hazard prediction database through several cross-validations to verify the reliability of the sample feature importance.

[0041] The mine disaster prediction algorithm library includes a machine learning algorithm library and a deep learning algorithm library. The algorithms in the mine disaster prediction algorithm library include at least: the support vector machine algorithm, random forest algorithm and extreme random tree algorithm in the machine learning algorithm library, and the BP neural network algorithm, CNN algorithm and LSTM algorithm in the deep learning algorithm library.

[0042] A computer device is also provided, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of any of the above methods when executing the computer program.

[0043] A computer-readable storage medium is also provided, on which a computer program is stored, and when the computer program is executed by a processor, the steps of any of the above methods are implemented.

[0044] Compared with the prior art, the present invention has the following beneficial effects:

[0045] (1) The present invention constructs and reconstructs a database of multiple categories of mine disasters. In order to address the data imbalance and outlier problems caused by the characteristics of mine disasters themselves, the Mahalanobis distance discriminant method is used to screen outliers in the mine disaster sample data, and Tomek Link is used to optimize the SMOTE algorithm, which significantly improves the imbalance of the mine disaster dataset and improves the accuracy of the overall prediction.

[0046] (2) This invention addresses the current problems of excessive number of indicators for predicting mine disaster classification, high correlation and mutual interference between some indicators, and the lack of a relatively unified prediction indicator system. Starting from the mine disaster sample data itself, the importance and correlation coefficient of the indicators are calculated. The mine disaster indicators are screened twice, and a mine disaster classification prediction indicator system based on data analysis is established to achieve data dimensionality reduction and enhance prediction accuracy.

[0047] (3) The present invention uses machine learning and deep learning algorithms to perform hierarchical predictions on the constructed mine disaster dataset, and uses parameters such as accuracy, precision, and F1 value as prediction evaluation indicators to analyze the hierarchical prediction results of different algorithms, further improving the prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 Schematic diagram of a flow chart of an embodiment of the present invention;

[0049] FIG2( a ) is a feature importance ranking diagram of the RF algorithm used to calculate feature importance in an embodiment of the present invention;

[0050] FIG2( b ) is a feature importance ranking diagram of the XGB algorithm used to calculate feature importance in an embodiment of the present invention;

[0051] FIG2( c ) is a feature importance ranking diagram of the ET algorithm used to calculate feature importance in an embodiment of the present invention;

[0052] FIG2( d ) is a feature importance ranking diagram of the ADA algorithm used to calculate feature importance in an embodiment of the present invention;

[0053] FIG2( e ) is a feature importance ranking diagram of the GBDT algorithm used to calculate feature importance in an embodiment of the present invention;

[0054] FIG2( f ) is a feature importance ranking diagram of the DT algorithm used to calculate feature importance in an embodiment of the present invention;

[0055] Figure 3 Schematic diagram of goaf stability evaluation index in an embodiment of the present invention;

[0056] Figure 4 Schematic diagram of the SMOTE algorithm improved by Tomek Link in an embodiment of the present invention;

[0057] Figure 5 Schematic diagram of ten-fold cross validation in an embodiment of the present invention. DETAILED DESCRIPTION

[0058] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely for the purpose of explaining the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below may be combined with each other as long as they do not conflict with each other.

[0059] The technical solution of the present invention is:

[0060] A multi-disaster graded prediction method based on artificial intelligence includes the following steps:

[0061] S1. Collect mining hazard data to form a mining hazard database; the mining hazard data at least includes hanging wall stability, slope stability, rock burst, pillar stability and goaf stability;

[0062] S2. Constructing a mine disaster prediction index system: Screening the data in the mine disaster database, eliminating indicators with high feature overlap and low feature importance, and forming a dimensionality-reduced mine disaster prediction index system based on the screened data; wherein, the method for eliminating indicators with high feature overlap is to calculate their Pearson correlation coefficients, then use the weighting method in Table 1 to determine the feature overlap, and then delete the features with high overlap. See Table 1 below;

[0063] S3. Constructing a mine disaster prediction database: reconstructing the data set in the mine disaster database, processing the outliers and missing values, obtaining a balanced data set and using it as the data in the mine disaster prediction database; using the median interpolation method to process the missing values;

[0064] S4. Select evaluation features based on the indicators in the mine disaster prediction index system, perform hierarchical prediction on the data in the mine multi-disaster prediction database according to each algorithm in the preset mine disaster prediction algorithm library, summarize and compare the prediction results of the hierarchical prediction of each algorithm under different preset evaluation indicators, obtain the algorithm with the best comprehensive evaluation index score of the prediction result, and use this algorithm as the algorithm of the optimal model for mine disaster prediction;

[0065] S5. Input the mine disaster data into the optimal mine disaster prediction model to conduct multi-disaster graded prediction.

[0066] There is a strong correlation between the indicators of mine disaster data. In order to reduce the dimension of disaster data, the Pearson correlation coefficient and feature importance are used to simplify the data indicators. The specific principles are as follows:

[0067] (1) Pearson correlation coefficient principle

[0068] The Pearson correlation coefficient, also known as Pearson correlation, is a statistical indicator used to measure the strength and direction of the linear relationship between two continuous variables. Its value range is between -1 and 1. The larger the absolute value, the stronger the correlation. If the value is positive, it is a positive correlation, and if the value is negative, it is a negative correlation. Its calculation formula is as follows:

[0069]

[0070] Where cov(X,Y) represents the covariance between the two variables X and Y, σ X σ Y is the standard deviation of the variable.

[0071] (2) Principles of feature importance assessment

[0072] Taking random forests as an example, variable importance measures (VIMs) are measured using the Gini coefficient. Assume there are J features X1, X2, X3, ..., XJ, I decision trees, and C classes. For each feature Xj, its contribution to each tree in the random forest needs to be calculated and then averaged to compare the importance of each feature.

[0073] When calculating the Gini index score, the impurity change of each node in each tree is considered, which is expressed as the Gini coefficient of the node. The Gini coefficient (GI) calculation formula of node q is:

[0074]

[0075] Where C represents the number of categories, represents the proportion of category c in node q.

[0076] The importance of feature Xj in the i-th tree node q, that is, the change in the Gini coefficient before and after the node q branch is:

[0077]

[0078] in, and Respectively represent the Gini coefficients of the two new nodes after branching.

[0079] First, the feature importance of a single disaster type is calculated based on the above feature importance. Based on the requirements of the prediction system, the features with lower performance are eliminated. Then, the Pearson correlation coefficient of the remaining features is calculated.

[0080] Taking goaf data as an example, we calculated its Pearson correlation coefficient. The absolute value of the value represents the strength of the correlation between features. A negative sign indicates a negative correlation, and vice versa. A correlation coefficient with an absolute value below 0.2 is defined as no correlation, 0.2-0.3 as a moderate correlation, 0.3-0.5 as a strong correlation, and 0.5 or above as a significant correlation. * indicates a moderate correlation, ** indicates a strong correlation, and *** indicates a significant correlation.

[0081] Table 1

[0082] X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 *** 8 10 6 6 3 5 6 5 6 7 6 ** 2 0 0 3 2 2 2 5 2 1 1 * 0 0 2 1 3 2 0 0 0 2 2 Scores 2.8 3 2 2.5 1.6 2.1 2.2 2.5 2.2 2.5 2.2

[0083] The correlations between each pair of feature variables were counted by frequency, and the weights of ***, **, and * were set to 0.3, 0.2, and 0.1, respectively. The correlations between each feature and the remaining features are shown in Table 1. The correlations between the 11 indicators affecting goaf stability are strong, so it is necessary to reduce the features. As shown in Table 1, X3, X5, and X6 have scores of 2, 1.6, and 2.1, respectively, and their correlations with other features are relatively low. Therefore, they can be considered independent features for selection.

[0084] The following six algorithms with the ability to calculate feature importance are selected: Random Forest (RF), Xgboost (XGB), ExtraTrees (ET), Adaboost (ADA), Gradient Boosting Decision Trees (GBDT), and DecisionTrees (DT) to rank the importance of features affecting goaf stability. The ranking results are as follows: Figure 2(a)-Figure 2(f) The gray shaded area in the figure represents the threshold of importance of 0.10 (the sum of the importance of all indicators is 1). When the importance of a feature in the corresponding algorithm exceeds 0.1, a √ mark is made at the corresponding position in Table 2.

[0085] As can be seen from Table 2, the importance of X4, X5, X6, X7, X9, and X10 is relatively high, while the importance of X1, X2, X8, and X11 is relatively low.

[0086] Table 2

[0087] X1 X2 X3 X4 X5 X6 X7 X8 X9 X10 X11 RF √ √ √ XGB √ √ √ √ ET √ √ √ √ ADA √ √ √ √ √ GBDT √ √ √ √ √ √ DT √ √ √ √

[0088] like Figure 3 As shown, these are the indicators represented by X1-X9.

[0089] Based on the above analysis of feature correlation and importance, the goaf indicators X3, X4, X5, X6, X7, X9, and X10 are selected.

[0090] The Tomek Link pair is used to improve the SMOTE algorithm to reduce the abnormal samples generated in the reconstruction of the data set. The Tomek Link algorithm helps to identify and remove the overlapping samples in adjacent classes, thus reducing the generation rate of abnormal samples.

[0091] The main idea of the Tomek Links algorithm is that if X and Y in the samples come from different classes respectively, and there is no other sample Z such that d(X,Z) < d(X,Y) or d(Y,Z) < d(X,Y) holds, then the above samples X and sample Y are called Tomek Link pairs. Here, d represents the Euclidean distance between two samples. In Figure 4 the part under the red shadow is the Tomek Link pair, and the data pairs under the red shadow are deleted to achieve the effect of cleaning the data at the classification boundary.

[0092] The Tomek Link method can be used to undersample the synthesized samples, thereby improving the quality of the training samples. The steps for reconstructing the Smote Tomek data set are as follows:

[0093] (1) After determining the minority class according to the sample set, for each minority class sample x, calculate the Euclidean distance between it and all minority class samples to obtain the sample set of its k nearest neighbors. Subsequently, the same operation is performed on other minority class samples in turn.

[0094] (2) According to the sample ratio of the majority class samples to the minority class samples, determine the number of samples to be generated. For each minority class sample k, randomly select the corresponding number of samples from its nearest neighbor set. Suppose the selected nearest neighbor is x.

[0095] (3) For each selected nearest neighbor sample x, construct a new sample according to the following formula:

[0096]

[0097] (4) According to Figure X, construct Tomek Link pairs for the newly generated samples and the original samples, and delete the paired synthesized samples, and clean each minority class sample in turn.

[0098] (5) Repeat steps (1) to (4) until the sample categories of the reconstructed data are basically balanced.

[0099] The main problems of the mine disaster data set mainly include the existence of outliers, data set imbalance, large differences in data dimensions, etc.

[0100] To properly handle outliers in mine disaster data, the Mahalanobis distance method was used to screen the data. Unlike the Euclidean distance, it considers the connections between various characteristics and is scale-independent, meaning it is independent of the measurement scale. Therefore, the Mahalanobis distance can be independent of the measurement unit and comprehensively reflect the correlation between multiple variables. The specific calculation rules are as follows:

[0101]

[0102] For a multivariate vector with mean μ and covariance matrix Σ, the Mahalanobis distance in n-dimensional space is calculated as follows, where d is the Mahalanobis distance between each individual sample Y and the average sample μ. Using the Mahalanobis distance method, a single indicator can be regressed with the remaining indicators, and the influence of each indicator on the indicator can be comprehensively considered to determine whether there are outliers in the indicator.

[0103] The dimensions of the indicator data in the mine disaster database are inconsistent, and the range of values ​​is large, which is not conducive to the algorithm's feature learning. In order not to affect the distribution of the samples, it is considered to use Formula 6 to normalize the rock burst sample data, and use the shuffle function in python to disrupt the order of the data set. In order to ensure the reliability of the sample importance obtained by the algorithm library analysis, a 10-fold cross-validation of all samples was used for algorithm analysis. The specific operation is to divide the database sample data into 10 parts, and each part of the data is called a fold. Then, each time cross-validation is performed, one of them is selected as the test set, and the remaining 9 are used as training sets. Next, the model is trained on the training set and predicted on the test set. Repeat this process 10 times, selecting a different test set each time. Finally, the accuracy of these 10 cross-validations is calculated, and the average is taken as the final cross-validation accuracy, as shown in the schematic diagram. Figure 5 This method can avoid the randomness caused by data set partitioning and ensure the reliability of the obtained sample importance.

[0104]

[0105] The team used support vector machines, random forests, and extreme random trees from the machine learning library, as well as BP neural networks, CNNs, and LSTMs from the deep learning library, to predict five types of mine hazards in a graded manner. The PSO algorithm was also used to optimize hyperparameters in the machine learning process. Using various evaluation metrics, such as accuracy, precision, and recall, the prediction results of the different algorithms for the five types of mine hazards were summarized and compared, resulting in an optimal model for graded mine prediction.

[0106] Using the aforementioned mining disaster dataset, we used algorithms from both the machine learning and deep learning libraries to perform a hierarchical prediction of five types of mining disasters. To achieve optimal results, we used the PSO algorithm to optimize its hyperparameters.

[0107] With the continuous development of artificial intelligence, the iteration speed of algorithms is gradually accelerating. In order to ensure the effectiveness of the prediction system, it is necessary to continuously iterate the algorithms in the algorithm library, transfer the newly developed algorithms into the mine disaster prediction algorithm library, and compare them with the optimal algorithms in the algorithm library through different evaluation indicators such as accuracy, precision, and recall rate to obtain the optimal algorithm in a timely manner.

[0108] The iteration of the database is that every time new data is predicted, the data is included in the database, and then the new database is used as the mine multi-disaster prediction database when the system is used next time, thus completing the iteration of the database.

[0109] (1) Construction and reconstruction of a database of multiple categories of mine disasters. In order to solve the problem of data imbalance and outliers caused by the characteristics of mine disasters themselves, the Mahalanobis distance discriminant method is used to screen outliers in the mine disaster sample data. The SMOTE algorithm is optimized using Tomek Link, which significantly improves the imbalance of the mine disaster dataset and improves the accuracy of the overall prediction.

[0110] (2) In view of the current problems of too many indicators for predicting mine disaster classification, high correlation and mutual interference between some indicators, and the lack of a relatively unified prediction indicator system, starting from the mine disaster sample data itself, the importance and correlation coefficient of the indicators are calculated. The mine disaster indicators are screened twice, and a mine disaster classification prediction indicator system based on data analysis is established to achieve data dimensionality reduction and enhance prediction accuracy.

[0111] (3) Use machine learning and deep learning algorithms to perform graded predictions on the constructed mine disaster dataset, and use parameters such as accuracy, precision, and F1 value as prediction evaluation indicators to analyze the graded prediction results of different algorithms. In the current mine disaster prediction database, the optimal machine learning algorithm is the extreme random tree algorithm, and the optimal algorithm is the deep forest algorithm. The accuracy of the deep forest algorithm in predicting the five levels of mine disasters is as follows: the goaf risk level prediction is 92.31%; the slope stability prediction is 96.77%; the rock burst level prediction is 92.50%; the pillar stability prediction is 91.67%; and the upper wall stability prediction is 95.00%.

[0112] It will be easily understood by those skilled in the art that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A multi-disaster classification prediction method based on artificial intelligence, characterized by: The following steps are involved: S1. Collect mining hazard data to form a mining hazard database; the mining hazard data at least includes hanging wall stability, slope stability, rock burst, pillar stability and goaf stability; S2. Constructing a mine disaster prediction index system: The data in the mine disaster database is screened based on the Pearson correlation coefficient and feature importance, and indicators with high feature overlap and low feature importance are eliminated. Based on the screened data, a dimensionality-reduced mine disaster prediction index system is formed. The feature importance is output by the random forest algorithm and measured by the Gini coefficient. S3. Constructing a mine multi-hazard prediction database: The data in the mine hazard database are reconstructed using the improved SmoteTomek algorithm, and outliers and missing values ​​are processed to obtain a balanced data set, which is used as the data in the mine multi-hazard prediction database. The Smote algorithm is improved using the Tomek Link algorithm to reduce outliers generated during data set reconstruction. S4. Select evaluation features based on the indicators in the mine disaster prediction index system, perform hierarchical prediction on the data in the mine multi-disaster prediction database according to each algorithm in the preset mine disaster prediction algorithm library, summarize and compare the prediction results of the hierarchical prediction of each algorithm under different preset evaluation indicators, obtain the algorithm with the best comprehensive evaluation index score of the prediction result, and use this algorithm as the algorithm of the optimal model for mine disaster prediction; S5. Input the mine disaster data into the optimal mine disaster prediction model to conduct multi-disaster graded prediction.

2. The multi-disaster classification prediction method based on artificial intelligence according to claim 1 is characterized in that: In S2, data indicators are screened based on the Pearson correlation coefficient and feature importance. The specific steps are as follows: Calculate the characteristic importance of any disaster type in the mine disaster data and eliminate the characteristic importance The low features are selected and the Pearson correlation coefficient is calculated for the remaining features. The remaining features are weighted according to their absolute values. The larger the absolute value of the Pearson correlation coefficient between the features, the greater the weight. The Pearson correlation coefficient is used to indicate the strength of the correlation between features. The larger the absolute value of the Pearson correlation coefficient between two features, the stronger the correlation. According to the preset different feature importance algorithms, the features corresponding to the absolute values ​​of the weighted Pearson correlation coefficients are ranked by importance, and the filtered data is obtained based on the importance ranking results and the correlation between the features.

3. The multi-disaster classification prediction method based on artificial intelligence according to claim 2 is characterized in that: The Pearson correlation coefficient is a statistical indicator used to measure the strength and direction of the linear relationship between two continuous variables. Its value range is between -1 and 1. The larger the absolute value, the stronger the correlation. If the value is positive, it is a positive correlation, and if the value is negative, it is a negative correlation. Its calculation formula is expressed as: （1） in, is the Pearson correlation coefficient, represents the covariance between the two variables X and Y. are the standard deviations of variables X and Y, respectively.

4. The multi-disaster classification prediction method based on artificial intelligence according to claim 2 is characterized in that: The feature importance is output by the random forest algorithm and measured by the Gini coefficient. The principle is as follows: suppose there are J features X1, X2, X3, ..., XJ, a total of I decision trees, and C categories; for each feature Xj, calculate its contribution to each tree in the random forest, and then take the average to compare the importance of the features; When calculating the Gini index score, the impurity change of each node of each tree is considered, which is expressed as the Gini coefficient of the node; among them, the Gini coefficient of node q is The calculation formula is expressed as: （2） Where C represents the number of categories, represents the proportion of category c in node q; The importance of feature Xj in the i-th tree node q, that is, the change in the Gini coefficient before and after the node q branch for: （3） in, and Respectively represent the Gini coefficients of the two new nodes after branching.

5. The multi-disaster classification prediction method based on artificial intelligence according to claim 4 is characterized in that: The specific steps for reconstructing the data set in the mine disaster database in S3 are: The Smote algorithm is improved by Tomek Link algorithm to reduce the outliers generated in the data set reconstruction; The steps of reconstructing the data set using the improved SmoteTomek algorithm are as follows: (1) The mine disaster database is used as a sample set, and the data in it is used as samples. After determining the minority class and the majority class based on the sample set, the Euclidean distance between each minority class sample and all minority class samples is calculated to obtain its k-nearest neighbor samples; then, the same operation is performed on other minority class samples in sequence; (2) Determine the number of samples to be generated based on the ratio of majority class samples to minority class samples. For each minority class sample k, randomly select a corresponding number of samples from its nearest neighbor set. Assume that the nearest neighbor selected is x. (3) For each selected neighbor sample x, a new sample is constructed according to the following formula : （4） (4) Construct Tomek Link pairs with the new samples and their original samples, delete the paired synthetic samples, and clean up each minority class sample in turn; (5) Repeat steps (1) to (4) until the sample categories of the reconstructed data are balanced.

6. The multi-disaster classification prediction method based on artificial intelligence according to claim 5 is characterized in that: The specific method for processing data outliers in the reconstructed dataset in S3 is: The Mahalanobis distance discriminant method is used to screen the data, and its calculation method is expressed as: （5） The above formula is for a mean of μ, the covariance matrix is The Mahalanobis distance of a multivariate vector in n-dimensional space; in, is the Mahalanobis distance between each individual sample Y and the average sample μ; T represents the transpose of the matrix.

7. The multi-disaster graded prediction method based on artificial intelligence according to claim 1 is characterized in that: The following steps are also included: The algorithm analysis of the mine multi-hazard prediction database was carried out through several cross-validations to verify the reliability of the sample feature importance.

8. The multi-disaster graded prediction method based on artificial intelligence according to claim 1 is characterized in that: The mine disaster prediction algorithm library includes a machine learning algorithm library and a deep learning algorithm library. The algorithms in the mine disaster prediction algorithm library include at least: the support vector machine algorithm, random forest algorithm and extreme random tree algorithm in the machine learning algorithm library, and the BP neural network algorithm, CNN algorithm and LSTM algorithm in the deep learning algorithm library.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 8 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 8 are implemented.

Images