Microseismic signal classification method based on multi-domain feature fusion and HOA optimized SVM

Abstract

The present application relates to a kind of microseismic signal classification method based on multi-domain feature fusion and HOA optimization SVM.The method of the present application is by the band-pass filtering and the pretreatment of removing trend to the original microseismic signal obtained, eliminate noise interference and baseline shift;Then extract fusion time-domain duration, frequency-domain main frequency, the four-dimensional physical feature vector of the strongest energy point time and the strongest energy point frequency obtained based on VMD-HHT time-frequency analysis, and based on training set statistical parameter carries out Z-score standardization;On the basis of preceding, adaptive optimization is carried out to the penalty factor C and kernel parameter gamma of SVM using hiking optimization algorithm (HOA), and the optimal SVM classification model is constructed.The method of the present application is applicable to mine microseismic real-time monitoring system, can effectively distinguish artificial blasting, natural earthquake and mine earthquake signal, keep higher classification precision and generalization ability in small sample scene, solve the problem of existing microseismic signal recognition method exists and is dependent on artificial experience selection and low classification precision.

Description

Technical Field

[0001] This invention belongs to the field of mine safety monitoring and geophysical signal processing technology, specifically involving a microseismic signal classification method based on the fusion of time-domain, frequency-domain and multi-domain features and the Hovering Optimization Algorithm (HOA) to optimize the Support Vector Machine (SVM). Background Technology

[0002] During deep mining in coal and non-coal mines, the stress distribution in the rock mass undergoes complex changes with increasing depth and intensity, easily inducing dynamic disasters such as fault slippage and roof fracture. Microseismic monitoring technology, as an important means of sensing changes in rock mass stability, analyzes vibration signals to identify events such as mine tremors, artificial blasting, and natural earthquakes caused by mining activities, playing a crucial role in ensuring safe mine production.

[0003] Current microseismic signal classification methods mainly rely on manual experience or traditional machine learning algorithms. However, existing technologies have the following significant drawbacks in practical applications:

[0004] (1) Single feature dimension and lack of multi-domain fusion: Existing methods mostly rely on a single feature for analysis and judgment, while microseismic signals have complex characteristics such as non-stationarity and nonlinearity, and a single feature is difficult to fully characterize the essential differences between different types of microseismic signals. In particular, natural earthquakes and artificial blasting have overlapping features in the frequency domain and time-frequency domain, and features from a single domain cannot be effectively distinguished.

[0005] (2) Classification model parameter selection relies on human experience: Support Vector Machine (SVM), as a classic supervised learning classifier, performs well in high-dimensional feature spaces and small sample scenarios, and is widely used in the field of seismic signal recognition. However, the classification performance of SVM is highly dependent on the selection of the penalty factor C and the kernel parameter gamma. Existing methods mostly use manual trial and error or grid search methods to adjust parameters. If the value of C is too large, it will lead to overfitting and a decrease in generalization ability, while if the value of C is too small, it will cause underfitting. The gamma value controls the influence range of the RBF kernel function, and its selection directly affects the shape of the classification boundary. Manual parameter tuning is not only inefficient and time-consuming, but also difficult to find the global optimal solution in complex, high-dimensional feature spaces, which can easily lead to fluctuations in model performance.

[0006] (3) Inconsistent feature scale affects the accuracy of kernel function calculation: The RBF kernel function is highly sensitive to feature scale. If the numerical ranges of different feature dimensions differ too much, some dimensions will have a much greater impact on distance calculation than other dimensions, thus distorting the classification boundary. Traditional methods often ignore this preprocessing step or use a simple normalization method, which makes it difficult to guarantee the calculation accuracy of the SVM kernel function in high-dimensional feature space.

[0007] Therefore, there is an urgent need to develop a microseismic signal classification method that integrates multi-domain features, has adaptive parameter optimization capabilities, and fully considers the characteristics of SVM. Summary of the Invention

[0008] This invention aims to address the problems of low classification accuracy caused by single feature dimensions, reliance on manual experience for SVM model parameter selection, and inconsistent feature scales in existing microseismic signal identification methods. It provides a microseismic signal classification method based on multi-domain feature fusion and Hovering Optimization (HOA) algorithm to optimize a Support Vector Machine (SVM). Furthermore, this invention also relates to a mine vibration event signal classification system and a computer-readable storage medium for implementing the above classification method.

[0009] To achieve the above objectives, the present invention employs the following technical solutions:

[0010] In a first aspect, the present invention provides a microseismic signal classification method based on multi-domain feature fusion and HOA-optimized support vector machine, the method comprising the following steps:

[0011] Step 1: Deploy a microseismic sensor array in the mine monitoring area to acquire the raw microseismic monitoring signal data to be identified;

[0012] Step 2: Preprocess the original microseismic monitoring signal data, including bandpass filtering and detrending processing, to obtain preprocessed microseismic signal data;

[0013] Step 3: Multi-domain feature extraction: Perform multi-dimensional feature extraction in the time domain, frequency domain, and time-frequency domain on the preprocessed microseismic signal to construct a multi-dimensional feature vector; the multi-dimensional feature vector includes time domain duration feature, frequency domain dominant frequency feature, time feature of the strongest energy point in the time-frequency domain, and frequency feature of the strongest energy point in the time-frequency domain.

[0014] Step 4: Construct a microseismic classifier based on support vector machine, using radial basis function (RBF) as the kernel function, and use the HOA algorithm to adaptively optimize the hyperparameter combination (C, gamma) of the support vector machine to obtain the optimal microseismic classification model;

[0015] Step 5: Input the multidimensional feature vector extracted in Step 3 into the optimal microseismic classification model trained in Step 4, and output the category label of the microseismic signal. The category label includes artificial blasting, mining tremor and natural earthquake.

[0016] Furthermore, in step 2, the filtering and denoising process employs a bandpass filter with a passband frequency range of 0.5Hz to 10Hz, eliminating low-frequency drift and high-frequency noise while preserving the main frequency components of the microseismic signal. Detrending processing is used to remove baseline shifts and long-term trend terms in the seismic waveform caused by factors such as instrument drift, environmental noise, and temperature changes.

[0017] Furthermore, after obtaining the multidimensional feature vector in step 3, the multidimensional feature vector of the training set is subjected to Z-score standardization; then, the same Z-score standardization method and calculation formula as the training data are used to standardize the samples to be classified. The formula for calculating the Z-score standardization is as follows:

[0018]

[0019] In the formula, x' is the standardized feature value; x is the original feature value; μ is the mean of the corresponding feature dimension of the training set; and σ is the standard deviation of the corresponding feature dimension of the training set.

[0020] Unlike simple max-min normalization, Z-score standardization preserves the original distribution of the data while ensuring that all features have the same scale (mean of 0 and standard deviation of 1), which is crucial for the RBF kernel function, which relies on distance calculations. Instead of recalculating μ and σ for the samples to be classified, μ and σ saved during the training phase are reused, avoiding inconsistencies in feature distribution between the training and online classification phases.

[0021] Furthermore, the method for extracting the time-domain duration feature in step 3 is as follows: determine the start and end times of the microseismic signal fluctuation, and use the time difference between the two as the duration feature T. duration The duration of the earthquake reflects the energy release process of the earthquake source and the signal propagation attenuation characteristics, and is one of the important time-domain criteria for distinguishing between natural earthquakes, artificial blasting, and mining earthquakes.

[0022] Furthermore, the method for extracting the dominant frequency feature in step 3 is as follows: A Fast Fourier Transform (FFT) is performed on the detrended microseismic signal to obtain the signal's amplitude spectrum. To prevent low-frequency ground pulsation interference from affecting the dominant frequency extraction, an effective frequency range Ω = {f | f>f} is set. th}, f th This is a preset low-frequency cutoff threshold. The frequency corresponding to the point of maximum amplitude within this effective frequency range is used as the dominant frequency feature.

[0023]

[0024] Physical significance: In the mining area monitoring data collected in this embodiment, natural earthquakes have a wider frequency band and more significant high-frequency components, artificial blasting signals have a slightly lower frequency band, and mine seismic signals have a narrower frequency band and a lower dominant frequency. Therefore, the dominant frequency characteristic is not used as a separate classification criterion, but is combined with the duration, the time of the strongest energy point, and the frequency of the strongest energy point to form a discriminative feature.

[0025] Furthermore, step 3 extracts the time and frequency features of the strongest energy point in the time-frequency domain based on VMD-HHT time-frequency analysis. The specific process is as follows: First, variational mode decomposition is performed on the microseismic signal to obtain several intrinsic mode components; then, Hilbert transform is performed on each intrinsic mode component to obtain the instantaneous frequency and instantaneous amplitude, and a Hilbert spectrum is constructed; finally, the time T_maxE and frequency F_maxE corresponding to the strongest energy point are extracted from the Hilbert spectrum as time-frequency domain features.

[0026] Furthermore, in step 4, the support vector machine uses the radial basis function (RBF) as its kernel function. Considering the non-stationary and nonlinear characteristics of microseismic signals in mining areas, linear kernel functions can only construct linear decision boundaries in the original feature space, making it difficult to characterize the complex nonlinear relationships inherent in the signal. The RBF kernel function can map nonlinearly distributed feature data in low-dimensional space to high-dimensional space, transforming it into linearly separable states. Its mathematical expression is:

[0027]

[0028] In the formula, x i and x j denoted as sample points; σ is a parameter controlling the width of the kernel function; and ||·| represents the Euclidean distance.

[0029] The advantage of the RBF kernel function lies in its ability to keenly capture the characteristics of data distribution and automatically adjust the mapping method accordingly, exhibiting excellent fitting ability for nonlinear classification boundaries. Furthermore, the RBF kernel function involves only one parameter to be optimized, σ (gamma = 1 / (2σ²)), which, compared to the polynomial kernel function (which requires setting the polynomial order) and the hyperbolic tangent (Sigmoid) kernel function (which requires setting two parameters), results in lower model complexity and higher parameter optimization efficiency.

[0030] The hyperparameters that the support vector machine needs to optimize are the penalty factor C and the kernel parameter gamma. Penalty factor C: Controls the model's tolerance for classification errors. The larger the C value, the more severely the model penalizes training errors, the more complex the decision boundary, and the easier it is to overfit; the smaller the C value, the higher the model's tolerance, the smoother the decision boundary, which is beneficial to improving generalization ability, but too small a C value will cause underfitting.

[0031] The kernel parameter gamma controls the influence range of the RBF kernel function. A larger gamma value results in a smaller influence range for a single sample, a more complex decision boundary, and may lead to overfitting to noise in the training data; a smaller gamma value results in a larger influence range and a smoother decision boundary, but if the value is too small, the model may not be able to effectively distinguish between different categories.

[0032] To address the issues of low efficiency and susceptibility to local optima in SVM hyperparameter selection using traditional manual trial-and-error and grid search methods, this invention employs the Walking Optimization (HOA) algorithm to adaptively optimize the (C, gamma) combination.

[0033] Furthermore, the adaptive optimization process of the HOA algorithm for SVM hyperparameters in step 4 is as follows:

[0034] (1) Initialization: Set the population size N=10 and the maximum number of iterations T=50. Randomly initialize the positions of N hikers in the hyperparameter search space, and each position vector X i = (C i gamma i ) represents a set of SVM hyperparameter combinations to be optimized. The hyperparameter search range is set as follows: C ∈ [0.1, 100], gamma ∈ [0.01, 1].

[0035] (2) Fitness evaluation: The SVM model was trained with the current hyperparameter combination for each hiker, and the classification accuracy of K-fold cross-validation was used as the fitness evaluation index. Cross-validation divides the training data into K mutually exclusive subsets, and K-1 subsets are used as training data and the remaining 1 subset is used as validation data. This process is repeated K times and the average value is taken, thereby reducing the evaluation bias caused by a single data partition and making the fitness evaluation results more robust.

[0036] (3) Determining the global optimum: Compare the fitness values ​​of all hikers and select the position of the hiker with the highest fitness as the current global optimum position X. target .

[0037] (4) Speed ​​and position update: The HOA algorithm simulates the behavior of hikers adjusting their speed according to changes in terrain. It determines the individual's pace based on the Tobler hiking function and updates the hyperparameter combination by combining the current hiker's position, the group's best position, and a random factor, so that the search process takes into account both global exploration and local optimization. The updated position is used to train the SVM and calculate the fitness value. If the fitness of the new position is better than the hiker's own historical best fitness, then its local optimal solution is updated. If the fitness of the new position is better than the global optimal fitness, then the global optimal solution X is updated. target .

[0038] (5) Boundary constraints: The updated position components are restricted to a preset range, i.e., C new = clip(C new (, 0.1, 100), gamma new = clip(gamma new , 0.01, 1), to ensure that the candidate hyperparameter combinations are always within the effective search space.

[0039] (6) Termination judgment: Repeat steps (2) to (5) until the maximum number of iterations is reached or the fitness converges. Output the globally optimal combination of hyperparameters (C). best gamma best The optimal microseismic classification model was obtained.

[0040] The key to the HOA-SVM optimization mechanism lies in the fact that, through Tobler function-driven step speed adjustment, the algorithm can dynamically adjust the movement range of the hyperparameter combination during the search process, thereby balancing global search and local optimization. Under the same data partitioning and parameter settings in this embodiment, compared with metaheuristic algorithms such as Particle Swarm Optimization (PSO) and Genetic Algorithm (GA), HOA-SVM exhibits a higher classification index on the test set.

[0041] Furthermore, in step 5, before inputting the multidimensional feature vector of the sample to be classified into the trained optimal microseismic classification model, it is necessary to use the same Z-score standardization parameters (i.e. the same mean μ and standard deviation σ) as the training data to standardize the sample to be classified, so as to ensure that the feature distribution is consistent with the training stage.

[0042] The present invention also relates to a mine vibration event signal classification system, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the mine vibration event signal classification method described above.

[0043] This invention solves the technical problems existing in the prior art, such as the single feature dimension, the reliance on manual experience in selecting SVM hyperparameters, and the impact of inconsistent feature scales on the accuracy of kernel function calculation. Compared with the prior art, the beneficial technical effects achieved are as follows:

[0044] (1) This invention constructs a four-dimensional multi-domain feature vector with clear physical meaning by jointly extracting the duration in the time domain, the dominant frequency in the frequency domain, and the time and frequency of the strongest energy point obtained based on VMD-HHT time-frequency analysis. This feature combination characterizes the microseismic signal from three perspectives: the energy release process, the spectral distribution law, and the energy accumulation characteristics in the time-frequency plane. This overcomes the shortcomings of a single feature, which can only reflect the local attributes of the signal and is difficult to comprehensively distinguish between artificial blasting, natural earthquakes, and mining seismic signals.

[0045] (2) This invention employs the HOA algorithm to adaptively optimize the penalty factor C and kernel parameter gamma of the SVM, using cross-validation classification accuracy as the fitness evaluation index, and obtaining a globally optimal combination of hyperparameters through iterative search. This approach reduces the reliance on empirical parameters from manual trial and error or grid search, and improves the automation of SVM model parameter selection and the stability of the classification model.

[0046] (3) Before classification modeling, this invention performs Z-score standardization on multidimensional features based on statistical parameters of the training set, and reuses the same mean and standard deviation in the test samples and the samples to be classified, so that each feature dimension is within a uniform scale range, avoiding the dominance of features with large numerical ranges in the RBF kernel function distance calculation, thereby improving the rationality of the support vector machine classification boundary construction and the reliability of the classification results.

[0047] (4) The present invention can address the main defects in the prior art from three aspects: feature characterization, model parameter optimization and feature scale unification. It can achieve automatic classification and identification of artificial blasting, natural earthquake and mine seismic signals without relying on manual experience to adjust parameters. Attached Figure Description

[0048] Figure 1 The optimal fitness curve for optimizing SVM hyperparameters using the HOA algorithm.

[0049] Figure 2 This is a schematic diagram illustrating the principle of the optimal classification hyperplane for Support Vector Machines (SVM).

[0050] Figure 3 The optimal fitness convergence curve for optimizing SVM hyperparameters using the HOA algorithm.

[0051] Figure 4 The confusion matrix and classification results of the HOA-SVM model on the test set are shown. Detailed Implementation

[0052] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0053] like Figure 1 As shown, the present invention provides a microseismic signal classification method based on multi-domain feature fusion and HOA-optimized support vector machine, which mainly includes five core steps: data acquisition, preprocessing and standardization, multi-domain feature extraction, HOA-SVM model construction and optimization, and online classification.

[0054] Step 1: Acquire raw microseismic signals

[0055] The data in this embodiment comes from actual monitoring data of the Datong Mine Seismic Research Demonstration Area in Shanxi Province. The research object is microseismic events of magnitude M1.5 to M3.0 within the monitoring range of the Datong mining area from 2021 to 2022. After dual verification using focal mechanism solutions and waveform characteristics, a total of 96 microseismic events were identified, including 34 natural earthquake events, 32 artificial blasting events, and 30 mine seismic events. From these events, 1589 microseismic waveform records were selected for feature analysis and classification, including 410 artificial blasting events, 586 natural earthquake events, and 593 mine seismic events, all with a signal-to-noise ratio greater than 2. The dataset was randomly divided into training and testing sets in a 7:3 ratio to ensure the independence and consistent distribution of the training and testing data.

[0056] Step 2: Data Preprocessing

[0057] The following three preprocessing operations are performed sequentially on the raw microseismic signal:

[0058] (1) Filtering and noise reduction: A bandpass filter is used with a passband frequency set from 0.5 Hz to 10 Hz to eliminate low-frequency drift and high-frequency noise while retaining the main frequency components of the seismic signal.

[0059] (2) Detrending processing: The waveform signal is detrending to highlight the effective microseismic waveform features and provide a stable basis for subsequent extraction of main frequency and time frequency features.

[0060] (3) Feature standardization: After completing the four-dimensional feature extraction, the mean μ and standard deviation σ of each feature dimension are calculated based on the training set, and then... Z-score standardization was performed to ensure that all features had the same scale (mean 0, standard deviation 1). The μ and σ values ​​from the training set were reused for both the test samples and the online samples to be classified; the standardization parameters were not recalculated separately. Table 1 shows some of the standardized feature vector data.

[0061] Table 1. Feature vector data after Z-score standardization.

[0062] 0.80600 1.14342 -0.74171 2.06790 0 (Mine Shock) 1.85604 -0.03880 0.18347 1.10058 0 (Mine Shock) 2.38760 1.23429 -0.03073 1.74546 0 (Mine Shock) -0.16726 1.72718 0.62747 -0.51160 0 (Mine Shock) 0.04705 1.24675 0.38857 -0.83404 0 (Mine Shock) -0.24151 -1.04476 1.91487 -1.15648 1 (natural earthquake) 0.12214 -1.14296 -0.02333 -1.02750 1 (natural earthquake) -0.12802 -0.79170 -0.29061 -0.83404 1 (natural earthquake) -0.38705 -0.87408 0.79186 -0.51160 1 (natural earthquake) -0.06095 -0.84866 1.90709 -0.70507 1 (natural earthquake) -1.33247 -0.89709 -1.70709 -0.83404 2 (Artificial blasting) -0.55622 0.86672 -1.36412 0.77815 2 (Artificial blasting) -1.48308 0.76331 -0.63729 0.00429 2 (Artificial blasting) -0.06010 -0.87170 -0.23080 0.58468 2 (Artificial blasting) -0.80218 -0.47190 -0.78766 0.13327 2 (Artificial blasting)

[0063] Step 3: Multi-domain feature extraction

[0064] This embodiment extracts four features with clear physical meaning and constructs a feature vector V = [T duration ,F dom , T maxE , F maxE ].

[0065] (1) Duration in the time domain (T)duration )

[0066] Duration refers to the time elapsed between the start and end of a microseismic signal fluctuation. This characteristic reflects the energy release process of the seismic source and the signal propagation attenuation characteristics: natural earthquakes usually have a longer duration, while the duration of mine-induced earthquakes falls between that of natural earthquakes and those caused by artificial blasting, with artificial blasting having the shortest duration due to concentrated energy release.

[0067] (2) Frequency domain dominant frequency (F) dom )

[0068] The detrended signal is subjected to a Fast Fourier Transform (FFT). To prevent low-frequency pulsation interference, a frequency mask—a low-frequency cutoff threshold f—is set. th In the effective frequency range Ω = {f | f>f} th The frequency corresponding to the point with the largest search amplitude within the area:

[0069]

[0070] Physical significance: In the mining area monitoring data collected in this embodiment, natural earthquakes have a wider frequency band and more significant high-frequency components, while artificial blasting events have a less pronounced frequency band, and mine seismic signals have a narrower frequency band and a lower dominant frequency. Therefore, the dominant frequency characteristic is not used as a separate classification criterion, but rather is combined with the duration, the time of the strongest energy point, and the frequency of the strongest energy point to form a discriminative feature.

[0071] (3) Time of the strongest energy point in the time-frequency domain (T) maxE ) and the frequency of the strongest energy point (F maxE )

[0072] This is the core feature extraction step of the present invention. First, variational mode decomposition is performed on the waveform to decompose the signal into several intrinsic mode components; then, Hilbert transform is performed on each intrinsic mode component to extract the instantaneous frequency and instantaneous amplitude, and a Hilbert spectrum is constructed; finally, the point with the strongest energy is searched in the Hilbert spectrum, and the time and frequency corresponding to this point are respectively taken as T. maxE and F maxE This feature can characterize the differences in energy distribution among different seismic sources in the combined time-frequency-energy space.

[0073] Step 4: HOA-SVM Model Construction and Optimization

[0074] We construct a support vector machine classifier based on the RBF kernel function and use the Hovering Optimization (HOA) algorithm to find the optimal combination of hyperparameters.

[0075] (1) SVM model configuration:

[0076] The binary SVM is extended to a ternary classifier using a one-vs-one strategy, with the RBF radial basis function chosen as the kernel function.

[0077] In the formula, x i and x j denoted as sample points; σ is a parameter controlling the width of the kernel function; and ||·| represents the Euclidean distance.

[0078] A schematic diagram of the optimal classification hyperplane principle of Support Vector Machine (SVM) in this invention is shown below. Figure 2 In the diagram, line H represents the optimal classification hyperplane, and H1 and H2 are the two boundary hyperplanes parallel to the optimal hyperplane. Solid circles and squares represent support vectors, the key edge samples that determine the classification boundary; hollow geometric shapes represent the remaining training samples. D1 and D2 are the geometric distances from the support vectors to the optimal hyperplane (satisfying D1 = D2 = 1 / ∥ω∥). The classification margin is the distance between the two boundary hyperplanes, expressed as margin = 2 / ∥ω∥. The core optimization objective of SVM is to maximize this classification margin while ensuring correct sample classification.

[0079] (2) Optimization target:

[0080] Penalty factor C: Search range [0.1, 100];

[0081] Kernel parameter gamma (=1 / (2σ²)): Search range [0.01, 1].

[0082] (3) HOA Implementation Steps:

[0083] Initialization: Set the population size N=10 and the maximum number of iterations T=50. Randomly initialize the positions (C0, T>, T>) of 10 hikers in the two-dimensional hyperparameter space. i gamma i ).

[0084] Fitness evaluation: The classification accuracy of 5-fold cross-validation is used as the fitness function f(·). 5-fold cross-validation divides the training set into five equal subsets, using four subsets as training data and one subset as validation data, repeating this training and validation process five times. The average of the five accuracies is taken as the fitness value for that hyperparameter combination. This evaluation strategy effectively reduces the evaluation bias caused by a single data partition, making the fitness evaluation results more robust and reliable.

[0085] Location Update: HOA simulates the behavior of hikers adjusting their speed according to terrain changes. It determines individual pace based on the Tobler hiking function and updates the hyperparameter combination by combining the current hiker's position, the group's optimal position, and a random factor. This mechanism allows the algorithm to dynamically adjust its movement within the search space, thus balancing global exploration and local optimization.

[0086] Termination condition: After reaching the maximum number of iterations (50), output the globally optimal parameter combination.

[0087] Experimental results: such as Figure 3 As shown, the optimal fitness of the HOA algorithm gradually increases with the number of iterations and tends to stabilize in the later stages. The optimal hyperparameter combination is C = 87.282 and gamma = 0.1839. This combination indicates that the model prefers a strong penalty (large C value, low tolerance for misclassification) and a moderate kernel function width (medium gamma value, moderate smoothness of decision boundary), reflecting the characteristic that the microseismic signal feature space has good inter-class separability but some local overlap.

[0088] Step 5: Online Classification and Performance Evaluation The trained HOA-SVM model was applied to the test set (477 samples in total) to output the classification results.

[0089] like Figure 4 As shown, the confusion matrix and classification results of the test set indicate that: 87 records of artificial blasting were correctly identified, while 36 records of artificial blasting were incorrectly identified as natural earthquakes; 147 records of natural earthquakes were correctly identified, while 6 records were incorrectly identified as artificial blasting and 23 records were incorrectly identified as mine tremors; and 178 records of mine tremors were correctly identified, with no misclassifications.

[0090] Depend on Figure 4 The confusion matrix shown can be recalculated to obtain the overall classification accuracy of the HOA-SVM model, which is 86.37%, the macro-average precision is approximately 87.48%, the macro-average recall is approximately 84.75%, and the macro-average F1 score is approximately 85.46%. Compared with the traditional SVM model, the HOA-optimized SVM shows significant improvements in both overall accuracy and stability in class recognition.

[0091] Of particular note is that 178 mine seismic signals were correctly identified in the test set, with no samples misclassified as other categories. This indicates that duration, dominant frequency, and the time-frequency characteristics of the strongest energy point in VMD-HHT have good characterization capabilities for mine seismic signals. SVM can clearly distinguish mine seismic signals from the other two types of signals in the standardized four-dimensional feature space.

[0092] In summary, this invention comprehensively characterizes the physical differences of different types of microseismic signals from three dimensions: time domain, frequency domain, and time-frequency domain, through a multi-domain feature fusion strategy. By introducing Z-score normalization, the computational accuracy of the RBF kernel function in the high-dimensional feature space is ensured. Furthermore, the Tobler function step rate adjustment mechanism of the HOA algorithm enables efficient adaptive optimization of SVM hyperparameters (C, gamma). This method improves the accuracy and robustness of three-class classification of microseismic signals while maintaining low computational complexity, making it suitable for deployment in real-time microseismic monitoring systems in mines.

[0093] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A microseismic signal classification method based on multi-domain feature fusion and HOA-optimized SVM, characterized in that, The method includes the following steps: Step 1: Deploy a microseismic sensor array in the mine monitoring area to acquire the raw microseismic monitoring signal data to be identified; Step 2: Preprocess the original microseismic monitoring signal data. The preprocessing includes bandpass filtering and detrending to obtain preprocessed microseismic signal data. Step 3: Perform multidimensional feature extraction on the preprocessed microseismic signal to construct a multidimensional feature vector containing physical meaning; the multidimensional features include time domain, frequency domain, and time-frequency domain; the multidimensional feature vector includes time domain duration feature, frequency domain dominant frequency feature, time feature of the strongest energy point in the time-frequency domain, and frequency feature of the strongest energy point in the time-frequency domain. Step 4: Construct a microseismic classifier based on SVM, using radial basis functions as kernel functions, and employing the HOA algorithm to adaptively optimize the hyperparameter combination of the SVM to obtain the optimal microseismic classification model; Step 5: Input the multidimensional feature vector extracted in Step 3 into the optimal microseismic classification model trained in Step 4, and output the category label of the microseismic signal; the category label includes artificial blasting, mining tremor and natural earthquake.

2. The microseismic signal classification method based on multi-domain feature fusion and HOA-optimized SVM according to claim 1, characterized in that, The bandpass filtering process in step 2 uses a bandpass filter with a passband frequency range of 0.5Hz to 10Hz; the detrending process is used to remove baseline offset and long-term trend terms from the original microseismic monitoring signal.

3. The microseismic signal classification method based on multi-domain feature fusion and HOA-optimized SVM according to claim 1, characterized in that, After obtaining the multidimensional feature vectors in step 3, Z-score standardization is performed on the multidimensional feature vectors of the training set; then, the same Z-score standardization method and calculation formula as the training data are used to standardize the samples to be classified. The formula for calculating the Z-score standardization is as follows: ， In the formula, x' is the standardized feature value; x is the original feature value; μ is the mean of the corresponding feature dimension of the training set; and σ is the standard deviation of the corresponding feature dimension of the training set.

4. The microseismic signal classification method based on multi-domain feature fusion and HOA-optimized SVM according to claim 1, characterized in that, The method for extracting the time-domain duration feature in step 3 is as follows: determine the start and end times of the microseismic signal fluctuation, and use the time difference between the two as the time-domain duration feature T. duration The duration characteristic is used to characterize the energy release process of the seismic source and the signal propagation attenuation characteristics.

5. The microseismic signal classification method based on multi-domain feature fusion and HOA-optimized SVM according to claim 1, characterized in that, The method for extracting the frequency domain dominant frequency feature in step 3 is as follows: A fast Fourier transform is performed on the detrended microseismic signal to obtain the signal amplitude spectrum. The frequency corresponding to the point of maximum amplitude within the effective frequency range is searched as the dominant frequency feature. ， In the formula, Fdom represents the dominant frequency characteristic in the frequency domain; FFT{·} represents the Fast Fourier Transform operation; Ω represents the effective frequency range, Ω = {f | f > f} th }, f th This is a preset low-frequency cutoff threshold used to filter out low-frequency ground pulsation interference.

6. The microseismic signal classification method based on multi-domain feature fusion and HOA-optimized SVM according to claim 1, characterized in that, Step 3 involves extracting the time and frequency features of the strongest energy point in the time-frequency domain. This method is based on VMD-HHT time-frequency analysis, and the specific process includes: Variational mode decomposition was performed on the preprocessed microseismic signal to obtain several intrinsic mode components; Hilbert transforms are performed on each intrinsic modal component to construct a Hilbert spectrum characterizing the relationship between time, frequency, and instantaneous energy. The strongest energy point is searched in the Hilbert spectrum, and the time and frequency corresponding to the strongest energy point are respectively used as time-frequency domain features T. maxE And F_maxE.

7. The microseismic signal classification method based on multi-domain feature fusion and HOA-optimized SVM according to claim 1, characterized in that, In step 4, the SVM uses a radial basis function as its kernel function. The expression for the radial basis function is: ， In the formula, x i and x j denoted as sample points; σ is a parameter controlling the width of the kernel function; ‖·‖ represents the Euclidean distance; The hyperparameters of the SVM include the penalty factor C and the kernel parameter gamma, where gamma = 1 / (2σ²).

8. The microseismic signal classification method based on multi-domain feature fusion and HOA-optimized SVM according to claim 7, characterized in that, Step 4 uses the HOA algorithm to adaptively optimize the hyperparameter combination (C, gamma) of the SVM. The specific process includes: (1) Initialization: Randomly initialize the positions of N hikers in the hyperparameter search space, each position representing a set of hyperparameter combinations (C, gamma); set the maximum number of iterations T; the hyperparameter search range is: C ∈ [0.1, 100], gamma ∈ [0.01, 1]; (2) Fitness evaluation: The SVM model was trained with the hyperparameter combination of each hiker's current position, and the classification accuracy of K-fold cross-validation was used as the fitness evaluation index to reduce the random error caused by a single data partitioning. (3) Global optimal update: Compare the fitness values ​​of all hikers and select the position of the hiker with the highest fitness as the current global optimal position X. target ; (4) Speed ​​and position update: Based on the Tobler hiking function, and combined with the current hiker position, the group's optimal position and random factor, the speed and position of each hiker are iteratively updated, so that the hyperparameter combination can be globally explored and locally optimized in the search space; (5) Boundary constraints: The updated position is restricted to the preset hyperparameter search range to avoid the penalty factor C and the kernel parameter gamma from exceeding the effective value range; (6) Termination judgment: Repeat steps (2) to (5) until the maximum number of iterations is reached or the fitness converges, and output the globally optimal hyperparameter combination (C). best gamma best ).

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 9.

10. A mine vibration event signal classification system, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method as described in any one of claims 1 to 9.

Images