A method and system for identifying deep mine microseismic signals or blasting signals

Abstract

The embodiment of the application relates to the technical field of mine microseismic monitoring and intelligent identification, and particularly relates to a deep mine microseismic signal or blasting signal identification method and system. The method comprises: inputting the collected mine vibration signal into a trained signal identification model to obtain a signal identification result; the signal identification result comprises: the mine vibration signal is a mine microseismic signal or a mine blasting signal. Through the improved multi-universe optimization algorithm set in the signal identification model, the mine vibration signal identification can not only improve the candidate solution quality and population diversity, but also accelerate the algorithm convergence speed and reduce the probability of falling into local optimization, while ensuring that the key features are effectively screened out and missed as little as possible, thereby improving the accuracy and stability of deep mine microseismic signal or blasting signal identification.

Description

Technical Field

[0001] This invention relates to the field of mine microseismic monitoring and intelligent identification technology, and in particular to a method and system for identifying microseismic signals or blasting signals in deep mines. Background Technology

[0002] With the continuous exploitation and gradual depletion of shallow mineral resources, mining operations are increasingly extending to deeper areas. Deep mines typically possess complex environmental characteristics such as high ground stress, deep wells, and high ground temperatures. Engineering disturbances such as drilling and blasting can easily induce disasters such as rock mass fracturing, goaf collapse, roof falls, and rock bursts. Therefore, how to monitor the stability of deep rock masses in real time and ensure safe mine production has become a key technical issue in deep mining.

[0003] Microseismic monitoring technology, by acquiring elastic wave signals released during rock deformation and fracturing, can effectively reflect the internal damage evolution process of rock masses and is an important means of monitoring and early warning of disasters in deep mines. However, in practical applications at the mine site, the signals acquired by microseismic monitoring systems are often mixed with blasting signals and noise signals, and microseismic signals and blasting signals overlap in some waveform characteristics and frequency distributions, making it difficult to accurately distinguish between the two. This affects the reliability of microseismic monitoring results for rock mass stability analysis and disaster early warning.

[0004] Currently, the identification of microseismic and blasting signals typically employs manual judgment, statistical analysis, or machine learning methods. However, each of these three categories of identification techniques has its own shortcomings. Specifically, manual judgment is inefficient and highly subjective; statistical analysis methods suffer from parameter selection dependence on experience; while machine learning methods can improve automatic identification capabilities, deep mine microseismic or blasting data are often characterized by high dimensionality, strong correlation, and numerous redundant features. Directly using raw features for classification can easily lead to low classification accuracy, poor generalization ability, and high computational complexity. Existing advanced identification technology is feature recognition based on swarm intelligence optimization. Although this technology can improve identification performance to some extent, its algorithms still suffer from slow convergence speed, insufficient solution quality, and a tendency to get trapped in local optima, resulting in the omission of key features and low identification accuracy during final signal identification. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for identifying microseismic signals or blasting signals in deep mines, which solves the problems of slow convergence speed, insufficient solution quality, and easy getting trapped in local optima in the existing technology, resulting in the omission of key features and low identification accuracy when finally identifying the signal.

[0006] To achieve the above objectives, the present invention provides a method for identifying microseismic signals or blasting signals in deep mines, the method comprising: The collected mine vibration signals are input into the trained signal recognition model to obtain the signal recognition results; The signal identification results include: the mine vibration signal is a mine micro-vibration signal or a mine blasting signal; The training methods for the signal recognition model include: The collected mine vibration signals are input into the trained signal recognition model to obtain the signal recognition results; The signal identification result is: the mine vibration signal is a mine micro-vibration signal or a mine blasting signal; The training methods for the signal recognition model include: S1: Extract features from the mine vibration signals in the acquired dataset to obtain the original feature set; S2: Determine the feature dimension and initial population size based on the original feature set; randomly initialize the initial population based on the feature dimension and the initial population size, and determine the optimal individual in the initial population; S3: Convert the individual vector of each individual in the initial population into a binary vector, calculate the individual fitness value of each individual, and obtain the binary feature subset corresponding to each individual in the initial population; S4: Perform global and local searches on the binary feature subset based on the multiverse optimization algorithm, and update the current individual's position for the first time based on the global and local search results; S5: After the initial position update, the position of the current individual is updated again using an orthogonal learning strategy; S6: After the current round of position update is completed, the initial population is updated based on the updated current individual using a quasi-reflective learning strategy; S7: Convert the individual vector of each individual in the updated population into a binary vector, calculate the individual fitness value of each individual vector, and obtain the binary feature subset corresponding to each individual in the updated population. S8: If the updated population meets the preset stopping condition, output the binary feature subset corresponding to the best individual in the updated population; otherwise, determine the updated population as the initial population and repeat S3-S7. The trained signal recognition model is obtained based on the binary feature subset corresponding to the optimal individual.

[0007] Beneficial Effects: This invention provides a method for identifying microseismic signals or blasting signals in deep mines. This method uses a signal recognition model to identify the collected mine vibration signals, distinguishing between microseismic signals and blasting signals. The signal recognition model incorporates an improved multi-verse optimization algorithm; this algorithm is a fusion algorithm obtained by integrating orthogonal learning, quasi-reflective learning, and binarized feature selection based on the traditional multi-verse optimization (MVO) algorithm. The orthogonal learning strategy updates the current individual; this strategy performs local searches near the current solution while simultaneously exploring other regions of the solution space using orthogonal design. This combination allows the algorithm to deeply explore potential optimal solutions near the current solution while avoiding getting trapped in local optima, thus constructing globally better candidate solutions and improving solution quality. The quasi-reflective learning strategy updates the population; this strategy utilizes information from the global optimal solution when generating candidate solutions, helping the algorithm to identify potential solution space regions more quickly, thereby accelerating convergence. Meanwhile, through reflection operations, new solutions different from both the current solution and the optimal solution can be generated. These new solutions may lie in regions of the solution space that have not been fully searched, thus increasing the diversity of the population. Binary feature selection, on the other hand, is a strategy that filters individual vectors after binary vector transformation; this strategy maps individuals in continuous space to binary individuals before selection, effectively filtering out the key features that contribute most to the classification of microseismic and blasting signals, and reducing interference from redundant features.

[0008] In summary, the improved multiverse optimization algorithm set in the signal recognition model for mine vibration signal identification can not only improve the quality and diversity of candidate solutions, but also accelerate the convergence speed of the algorithm, reduce the probability of getting trapped in local optima, and ensure the effective screening of key features, minimizing their omission, thereby improving the accuracy and stability of deep mine microseismic signals or blasting signals identification.

[0009] Furthermore, the method of updating the current individual using an orthogonal learning strategy includes: Randomly select at least two individuals from the initial population, and construct a guiding vector based on the at least two individuals; A search interval is constructed based on each of the at least two individuals and the guiding vector. At least two candidate solutions are constructed using an orthogonal array. The candidate solution with better fitness is retained as the updated current individual based on the factor analysis results. The orthogonal array is determined based on the number of levels, factors, and candidate solutions set in the orthogonal learning strategy; The factor is a feature dimension of the individual vector; the number of factors is the number of the feature dimensions. The level is the discrete value of the factor within the search interval; the number of levels is the number of the discrete values. The discrete values ​​are obtained by discretizing the search interval; The number of candidate solutions is determined based on the structure of the orthogonal array and is consistent with the number of rows in the orthogonal array; The methods for obtaining the factor analysis results include: Based on the fitness values ​​of each candidate solution in each factor and the fitness performance of each factor at each level, the effect value of each factor at each level is obtained, and the effect value is determined as the result of factor analysis.

[0010] Furthermore, the method of updating the population using a quasi-reflexive learning strategy includes: Determine the quasi-reflection point and its fitness value of the initial population, compare the fitness value of the current individual with the fitness value of the quasi-reflection point, retain the individual with the better fitness value to update the initial population, and obtain the updated population.

[0011] Further, calculating the individual fitness value of each individual includes: The collected mine vibration signals are filtered based on the binary feature subset corresponding to each individual, and the filtering results are input into the fuzzy K-nearest neighbor model for classification. Based on the distance relationship between training samples and test samples in the fuzzy K-nearest neighbor model, the membership degree of the test sample to the mine microseismic signal and the mine blasting signal is calculated, and the calculation results are obtained. Based on the maximum membership principle and the calculation results, the test sample is identified as a mine microseismic signal or a mine blasting signal. The identification result is compared with the actual signal type of the test sample. If they match, the identification is considered correct. The classification accuracy is obtained by dividing the number of correctly identified test samples by the total number of test samples. The classification error rate is obtained by using the fuzzy K-nearest neighbor model based on the difference between 1 and the classification accuracy. The classification error rate and the feature ratio are weighted and summed according to a predetermined weight coefficient to obtain the individual fitness value of each individual; the weight coefficient includes a classification error rate weight coefficient and a feature ratio weight coefficient; the classification error rate weight coefficient is greater than the feature ratio weight coefficient.

[0012] Furthermore, it also includes: The maximum number of iterations is determined based on the original feature set; If the number of iterations corresponding to the updated population is greater than the maximum number of iterations, then the updated population is determined to satisfy the stopping condition.

[0013] Further, the step of converting each individual vector in the current population into a binary vector includes: The transformation function value is obtained by combining the individual vector of each individual in the current population with the transformation function; the transformation function is: in, The value of the transformation function; The first in the current population Individual; For the first The first individual Dimensional value; If the value of the transformation function is greater than the set threshold, then the binary vector of the current individual is determined to be 1; If the value of the transformation function is less than or equal to the set threshold, then the binary vector of the current individual is determined to be 0.

[0014] Furthermore, a global and local search is performed on the binary feature subset based on the multiverse optimization algorithm, and the current individual's position is updated for the first time based on the global and local search results, including: The white hole or black hole mechanism in the multiverse optimization algorithm drives the exchange of positional component information between different individuals in the current population, so as to perform global updates between each binary feature subset; each binary feature subset is determined according to each candidate solution and corresponds one-to-one with each candidate solution. The wormhole mechanism in the multiverse optimization algorithm is used to perturb and update the current individual based on the position of the optimal individual, so that some dimensions of the current individual move closer to the optimal individual or are finely adjusted within the neighborhood of the optimal individual, thereby updating the local position of the binary feature subset.

[0015] Furthermore, the original feature set includes: color moment features, gray-level co-occurrence matrix texture features, local texture features, and shape features.

[0016] Furthermore, the method also includes: Extract the number of co-occurrences of gray values ​​of two pixels separated by a given distance and a given direction in a statistical grayscale image, and obtain the gray-level co-occurrence matrix texture features based on the number of occurrences. The local texture features are extracted using a local binary mode. The shape features are extracted using shape feature parameters, including area, perimeter, expansion, and eccentricity. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating the overall process of identifying microseismic signals or blasting signals in deep mines, as described in an embodiment of the present invention. Figure 2 This is a flowchart illustrating the method for identifying microseismic signals or blasting signals in deep mines, as described in an embodiment of the present invention. Figure 3 Box plots show the performance indicators of various algorithms in the embodiments of the method for identifying microseismic signals or blasting signals in deep mines according to the present invention. Detailed Implementation

[0019] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, the accompanying drawings show only the parts relevant to the present invention, and not all of the structures.

[0020] Implementation of methods for identifying microseismic signals or blasting signals in deep mines This embodiment presents a technical solution for identifying microseismic signals or blasting signals in deep mines. The method inputs the collected mine vibration signals into a trained signal recognition model to output signal category determination results. The model training process is as follows: First, features are extracted from the mine vibration signals in the training set to construct an initial feature set, thereby determining the initial population size and optimal individuals. Then, individual vectors are mapped to binary vectors, and fitness is calculated to obtain the corresponding binary feature subsets. Next, a multiverse optimization algorithm combining white hole / black hole and wormhole mechanisms is used to update the individual's initial position. Orthogonal learning and quasi-reflective learning strategies are then introduced for secondary updates and a global population update. In each iteration, the updated individual vectors are converted to binary vectors, and fitness is recalculated. If a preset stopping condition is met, the binary feature subset corresponding to the optimal individual is output as the optimal feature combination; otherwise, the updated population is used as the new initial population for further iteration until convergence. Finally, the trained signal recognition model is constructed based on the selected optimal feature subset.

[0021] Figure 1 This is a flowchart illustrating the overall process for identifying microseismic or blasting signals in deep mines. Figure 1As shown, in this embodiment, the method includes:

[0022] The collected mine vibration signals are input into the trained signal recognition model to obtain the signal recognition results; Signal identification results include: mine vibration signals are either mine micro-seismic signals or mine blasting signals; The training methods for signal recognition models include: The collected mine vibration signals are input into the trained signal recognition model to obtain the signal recognition results; The signal identification result is: the mine vibration signal is either a mine micro-seismic signal or a mine blasting signal; The training methods for signal recognition models include: S1: Extract features from the mine vibration signals in the acquired dataset to obtain the original feature set; S2: Determine the feature dimension and initial population size based on the original feature set; randomly initialize the initial population based on the feature dimension and initial population size, and determine the optimal individual in the initial population; S3: Convert the individual vector of each individual in the initial population into a binary vector, calculate the individual fitness value of each individual, and obtain the binary feature subset corresponding to each individual in the initial population; S4: Perform global and local searches on binary feature subsets based on multi-universe optimization algorithm, and update the position of the current individual based on the global and local search results; S5: After the initial position update, the position of the current individual is updated again using an orthogonal learning strategy; S6: After the current position update is completed, the initial population is updated based on the updated current individual using a quasi-reflective learning strategy; S7: Convert the individual vector of each individual in the updated population into a binary vector, calculate the individual fitness value of each individual vector, and obtain the binary feature subset corresponding to each individual in the updated population. S8: If the updated population meets the preset stopping condition, output the binary feature subset corresponding to the best individual in the updated population; otherwise, determine the updated population as the initial population and repeat S3-S7. The trained signal recognition model is obtained from the binary feature subset corresponding to the optimal individual.

[0023] In a preferred embodiment, the method includes the following steps: Step 1: Construct the dataset; This dataset is a dataset of mine microseismic signals or blasting signals, including mine microseismic signal samples and mine blasting signal samples, which can be constructed from the waveforms of the original signals (i.e., the collected mine vibration signals) collected by the field monitoring system.

[0024] Step 2: Feature extraction (this step corresponds to S1); Figure 2 This is a flowchart illustrating the method for identifying microseismic signals or blasting signals in deep mines, as described in this embodiment. See also... Figure 2 This step involves preprocessing the samples in the dataset to generate waveform images or feature matrices; extracting color features using color moment techniques (including first, second, and third moments); and extracting texture features using gray-level co-occurrence matrix (GLCM) techniques, which involves selecting any point in the grayscale image. and the distance to that point is , direction is Another point The co-occurrence frequency of gray values ​​between two points is counted to form a gray-level co-occurrence matrix; further, local texture features and shape features (used to characterize the texture distribution characteristics of the image) are extracted to generate the original feature set. Therefore, the original feature set includes: color moment features, gray-level co-occurrence matrix texture features, local texture features, and shape features. The co-occurrence frequency of gray values ​​between two pixels separated by a given distance and in a given direction in a grayscale image is extracted and counted, and the gray-level co-occurrence matrix texture features are obtained based on the frequency.

[0025] Local Binary Patterns (LBP) are used to extract local texture features; Shape features are extracted using shape feature parameters, including area, perimeter, spread, and eccentricity.

[0026] Step 3: Data preprocessing and partitioning; All sample features in the dataset are normalized to the interval [-1,1]. Then, the samples are divided into 10 parts using a 10x cross-validation method, with one part serving as the test set and the rest as the training set.

[0027] Step 4: Initialize algorithm parameters and population (this step corresponds to S2); Based on the original feature set, the initial population size N and the optimal individual Xbest in the initial population are determined within the parameter range. In addition, the feature dimension dim and the maximum number of iterations Maxiter are also determined. The length of each individual in the initial population is consistent with the original feature dimension, and each component of the dimension represents the tendency for the corresponding feature to be selected.

[0028] For example, a multiverse population It can be represented as: in, Indicates the total number of individuals in the population. This represents the total dimension of the original features.

[0029] Step 5: Binarization and Fitness Calculation (This step corresponds to S3); Based on S3, the vector of each individual in the current population is converted into a binary vector, including: The transformation function value is obtained by combining the individual vector of each individual in the current population with the transformation function; the transformation function is: in, To convert function values; The first in the current population Individual; For the first The first individual Dimensional value, that is, the result after binaryization.

[0030] If the value of the transformation function is greater than the set threshold, then the binary vector of the current individual is determined to be 1; If the value of the transformation function is less than or equal to the set threshold, then the binary vector of the current individual is determined to be 0.

[0031] Specifically, the above binarization transformation has the following expression: in, This is a preset threshold. If... , indicates the selection of the corresponding feature; if , indicates that the corresponding feature is not selected.

[0032] In this embodiment, the individual fitness value of each individual is calculated, including: The collected mine vibration signals are filtered based on the binary feature subset corresponding to each individual, and the filtering results are input into the fuzzy K-nearest neighbor model for classification. Based on the distance relationship between training samples and test samples in the fuzzy K-nearest neighbor model, the membership degree of the test samples to the mine microseismic signal and the mine blasting signal is calculated, and the calculation results are obtained. Based on the principle of maximum membership and the calculation results, the test samples were identified as mine microseismic signals or mine blasting signals. The identification result is compared with the actual signal type of the test sample. If they match, the identification is considered correct. The classification accuracy is obtained by dividing the number of correctly identified test samples by the total number of test samples. The classification error rate is obtained by using the fuzzy K-nearest neighbor model based on the difference between 1 and the classification accuracy. The individual fitness value of each individual is obtained by weighting and summing the classification error rate and feature ratio according to the predetermined weight coefficients; the weight coefficients include the classification error rate weight coefficient and the feature ratio weight coefficient; the classification error rate weight coefficient is greater than the feature ratio weight coefficient.

[0033] Specifically, the classification accuracy mentioned above is calculated based on the classification of samples using the Fuzzy K-Nearest Neighbor (FKNN) model. After obtaining the binary feature subset corresponding to each individual, this feature subset is used to filter the original data (i.e., the collected mine vibration signals), and the filtered data is then input into the FKNN model for classification. During the classification process, the membership degree of the test sample to each category is calculated based on the distance relationship between the training samples and the test samples, and the predicted category (i.e., the recognition result) of the test sample is determined according to the principle of maximum membership degree. The predicted category is compared with the true category of the test sample, and the ratio of the number of correctly classified samples to the total number of test samples is calculated to obtain the classification accuracy. The calculation basis for the classification accuracy is: the classification accuracy equals the proportion of the number of correctly classified samples to the total number of samples, that is, it is obtained by comparing the predicted label output by the FKNN model with the true label of the sample; based on this, the classification error rate is obtained by "1 - classification accuracy", which can be used to evaluate the classification performance of the current feature subset.

[0034] The classification error rate is calculated using the trained FKNN model, and then the individual fitness value is calculated as follows: First, calculate the classification error rate: Then calculate the individual fitness value based on the classification error rate: in, Indicates classification accuracy; Indicates the classification error rate; This represents the weighting coefficient for the classification error rate; Indicates the characteristic ratio; This represents the feature ratio weight coefficient, and .

[0035] In a preferred embodiment, and The values ​​were set to 0.99 and 0.01 respectively to ensure that the algorithm prioritizes classification performance while also taking into account feature compression capabilities.

[0036] Step 6: Use BQOLMVO (i.e., the improved multiverse optimization algorithm) to find the optimal feature subset (this step corresponds to S4); A multi-universe optimization algorithm is used to perform global and local searches on binary feature subsets. Based on the global and local search results, the current individual's position is updated for the first time, including: The white hole or black hole mechanism in the multiverse optimization algorithm drives the exchange of positional component information between different individuals in the current population, so as to perform global updates between each binary feature subset; each binary feature subset is determined according to each candidate solution and corresponds one-to-one with each candidate solution. By using the wormhole mechanism in the multiverse optimization algorithm, the current individual is perturbed and updated according to the position of the optimal individual, so that some dimensions of the current individual are closer to the optimal individual or finely adjusted in the neighborhood of the optimal individual, so as to update the local position of the binary feature subset.

[0037] Specifically, the white hole or black hole mechanism in the multiverse optimization algorithm is used for global updates. The global update involves, within the current population, driving the exchange of positional components between different individuals through the white hole or black hole mechanism. This allows individuals with higher fitness to output partial dimensional information to other individuals with a higher probability, and other individuals receive the corresponding dimensional information, thereby achieving a combined update of binary feature subsets (i.e., a global update). The expression for this search update process is as follows:

[0038] in, Indicates the number of the current population The normalized inflation rate of each individual, for Random numbers.

[0039] After the global search update is completed, a local update is performed using the wormhole mechanism. The specific content of the local update is as follows: For each current individual in the current population, a neighborhood search is performed around the current best individual. That is, each current individual is perturbed and updated based on the position of the current best individual, causing some dimensions to move closer to the current best individual or to be fine-tuned within the neighborhood of the current best individual, thereby locally updating the binary feature subset. The expression for this search update process is as follows:

[0040] in, This indicates the probability that a wormhole exists. Indicates the distance traveled. and They represent the first The upper and lower bounds of the dimensional search space.

[0041] Step 7: Update the current individual using an orthogonal learning strategy (this step corresponds to S5); In this step, the current individual is updated using an orthogonal learning strategy, including: Randomly select at least two individuals from the initial population, and construct a guiding vector based on at least two individuals; Construct a search interval based on each individual and the guiding vector from at least two individuals, construct at least two candidate solutions using an orthogonal array, and retain the candidate solution with better fitness as the updated current individual based on the factor analysis results. It is important to note that the orthogonal array is determined based on the number of levels, factors, and candidate solutions set in the orthogonal learning strategy; Factors are the feature dimensions of an individual vector; the number of factors is the number of feature dimensions. The level represents the discrete value of the factor within the search interval; the number of levels represents the number of discrete values. Discrete values ​​are obtained by discretizing the search interval; The number of candidate solutions is determined by the structure of the orthogonal array and is consistent with the number of rows in the orthogonal array; The above factor analysis results were obtained through the following methods: Based on the fitness values ​​of each candidate solution in each factor and the fitness performance of each factor at each level, the effect values ​​of each factor at each level are obtained, and the effect values ​​are determined as the results of factor analysis.

[0042] Specifically, the current candidate solution is further optimized using an orthogonal learning strategy (OLS). First, multiple individuals are randomly selected to construct a guiding vector:

[0043] in, , , For randomly selected individuals, for The random number is generated. Then, a search interval is constructed based on the current individual and the guiding vector. Multiple candidate solutions are constructed using an orthogonal array. Based on the factor analysis results, the candidate solution with better fitness is retained as the updated current individual.

[0044] The aforementioned orthogonal array is determined based on the number of levels Q, the number of factors F, and the number of candidate solutions M set in the orthogonal learning strategy, and is used to construct multiple candidate solutions. Here, a factor is a feature dimension of an individual vector, with each dimension corresponding to one factor; the number of factors is the number of feature dimensions; the level is the discrete value of the factor within the search interval; the number of levels Q represents the number of discrete values ​​that each factor can select within the constructed search interval, and these discrete values ​​are obtained by discretizing the search interval jointly determined by at least two individuals and the guiding vector; the number of candidate solutions M is the number of rows in the orthogonal array used (i.e., the number of candidate solutions M is consistent with the number of rows in the orthogonal array), which is predetermined by the orthogonal array structure and corresponds to the number of candidate solutions generated through orthogonal array combination.

[0045] The above factor analysis results were obtained based on the fitness values ​​of candidate solutions during orthogonal learning. That is, after constructing multiple candidate solutions using the orthogonal array, the fitness value of each candidate solution was calculated. Based on the fitness performance of each candidate solution at each factor and its different levels, the effect value of each factor at different levels was statistically obtained, thereby determining the optimal level for each factor. Then, new candidate solutions were formed by combining the optimal levels of each factor, and their fitness was compared with the aforementioned candidate solutions. The candidate solution with the better fitness was selected as the updated current individual.

[0046] Step 8: Update the population using a quasi-reflexive learning strategy (this step corresponds to S6). In this step, the population is updated using a quasi-reflexive learning strategy, including: Determine the quasi-reflection point and its fitness value of the initial population. Compare the fitness value of the current individual with the fitness value of the quasi-reflection point. Retain the individual with the better fitness value to update the initial population, thus obtaining the updated population.

[0047] Specifically, after each iteration, the population is updated using a quasi-reflection learning strategy (QLS). Quasi-reflection points satisfy the following condition:

[0048] in, This represents the lower bound of the solution space. This represents the upper bound of the solution space. This indicates the current position of the individual. By comparing the fitness value of the current individual with that of the quasi-reflection point, those with better fitness are retained for the next generation of the population.

[0049] Step 9: Binarization and fitness recalculation (this step corresponds to S7); The individual vector of each individual in the updated population (i.e. the current population) is converted into a binary vector through a transformation function, and the fitness value is calculated by calling the FKNN classifier to obtain the binary feature subset corresponding to each individual in the updated population.

[0050] Step 10: Determine the stopping condition (this step corresponds to S8); If the updated population meets the preset stopping condition, output the binary feature subset corresponding to the best individual in the updated population; otherwise, determine the updated population as the initial population and repeat steps five to nine.

[0051] In this embodiment, the maximum number of iterations is determined based on the original feature set described above. If the number of iterations corresponding to the updated population is greater than the maximum number of iterations, then the updated population is determined to meet the stopping condition. In a specific embodiment, the preset stopping condition is "the number of iterations corresponding to the updated population is greater than the maximum number of iterations, 50". Therefore, the stopping condition determination in this embodiment can be presented as: if the number of iterations corresponding to the updated population is 51, which is greater than the maximum number of iterations, 50, then the updated population is determined to meet the stopping condition.

[0052] Step 11: Model Building and Signal Recognition; A trained signal recognition model is obtained based on the binary feature subset corresponding to the optimal individual. This trained model exhibits high recognition accuracy, real-time performance, and stability. Simply input the collected mine vibration signals into the model to obtain the signal recognition results, which ultimately determine whether the mine vibration signal is a mine microseismic signal or a mine blasting signal.

[0053] In summary, the improvements of this algorithm over the Many Universes Optimization (MVO) algorithm include the following: 1) Use the orthogonal learning strategy OLS to update the current individual; 2) Use the quasi-reflective learning strategy QLS to update the current population; 3) The continuous spatial individuals are mapped to binary vectors representing feature subsets through binarization transformation.

[0054] Figure 3 Box plots showing the performance metrics of each algorithm in this embodiment. For example... Figure 3As shown, the recognition method provided in this embodiment is compared with recognition technologies based on various algorithms in the following experiments. Among them, BMFO represents the Binary Moth-Flame Optimization algorithm; BGWO represents the Binary Grey Wolf Optimizer algorithm; BGSA represents the Binary Gravitational Search Algorithm; BPSO represents the Binary Particle Swarm Optimization algorithm; BALO represents the Binary Ant Lion Optimizer algorithm; BBA represents the Binary Bat Algorithm; BSSA represents the Binary Salp Swarm Algorithm; BWOA represents the Binary Whale Optimization Algorithm; BHHO represents the Binary Harris Hawks Optimization algorithm; BCS represents the Binary Cuckoo Search algorithm; and BMVO represents the Binary Multi-Verse Optimizer algorithm. Among them, BMVO is the binary basic version of the multiverse optimization algorithm used in this invention, used to illustrate the performance improvement effect of introducing the orthogonal learning strategy OLS and the quasi-reflective learning strategy QLS. The other comparative algorithms are all binary swarm intelligent optimization methods commonly used in feature selection, and have shown good performance in multiple feature selection tasks. Therefore, these algorithms were selected as comparison objects to verify the effectiveness and superiority of the method of this invention in the task of identifying microseismic signals or blasting signals in deep mines. To ensure the impartiality of the verification results, each method was run independently 10 times, and 10-fold cross-validation was used. Evaluation metrics included accuracy, sensitivity, Matthews correlation coefficient, and F-metric.

[0055] Implementation of a system for identifying microseismic or blasting signals in deep mines This embodiment provides a technical solution for a system for identifying microseismic signals or blasting signals in deep mines. The system includes a processor containing executable program instructions. These instructions are used to implement the method for identifying microseismic signals or blasting signals in deep mines as described in the above-described embodiment. Since the specific working method and principle of the system for identifying microseismic signals or blasting signals in deep mines have been described in detail in the above-described embodiment, they will not be repeated here.

[0056] Note that the above description is merely a preferred embodiment of the present invention and the technical principles employed. Those skilled in the art will understand that the present invention is not limited to the specific embodiments described herein, and various obvious changes, readjustments, and substitutions can be made without departing from the scope of protection of the present invention. Therefore, although the present invention has been described in detail through the above embodiments, the present invention is not limited to the above embodiments, and may include many other equivalent embodiments without departing from the concept of the present invention, the scope of which is determined by the scope of the appended claims.

Claims

1. A method for identifying microseismic signals or blasting signals in deep mines, characterized in that, include: The collected mine vibration signals are input into the trained signal recognition model to obtain the signal recognition results; The signal identification result is: the mine vibration signal is a mine micro-vibration signal or a mine blasting signal; The training methods for the signal recognition model include: S1: Extract features from the mine vibration signals in the acquired dataset to obtain the original feature set; S2: Determine the feature dimension and initial population size based on the original feature set; randomly initialize the initial population based on the feature dimension and the initial population size, and determine the optimal individual in the initial population; S3: Convert the individual vector of each individual in the initial population into a binary vector, calculate the individual fitness value of each individual, and obtain the binary feature subset corresponding to each individual in the initial population; S4: Perform global and local searches on the binary feature subset based on the multiverse optimization algorithm, and update the current individual's position for the first time based on the global and local search results; S5: After the initial position update, the position of the current individual is updated again using an orthogonal learning strategy; S6: After the current round of position update is completed, the initial population is updated based on the updated current individual using a quasi-reflective learning strategy; S7: Convert the individual vector of each individual in the updated population into a binary vector, calculate the individual fitness value of each individual vector, and obtain the binary feature subset corresponding to each individual in the updated population. S8: If the updated population meets the preset stopping condition, output the binary feature subset corresponding to the best individual in the updated population; otherwise, determine the updated population as the initial population and repeat S3-S7. The trained signal recognition model is obtained based on the binary feature subset corresponding to the optimal individual.

2. The method for identifying microseismic signals or blasting signals in deep mines according to claim 1, characterized in that, The method of updating the current individual using an orthogonal learning strategy includes: Randomly select at least two individuals from the initial population, and construct a guiding vector based on the at least two individuals; A search interval is constructed based on each of the at least two individuals and the guiding vector. At least two candidate solutions are constructed using an orthogonal array. The candidate solution with better fitness is retained as the updated current individual based on the factor analysis results. The orthogonal array is determined based on the number of levels, factors, and candidate solutions set in the orthogonal learning strategy; The factor is a feature dimension of the individual vector; the number of factors is the number of the feature dimensions. The level is the discrete value of the factor within the search interval; the number of levels is the number of the discrete values. The discrete values ​​are obtained by discretizing the search interval; The number of candidate solutions is determined based on the structure of the orthogonal array and is consistent with the number of rows in the orthogonal array; The methods for obtaining the factor analysis results include: Based on the fitness values ​​of each candidate solution in each factor and the fitness performance of each factor at each level, the effect value of each factor at each level is obtained, and the effect value is determined as the result of factor analysis.

3. The method for identifying microseismic signals or blasting signals in deep mines according to claim 1, characterized in that, The method of updating the population using a quasi-reflexive learning strategy includes: Determine the quasi-reflection point and its fitness value of the initial population, compare the fitness value of the current individual with the fitness value of the quasi-reflection point, retain the individual with the better fitness value to update the initial population, and obtain the updated population.

4. The method for identifying microseismic signals or blasting signals in deep mines according to claim 1, characterized in that, The calculation of the individual fitness value for each individual includes: The collected mine vibration signals are filtered based on the binary feature subset corresponding to each individual, and the filtering results are input into the fuzzy K-nearest neighbor model for classification. Based on the distance relationship between training samples and test samples in the fuzzy K-nearest neighbor model, the membership degree of the test sample to the mine microseismic signal and the mine blasting signal is calculated, and the calculation results are obtained. Based on the maximum membership principle and the calculation results, the test sample is identified as a mine microseismic signal or a mine blasting signal. The identification result is compared with the actual signal type of the test sample. If they match, the identification is considered correct. The classification accuracy is obtained by dividing the number of correctly identified test samples by the total number of test samples. The classification error rate is obtained by using the fuzzy K-nearest neighbor model based on the difference between 1 and the classification accuracy. The classification error rate and the feature ratio are weighted and summed according to a predetermined weight coefficient to obtain the individual fitness value of each individual; the weight coefficient includes a classification error rate weight coefficient and a feature ratio weight coefficient; the classification error rate weight coefficient is greater than the feature ratio weight coefficient.

5. The method for identifying microseismic signals or blasting signals in deep mines according to claim 1, characterized in that, Also includes: The maximum number of iterations is determined based on the original feature set; If the number of iterations corresponding to the updated population is greater than the maximum number of iterations, then the updated population is determined to satisfy the stopping condition.

6. The method for identifying microseismic signals or blasting signals in deep mines according to claim 1, characterized in that, The step of converting the vector of each individual in the current population into a binary vector includes: The transformation function value is obtained by combining the individual vector of each individual in the current population with the transformation function; the transformation function is: in, The value of the transformation function; The first in the current population Individual; For the first The first individual Dimensional value; If the value of the transformation function is greater than the set threshold, then the binary vector of the current individual is determined to be 1; If the value of the transformation function is less than or equal to the set threshold, then the binary vector of the current individual is determined to be 0.

7. The method for identifying microseismic signals or blasting signals in deep mines according to claim 1, characterized in that, A global and local search is performed on the binary feature subset based on a multi-universe optimization algorithm. The initial position update of the current individual is then performed based on the global and local search results, including: The white hole or black hole mechanism in the multiverse optimization algorithm drives the exchange of positional component information between different individuals in the current population, so as to perform global updates between each binary feature subset; each binary feature subset is determined according to each candidate solution and corresponds one-to-one with each candidate solution. The wormhole mechanism in the multiverse optimization algorithm is used to perturb and update the current individual based on the position of the optimal individual, so that some dimensions of the current individual move closer to the optimal individual or are finely adjusted within the neighborhood of the optimal individual, thereby updating the local position of the binary feature subset.

8. The method for identifying microseismic signals or blasting signals in deep mines according to claim 1, characterized in that, The original feature set includes: color moment features, gray-level co-occurrence matrix texture features, local texture features, and shape features.

9. The method for identifying microseismic signals or blasting signals in deep mines according to claim 8, characterized in that, The method further includes: Extract the number of co-occurrences of gray values ​​of two pixels separated by a given distance and a given direction in a statistical grayscale image, and obtain the gray-level co-occurrence matrix texture features based on the number of occurrences. The local texture features are extracted using a local binary model; The shape features are extracted using shape feature parameters, including area, perimeter, expansion, and eccentricity.

10. A system for identifying microseismic signals or blasting signals in deep mines, comprising a processor, characterized in that, The processor is used to execute a computer program to implement the steps of the method for identifying microseismic signals or blasting signals in deep mines as described in any one of claims 1-9.

Images