ISCSO-KELM-based mine water inrush source accurate identification method and system

Abstract

The invention discloses an ISCSO-KELM-based mine water inrush source accurate identification method and system, relates to the technical field of coal mine safety, and takes water chemical indexes as input of a KELM model. On the basis of chaotic mapping initialization, a mutualism strategy and a Levy flight strategy are introduced. And finally, performing automatic optimization on two key parameters, namely a kernel function parameter and a regularization coefficient of the KELM by using an ISCSO algorithm, and establishing a discrimination model based on ISCSO-KELM. According to the improved sodat swarm optimization algorithm, chaotic mapping is introduced to improve the distribution diversity of population initialization, a mutualism strategy and a Levy flight strategy are introduced, information exchange between individuals and optimal individuals is enhanced, the carelessness of mining is expanded, and the optimization precision and speed of the algorithm are improved. According to the method, the problem of the suitability of parameter selection in the KELM model is solved, and the accuracy of mine water inrush source identification is improved.

Description

Technical Field

[0001] The present invention relates to the field of coal mine safety technology, and more particularly to an ISCSO-KELM-based method and system for accurately identifying the source of water inrush in a mine. Background Art

[0002] Coal is not only China's primary energy source but also plays a vital role as a key industrial raw material. With the continuous depletion of shallow mineral resources, mining operations are increasingly shifting to deeper mines. The hydrogeological conditions of deep mines are extremely complex, especially in deep, multi-aquifer systems, which are susceptible to mine water inrush disasters due to mining disturbances. Therefore, water inrush source identification plays a vital role in mine production. Accurately and rapidly identifying the source type of water inrush in mines provides a technical basis for rescue efforts and valuable decision-making support for water resource management and sustainable ecological and environmental development.

[0003] Many scholars have conducted fruitful research on water inrush source identification. Currently, given the unique physical and chemical properties of different water inrush sources, researchers can conduct in-depth investigations into the hydrogeological properties and migration mechanisms of various water sources through hydrodynamic modeling, geological, and geochemical analysis. Water inrush source identification based on hydrochemical characteristics offers significant advantages in speed, accuracy, and cost-effectiveness, and has proven effective in practical applications. With the development of multivariate statistical analysis and computing technology, various mathematical methods (multivariate statistics, gray systems, and fuzzy mathematics) have emerged as discriminant models, along with computer-based methods such as principal component analysis, backpropagation neural networks, Bayesian discriminant methods, hierarchical analysis, multivariate logistic regression, Fisher discriminant, and support vector machines. Furthermore, various combinations of these methods are widely used for water inrush source identification. However, due to the complex hydraulic connections between groundwater aquifers and the similarity of hydrochemical properties between adjacent aquifers, these mathematical models remain challenging to accurately identify. For example, the main disadvantages of multivariate statistics, grey systems and fuzzy mathematics are that a large amount of water geochemical information may be lost; artificial neural networks require too many training samples, and the selection of training samples has a great impact on the evaluation results; the support vector machine method is very sensitive to missing values, and the modified parameters are fuzzy.

[0004] With the development of machine learning theory and technology, its strong nonlinear processing capabilities, automatic feature extraction, and robustness have gained widespread favor, showing great potential in the field of parameter identification. For example, a mine water inrush source identification model based on the particle swarm optimization (PSO)-optimized extreme gradient boosting regression tree (XGBoost) and the moth-flame algorithm (MFO)-optimized least squares support vector machine (LSSVM) has been proposed. The results demonstrate strong robustness, good real-time performance, and high accuracy. KELM (Kernel Extreme Learning Machine) is a kernel-based extreme learning machine method used for regression prediction problems. Compared with the ELM algorithm, KELM demonstrates better robustness and prediction accuracy. KELM improves identification accuracy by introducing a kernel function into the ELM algorithm. Using the kernel technique, KELM maps the input data into a high-dimensional feature space, making the originally linearly inseparable problem linearly separable. Therefore, the kernel extreme learning machine is more powerful in handling nonlinear problems.

[0005] Compared to the ELM, the KELM only requires selecting appropriate kernel parameters and regularization coefficients during training. However, the predictive performance of the KELM algorithm is constrained by these parameters. Currently, many algorithms exist for optimizing KELM model parameters, such as the Whale Optimization Algorithm (WOA), the Sand Cat Swarm Optimization (SCSO), the Dung Beetle Optimization (DBO), and the Northern Goshawk Optimization (NGO). Among them, the SCSO algorithm, introduced in 2022 as a new biomimetic population search method, offers significant advantages. The SCSO algorithm is inspired by the hearing abilities of cats and their predation and hunting habits. Using the SCSO algorithm preserves information about the search space during iterations. During the search process, all operations are performed with minimal input parameters, memory space, and processing cost. However, another advantage of this algorithm lies in its balanced occurrence rates between the exploration and exploitation phases, resulting in high conversion efficiency. However, the SCSO algorithm also exhibits a conflict between exploration and exploitation during its computational process.

[0006] Therefore, how to provide a method and system for accurately identifying the source of mine water inrush based on ISCSO-KELM is an urgent problem that needs to be solved by those skilled in the art. Summary of the Invention

[0007] In view of this, the present invention provides a method and system for accurately identifying the source of mine water inrush based on ISCSO-KELM to solve the problems in the background technology.

[0008] In order to achieve the above object, the present invention adopts the following technical solutions:

[0009] On the one hand, the present invention discloses a method for accurately identifying the source of mine water inrush based on ISCSO-KELM, comprising:

[0010] The water quality variable data is used as the input signal of the water source discrimination model, divided into a training sample set and a test sample set in proportion, and the data is normalized;

[0011] Construct a sand cat swarm optimization algorithm, initialize the basic parameters of KELM to obtain the regularization parameters and kernel function parameters, then use the chaotic map to initialize the population and calculate the fitness value;

[0012] The sand cat begins searching for prey, updates its position and calculates its fitness. Then, the mutualism strategy and Levy flight theory are introduced, and the positions of the sand cats using the two methods are updated and their corresponding fitness values ​​are calculated.

[0013] Compare the fitness values ​​of the three positions, select the better solution, and determine whether the maximum number of iterations has been reached. If so, retain the optimal regularization coefficient and kernel function parameters. If not, return to search again.

[0014] The ISCSO-KELM discriminant model established with the optimal parameters is used, and the divided training set and test set are used as input data to finally obtain the water sample classification results.

[0015] Preferably, the specific steps for obtaining the regularization parameter and the kernel function parameter are as follows:

[0016] Based on the extreme learning machine, the kernel function is introduced; the kernel function is defined as K(x, x1) and replaces the feature map h(x) corresponding to the hidden layer; through improvement, the model formula is:

[0017]

[0018] If h(x) is unknown, define the kernel matrix as follows;

[0019]

[0020] The RBF kernel function is shown in the formula:

[0021] K(x,x1)=exp(-γ‖rs‖ 2 ),γ>0;

[0022] C is the regularization parameter; γ is the kernel function parameter; the target value is T, I is the identity matrix, h(x) is defined as H, and H is the feature mapping matrix of the hidden layer.

[0023] Preferably, the specific steps of the sand cat swarm optimization algorithm are as follows:

[0024] The sensitivity range of each desert cat is defined as 0 to 2kHz. Assuming that the sensitivity of the desert cat decreases linearly from 2kHz to 0 during the process of searching for prey, the process can be expressed by the following formula:

[0025]

[0026] Among them, S M Inspired by the auditory characteristics of desert cats, t is the current iteration number and T is the maximum iteration number;

[0027] During the search process, in order to prevent the algorithm from falling into a local optimal solution, the sensitivity of each desert cat is expressed by the following formula:

[0028]

[0029] Introducing vectors Its expression is

[0030]

[0031] When |R| is greater than 1, the prey will be searched for global exploration; otherwise, the prey will be attacked for local development;

[0032] (1) Exploration: Searching for prey. When |R| is greater than 1, the desert cat will search for prey based on the current optimal position. Its current location and sensitive range Update its position to search for other better prey. The updated current position is between the current position and the prey position. The specific expression is as follows:

[0033]

[0034] (2) Development: Attacking prey. When |R| is less than or equal to 1, the desert cat will update its position by attacking the prey. The desert cat's sensitive range is represented by a circle, so a random angle θ is introduced to guide the desert cat's movement direction; the angle range is from 0 to 360 degrees; the random angle θ of each desert cat is determined by a roulette wheel selection strategy to ensure that the desert cat is close to the prey to be attacked; the update process is expressed by the following formula:

[0035]

[0036] Where, represents the optimal position, Indicates a random position.

[0037] Preferably, the improved formula for initializing the population using chaotic mapping is as follows:

[0038]

[0039] Where, represents the d-th dimension position of the i-th sand cat; ub and lb represent the search range of the population; C id Indicated by C k+1 The resulting chaotic sequence.

[0040] Preferably, the specific steps of the mutualistic symbiosis strategy are as follows:

[0041] After introducing the mutualistic symbiosis strategy, the improved formula is as follows:

[0042]

[0043] Where, is the updated position; is the position of the optimal individual; is the position of the random individual; bf is the benefit factor, indicating that it may benefit partially or completely; R MV Represents the information exchange between the optimal individual and the random individual.

[0044] Through the above technical solution, the technical effects of the present invention are:

[0045] (1) Enhanced population collaboration: The mutualism strategy simulates the collaborative relationship between biological populations, forcing individuals to exchange information with the optimal individual, avoiding the local optimal trap caused by isolated search, and accelerating global convergence. (2) Computational efficiency advantage: Compared with multi-position fitness comparison (which requires simultaneous evaluation of multiple candidate solutions), mutualism only requires adjusting the position when individuals are updated, which has lower computational complexity and is particularly suitable for high-dimensional optimization problems. (3) Experimental verification of effectiveness: This strategy improves the convergence speed of ISCSO and is effective on standard test functions.

[0046] Preferably, the specific steps of Levy flight theory are as follows:

[0047] The Levy flight strategy is introduced, and the improved formula is as follows:

[0048]

[0049] Where, is the updated individual position; represents the position of the optimal individual; represents the Levy flight position; l is a random number (0, 1); represents the position of a random individual; μ is subject to A random number, where Γ is the gamma function and β is a random number in (0, 2).

[0050] Through the above technical solution, the technical effects of the present invention are:

[0051] 1) Balance between exploration and exploitation: The long-step jumps (low-frequency, large-scale search) and short-step, intensive search (high-frequency, local exploitation) characteristics of Levy flight effectively compensate for the insufficient exploration capability of the original algorithm due to the fixed step size.

[0052] (2) Probability distribution advantage: The heavy-tailed distribution of Levy flight is more efficient than uniform random search, can cover a larger solution space at a lower computational cost, and reduce the number of invalid iterations.

[0053] (3) Adaptability: In high-dimensional complex problems such as path planning, Lévy flight can jump out of the local optimum through random long steps and fine-tune the path through short steps, which is better than the traditional fixed step length strategy.

[0054] In summary, the complementarity of the mutualistic symbiosis strategy and the Lévy flight strategy: information sharing + random jumping: mutualistic symbiosis focuses on collaboration within the population (global information sharing), while Lévy flight focuses on individual dynamic search (local random exploration). The two synergistically cover the "macro-micro" dual optimization dimensions.

[0055] On the other hand, the present invention discloses a mine water inrush source accurate identification system based on ISCSO-KELM, comprising:

[0056] The input module uses the water quality variable data as the input signal of the water source discrimination model, divides it into a training sample set and a test sample set in proportion, and normalizes the data;

[0057] The KELM module builds a sand cat swarm optimization algorithm, initializes the basic KELM parameters to obtain the regularization parameters and kernel function parameters, and then uses the chaotic map to initialize the population and calculate the fitness value;

[0058] In the ISCSO module, the sand cat begins searching for prey, updates its position, and calculates its fitness. It then introduces the mutualism strategy and Lévy flight theory, updates the positions of the sand cats using both methods, and calculates their corresponding fitness values.

[0059] The solution module compares the fitness values ​​of the three positions, selects the better solution, and determines whether the maximum number of iterations has been reached. If so, the optimal regularization coefficient and kernel function parameters are retained. If not, the search is returned and repeated.

[0060] The prediction module uses the ISCSO-KELM discriminant model established with optimal parameters, takes the divided training set and test set as input data, and finally obtains the water sample classification results.

[0061] Through the above technical solution, it can be seen that compared with the prior art, the present invention discloses a method and system for accurately identifying the source of mine water inrush based on ISCSO-KELM. In order to make the sand cat population evenly distributed in the search space, the improved sand cat swarm optimization algorithm (ISCSO) introduces chaotic mapping to improve the distribution diversity of population initialization, calculates the fitness value, introduces the mutualistic symbiosis strategy and the Levy flight strategy, updates the sand cat positions of the two methods and calculates their fitness values, strengthens the information exchange between individuals and the optimal individuals, expands the thoroughness of mining, improves the optimization accuracy and speed of the algorithm, and uses the ISCSO algorithm to optimize the kernel parameters and regularization coefficient. Compare the fitness values ​​of the three positions, select the better solution, and judge whether the maximum number of iterations is reached. If so, retain the optimal regularization coefficient and kernel function parameters. If not, return to search again; the ISCSO-KELM discrimination model established using the optimal parameters is used, and the divided training set and test set are used as input data to finally obtain the water sample classification results. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are merely embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0063] Figure 1 The ELM structure diagram provided by the present invention;

[0064] Figure 2 The optimization curve of the standard test function provided by the present invention;

[0065] Figure 3 Flowchart of the ISCSO-KELM discrimination model provided by the present invention;

[0066] Figure 4 A location map of the study area provided for this invention;

[0067] Figure 5 A vertical schematic diagram of a water-blocking layer provided by the present invention;

[0068] Figure 6 The model classification prediction results and confusion matrix diagram provided by the present invention;

[0069] Figure 7 The discrimination results of the ELM, KELM and SCSO-KELM algorithms provided by the present invention;

[0070] Figure 8This is a comparison chart of the ROC curves of the ELM, KELM, SCSO-KELM and ISCSO-KELM algorithms provided by the present invention. DETAILED DESCRIPTION

[0071] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0072] The embodiment of the present invention discloses a method for accurately identifying the source of water inrush in a mine based on ISCSO-KELM, comprising:

[0073] The water quality variable data is used as the input signal of the water source discrimination model, divided into a training sample set and a test sample set in proportion, and the data is normalized;

[0074] Construct a sand cat swarm optimization algorithm, initialize the basic parameters of KELM to obtain the regularization parameters and kernel function parameters, then use the chaotic map to initialize the population and calculate the fitness value;

[0075] The sand cat begins searching for prey, updates its position and calculates its fitness. Then, the mutualism strategy and Levy flight theory are introduced, and the positions of the sand cats using the two methods are updated and their corresponding fitness values ​​are calculated.

[0076] Compare the fitness values ​​of the three positions, select the better solution, and determine whether the maximum number of iterations has been reached. If so, retain the optimal regularization coefficient and kernel function parameters. If not, return to search again.

[0077] The ISCSO-KELM discriminant model established with the optimal parameters is used, and the divided training set and test set are used as input data to finally obtain the water sample classification results.

[0078] What you need to know is:

[0079] Extreme Learning Machine (ELM) is a learning mechanism that is faster than other similar methods. As a learning algorithm for a single hidden layer feedforward neural network (SLFN), the bias selection of hidden neurons in ELM does not require training, and the weights of these neurons are randomly set ( Figure 1In the ELM framework, although the activation function type of hidden layer neurons can be specified, the activation function of output layer neurons must be linear. The output weights are calculated using two analytical methods: normed least squares solution and Moore-Penrose pseudoinverse of general linear systems. These methods are adopted because ELM belongs to the SLFN algorithm, and the SLFN model is shown in Equation (1).

[0080] f(x)=h(x)β=Hβ (1)

[0081] In the SLFN model, f(x) represents the output of the neural network. The input sample is x, and h(x) is also defined as H, which is set as the feature map matrix of the hidden layer. Finally, β represents the weight connecting the hidden layer and the output layer, and its value is calculated as follows:

[0082]

[0083] In formula (2), the target value is T, I is the identity matrix, and C is the parameter used for regularization. Therefore, ELM can be expressed as the following formula:

[0084]

[0085] Finally, by classifying f(x), the predicted category of a sample can be obtained.

[0086] To further optimize the above technical solution, the Kernel Extreme Learning Machine (KELM) is a variant of the ELM. It is an improved version of the ELM that introduces a kernel function (K). The kernel function is defined as K(x, x1) and replaces the feature map h(x) corresponding to the hidden layer. By introducing K, the KELM becomes a learning mechanism with fewer internal parameters, with better convergence and improved generalization performance. With this modification, Equation (1) can be reformulated as Equation (4).

[0087]

[0088] If h(x) is unknown, the kernel matrix can be defined according to formula (5).

[0089]

[0090] In the relevant literature, a variety of kernel functions are commonly found. The most commonly used include the radial basis function (RBF) kernel function, the linear kernel function, and the polynomial kernel function. However, the RBF kernel function is the most commonly used, also known as the Gaussian kernel function. The popularity of the RBF kernel function is closely related to its ability to map the input space to an infinite-dimensional space. Due to the undefined constraints of the data, the RBF kernel function is an ideal choice for handling this problem. The RBF kernel function is shown in Equation (6).

[0091] K(x,x1)=exp(-γ‖rs‖ 2 ),γ>0 (6)

[0092] Finally, the regularization parameter C and the kernel parameter γ need to be carefully tuned because they affect the performance of KELM. In this case, an accurate search mechanism must be used to optimize these two parameters.

[0093] It's important to understand that the Sand Cat Swarm Optimization (SCSO) algorithm is characterized by its low number of adjustable parameters and powerful optimization capabilities. The SCSO algorithm's optimization performance surpasses seven other intelligent algorithms, including the Desert Whiptail Fish Swarm Algorithm and the Cat Swarm Optimization Algorithm. Desert cats are a species of felid mammal that inhabit the harsh environments of the Central Asian desert, the Sahara Desert in Africa, and the Arabian Peninsula, which are characterized by numerous sandy and rocky areas. Desert cats possess acute hearing, capable of detecting noises below 2kHz, and the ability to dig rapidly. These unique abilities enable desert cats to detect noise (prey movement), track prey, and successfully attack, thereby ensuring their exploration and hunting capabilities within the search space.

[0094] The algorithm simulates the predation behavior of desert cats in nature to achieve global search. Its main process is as follows.

[0095] The SCSO algorithm uses the fact that desert cats can detect noise with a frequency lower than 2kHz and defines the sensitivity range of each desert cat as 0 to 2kHz. Therefore, assuming that the sensitivity of a desert cat decreases linearly from 2kHz to 0 during the process of searching for prey, the process can be expressed by formula (7):

[0096]

[0097] Among them, S M Inspired by the auditory characteristics of desert cats, we assume its value is 2, t is the current number of iterations, and T is the maximum number of iterations.

[0098] During the search process, in order to prevent the algorithm from falling into a local optimal solution, the position of each desert cat is random, which makes the sensitivity of each desert cat to prey also random. Therefore, the sensitivity of each desert cat can be expressed by formula (8):

[0099]

[0100] The global search capability and local development capability of intelligent algorithms are crucial. Its expression is formula (9). Due to this adaptive strategy, the transition between the two stages is more balanced.

[0101]

[0102] When |R| is greater than 1, the agent will search for prey to achieve global exploration; otherwise, it will attack the prey for local exploitation.

[0103] (1) Searching for prey (exploration) When |R| is greater than 1, the desert cat will Its current location and sensitive range Update its position to search for other better prey, and the updated current position is between the current position and the prey position. The specific expression is shown in formula (10):

[0104]

[0105] (2) Attacking prey (exploitation) When |R| is less than or equal to 1, the desert cat will update its position by attacking the prey. The desert cat's sensitive range can be represented by a circle, so a random angle θ is introduced to guide the desert cat's movement direction. The angle range is from 0 to 360 degrees. The random angle θ of each desert cat is determined by a roulette wheel selection strategy to ensure that the desert cat is close to the prey it is about to attack. The update process can be expressed by formula (11):

[0106]

[0107] Where, represents the optimal position, Indicates a random position.

[0108] To further optimize the aforementioned technical solution, the Sand Cat Swarm Optimization algorithm faces the challenge of balancing exploration and exploitation during its operation. To ensure a uniform distribution of sand cats within the search space, the Improved Sand Cat Swarm Optimization (ISCSO) algorithm employs chaotic mapping techniques to enhance the diversity of the initial population distribution. Furthermore, the algorithm incorporates mutualism and the Lévy flight strategy to promote more efficient information exchange between individuals and the optimal individual, enhancing the meticulousness of exploration and ultimately accelerating the algorithm's optimization accuracy and speed.

[0109] Chaotic mapping has excellent randomness, regularity, and ergodicity. It can not only enhance the diversity of the population, but also significantly improve the algorithm's performance in terms of global search efficiency, convergence speed, and accuracy. The improved formula is as follows:

[0110]

[0111] In the formula, a is generally taken as 0.1; represents the d-th dimension position of the i-th sand cat; ub and lb represent the search range of the population; C id Represents the chaotic sequence generated by formula 12.

[0112] To further optimize the above technical solution, the sand cat swarm optimization algorithm uses a roulette wheel method to randomly determine the attack angle during prey attacks. However, this method is highly random and prone to falling into local optima. By introducing a mutualistic symbiosis strategy, by enhancing information exchange between individuals and the optimal individual, the adverse effects of the attack process can be effectively avoided, thereby improving the sand cat swarm's optimization accuracy and convergence speed. The improved formula is as follows:

[0113]

[0114] Where, is the updated position; is the position of the optimal individual; is the position of a random individual; bf is the benefit factor, randomly selected as 1 or 2, indicating that partial or full benefit may be obtained; R MV Represents the information exchange between the optimal individual and the random individual.

[0115] To further optimize the above technical solution, the Sand Cat Swarm Optimization Algorithm typically uses random angles when searching and attacking prey. This can cause some high-quality solutions to be missed, thereby reducing the algorithm's convergence accuracy. To improve the comprehensiveness of the search and reduce the negative impact of local optimal solutions, the present embodiment introduces the Levy flight strategy. This strategy not only enhances local search capabilities but also ensures the algorithm's thoroughness during the global search process. The improved formula is as follows:

[0116]

[0117] Where, is the updated individual position; represents the position of the optimal individual; represents the Levy flight position; l is a random number (0, 1); represents the position of a random individual; μ is subject to A random number, where Γ is the gamma function and β is a random number in (0, 2).

[0118] Improved Sand Cat Swarm Algorithm Performance Test and Analysis:

[0119] In order to test the overall optimization performance of the ISCSO algorithm in terms of solution accuracy and convergence speed, the traditional WOA (whale optimization algorithm), SCSO, DBO (dung beetle optimization algorithm) and NGO (northern goshawk optimization algorithm) were compared with it on 5 types of standard test functions. The benchmark functions can be divided into unimodal, multimodal and fixed-dimensional multimodal. Unimodal functions have only one global optimal solution and no local optimal solutions. On the other hand, multimodal functions also have a single global optimal solution, but may contain multiple local optimal solutions. Fixed-dimensional multimodal functions have a fixed function dimension that cannot be adjusted. The optimization algorithm uses a population size of 30 and a number of iterations of 500. Table 1 shows the relevant information of the test function, f min is the theoretical optimal value. To eliminate randomness, 50 experiments were conducted to ensure accuracy. Table 2 shows the average, maximum, and minimum values ​​of the optimal solutions of the five optimization algorithms.

[0120] Table 1 Standard test function information

[0121]

[0122]

[0123] Table 2 lists the optimization results of SCSO, WOA, DBO, NGO and ISCSO algorithms in different dimensions of five standard test functions. The ISCSO algorithm proposed in this embodiment has better overall performance than other algorithms. The most obvious effect is the multimodal test function F10, where the ISCSO algorithm reaches the theoretical optimal value of 0. This is because the ISCSO algorithm improves the sand cat position update mechanism. When the sand cat update position falls into a local optimum, the ISCSO algorithm can escape this state through an adaptive method. The diversity of the algorithm is increased, making it easier to escape the local optimal solution. In addition, the nonlinear attenuation of the sand cat sensitivity with iteration enables the parameters to effectively balance the global search and local development, thereby accelerating the convergence speed.

[0124] In summary, the ISCSO algorithm effectively coordinates global optimization and local exploitation, and improves the convergence accuracy of complex high-dimensional functions.

[0125] In order to observe the optimization performance of the algorithm more intuitively, Figure 2 The convergence process of the five algorithms for optimizing some benchmark functions is shown. Compared with other algorithms, the ISCSO algorithm has the highest convergence accuracy and speed for both unimodal and multimodal test functions. In summary, the ISCSO algorithm has advantages in convergence speed and optimization effect.

[0126] Table 2 Average, optimal and worst values ​​of different algorithms

[0127]

[0128] Model evaluation metrics:

[0129] The model selection confusion matrix and receiver operating characteristic (ROC) curve are used to evaluate classification prediction results. The basic principle is that the confusion matrix, also known as the error matrix in analysis, is a metric for assessing the accuracy of classification results. In a binary classification model, the only requirement is to determine whether the result is 0 or 1, i.e., positive or negative. The predicted value obtained by model establishment is the predicted value, while the true value recorded in the actual sample is the true value. During classification training, the actual classification results of the sample are compared with the predicted classification results of the model training. The columns represent the predicted values, and the rows represent the true values. There are generally four types of classifications: Class I: True Positives (TP): true positive classification (predicted by the model); Class II: False Positives (FP): true positive classification (predicted by the model); Class III: False Negatives (FN): true negative classification (predicted by the model); and Class IV: True Negatives (TN): true negative classification (predicted by the model). TP, FN, TN, and FP are the basic metrics used to evaluate confusion matrix models. In the confusion matrix, the horizontal direction represents the true value and the vertical direction represents the predicted value.

[0130] Based on this confusion matrix, the following related metrics can be derived: Accuracy represents the probability that the model correctly predicts a sample class, relative to the overall classification. It is calculated as the sum of all correctly predicted samples (diagonal sum) / total number of samples. Precision is specific to a particular class of interest. It focuses on the proportion of samples predicted to be of that class whose true values ​​also belong to that class, calculated as TP / (TP+FP). Similarly, recall is also specific to a particular class of interest. It focuses on the proportion of samples whose true values ​​are of that class that the model correctly predicts to be of that class, calculated as TP / (TP+FN). The F1 score is the harmonic mean of precision and recall.

[0131] ROC stands for Receiver Operating Characteristic. The ROC curve is a graph that allows you to visually observe the performance of a classifier. The horizontal axis of the ROC represents the false positive rate (FP / (FP+TN)), while the vertical axis represents the true positive rate (TP / (TP+FN)). The closer the ROC curve is to the upper left corner, the better the classifier's performance. Similarly, the AUC (Area Under the Curve) refers to the area under the curve, which is the sum of countless small rectangles. The larger the AUC value, the better the model's prediction performance.

[0132] Application examples:

[0133] (1) The study area is located in the Linhuan mining area of ​​Huaibei Coalfield ( Figure 4 The area includes the Qingdong, Linhuan, Wugou and Xutuan mines, which are rich in coal resources. It is about 26-32km long from east to west and about 15-18km wide from north to south, covering an area of ​​about 133.61km. 2 . The Huaibei area is an alluvial plain with thick, loose sediments covering the bedrock, and its paleotopography gradually descends from northeast to southwest. The region has cold and windy winters, hot and rainy summers, and mild spring and autumn, with most precipitation occurring in July and August. The average annual evaporation is about 832.4 mm. The tributaries of the Huaihe River (Dongsha River and Beifei River) are the main rivers flowing through the area. River levels are generally low in autumn and winter, and higher during spring and summer. Shallow groundwater resources are abundant and sufficient to meet the water needs of most of the population.

[0134] According to the regional strata, groundwater storage space and rock type of the water-bearing medium, the main aquifers affected by coal mining include the Quaternary porous confined aquifer (QA), the Permian fractured sandstone aquifer (PA), the Carboniferous Taiyuan Formation limestone aquifer (CFA) and the deep Ordovician limestone aquifer (OFA) ( Figure 5 ).

[0135] The minerals in the bedrock aquifers in the study area primarily include calcite, dolomite, gypsum, and some clay / silicate minerals. Coal energy primarily comes from Carboniferous-Permian strata. The QA water-bearing lithology is primarily brownish-yellow gravel, clay gravel, fine sand, and silty clayey sand, with a water table elevation of 3.99-27.57 m. The PA is composed of sandstone, mudstone, coal seams, and siltstone, and its water-yielding properties are relatively weak. The CFA aquifer is composed of limestone, mudstone, siltstone, sandstone, and thin coal seams, with limestone being the primary component. The water table elevation is 20.6-21.05 m. The OFA water-bearing lithology is thick-bedded limestone in light gray, brown, and dark gray. It is relatively well-developed in shallow aquifers and weakens toward deeper aquifers, with a static water table elevation of 25.40 m.

[0136] (2) Data collection and processing: Compared with surface water, the circulation rate of groundwater is relatively slow and is mainly controlled by long-term water-rock interaction. Therefore, the embodiment of the present invention collected 366 water samples from the Linhuan mining area between 2017 and 2022. Among them, there are 31 QA water samples, 180 PA water samples and 155 CFA water samples. All samples were obtained from surface or underground water level observation wells. Six parameters were measured in all samples, including Na + +K + , Ca 2+ Mg 2+ 、Cl - 、SO4 2- and HCO3 -All water samples were filtered on-site through a 0.45 μm membrane before testing at the Anhui University of Science and Technology Testing Center. Samples for cation analysis were stored in 550 mL polypropylene bottles and acidified with HNO3 to a pH value less than 2. Major anions, such as Cl, were analyzed by ion chromatography (Dionex 120, ThermoFisher Scientific, USA). - 、SO4 2- and HCO3 - The concentrations of major cations, such as Na + +K + , Ca 2+ and Mg 2+ The concentration of K in water samples + The concentration of Na + and K + The concentration of Na + +K + ) were analyzed. All water samples were analyzed for major ions within 24 hours of collection. The charge balance of each sample was calculated using the Aq·QA software package, and all water samples maintained good charge balance, with an error of less than 5%.

[0137] (3) Water source discrimination results: The ISCSO algorithm is used to optimize the optimal regularization coefficient and kernel function parameters of the KELM classifier, and the confusion matrix is ​​used to judge the classification accuracy of the evaluation model. At the same time, the ROC curve is used to assist in the evaluation of the model prediction results. Based on the ISCSO-KELM classification prediction model, the test set is discriminated after learning the training set. The results are as follows Figure 6 (a) and Figure 6 (b) shows the model's accuracy on the training set is 98.1%, and on the test set it is 97.3%. Based on the confusion matrix results, the three-class classification problem can be converted into three two-class classification problems: class 1 vs. other classes, class 2 vs. other classes, and class 3 vs. other classes.

[0138] For the training set, Figure 6 As shown in (c), there are 19 samples predicted to be classified as Class 1 water, and 22 samples actually classified as Class 1 water. The recall rate of Class 1 water is 100% and the precision rate is 86.4%; there are 130 samples predicted to be classified as Class 2 water, and 126 samples actually classified as Class 2 water. The recall rate of Class 2 water is 96.9% and the precision rate is 100%; there are 108 samples predicted to be classified as Class 3 water, and 109 samples actually classified as Class 3 water. The recall rate of Class 3 water is 99.1% and the precision rate is 98.2%.

[0139] For the test set, Figure 6 As shown in (d), there are 8 samples predicted to be classified as Class 1 water, and 9 samples actually classified as Class 1 water. The recall rate of Class 1 water is 100% and the precision rate is 88.9%; there are 55 samples predicted to be classified as Class 2 water, and 54 samples actually classified as Class 2 water. The recall rate of Class 2 water is 96.4% and the precision rate is 98.1%; there are 46 samples predicted to be classified as Class 3 water, and 46 samples actually classified as Class 3 water. The recall rate of Class 3 water is 97.8% and the precision rate is 97.8%.

[0140] Table 3 shows that in the test set, the precision for the water category 2 classification was the highest, at 98.1%. The precision for the water category 1 classification was lower, likely due to the smaller sample size. The recall rates for each category were similar. Furthermore, the three-level F1 score shows that the minimum F1 score was 94.1%, indicating that the model's overall prediction performance was very good.

[0141] Table 3 Confusion matrix analysis of water source discrimination in the test set

[0142]

[0143] Comparison of different model discrimination results:

[0144] In order to further verify the performance of the ISCSO-KELM model in water inrush source identification research, it is compared with the KELM, ELM and SCSO-KELM models. All samples are divided into training sets and prediction sets according to a ratio of 7:3, and are respectively introduced into the three models for classification training. The classification results are as follows: Figure 7 As shown in the figure, the ELM model achieved 99.6% accuracy on the training set and 93.6% accuracy on the test set. The KELM model achieved 98.8% accuracy on the training set and 95.4% accuracy on the test set. The SCSO-KELM model achieved 97.2% accuracy on the training set and 96.3% accuracy on the test set.

[0145] The discrimination accuracy of the ISCSO-KELM model training set is 98.05%, and the discrimination accuracy of the test set is 97.25%. By comparison, the ISCSO-KELM model proposed in the embodiment of the present invention has the best discrimination effect on the test set.

[0146] In addition, the classification effects of each model were compared by the AUC value in the ROC curve, such as Figure 8As shown in the ROC curve, a higher false positive rate indicates a higher true positive rate, indicating that the model training results closely match the actual results. The AUC values ​​for ISCSO-KELM, ELM, KELM, and SCSO-KELM are 0.97, 0.75, 0.82, and 0.915, respectively. As the larger the AUC value, the better the model's prediction performance, the ISCSO-KELM model performs best.

[0147] In summary, considering the hydrogeological conditions of the Su'nan mining area, 366 water samples were collected from three different aquifers. Six water chemical characteristic factors were selected as identification indicators, mainly including K + +Na + , Ca 2+ Mg 2+ 、Cl - 、SO4 2- and HCO3 - These characteristic indicators are used as parameters for water inrush source identification and as inputs to the KELM model. The ISCSO algorithm is then used to optimize the key parameters of the KELM model. A water source discrimination model for multiple aquifers is proposed, with the ultimate goal of improving the classification capabilities of the KELM model to quickly and accurately identify water inrush sources within mines. Finally, the model's performance is evaluated using confusion matrix-based accuracy, precision, recall, F1 score, and receiver operating characteristic (ROC) curves. A comparative analysis with other models highlights its advantages. This not only addresses the issue of appropriate parameter selection in the KELM model but also improves the accuracy of mine water inrush source identification.

[0148] (1) Based on water sample data, the SCSO algorithm was improved by introducing the mutualistic symbiosis strategy and the Levy flight strategy on the basis of chaotic map initialization. The ISCSO algorithm was compared with other optimization algorithms using the standard test function, and its performance was better than other algorithms. The KELM classifier was optimized by combining the ISCSO algorithm, and an ISCSO-KELM model suitable for the classification and prediction of water inrush sources in multi-aquifer mines was constructed, which provides a new idea for the identification of water inrush sources in mines.

[0149] (2) The ISCSO optimization algorithm was used to rapidly optimize the key parameters of the KELM model, and a mine water inrush source identification model based on the ISCSO-KELM was established. The model achieved a discrimination accuracy of 98.1% for the training samples and 97.3% for the test samples. This result highlights that the model not only has good recognition ability for the training samples, but also has good predictive performance for new samples.

[0150] (3) Compared with the traditional ELM model, the traditional KELM model, and the SCSO-KELM model, the discrimination performance of the ISCSO-KELM discriminant model is significantly improved, which proves that the model effectively utilizes the advantages of the SCSO algorithm in fast parameter optimization and the advantages of the KELM algorithm over the ELM algorithm.

[0151] Another embodiment of the present invention discloses a system for accurately identifying the source of water inrush in a mine based on ISCSO-KELM, comprising:

[0152] The input module uses the water quality variable data as the input signal of the water source discrimination model, divides it into a training sample set and a test sample set in proportion, and normalizes the data;

[0153] The KELM module builds a sand cat swarm optimization algorithm, initializes the basic KELM parameters to obtain the regularization parameters and kernel function parameters, and then uses the chaotic map to initialize the population and calculate the fitness value;

[0154] In the ISCSO module, the sand cat begins searching for prey, updates its position, and calculates its fitness. It then introduces the mutualism strategy and Lévy flight theory, updates the positions of the sand cats using both methods, and calculates their corresponding fitness values.

[0155] The solution module compares the fitness values ​​of the three positions, selects the better solution, and determines whether the maximum number of iterations has been reached. If so, the optimal regularization coefficient and kernel function parameters are retained. If not, the search is returned and repeated.

[0156] The prediction module uses the ISCSO-KELM discriminant model established with optimal parameters, takes the divided training set and test set as input data, and finally obtains the water sample classification results.

[0157] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Reference can be made to the common and similar parts between the various embodiments. For the devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method description.

[0158] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be readily apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not limited to the embodiments shown herein but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. The method for accurately identifying the source of mine water inrush based on ISCSO-KELM is characterized by: include: The water quality variable data is used as the input signal of the water source discrimination model, divided into a training sample set and a test sample set in proportion, and the data is normalized; Construct a sand cat swarm optimization algorithm, initialize the basic parameters of KELM to obtain the regularization parameters and kernel function parameters, then use the chaotic map to initialize the population and calculate the fitness value; The sand cat begins searching for prey, updates its position and calculates its fitness. Then, the mutualism strategy and Levy flight theory are introduced, and the positions of the sand cats using the two methods are updated and their corresponding fitness values ​​are calculated. Compare the fitness values ​​of the three positions, select the better solution, and determine whether the maximum number of iterations has been reached. If so, retain the optimal regularization coefficient and kernel function parameters. If not, return to search again. The ISCSO-KELM discriminant model established with the optimal parameters is used, and the divided training set and test set are used as input data to finally obtain the water sample classification results.

2. The method for accurately identifying the source of mine water inrush based on ISCSO-KELM according to claim 1, characterized in that: The specific steps to obtain the regularization parameters and kernel function parameters are as follows: Based on the extreme learning machine, the kernel function is introduced; the kernel function is defined as K(x, x1) and replaces the feature map h(x) corresponding to the hidden layer; through improvement, the model formula is: If h(x) is unknown, define the kernel matrix as follows; The RBF kernel function is shown in the formula: K(x,x1)=exp(-γ‖rs‖ 2 ),γ>0; C is the regularization parameter; γ is the kernel function parameter; the target value is T, I is the identity matrix, and h(x) is defined as H, the feature mapping matrix of the H hidden layer.

3. The method for accurately identifying the source of water inrush in mines based on ISCSO-KELM according to claim 1, characterized in that: The specific steps of the sand cat swarm optimization algorithm are as follows: The sensitivity range of each desert cat is defined as 0 to 2kHz. Assuming that the sensitivity of the desert cat decreases linearly from 2kHz to 0 during the process of searching for prey, the process can be expressed by the following formula: Among them, S M Inspired by the auditory characteristics of desert cats, t is the current iteration number and T is the maximum iteration number; During the search process, in order to prevent the algorithm from falling into a local optimal solution, the sensitivity of each desert cat is expressed by the following formula: Introducing vectors Its expression is When |R| is greater than 1, the prey will be searched for global exploration; otherwise, the prey will be attacked for local development; (1) Exploration: Searching for prey. When |R| is greater than 1, the desert cat will search for prey based on the current optimal position. Its current location and sensitive range Update its position to search for other better prey. The updated current position is between the current position and the prey position. The specific expression is as follows: (2) Development: Attacking prey. When |R| is less than or equal to 1, the desert cat will update its position by attacking the prey. The desert cat's sensitive range is represented by a circle, so a random angle θ is introduced to guide the desert cat's movement direction; the angle range is from 0 to 360 degrees; the random angle θ of each desert cat is determined by a roulette wheel selection strategy to ensure that the desert cat is close to the prey to be attacked; the update process is expressed by the following formula: Where, represents the optimal position, Indicates a random position.

4. The method for accurately identifying the source of water inrush in mines based on ISCSO-KELM according to claim 3, characterized in that: The improved formula for initializing the population using chaotic mapping is as follows: Where, represents the d-th dimension position of the i-th sand cat; ub and lb represent the search range of the population; C id Indicated by C k+1 The resulting chaotic sequence.

5. The method for accurately identifying the source of water inrush in mines based on ISCSO-KELM according to claim 3, characterized in that: The specific steps of the mutualistic symbiosis strategy are as follows: After introducing the mutualistic symbiosis strategy, the improved formula is as follows: Where, is the updated position; is the position of the optimal individual; is the position of the random individual; bf is the benefit factor, indicating that it may benefit partially or completely; R MV Represents the information exchange between the optimal individual and the random individual.

6. The method for accurately identifying the source of water inrush in a mine based on ISCSO-KELM according to claim 3, characterized in that: The specific steps of Levy's flight theory are as follows: The Levy flight strategy is introduced, and the improved formula is as follows: Where, is the updated individual position; represents the position of the optimal individual; represents the Levy flight position; l is a random number (0, 1); Represents the position of a random individual; μ is subject to μ~N A random number, where Γ is the gamma function and β is a random number in (0, 2).

7. The mine water inrush source accurate identification system based on ISCSO-KELM is characterized by: include: The input module uses the water quality variable data as the input signal of the water source discrimination model, divides it into a training sample set and a test sample set in proportion, and normalizes the data; The KELM module builds a sand cat swarm optimization algorithm, initializes the basic KELM parameters to obtain the regularization parameters and kernel function parameters, and then uses the chaotic map to initialize the population and calculate the fitness value; In the ISCSO module, the sand cat begins searching for prey, updates its position, and calculates its fitness. It then introduces the mutualism strategy and Lévy flight theory, updates the positions of the sand cats using both methods, and calculates their corresponding fitness values. The solution module compares the fitness values ​​of the three positions, selects the better solution, and determines whether the maximum number of iterations has been reached. If so, the optimal regularization coefficient and kernel function parameters are retained. If not, the search is returned and repeated. The prediction module uses the ISCSO-KELM discriminant model established with optimal parameters, takes the divided training set and test set as input data, and finally obtains the water sample classification results.

Images