Mine water inrush source discrimination method based on width learning integration algorithm

By combining the width learning system and Adaboost integrated learning algorithm, the accuracy and time deficiency of the mine inrush water source discrimination method is solved, and efficient and accurate water source discrimination is achieved, which is suitable for the rapid monitoring needs of complex mining areas.

CN118981704APending Publication Date: 2024-11-19TAIYUAN UNIVERSITY OF TECHNOLOGY
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410996085.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-24
Publication Date
2024-11-19

AI Technical Summary

Technical Problem

The existing mine water source identification methods are insufficient in accuracy and calculation time, especially in complex mining areas, and it is difficult to accurately determine the characteristics of water source. The traditional method is inefficient in training, making it difficult to meet the needs of fast and accurate water source identification.

Method used

Using a method based on width learning ensemble algorithm, the width learning system (BLS) is combined with Adaboost integrated learning. Through the incremental learning of BLS and the optimization and weighted classification of Adaboost, an efficient mine water source discrimination prediction model is constructed, and the The accuracy and time deficiency of traditional methods improve the discriminant accuracy of water source characteristics in complex mining areas and reduce model training and prediction time.

Benefits of technology

It significantly improves the accuracy of water source discrimination and the robustness of the model, optimizes the randomness of parameter selection, reduces the risk of overfitting, and makes the model more adaptable and fast in a variable water quality environment, suitable for mine safety monitoring system.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118981704A_ABST
    Figure CN118981704A_ABST
Patent Text Reader

Abstract

The invention discloses a mine water inrush source discrimination method based on a width learning integration algorithm, and the method comprises the following steps: collecting and arranging influence factors of water source discrimination, and obtaining a relatively complete data set of each factor; dividing the primarily processed data set into a test set and a training set; and selecting all training sample data, establishing a BLS model, outputting an optimized water source discrimination result through the model, optimizing the BLS model by using an Adaboost algorithm, and outputting an optimized water source discrimination result through the model. The model is higher in generalization ability under the condition that the iteration speed is guaranteed, and the robustness and accuracy of the model can be further improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to a method for distinguishing a mine water inrush source based on a width learning integrated algorithm, and belongs to the technical field of coal mining. Background Art

[0002] Coal is an important part of my country's energy. With the increasing intensity and depth of coal mining, mine water inrush accidents may occur, which are serious hazards. It is of great significance to enrich and improve the theory of mine water hazard prevention and control, realize efficient analysis and evaluation of mine water quality detection data, grasp the current status of water quality, analyze the causes of accidents, and better prevent mine water inrush accidents to ensure safe production in coal mines. The most priority step is to identify the source of water inrush.

[0003] Since the beginning of resource exploitation, people have proposed many feasible methods for identifying mine water inrush. Water temperature and water level method, water chemical characteristic ion analysis method and isotope analysis method, Fisher discriminant method, cluster analysis method, BP neural network, laser induced fluorescence technology, grey correlation method and other methods are widely used by scholars in water source identification. Among them, water temperature and water level method is only applicable to single-layer water inrush with simple hydrogeological conditions, and has poor effect on multi-aquifers with complex geological conditions; water chemical characteristic ion and isotope analysis method is applicable to water with certain differences in water quality characteristics. Mining areas, but it cannot produce good analysis results for mining areas with small differences in aquifer water quality characteristics and complex water sources; Fisher discriminant method is easy to operate and has a high success rate, but its discrimination is greatly limited by the sample itself; BP neural network has the advantages of discrimination persistence and real-time prediction, but for small sample data, the discrimination results are somewhat unsatisfactory; laser induced fluorescence technology has the advantages of online prediction, short time consumption, and accurate results in discrimination, but obtaining a discrimination model may require a large amount of standard aquifer water sample data for training, which has certain limitations.

[0004] Machine learning is the most intelligent branch of artificial intelligence, and it is widely used in various fields. At present, machine learning methods such as support vector machines and convolutional neural networks are widely used in mine water inrush source identification, but to a certain extent, there is still a problem of low accuracy. Taking a typical production mine in the central part of Fenxi Mining Area as the research object, a machine learning water source identification model based on width learning is established. This is a new learning system proposed by CLIPhilip Chen, called the Broad Learning System. It is different from deep learning, which requires adding grid layers again and again to increase the complexity of the model. The core of width learning lies in the incremental learning algorithm, which improves accuracy by changing the mapping features and the number of enhanced nodes, and then improves the model speed by solving pseudo-inverse instead of gradient descent. Therefore, this method has high training efficiency and good classification effect. BLS has been proven to have comparable performance to deep learning methods such as stacked autoencoders, deep belief networks, deep Boltzmann machines, and multi-layer perception on MNIST data, and has the fastest training speed. At present, the width learning method has not been applied to the identification of the source of water inrush in mines. However, in theory, width learning has great potential in this field. However, width learning has defects such as difficulty in parameter selection, so it is also necessary to further improve the above defects and apply them to the identification of the source of water inrush in mines.

[0005] Currently, traditional water source identification methods, such as random forest (RF), decision tree (CART) and other machine learning methods, are applied to mine water inrush source identification, but there are still problems such as low accuracy and long time. Summary of the invention

[0006] Aiming at the defects of long running time and poor accuracy of the model for distinguishing the source of water inrush in mines, the present invention provides a method for distinguishing the source of water inrush in mines based on a width learning ensemble algorithm, and establishes a prediction model for distinguishing the source of water inrush in mines based on a combination of a width learning system (BLS) and Adaboost ensemble learning. The model combines the high efficiency and incremental learning ability of BLS with the optimization and weighted classification ability of Adaboost for the first time, overcoming the shortcomings of traditional methods such as random forest (RF) and decision tree (CART) in terms of accuracy and calculation time; through the iterative learning of Adaboost, the present invention significantly improves the accuracy of distinguishing the water source characteristics of complex mining areas, and at the same time utilizes the fast learning characteristics of BLS to reduce the time required for model training and prediction. In addition, the ensemble method optimizes the adaptability and robustness of the model in a variable water quality environment, and is particularly suitable for mine safety monitoring systems that require fast and accurate water source discrimination.

[0007] The concept of the present invention is as follows: the present invention adopts a method combining BLS with Adaboost integrated learning, so as to obtain a prediction model with shorter time and higher accuracy. Among them, BLS, as a new machine learning method, can overcome the limitations and precision of other machine learning algorithms in theoretical sense by changing the mapping features and enhancing the nodes to improve the accuracy. However, since the selection parameters of the BLS model are too random, if the parameters are improperly adjusted, overfitting will be caused. The present invention adopts the Adaboost integrated algorithm to optimize the BLS model for the disadvantage of difficult parameter adjustment, so that the model prediction results are closer to the actual situation. The Adaboost algorithm, also known as reinforcement learning, can enhance weak learners with low prediction accuracy to strong learners with high prediction accuracy. This provides an effective new idea and new method for the design of the algorithm when it is very difficult to directly construct a strong learner. The algorithm principle is to adjust the sample weight and the weak classifier weight according to the prediction results of each learning machine, and select the weak classifier with the smallest weight coefficient from the trained weak classifiers to combine into a strong classifier. Weak classifiers are trained based on the training set, and each next weak classifier is trained on different weight sets of samples. The difficulty of classifying each sample determines the weight, and the difficulty of classification is estimated by the output of the classifier in the previous step.

[0008] The present invention provides a method for distinguishing the source of water inrush in a mine based on a width learning integrated algorithm, comprising the following steps: collecting and arranging the influencing factors of water source discrimination (such as calcium ions, magnesium ions, sodium ions and potassium ions, bicarbonate, sulfate, chloride ions, carbonate ions and dissolved total solids, etc.) to obtain a relatively complete data set of each factor; then dividing the data set after preliminary processing into a test set and a training set; selecting all training sample data, establishing a BLS model, and outputting the optimized water source discrimination result through the model: optimizing the BLS model using the Adaboost algorithm, and outputting the optimized water source discrimination result through the model. In addition, the present invention also verifies the effect of the above-mentioned newly created model: establishing an RF model, outputting the unoptimized water source discrimination result after determining the relevant parameters; then optimizing the RF model using the sparrow optimization algorithm to obtain the RF-SSA model, and outputting the optimized water source discrimination result through the model, and comparing the accuracy of the four models and the ROC curve to determine whether the width learning integrated algorithm is effective.

[0009] The above method specifically comprises the following steps:

[0010] S1: Data collection and data processing:

[0011] According to the previous research conclusions and data collected in the past seven years, the actual water sample data obtained by chemical analysis were analyzed. After analysis, it was found that 8 water quality indicators such as calcium ion, magnesium ion, sodium ion and potassium ion, bicarbonate, sulfate, chloride ion, carbonate ion and total dissolved solids played a great role in distinguishing water sources, so these 8 indicators were selected as identification factors; four categories were selected: 1 surface water, 2 sandstone water, 3 space water, 4 Ordovician ash water as the identification set;

[0012] S2: Building a BLS model

[0013] Extract features from the data, and randomly generate bias values ​​and corresponding weight matrices based on the data matrix obtained by S1. The bias values ​​are located in the last column of the data matrix. The two matrices are multiplied to obtain the various feature nodes of the model. Then generate enhanced nodes, introduce enhanced nodes to enhance the nonlinear factors in the network, and assign activation function tansig to the previously generated feature nodes, so that the model can better process data information, thereby having a better prediction effect on water source discrimination. Finally, the pseudo-inverse is calculated based on the obtained feature nodes, enhanced nodes, and target outputs, that is, the mapping from input to output is completed, so as to evaluate the susceptibility of water source discrimination in the region.

[0014] S3: Combined with the Adaboost ensemble algorithm: The data is first input into a weak learner BLS for classification, where the discriminant data of the classification with larger errors is assigned larger weights, and the discriminant data of the classification with smaller errors is assigned smaller weights; then the input data is updated according to the obtained weights and input into the next BLS; this process is repeated until a certain number of BLS learners are obtained; finally, the predictions of all BLS are combined into a model for linear superposition, and weighted according to the errors of their prediction results for the training set; and then a BLS model optimized by the Adaboost ensemble algorithm is constructed.

[0015] S4: Water source identification:

[0016] The water sample data to be identified is preprocessed. The preprocessed water sample data is input into the trained BLS-Adaboost integrated model. The model processes the input data and outputs the result of water source identification. The result will be classified according to the four pre-set categories (surface water, sandstone water, space water, and Ordovician ash water) to identify the specific water source type.

[0017] The following steps S5-S7 use a comparison model to further verify the effect of the technical solution of the present invention:

[0018] S5: Build RF model:

[0019] The model is defined including parameters such as the number of trees and the minimum number of leaf nodes. The default values ​​are 10 for the number of trees and 1 for the leaf node. Then, the relationship between each influencing factor and water source discrimination is obtained through the training set data, and the model is obtained.

[0020] S6: Establish SSA-RF model:

[0021] The number of trees and the minimum number of leaf points of each influencing factor in the RF model are optimized through the SSA sparrow search algorithm, and the relationship between each influencing factor and water source discrimination is obtained through the training set data to obtain the optimized model;

[0022] S7: Evaluation and analysis:

[0023] The data set (i.e., water sample data collected in the past six years) is substituted into BLS, and the prediction results obtained are compared with the RF results and SSA-RF results to evaluate and analyze whether BLS is feasible in the field of water source identification; then BLS is combined with the Adaboost integrated algorithm to obtain a set of results, and the prediction accuracy of BLS combined with the Adaboost integrated algorithm is compared. The prediction effects of the four are evaluated to determine whether BLS has better results after being optimized by the Adaboost integrated algorithm than the other three algorithms.

[0024] Beneficial effects of the present invention:

[0025] (1) The present invention adopts the breadth learning model (BLS) for the first time in the prediction of water source discrimination. On the basis of the BLS model, the Adaboost integrated algorithm is used to optimize the model, so that the model has stronger generalization ability while ensuring the iteration speed, which can further improve the robustness and accuracy of the model.

[0026] (2) The BLS model was optimized and improved using the Adaboost integrated algorithm, which improved the problem of over-fitting caused by random selection of parameters in the BLS model and improper parameter adjustment; the fitting results of the model in water source identification and prediction were made closer to the actual values.

[0027] (3) By comparing with RF and SSA-RF, the maximum AUC value of the ensemble model based on BLS is 0.91, indicating the highest accuracy and best prediction effect. This shows that BLS and the ensemble learning model based on BLS are feasible in the field of water source identification, and the time required for the ensemble model based on BLS is only slightly improved compared with BLS, and the ensemble width learning method can be implemented on the pre-trained BLS network. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 Represents the BLS network model structure;

[0029] Figure 2Represents the BLS network model process;

[0030] Figure 3 Represents the Adaboost network model structure;

[0031] Figure 4 Represents the BLS-Adaboost integrated model process;

[0032] Figure 5 Represents the BLS-Adaboost integrated model structure;

[0033] Figure 6 Represents the RF model structure;

[0034] Figure 7 Represents the SSA-RF model process;

[0035] Figure 8 Represents the ROC curve corresponding to each model. DETAILED DESCRIPTION

[0036] The present invention is further illustrated by the following examples, but is not limited to the following examples.

[0037] The present invention will be described in detail below in conjunction with the accompanying drawings and specific embodiments. The exemplary embodiments and descriptions of the present invention are used to explain the present invention but are not intended to limit the present invention.

[0038] The method for identifying the source of water inrush in a mine based on a width learning integrated algorithm provided by the embodiment of the present invention has the following process: Figure 4 As shown, the following steps are included:

[0039] Step 1: According to the previous research conclusions and data, the water sample data from 2016 to 2021, as well as the water samples collected in the study area in 2022, and the water sample data obtained by chemical analysis were analyzed. After analysis, it was found that 8 water quality indicators, including calcium ions, magnesium ions, sodium ions and potassium ions, bicarbonate, sulfate, chloride ions, carbonate ions and total dissolved solids, played a huge role in distinguishing water sources, so these 8 indicators were selected as identification factors. Four categories, 1 surface water, 2 sandstone water, 3 space water, and 4 Ordovician ash water, were taken as the identification set.

[0040] Step 2: Eliminate the dimension of the data collected in step 1 and divide it into training set D train and the test set D test .

[0041] Step 3: Build the first BLS base learner

[0042] The training set D train The input matrix R r×8First, the input data X' is processed to obtain the feature node Z, and then the enhanced node H is obtained from the feature node. The feature node and the enhanced node are combined into the input A = [Z|H]. The process from X to map features and the process from map features to enhanced nodes can be expressed as follows.

[0043]

[0044] Among them, Z i : the i-th feature of the feature node; X': the preprocessed input data matrix; Mapping function; W ei : The weight matrix from the input layer to the feature node layer; β ei : Bias of feature node layer; H j : The jth enhanced feature of the enhanced node; Z n : feature node matrix; ξ: activation function; W hj : The weight matrix from the feature node layer to the enhanced node layer; β hj : Enhance the bias of the node layer;

[0045] n groups of map features form a feature node layer, denoted as Z n =[Z 1 ,Z 2 ,L,Z n ], m groups of enhancement nodes form an enhancement layer denoted as H m =[H 1 ,H 2 ,L,H m The weights and biases corresponding to the map features and enhancement nodes are randomly generated. The weight between the output and A is defined as W n , the output can be expressed as:

[0046] Y=[Z 1 ,L,Z n |H 1 ,L,H m ]W n =[Z n |H m ]W n =AW n

[0047] Among them, W n =[Z n |H m ] + Y=A + Y is the result that needs to be obtained through training and is the key to prediction. n It can be quickly calculated using the following equation.

[0048]

[0049] Where A: input matrix; A + : Pseudo-inverse matrix, used to calculate the weight matrix W n ; λ: regularization parameter; I: identity matrix; A T : Transposed matrix of input matrix A.

[0050] Step 3: Build RF model

[0051] A data set of size K is randomly sampled without replacement N times, and one sample is selected each time. Then, the sampled data is used to form a training sample set using the selected N training samples. When splitting each node of the decision tree, m attributes are randomly selected from the M attributes of each sample, satisfying the condition m<<M, which helps to avoid excessive attention to noise during model training, thereby improving the model's predictive ability for new data; select a feature A from the randomly selected m attributes for partitioning. Feature A refers to the specific attribute used to split the data set into two subsets; for the selected feature A, select the best partition point from all possible partition points s; then select a partition point s from the m attributes as the node's splitting attribute according to the following formula;

[0052] min[min∑(y i -v 1 ) 2 +min∑(y i -v 2 ) 2 ]

[0053] Among them, y i Indicates the specific value of sample i on a certain attribute; v 1 represents the average value of all samples in the subset to the left of the current split point on this attribute; v 2 Represents the average value of this attribute for all samples in the subset to the left of the current split point.

[0054] In the process of forming the decision tree, each node is split according to the above steps until it cannot be split. It is worth noting that no pruning is performed during the formation of the entire decision tree. A large number of decision trees are established as required to form a random forest, and the results of all decision trees are averaged to determine the final result. The number of trees and the minimum number of leaf points in the RF random forest are used as the parameters to be optimized for the SSA algorithm in step 4.

[0055] Step 4: Establish the SSA-RF model.

[0056] In the SSA algorithm, the mean square error between the predicted value and the true value obtained by the RF model is used as the fitness of the optimization algorithm. The process of repeatedly calculating the fitness ~ individual historical optimal solution and group historical optimal solution is continuously updated through iteration. When a suitable number of iterations is set, the position of each sparrow will change with the iteration of the sparrow group according to the individual historical optimal value and the global optimal value, so that the sparrow population gradually approaches the optimal solution. The collective algorithm process is as follows Figure 2 .

[0057] The number of trees and the minimum number of leaf points obtained by the sparrow algorithm are transferred into the RF model, and then the test set D test Input into the SSA-RF model to obtain the optimized prediction results.

[0058] Step 5: Construct the Adaboost ensemble algorithm based on BLS

[0059] Water source identification is a four-classification problem, the output is {1, 2, 3, 4}, and the weak classifier of the Kth round is G k (x).

[0060] S51: Calculate the classification error of the weak classifier

[0061] The weighted error on the training set is: The symbol I represents the indicator function.

[0062] S52: Calculation of weak learning weight coefficient α:

[0063] From this formula, we can see that the smaller the learner error rate, the larger the corresponding weight coefficient.

[0064] S53: Weight update of the next round of samples

[0065] Where Z k is the normalization factor that makes the sum of the sample weights of each round of training data set equal to 1.

[0066] When the sample is misclassified, G k (x)y(i)=-1, then w k+1,i >w k,i , the weight of the error classification sample increases; when the sample is in the correct classification situation, G k (x)y(i)=1, then w k+1,i >w k,i , the weight of the correctly classified sample decreases.

[0067] S54: Perform weighted superposition on all classifiers to construct the final classifier: Then substitute the data to get the final classification result.

[0068] Step 6: Use the prediction results of the four model test sets to compare the accuracy and ROC curve with the true value to evaluate the four groups of results; accuracy refers to the ratio of the number of correctly classified by the model to the total number of samples. It can be expressed as the following formula:

[0069]

[0070] AUC is also a common method in evaluation. The ROC curve is obtained by plotting the negative prediction rate on the x-axis and the correct positive prediction rate on the y-axis. The ROC curve reflects the change in learner performance after threshold adjustment. The area under the ROC curve (AUC) is an indicator of model reliability. The calculation formula of AUC is as follows:

[0071]

[0072] According to the above formula, this study used AUC value and ROC curve to estimate the performance of BLS and BLS-Adaboost in the field of water source identification, and then verified it in the examples.

[0073] Embodiment:

[0074] Description of the study area

[0075] The Fenxi mining area is located in the central part of Shanxi Province. The three coal mines, Zhengyi Coal Mine, Jiaguo Coal Mine and Heshan Coal Mine, are studied. The mining areas are relatively close to each other. Based on the working face profile and geological conditions of the coal mines, this method is used to identify water sources. The specific steps are as follows:

[0076] Step 1) According to the selected working surface and previous research conclusions, this patent has found through analysis that 8 water quality indicators such as calcium ions, magnesium ions, sodium ions and potassium ions, bicarbonate, sulfate, chloride ions, carbonate ions, and total dissolved solids have a huge effect on distinguishing water sources, so these 8 indicators are selected as identification factors. Take four categories 1 surface water, 2 sandstone water, 3 space water, and 4 Ordovician ash water as the identification set. A total of 51 sets of sample data for water source identification were collected. Among them, 39 sets of data were used as training sets, and the remaining 12 sets of data were used as test sets.

[0077] Step 2) BLS model

[0078] First, extract the features of the input data to obtain the feature layer Z, and use the feature layer to directly calculate the enhancement layer H. The feature layer and all nodes of the enhancement layer are combined into the input A = [Z|H]. Since the training data Y is known, W n =A + Y is the result that needs to be obtained through training and is the key to prediction. nIt can be quickly calculated using the following equation.

[0079]

[0080] Step 3) RF model.

[0081] First, the RF model structure is initialized, the number of trees is set to 20, the minimum number of leaf points is randomly generated, and the mean square error between the predicted value and the true value is used as the fitness of the optimization algorithm.

[0082] Step 4) SSA-RF model

[0083] Repeat the process of calculating fitness ~ individual historical optimal solution and group historical optimal solution, and continuously adjust the position of each sparrow until 30 iterations are reached or the fitness value reaches the termination threshold. After the termination condition is met, the position of the current optimal sparrow is output as the number of trees and the estimated position of the minimum number of leaf points of this influencing factor, and the optimal value is output. That is, the parameters of the random forest are updated to obtain the optimized SSA-RF model, and 12 predicted water source discrimination results are output, and the results are reverse normalized to obtain the final result.

[0084] Step 5) Construct an Adaboost ensemble algorithm based on BLS as the base learner

[0085] BLS is used as a base learner in Adaboost, and all BLS predictions are combined into a model for linear superposition, and weighted according to the errors of their prediction results for the training set. Then a BLS model optimized by the Adaboost ensemble algorithm is constructed. The model steps are as follows: Figure 7 As shown in the figure, the Adaboost ensemble algorithm model with BLS as the base learner is obtained, and the corresponding prediction results are obtained by inserting the training set into the model.

[0086] Step 6) Model performance evaluation

[0087] Water source identification is a four-classification problem. This paper uses binary to convert the four classifications into two categories, that is, class one is recorded as [1, 0, 0, 0], class two is recorded as [0, 1, 0, 0], class three is recorded as [0, 0, 1, 0], and class four is recorded as [0, 0, 0, 1]. The model is used to calculate the probability of the classification. By adjusting the threshold, the corresponding TP, TN, FN, and FP values ​​are calculated using the classification predicted by each model, and then the accuracy, ROC curve and AUC value are calculated to evaluate the model performance. The calculation formula is as follows:

[0088]

[0089] In this study, we use the accuracy rate to test the test set D testThe evaluation results of the four models are shown in Table 3. In terms of overall accuracy, BLS-Adaboost performs best (accuracy is 91.7%), followed by BLS (accuracy is 83.3%) and SSA-RF (accuracy is 83.3%), and finally RF (accuracy is 75%).

[0090] Table 1. Some water source identification samples

[0091]

[0092] Table 2. Classification of water sources

[0093] Water source type Surface water Sandstone Water Space Water Aohui Water Serial number 1 2 3 4

[0094] Table 3. Model index evaluation results

[0095] Model BLS-Adaboost BLS SSA-RF RF Accuracy 91.7% 83.3% 83.3% 75% AUC value 90.4% 85.7% 87.12% 83.7%

[0096] ROC-AUC is another evaluation method. Figure 8 As shown. Figure 8 In the figure, the ROC curves of the BLS model and the BLS-Adaboost model on the validation data set are represented by green and red, respectively, and the ROC curves of the RF model and the SSA-RF model on the validation data set are represented by orange and black, respectively. It is concluded that the AUC value of BLS-Adaboost (AUC-0.904) is the highest, followed by SSA-RF (AUC-0.8712) and BLS (AUC-0.857), and finally RF (AUC-0.837).

[0097] In this study, we proposed a BLS-based and a BLS-based integrated learner. Both the individual BLS learner and the BLS-Adaboost learner outperformed the sparrow-optimized random forest (SSA-RF) and random forest (RF) based on statistical evaluation and ROC curve evaluation. The results show that the BLS integrated algorithm has great potential in the identification of water inrush sources in mines.

[0098] The conclusions of this embodiment can be summarized as follows: (1) After a mine water inrush accident occurs, it is crucial to quickly grasp the source and channel of the water inrush for the formulation of a post-disaster management plan. In view of the complexity of the water inrush mechanism, the Adaboost integrated algorithm recognition model based on BLS provides a solution to this problem; (2) This study found that eight water quality indicators, including calcium ions, magnesium ions, sodium ions and potassium ions, bicarbonate, sulfate, chloride ions, carbonate ions and total dissolved solids, play a huge role in distinguishing water sources, so these eight indicators are selected as identification factors; (3) Statistical methods and AUC are used to evaluate and compare the four models of RF, SSA-RF, BLS and BLS-Adaboost. All four methods have good results, but it is obvious that the integrated BLS method proposed in this paper has the best performance compared with the other three.

[0099] In this study, a total of 8 influencing factors were used for prediction based on the results of the importance evaluation of influencing factors. All four models performed well on the validation dataset, especially the integrated model based on BLS as the learner had the highest AUC value of 0.904, which was significantly improved compared with BLS, SSA-RF, and RF. These results show that BLS and the integrated model based on BLS are effective methods for evaluating and identifying water source types. In addition, the time required for the integrated model based on BLS is only slightly more than that of BLS, and the integrated BLS model can be easily implemented on the pre-trained BLS network.

Claims

1. A method for identifying the source of water inrush in a mine based on a width learning ensemble algorithm, characterized in that The following steps are involved: Collect and organize the influencing factors of water source identification to obtain a more complete data set of each factor; After that, the preliminarily processed data set is divided into a test set and a training set; All training sample data are selected to establish a BLS model, and the optimized water source identification results are output through the model; the BLS model is optimized using the Adaboost algorithm, and the optimized water source identification results are output through the model.

2. The method for identifying the source of water inrush in a mine based on a width learning ensemble algorithm according to claim 1 is characterized in that The steps include: S1: Data collection and data processing: Based on previous research conclusions and information, we collected actual water sample data from the past seven years, conducted chemical analysis, and conducted in-depth analysis of the water sample data obtained; S2: Building a BLS model The data is feature extracted. The bias value and the corresponding weight matrix are randomly generated according to the data matrix obtained by S1. The bias value is located in the last column of the data matrix. The two matrices are multiplied to obtain the characteristic nodes of the model. Then, the enhanced nodes are generated, and the nonlinear factors in the enhanced network are enhanced by introducing the enhanced nodes. The activation function tansig is assigned to the previously generated characteristic nodes, so that the model can better process the data information, thereby having a better prediction effect on water source discrimination. Finally, the pseudo-inverse is obtained according to the obtained characteristic nodes, enhanced nodes and target outputs, that is, the mapping from input to output is completed, so as to evaluate the susceptibility of water source discrimination in the region. S3: Combined with Adaboost ensemble algorithm: The data is first input into a weak learner BLS for classification, where the discriminant data of the classification with larger errors is assigned larger weights, and the discriminant data of the classification with smaller errors is assigned smaller weights; then the input data is updated according to the obtained weights and input into the next BLS; this process is repeated until a certain number of BLS learners are obtained; finally, the predictions of all BLS are combined into a model for linear superposition, and weighted according to the errors of their prediction results for the training set; and then a BLS model optimized by the Adaboost ensemble algorithm is constructed; S4: Water source identification: The water sample data to be identified are preprocessed; the preprocessed water sample data are input into the trained BLS-Adaboost integrated model; the model processes the input data and outputs the result of water source identification; the result will be classified according to the four pre-set types: surface water, sandstone water, space water, and Ordovician ash water to identify the specific water source type.

3. The method for identifying the source of water inrush in a mine based on a width learning ensemble algorithm according to claim 2 is characterized in that In step S1, after analysis, it was found that eight water quality indicators, including calcium ions, magnesium ions, sodium ions and potassium ions, bicarbonate, sulfate, chloride ions, carbonate ions and total dissolved solids, played a huge role in identifying water sources. Therefore, these eight indicators were selected as identification factors; four categories were taken: 1 for surface water, 2 for sandstone water, 3 for space water, and 4 for Ordovician ash water as the identification set.

4. The method for identifying the source of water inrush in a mine based on a width learning ensemble algorithm according to claim 2 is characterized in that The specific method of step S2 is: The data collected in step S1 is dimensionally eliminated and divided into training set D train and the test set D test ; The training set D train The input matrix R r×8 ; First, the input data X′ is processed to obtain the feature node Z, and then the enhanced node H is obtained from the feature node; the feature node and the enhanced node are combined into the input A = [Z|H], and the process from X to map features and from map features to enhanced nodes are respectively expressed by the following formulas: H j =ξ(Z n W hj +β hj ),j=1,2L,m Among them, Z i : the i-th feature of the feature node; X': the preprocessed input data matrix; Mapping function; W ei : The weight matrix from the input layer to the feature node layer; β ei : Bias of feature node layer; H j : The jth enhanced feature of the enhanced node; Z n : feature node matrix; ξ: activation function; W hj : The weight matrix from the feature node layer to the enhanced node layer; β hj : Enhance the bias of the node layer; n groups of map features form a feature node layer, denoted as Z n =[Z1,Z2,L,Z n ], m groups of enhancement nodes form an enhancement layer, denoted as H m =[H1,H2,L,H m ]; the weights and biases corresponding to the map features and enhancement nodes are randomly generated; the weight between the output and A is defined as W n , the output is expressed as: Y=[Z1,L,Z n |H1,L,H m ]IN n =[From n |H m ]IN n =AW n Among them, W n =[Z n |H m ] + Y=A + Y is the result that needs to be obtained through training and is the key to prediction. n It can be quickly calculated using the following equation: Where A: input matrix; A + : Pseudo-inverse matrix, used to calculate the weight matrix W n ; λ: regularization parameter; I: identity matrix; A T : Transposed matrix of input matrix A.

5. The method for identifying the source of water inrush in a mine based on a width learning ensemble algorithm according to claim 2 is characterized in that The specific method of step S3 is: Construct an Adaboost ensemble algorithm based on BLS: Water source identification is a four-classification problem, the output is {1, 2, 3, 4}, and the weak classifier of the Kth round is G k (x); (1) Calculate the classification error of the weak classifier The weighted error on the training set is: Wherein, the symbol I represents the indicator function; k : The classification error of the k-th round weak classifier; P(G k (x i )≠y i ): probability of classification error; w ki : The weight of the i-th sample in the k-th round; G k : kth weak classifier; x i : the i-th training sample; y i : The true label of the i-th training sample; (2) Calculation of weak learning weight coefficient α: From this formula, we can see that the smaller the error rate of the learner, the larger the corresponding weight coefficient; (3) Weight update of the next round of samples Where Z k is the normalization factor, which makes the sum of the sample weights of each round of training data set equal to 1; When the sample is misclassified, G k (x)y(i)=-1, then w k+1,i >w k,i , the weight of the error classification sample increases; when the sample is in the correct classification situation, G k (x)y(i)=1, then w k+1,i >w k,i , then the weight of the correctly classified sample decreases; (4) Perform weighted superposition on all classifiers to construct the final classifier: Then substitute the data to get the final classification result.