Engineering rock mass fracturing induced shock multi-information fusion early warning method and system

Abstract

The invention discloses an engineering rock mass fracturing induced shock multi-information fusion early warning method and system, and the method comprises the steps: arranging a microseismic sensor to collect rock fracturing induced shock original waveform data, and extracting waveform characteristic parameters and seismic source mechanism parameters as a data set after wavelet transform filtering processing; and inputting the data set into the TFT regression model to obtain fracturing induced time-space intensity multi-dimensional parameter early warning index data. And inputting the characteristic variables in the data set and the target variables about whether the earthquake occurs or not into the random forest classification model to obtain the predicted fracturing earthquake probability. And finally, time-space strong multi-dimensional data obtained through TFT regression prediction and the seismic probability obtained through random forest classification are integrated into a new data set, multi-layer perception MLP and Transform fusion models are used for training evaluation, a rock fracturing seismic multi-parameter fusion early warning result is obtained through weighted fusion, and time-space strong multi-dimensional advanced early warning of reservoir hydraulic fracturing induced earthquakes is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of hydraulic fracturing induced earthquake early warning, and particularly to a multi-information fusion early warning method and system for hydraulic fracturing induced earthquake in engineering rock mass. Background Art

[0002] As a renewable energy source, hot dry rock has great development potential. By using hydraulic fracturing technology to extract heat from hot dry rock, heat energy can be provided for power generation. However, the hydraulic fracturing process may induce seismic activities. Especially during the fracturing of hot dry rock, the injection of high-pressure water will change the stress state of underground rocks, leading to the occurrence of earthquakes. This induced earthquake phenomenon not only poses a threat to the safety of equipment and personnel, but also may affect the environmental stability of the surrounding areas. Therefore, researching and developing a method and system that can effectively early warn of earthquakes induced by hot dry rock fracturing is of great significance for ensuring the safety and stability of hot dry rock development.

[0003] Microseismic / acoustic emission monitoring technology is of irreplaceable importance in the monitoring of earthquakes induced by hot dry rock. These technologies can real-time monitor the generation and propagation process of microcracks in rock mass, and provide early warning signals for seismic activities. Microseismic / acoustic emission monitoring technology captures tiny vibration signals through sensors arranged underground or on the surface, and through data analysis, identifies and locates seismic events. Compared with traditional seismic monitoring methods, microseismic / acoustic emission technology has higher sensitivity and resolution, and can detect extremely small-scale rock fracture events, which is an ideal choice for early warning of earthquakes induced by hot dry rock. For example, Patent CN114779330A discloses a method for analyzing and predicting the main fracture orientation of an excavation working face based on microseismic monitoring; Patent CN115983465A discloses a method for constructing a time series prediction model of rock burst based on small sample learning; Patent CN116307683A provides an evaluation method for earthquakes induced by hydraulic fracturing activating faults, etc. The above achievements provide reference for microseismic / acoustic emission monitoring and early warning of earthquakes induced by hot dry rock fracturing.

[0004] However, most of the above methods for monitoring and early warning of seismic activities induced by rock fracturing have the following deficiencies: First, the analysis method mainly relies on a single parameter for early warning, and fails to fully consider the multi-parameter characteristics of seismic activities, resulting in low accuracy and reliability of early warning. Second, when dealing with seismic data, the existing methods often ignore the complex dependence relationship of time series data and fail to fully utilize the potential information of time series data. In addition, most of the existing early warning systems have not achieved multi-model fusion, and it is difficult to comprehensively utilize the advantages of different models to improve the overall performance of the early warning system.

[0005] Based on the above problems, there is an urgent need for an intelligent early warning method and system for multi-information fusion of induced earthquakes in engineering rock masses, which aims to achieve accurate early warning of rock fracturing-induced earthquakes through multi-parameter fusion, multi-model collaboration, and in-depth mining of time series data, improve the reliability and stability of the early warning system, and provide strong support for ensuring the safety and sustainability of dry hot rock development. Summary of the Invention

[0006] In response to the problems and requirements raised above, the present solution proposes an early warning method and system for multi-information fusion of induced earthquakes in engineering rock masses. Due to the following technical features, it can achieve the above technical objectives and bring many other technical effects.

[0007] An object of the present invention is to propose an early warning method for multi-information fusion of induced earthquakes in engineering rock masses, including the following steps:

[0008] S10: Collect the original waveform data induced during the process of rock fracturing-induced earthquake, and perform filtering processing on the obtained original waveform data to obtain effective waveform data;

[0009] S20: Extract characteristic coefficients and invert the focal mechanism for the effective waveform data to obtain spatio-temporal-intensity multi-dimensional early warning index data of waveform characteristic parameters, focal location, and focal mechanism parameters. Use the original waveform data induced by fracturing and the spatio-temporal-intensity multi-dimensional early warning index data as a data set;

[0010] S30: Input the data set into the TFT regression model for regression analysis, and perform preprocessing by means of normalization and filling missing values, train the model and evaluate the model performance to obtain predicted spatio-temporal-intensity multi-dimensional data;

[0011] S40: Extract characteristic variables from the spatio-temporal-intensity multi-dimensional early warning index data set, and obtain the target variable regarding whether an earthquake is induced through the analysis of historical microseismic data; Preprocess the characteristic variables extracted from the data set and the target variable regarding whether an earthquake is induced, and then input them into the random forest classification model for classification analysis. Use the method of integrating multiple decision trees for hazard classification, and train and evaluate the performance of the preprocessed data through the random forest classification model to obtain the predicted earthquake-induced probability;

[0012] S50: Use the predicted spatio-temporal-intensity multi-dimensional data and the earthquake-induced probability obtained from hazard classification as input data, reorganize and establish a new data set, divide it into a training set and a test set, use the multi-layer perceptron MLP and Transformer fusion model to train and evaluate the training set data, and perform weighted fusion on the outputs of the multi-layer perceptron MLP and Transformer using the test set data to obtain the multi-parameter fusion early warning result of rock fracturing-induced earthquake.

[0013] In addition, the multi-information fusion early warning method for induced earthquake in engineering rock mass according to the present invention may further have the following technical features:

[0014] In an example of the present invention, in step S10, the acquired original waveform data is filtered to obtain effective waveform data, which specifically includes the following steps:

[0015] S11: Decompose the original waveform data to determine the wavelet basis function and the decomposition level;

[0016] S12: Perform threshold processing on the decomposed signal to remove the noise components;

[0017] S13: Reconstruct the signal after removing the noise to obtain the filtered effective waveform data.

[0018] In an example of the present invention, in step S20, feature coefficients are extracted from the effective waveform data and the focal mechanism is inverted to obtain spatio-temporal-intensity multi-dimensional early warning index data of waveform characteristic parameters, focal location and focal mechanism parameters, including:

[0019] S21: Perform feature recognition on the induced waveform to extract waveform feature coefficients, and the feature coefficients include ring count, energy release rate, frequency band energy distribution, fractal feature, origin time and first motion amplitude;

[0020] S22: Invert the focal mechanism parameters, and the focal mechanism parameters include rupture occurrence position index and rupture occurrence intensity index, where the rupture occurrence position index includes spatial concentration factor, spatial distribution intensity coefficient, error spatial form, tension-shear attribute, rupture plane attitude, extension azimuth, and the rupture occurrence intensity index includes seismic moment magnitude, b value, stress drop, corner frequency;

[0021] S23: Establish a data set according to the waveform feature coefficient data and the focal mechanism parameter data.

[0022] In an example of the present invention, step S30 specifically includes the following steps:

[0023] S31: Construct a regression data set, and combine the waveform feature coefficients and the focal mechanism parameters in the effective waveform data to form a regression data set;

[0024] S32: Construct a TFT regression model, and input the regression data set into the TFT regression model for regression prediction analysis;

[0025] S33: Performance evaluation and result prediction, use the mean absolute scaled error MASE to quantitatively evaluate the prediction result of the model; output multi-parameter prediction data and prediction intervals through the trained model to obtain the future seismic activity trend.

[0026] In an example of the present invention, the TFT regression model includes:

[0027] A data processing and feature selection module, configured to select the most significant features from the input using a variable selection network (VSN), and encode the context vector of the static covariates through a static covariate encoder (SCE);

[0028] An information flow modeling module, configured to ensure effective information flow using a gated residual network (GRN), learn the long-term dependencies of time series data using a temporal self-attention layer (TSL), and perform additional non-linear processing on the output of the self-attention layer through a position-guided feed-forward layer (PFL);

[0029] A decoding and prediction module, configured to enhance the temporal features using a static enrichment layer (SEL) in a temporal fusion decoder (TFD);

[0030] A training optimization module, configured to use the mean squared error (MSE) as the loss function during the training process, optimize the parameters using an Adam optimizer, and iteratively update the model parameters through a mini-batch stochastic gradient descent algorithm.

[0031] In an example of the present invention, in S40, the feature variables extracted from the dataset and the target variable regarding whether an earthquake occurs are preprocessed, and then input into a random forest classification model for classification analysis, which specifically includes the following steps:

[0032] S41: Dataset division and data preprocessing: Combine the feature variables extracted from the spatio-temporal strong multi-dimensional early warning index dataset and the target variable regarding whether an earthquake occurs to form a classification dataset;

[0033] S42: Construct a random forest model: Generate multiple sub-samples through bootstrap sampling from the classification dataset, train a decision tree for each sub-sample, reduce the overfitting risk by integrating multiple trees, and finally determine the final category through the majority voting method;

[0034] S43: Model training and evaluation: Use the cross-entropy loss function to optimize the model, calculate the classification effect of each tree during the training process; evaluate the performance of the model through evaluation metrics, and analyze the classification performance of the model through a confusion matrix; wherein, the evaluation metrics include: accuracy, recall rate, and comprehensive accuracy rate;

[0035] S44: Use the trained random forest model to classify and predict future earthquake data, and output the probability of an earthquake occurring at each time point.

[0036] In an example of the present invention, the accuracy rate A, recall rate R, and comprehensive accuracy rate F 1 , and the specific calculation formulas are as follows:

[0037]

[0038] Wherein, TP is the assumed true positive: the actual situation is earthquake-causing and the model recognition result is earthquake-causing, that is, the earthquake-causing phenomenon is correctly recognized; FP is the false positive: the actual situation is non-earthquake-causing but the model recognition result is earthquake-causing, that is, the non-earthquake-causing phenomenon is misrecognized; TN is the true negative: the actual situation is non-earthquake-causing and the model recognition result is non-earthquake-causing, that is, the non-earthquake-causing phenomenon is correctly recognized; FN is the false negative: the actual situation is earthquake-causing but the model recognition result is non-earthquake-causing, that is, the earthquake-causing phenomenon is misrecognized.

[0039] In an example of the present invention, in the step S50, a multi-layer perceptron MLP and a Transformer fusion model are used to train and evaluate the training set data, and the specific classification steps are as follows:

[0040] S51: Construct a data set: Integrate the output data of the TFT regression model and the random forest classification model, that is, the predicted spatio-temporal strong multi-dimensional data and the earthquake-causing probability, into a new data set;

[0041] S52: Construct an MLP and Transformer fusion model: Train the MLP model and the Transformer model respectively, automatically select the MLP or Transformer model according to the task characteristics of regression or classification, determine the best hyperparameter combination, and further perform hyperparameter tuning, calculate the errors of the training set and the validation set, judge whether the model is overfitting or underfitting, and evaluate the model performance;

[0042] S53: Output the multi-parameter fusion warning result of rock fracturing earthquake-causing: Weightedly fuse the predicted spatio-temporal strong multi-dimensional data output by the MLP and Transformer models and the classified and predicted earthquake-causing probability, and finally generate the multi-dimensional parameter fusion warning result of microseismic monitoring related to the waveform difference characteristics, earthquake source occurrence time, location, and intensity of rock fracturing earthquake-causing.

[0043] In an example of the present invention, constructing an MLP and Transformer fusion model includes the following steps:

[0044] S521: Model training: When training the MLP, use the mean square error MSE as the loss function and the Adam optimizer to optimize the parameters, and iteratively update the model parameters through the mini-batch stochastic gradient descent algorithm; when training the Transformer, use the cross-entropy loss function, optimize with the Adam optimizer, and capture the complex relationships between features through the multi-head attention mechanism;

[0045] S522: Model Selection and Hyperparameter Tuning: Automatically select an MLP or Transformer model according to the task characteristics of regression or classification, determine the optimal hyperparameter combination through cross-validation, and further tune the hyperparameters to optimize the model performance;

[0046] S523: Model Evaluation: During training, calculate the errors of the training set and the validation set through MSE and cross-entropy to determine whether the model is overfitting or underfitting. For regression tasks, use the root mean square error (RMSE) to evaluate the model performance, and for classification tasks, use accuracy, precision, and recall for evaluation.

[0047] Another object of the present invention is to propose an engineering rock mass fracturing-induced earthquake multi-information fusion early warning system, including:

[0048] An acquisition module configured to collect the original waveform data induced by the rock fracturing-induced earthquake process, and perform filtering processing on the obtained original waveform data to obtain effective waveform data;

[0049] A processing module configured to extract characteristic coefficients and invert the focal mechanism from the effective waveform data to obtain spatio-temporal-intensity multi-dimensional early warning index data of waveform characteristic parameters, focal location, and focal mechanism parameters, and use the original waveform data induced by fracturing and the spatio-temporal-intensity multi-dimensional early warning index data as a data set;

[0050] A regression model module configured to input the data set into a TFT regression model for regression analysis, and perform preprocessing by means of normalization and filling missing values, train the model and evaluate the model performance to obtain predicted spatio-temporal-intensity multi-dimensional data;

[0051] A classification model module configured to extract characteristic variables from the spatio-temporal-intensity multi-dimensional early warning index data set, and obtain target variables regarding whether an earthquake is induced through the analysis of historical microseismic data; perform preprocessing on the characteristic variables extracted from the data set and the target variables regarding whether an earthquake is induced, and then input them into a random forest classification model for classification analysis, use the method of multi-decision tree integration for hazard classification, and train and evaluate the performance of the preprocessed data through the random forest classification model to obtain the predicted earthquake-induced probability;

[0052] A weighted fusion module configured to use the predicted spatio-temporal-intensity multi-dimensional data and the earthquake-induced probability obtained from the hazard classification as input data, re-integrate to establish a new data set, divide it into a training set and a test set, use a multi-layer perceptron (MLP) and a Transformer fusion model to train and evaluate the training set data, and perform weighted fusion on the outputs of the multi-layer perceptron (MLP) and the Transformer using the test set data to obtain the multi-parameter fusion early warning result of rock fracturing-induced earthquake.

[0053] This technical solution has the following beneficial effects compared with the prior art:

[0054] This technical solution obtains and processes the induced waveform data of hydraulic fracturing-induced earthquakes in hot dry rocks, uses wavelet transform filtering technology to obtain effective waveform data, and further extracts characteristic coefficients and focal mechanism parameters. Inputting these multi-parameter data into the TFT regression model for regression analysis and combining with the random forest classification model for classification analysis, through the multi-parameter prediction and earthquake-induced probability analysis of the fusion model, the possibility of future earthquakes can be evaluated more accurately. This technical solution effectively solves the data noise and nonlinear problems in traditional methods, significantly improves the overall performance of the early warning system, and provides a scientific reference basis for the early warning of hydraulic fracturing-induced earthquakes in hot dry rocks.

[0055] This technical solution comprehensively utilizes various technical means such as wavelet transform filtering, characteristic coefficient extraction, and focal mechanism parameter inversion to perform multi-dimensional analysis on the seismic waveform data induced by hydraulic fracturing in hot dry rocks. By jointly using the regression model and the classification model, it can not only predict multi-parameter data such as future magnitudes and energy release rates but also evaluate the probability of future earthquake induction. The fusion model further integrates the results of regression prediction and classification analysis to achieve a comprehensive early warning of hydraulic fracturing-induced earthquakes in hot dry rocks. This technical solution has significant advantages in data processing, model training, and prediction evaluation, can effectively improve the accuracy and stability of earthquake early warning, and provides important technical support for seismic safety monitoring during the exploitation of hot dry rocks.

[0056] The fusion model of this technical solution utilizes the advantages of MLP and Transformer to capture the nonlinear features and time-dependent relationships in the data, improving the accuracy and stability of prediction. Through the late-stage fusion model, the prediction results of the regression and classification models are comprehensively utilized to improve the overall performance of the early warning system. The MLP and Transformer models are respectively suitable for different data characteristics and task requirements, providing more flexible choices. The late-stage fusion model can reduce the bias and error of a single model, enhance the stability and accuracy of prediction, and finally achieve multi-parameter fusion early warning of hydraulic fracturing-induced earthquakes in hot dry rocks through the fusion of regression prediction and hazard classification results.

[0057] In the following text, the optimal embodiments of implementing the present invention will be described in more detail in conjunction with the accompanying drawings, so as to easily understand the features and advantages of the present invention. Brief Description of the Drawings

[0058] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings of the embodiments of the present invention will be briefly introduced below. Among them, the accompanying drawings are only used to show some embodiments of the present invention, rather than limiting all embodiments of the present invention thereto.

[0059] Figure 1Flow chart of a multi - information fusion early warning method for engineering rock mass fracturing - induced earthquake according to an embodiment of the present invention;

[0060] Figure 2 Wavelet transform filtering flow chart according to an embodiment of the present invention;

[0061] Figure 3 Schematic diagram of the FTF regression model structure according to an embodiment of the present invention;

[0062] Figure 4 Schematic diagram of the random forest model structure according to an embodiment of the present invention;

[0063] Figure 5 Relationship diagram of space - time - intensity multi - dimensional parameter early warning indicators according to an embodiment of the present invention. Detailed implementation manners

[0064] In order to make the objectives, technical solutions and advantages of the technical solutions of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings of the specific embodiments of the present invention. The same reference numerals in the drawings represent the same components. It should be noted that the described embodiments are part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the described embodiments of the present invention without creative efforts fall within the scope of protection of the present invention.

[0065] Unless otherwise defined, the technical terms or scientific terms used herein shall have the ordinary meanings understood by those of ordinary skill in the art to which the present invention pertains. The "first", "second" and similar terms used in the specification and claims of this patent application of the present invention do not denote any order, quantity or importance, but are only used to distinguish different components. Similarly, terms such as "a" or "an" do not necessarily denote a quantity limitation. The terms "including" or "comprising" and similar terms mean that the elements or items appearing before this term cover the elements or items listed after this term and their equivalents, without excluding other elements or items. The terms "connected" or "coupled" and similar terms are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. The terms "upper", "lower", "left", "right", etc. are only used to represent relative positional relationships, and when the absolute position of the object being described changes, the relative positional relationship may also change accordingly.

[0066] A multi - information fusion early warning method for engineering rock mass fracturing - induced earthquake according to the first aspect of the present invention, as Figure 1 shown, includes the following steps:

[0067] S10: Arrange microseismic sensors according to the actual situation, collect the original waveform data induced by the rock fracturing earthquake process, and perform filtering processing on the obtained original waveform data to obtain effective waveform data; for example, filtering processing can be performed through wavelet transform;

[0068] S20: Extract characteristic coefficients and invert the source mechanism for the effective waveform data to obtain multi-dimensional early warning index data of waveform characteristic parameters, source location, and source mechanism parameters in terms of space, time, and intensity. Take the original waveform data induced by fracturing and the multi-dimensional early warning index data in terms of space, time, and intensity as a data set;

[0069] S30: Input the data set into the TFT regression model for regression analysis, and perform preprocessing by means of normalization and filling missing values, train the model and evaluate the model performance, and obtain the predicted multi-dimensional data in terms of space, time, and intensity to provide data for subsequent early warning of future seismic activity trends;

[0070] S40: Extract characteristic variables from the multi-dimensional early warning index data set in terms of space, time, and intensity, and obtain the target variable regarding whether an earthquake is induced through the analysis of historical microseismic data; perform preprocessing on the characteristic variables extracted from the data set and the target variable regarding whether an earthquake is induced, and then input them into the random forest classification model for classification analysis, and use the method of integrating multiple decision trees for hazard classification to solve data noise and nonlinear problems. Through the training and performance evaluation of the random forest classification model on the preprocessed data, obtain the predicted earthquake-induced probability to provide data for subsequent early warning of the possibility of future earthquakes;

[0071] S50: Take the predicted multi-dimensional data in terms of space, time, and intensity and the earthquake-induced probability obtained from hazard classification as input data, reorganize and establish a new data set, and divide it into a training set and a test set. Use the multi-layer perceptron MLP and Transformer fusion model to train and evaluate the training set data, and perform weighted fusion on the outputs of the multi-layer perceptron MLP and Transformer using the test set data to obtain the multi-parameter fusion early warning result of rock fracturing earthquake-induced.

[0072] This method obtains and processes the induced waveform data of hot dry rock fracturing earthquake, uses wavelet transform filtering technology to obtain effective waveform data, and further extracts characteristic coefficients and source mechanism parameters. Input these multi-parameter data into the TFT regression model for regression analysis, combine with the random forest classification model for classification analysis, and through the multi-parameter prediction and earthquake-induced probability analysis of the fusion model, can more accurately evaluate the possibility of future earthquakes. This method effectively solves the data noise and nonlinear problems in traditional methods, significantly improves the overall performance of the early warning system, and provides a scientific reference basis for the early warning of hot dry rock fracturing-induced earthquakes.

[0073] This method comprehensively utilizes various technical means such as wavelet transform filtering, feature coefficient extraction, and seismic source mechanism parameter inversion to perform multi-dimensional analysis on the seismic waveform data induced by hot dry rock fracturing. By jointly using regression models and classification models, it can not only predict multi-parameter data such as future magnitudes and energy release rates, but also evaluate the probability of future earthquake occurrence. The fusion model further integrates the results of regression prediction and classification analysis to achieve a comprehensive early warning of earthquakes induced by hot dry rock fracturing. This method has significant advantages in data processing, model training, and prediction evaluation, can effectively improve the accuracy and stability of earthquake early warning, and provides important technical support for seismic safety monitoring during the exploitation of hot dry rock.

[0074] The fusion model of this method utilizes the advantages of MLP and Transformer to capture the non-linear features and time-dependent relationships in the data, improving the accuracy and stability of prediction. Through the late fusion model, the prediction results of regression and classification models are comprehensively utilized to improve the overall performance of the early warning system. The MLP and Transformer models are respectively suitable for different data characteristics and task requirements, providing more flexible choices. The late fusion model can reduce the bias and error of a single model, enhance the stability and accuracy of prediction, and finally achieve multi-parameter fusion early warning of earthquakes induced by hot dry rock fracturing through the fusion of regression prediction and hazard classification results.

[0075] In an example of the present invention, in step S10, the acquired original waveform data is subjected to wavelet transform filtering to obtain effective waveform data, as Figure 2 shown, and specifically includes the following steps:

[0076] S11: Decompose the original waveform data to determine the wavelet basis function and the decomposition level; for example, select Daubechies db4 and db6 wavelets as the basis functions, and the formula is as follows:

[0077]

[0078] where ψ(t) is the Daubechies wavelet function, including db4 and db6 wavelet functions, φ(t) is the approximation function, b k is the filtering coefficient of ψ(t), k = 4, M = 4 for db4 wavelet, k = 6, M = 4 for db6 wavelet. Select 3 to 6 levels of wavelet decomposition to decompose the original signal into different frequency band components.

[0079] S12: Threshold process the decomposed signal to remove the noise components; specifically, perform threshold processing on the high-frequency components, set the threshold according to the standard deviation of the noise, and finally combine all components to obtain the signal after removing the noise.

[0080] S13: Reconstruct the signal after noise removal to obtain the effective waveform data after filtering. Specifically, combine the low-frequency and high-frequency components of the signal after noise removal and recover the effective waveform data after filtering through inverse wavelet transform.

[0081] In an example of the present invention, in step S20, perform feature coefficient extraction and focal mechanism inversion on the effective waveform data to obtain spatio-temporal-intensity multi-dimensional warning index data of waveform characteristic parameters, focal location, and focal mechanism parameters, including:

[0082] S21: Perform feature recognition on the induced waveform, extract waveform feature coefficients, and the feature coefficients include ring count, energy release rate, band energy distribution, fractal feature, origin time, and initial motion amplitude; among them, the rupture occurrence time index includes origin time and initial motion amplitude;

[0083] Specifically, ring count: the number of times exceeding the preset threshold; energy release rate: judge the driving force for crack propagation, and the formula is as follows: The formula for band energy distribution is as follows: P(f) = |N(f)| 2 , where P(f) is the energy density at frequency f, and X(f) is the Fourier transform value of the signal at frequency f; fractal feature: described by the box dimension in the fractal dimension, and the formula is: Among them, N(x) is the number of boxes with a radius of x required to cover the waveform signal; origin time: the time when the first peak or trough exceeding the set threshold appears; initial motion amplitude: the amplitude of the first peak or trough in the initial stage of the waveform signal.

[0084] S22: Invert the focal mechanism parameters, and the focal mechanism parameters include rupture occurrence position indicators and rupture occurrence intensity indicators. Among them, the rupture occurrence position indicators include spatial concentration factor, spatial distribution intensity coefficient, error spatial form, tension-shear attribute, rupture plane attitude, and extension azimuth, and the rupture occurrence intensity indicators include seismic moment magnitude, b value, stress drop, and corner frequency; The spatio-temporal-intensity multi-dimensional parameter warning index relationship diagram is as Figure 5 shown,

[0085] Specifically, (1) The rupture occurrence position indicators are as follows:

[0086] Spatial concentration factor: the degree of spatial concentration of the focal stress distribution, and the formula is as follows:

[0087]

[0088] Among them, N is the number of rupture sources in the unit volume V;

[0089] Spatial distribution intensity coefficient: the spatial intensity of the focal stress distribution, and the formula is as follows:

[0090]

[0091] Among them, X i is the distance from the rupture source to the center of the volume within the unit volume V, and g i is the energy of the rupture source i;

[0092] Error space morphology: The spatial difference between the theoretical calculated value and the actual observed value during the inversion process,

[0093]

[0094] Among them, The observed value of the jth component of the ith station, is the corresponding theoretical value, n is the number of stations, and m is the number of components;

[0095] Tensile-shear property: The relative action of tension and shear force during the rupture process. The formula is as follows:

[0096]

[0097] Among them, L is the tensile-shear ratio: when L > 1, the tensile rupture property is dominant; when L < 1, the corresponding shear rupture property is dominant;

[0098] Rupture scale: The scale of the rupture. The formula is as follows:

[0099]

[0100] In the formula, ΔA is the rupture area, and M 1 、M 3 are the eigenvalues of the moment tensor, μ is the Lame constant, and ξ is the proportionality coefficient between the rupture area and the displacement;

[0101] Rupture plane attitude: The spatial orientation of the rupture plane. The formula is as follows:

[0102]

[0103] n = (cosβ, 0, ±sinβ)

[0104] Among them, β is the angle between the movement direction and the normal direction of the earthquake source rupture plane, and M 1 、M 2 、M 3 are the eigenvalues of the moment tensor, μ and λ are the Lame constants, and n is the spatial orientation;

[0105] Extension azimuth: It is the movement direction of the rupture plane. The formula is as follows:

[0106]

[0107] Among them, β is the angle between the movement direction and the normal direction of the earthquake source rupture plane, and b is the movement direction;

[0108] (2) The rupture occurrence intensity indicators are as follows:

[0109] Seismic moment magnitude: Describes the size of an earthquake, and the formula is as follows:

[0110]

[0111] Where, M 0 is the seismic moment,

[0112] b value: Describes the frequency - magnitude distribution of earthquake activities, and the formula is as follows:

[0113] b = log 10 (N) / M

[0114] Where, N is the number of earthquakes with a magnitude greater than or equal to a certain magnitude M;

[0115] Stress drop: The change value of the shear stress at a point on the rupture surface in the final state before and after rupture, and the formula is as follows:

[0116]

[0117] Where, M 0 is the seismic moment, and R is the source radius;

[0118] Corner frequency: Obtained by fitting the high - frequency part of the seismic wave spectrum, and its relationship with the seismic moment M 0 and the stress drop Δσ is as follows, and the formula is as follows:

[0119] fc = 1 / 2π(Δσ / ρV s 3 ) 1 / 2 ,

[0120] Where ρ is the rock density and Vs is the shear wave velocity.

[0121] S23: Establish a data set based on the waveform characteristic coefficient data and the source mechanism parameter data.

[0122] In an example of the present invention, the step S30 specifically includes the following steps:

[0123] S31: Construct a regression data set, combine the waveform characteristic coefficients and source mechanism parameters in the effective waveform data to form a regression data set; the data set uses a table form, and each row represents the data record of a time point, and each record contains all the characteristics of a time point. Use the Pandas library to load the data set, normalize the feature variables, and process the missing data points through methods such as linear interpolation and mean filling to ensure data integrity and consistency.

[0124] S32: Construct a TFT regression model and input the regression dataset into the TFT regression model for regression prediction analysis;

[0125] The TFT regression model is selected because the TFT model enhances the ability to capture time features through the Time Fusion Decoder (TFD), and is particularly suitable for the long-term sequence prediction of hot dry rock seismic activities.

[0126] Among them, the configuration of constructing the TFT regression model includes:

[0127] The input multi-parameter data is 17-dimensional multi-parameter data including ring count, energy release rate, duration, band energy distribution, fractal feature, origin time, initial motion amplitude, spatial concentration factor, spatial distribution intensity coefficient, error spatial morphology, tension-shear attribute, rupture scale, rupture plane attitude, seismic moment magnitude, corner frequency, b value, and extension azimuth; the number of neurons in the hidden layer is determined to be 256 through grid search or Bayesian optimization; the number of heads of the multi-head attention mechanism is 8; the number of layers of the Transformer is 4; the output is 17-dimensional multi-parameter data at future time points.

[0128] S33: Performance evaluation and result prediction. The mean absolute scaled error MASE is used to quantitatively evaluate the prediction results of the model; the trained model outputs multi-parameter prediction data and prediction intervals to obtain the future seismic activity trend, providing data support for subsequent earthquake early warnings. The mean absolute scaled error MASE is defined as follows:

[0129] Assume that y represents the true value, represents the predicted value, and n is the number of samples. Then the calculation formula of MASE is:

[0130]

[0131] When the MASE value is smaller, it means that the prediction performance of the model is better.

[0132] In an example of the present invention, as Figure 3 shown, the TFT regression model includes:

[0133] A data processing and feature selection module, configured to select the most significant features from the input using a variable selection network VSN and encode the context vector of the static covariates through a static covariate encoder SCE;

[0134] An information flow modeling module, configured to ensure effective information flow using a gated residual network GRN, learn the long-term dependencies of time series data using a time self-attention layer TSL, and perform additional non-linear processing on the output of the self-attention layer through a position-guided feed-forward layer PFL;

[0135] A decoding and prediction module, configured to enhance temporal features by using a static enrichment layer (SEL) in a temporal fusion decoder (TFD);

[0136] A training optimization module, configured to use the mean squared error (MSE) as the loss function during the training process, and the Adam optimizer for parameter optimization, and iteratively update the model parameters through the mini-batch stochastic gradient descent algorithm.

[0137] In an example of the present invention, in S40, the feature variables extracted from the dataset and the target variable regarding whether an earthquake occurs are preprocessed, and then input into a random forest classification model for classification analysis, which specifically includes the following steps:

[0138] S41: Dataset division and data preprocessing: Combine the feature variables extracted from the spatio-temporal strong multi-dimensional early warning index dataset and the target variable regarding whether an earthquake occurs to form a classification dataset; wherein, the dataset is in tabular form, and each row represents a data record at a time point, including feature variables and target variables. Use the Pandas library to load the classification dataset, and normalize the feature variables to ensure that each feature quantity is on the same scale, improve the model training effect, and use the mean filling or linear interpolation method to ensure data integrity.

[0139] S42: Construct a random forest model: As Figure 4 shown, generate multiple subsamples from the classification dataset through bootstrap sampling. For each subsample, train a decision tree, and reduce the overfitting risk by integrating multiple trees. Finally, determine the final class through the majority voting method;

[0140] For the dangerous classification problem of whether an earthquake occurs, determine the class through voting. The random forest integrates multiple trees to reduce the overfitting risk, improve the classification accuracy and stability, and evaluate the feature importance. The random forest realizes a strong classifier by constructing multiple weak classifiers and synthesizing the results, improves the classification performance, and uses the bootstrap sampling method to generate multiple subsamples and train multiple decision trees.

[0141] Among them, the configuration for constructing the random forest model includes:

[0142] The input is multi-parameter data including ring count, energy release rate, duration, frequency band energy distribution, fractal feature, earthquake occurrence time, initial motion amplitude, spatial concentration factor, spatial distribution intensity coefficient, error spatial morphology, tension-shear attribute, rupture scale, rupture plane attitude, seismic moment magnitude, corner frequency, b value, and extension azimuth; determine the number of trees to be 64 through the grid search method or Bayesian optimization; the maximum depth is 7 layers; the output is the earthquake occurrence probability at the current time node.

[0143] S43: Model Training and Evaluation: Use the cross-entropy loss function to optimize the model, and calculate the classification effect of each tree during the training process; evaluate the performance of the model through evaluation metrics, and analyze the classification performance of the model through the confusion matrix; among them, the evaluation metrics include: accuracy, recall rate, and comprehensive accuracy rate;

[0144] The definitions of accuracy, recall rate, and F1-score are as follows:

[0145] Assume true positive (TP): The actual situation is earthquake-causing and the model's recognition result is earthquake-causing, that is, the earthquake-causing phenomenon is correctly recognized; false positive (FP): The actual situation is non-earthquake-causing but the model's recognition result is earthquake-causing, that is, the non-earthquake-causing phenomenon is misrecognized; true negative (TN): The actual situation is non-earthquake-causing and the model's recognition result is non-earthquake-causing, that is, the non-earthquake-causing phenomenon is correctly recognized; false negative (FN): The actual situation is earthquake-causing but the model's recognition result is non-earthquake-causing, that is, the earthquake-causing phenomenon is misrecognized.

[0146] The three evaluation criteria are accuracy rate A, recall rate R, and comprehensive accuracy rate F 1 , and the specific calculation formulas are as follows:

[0147]

[0148] S44: Use the trained random forest model to classify and predict future earthquake data, output the probability of earthquake occurrence at each time point, and the prediction results evaluate the possibility of future earthquakes, providing a reference basis for earthquake early warning.

[0149] In an example of the present invention, the accuracy rate A, recall rate R, and comprehensive accuracy rate F 1 , and the specific calculation formulas are as follows:

[0150]

[0151] In the formula, TP is the assumed true positive: The actual situation is earthquake-causing and the model's recognition result is earthquake-causing, that is, the earthquake-causing phenomenon is correctly recognized; FP is the false positive: The actual situation is non-earthquake-causing but the model's recognition result is earthquake-causing, that is, the non-earthquake-causing phenomenon is misrecognized; TN is the true negative: The actual situation is non-earthquake-causing and the model's recognition result is non-earthquake-causing, that is, the non-earthquake-causing phenomenon is correctly recognized; FN is the false negative: The actual situation is earthquake-causing but the model's recognition result is non-earthquake-causing, that is, the earthquake-causing phenomenon is misrecognized.

[0152] In an example of the present invention, in the step S50, use a multi-layer perceptron MLP and Transformer fusion model to train and evaluate the training set data, and the specific classification steps are as follows:

[0153] S51: Construct a dataset: Integrate the output data of the TFT regression model and the random forest classification model, that is, the predicted spatio-temporal strong multi-dimensional data and the earthquake occurrence probability, into a new dataset; the integrated dataset is in tabular form, and each row contains the multi-parameter data predicted by regression and the earthquake occurrence probability predicted by classification. Divide the integrated dataset into a training set and a test set; the division ratio is 70% for training and 30% for testing to ensure the fairness of model training and evaluation.

[0154] S52: Construct an MLP and Transformer fusion model: Train the MLP model and the Transformer model respectively, automatically select the MLP or Transformer model according to the task characteristics of regression or classification, determine the best hyperparameter combination, and further optimize the hyperparameters, calculate the errors of the training set and the validation set, judge whether the model is overfitting or underfitting, and evaluate the model performance;

[0155] Among them, the configuration of the MLP and Transformer models includes:

[0156] The MLP configuration is as follows: The input is the multi-parameter data predicted by regression and the earthquake occurrence probability predicted by classification, the number of neurons in the hidden layer is two layers, with 128 neurons in each layer, the activation function is ReLU, and the output is the final earthquake occurrence probability.

[0157] The Transformer configuration is as follows: The input is the multi-parameter data predicted by regression and the earthquake occurrence probability predicted by classification, the number of attention heads is 8, the number of neurons in each hidden layer is 256, the number of layers is 4 layers, and the output is the final earthquake occurrence probability.

[0158] S53: Output the multi-parameter fusion early warning result of rock fracturing-induced earthquake: Weightedly fuse the predicted spatio-temporal strong multi-dimensional data output by the MLP and Transformer models and the earthquake occurrence probability predicted by classification, and finally generate the multi-dimensional parameter fusion early warning result of microseismic monitoring related to the waveform difference characteristics, earthquake source occurrence time, location, and intensity of rock fracturing-induced earthquake.

[0159] In an example of the present invention, constructing an MLP and Transformer fusion model includes the following steps:

[0160] S521: Model training: When training the MLP, use the mean squared error MSE as the loss function and the Adam optimizer to optimize the parameters, and iteratively update the model parameters through the mini-batch stochastic gradient descent algorithm; when training the Transformer, use the cross-entropy loss function, optimize with the Adam optimizer, and capture the complex relationships between features through the multi-head attention mechanism;

[0161] S522: Model Selection and Hyperparameter Tuning: Automatically select an MLP or Transformer model according to the task characteristics of regression or classification, determine the optimal hyperparameter combination through cross-validation, and further tune the hyperparameters to optimize the performance of the model;

[0162] S523: Model Evaluation: During training, calculate the errors of the training set and the validation set through MSE and cross-entropy to determine whether the model is overfitting or underfitting. For regression tasks, use the root mean square error (RMSE) to evaluate the model performance, and for classification tasks, use accuracy, precision, and recall for evaluation. Among them, the root mean square error is defined as follows:

[0163]

[0164] where n is the number of data points, y i is the i-th actual value, is the i-th predicted value.

[0165] An engineering rock mass fracturing-induced earthquake multi-information fusion early warning system according to the second aspect of the present invention includes:

[0166] A collection module configured to arrange microseismic sensors according to actual conditions, collect the original waveform data induced by the rock fracturing earthquake process, and perform wavelet transform filtering on the obtained original waveform data to obtain effective waveform data;

[0167] A processing module configured to extract characteristic coefficients and invert the focal mechanism for the effective waveform data to obtain spatio-temporal intensity multi-dimensional warning index data of waveform characteristic parameters, focal location, and focal mechanism parameters, and use the original waveform data induced by fracturing and the spatio-temporal intensity multi-dimensional warning index data as a data set;

[0168] A regression model module configured to input the data set into a TFT regression model for regression analysis, and perform preprocessing by means of normalization and filling missing values, train the model and evaluate the model performance, and obtain the predicted spatio-temporal intensity multi-dimensional data to provide data for subsequent early warning of future earthquake activity trends;

[0169] A classification model module configured to extract characteristic variables from the spatio-temporal intensity multi-dimensional warning index data set, and obtain target variables regarding whether an earthquake is induced through the analysis of historical microseismic data; preprocess the characteristic variables extracted from the data set and the target variables regarding whether an earthquake is induced, and then input them into a random forest classification model for classification analysis, and use the method of integrating multiple decision trees for hazard classification to solve data noise and nonlinear problems. Through the training and performance evaluation of the preprocessed data by the random forest classification model, obtain the predicted earthquake-induced probability to provide data for subsequent early warning of future earthquake possibilities;

[0170] The weighted fusion module is configured to use the predicted spatio-temporal strong multi-dimensional data and the earthquake-causing probability obtained from hazard classification as input data, re-integrate to establish a new data set, divide it into a training set and a test set, use a multi-layer perceptron MLP and a Transformer fusion model to train and evaluate the training set data, and perform weighted fusion on the outputs of the multi-layer perceptron MLP and the Transformer using the test set data to obtain the multi-parameter fusion warning result for rock fracturing-induced earthquakes.

[0171] This system obtains and processes the induced waveform data of hot dry rock fracturing-induced earthquakes, uses wavelet transform filtering technology to obtain effective waveform data, and further extracts characteristic coefficients and focal mechanism parameters. Input these multi-parameter data into the TFT regression model for regression analysis, and combine with the random forest classification model for classification analysis. Through the multi-parameter prediction and earthquake-causing probability analysis of the fusion model, the possibility of future earthquakes can be evaluated more accurately. This system effectively solves the data noise and non-linearity problems in traditional methods, significantly improves the overall performance of the warning system, and provides a scientific reference basis for the warning of hot dry rock fracturing-induced earthquakes.

[0172] This system comprehensively utilizes various technical means such as wavelet transform filtering, characteristic coefficient extraction, and focal mechanism parameter inversion to perform multi-dimensional analysis on the seismic waveform data induced by hot dry rock fracturing. By jointly using the regression model and the classification model, it can not only predict multi-parameter data such as future magnitudes and energy release rates, but also evaluate the probability of future earthquake occurrence. The fusion model further integrates the results of regression prediction and classification analysis to achieve a comprehensive warning of hot dry rock fracturing-induced earthquakes. This system has significant advantages in data processing, model training, and prediction evaluation, can effectively improve the accuracy and stability of earthquake warnings, and provides important technical support for seismic safety monitoring during hot dry rock mining.

[0173] The fusion model of this system utilizes the advantages of MLP and Transformer to capture the non-linear features and time-dependent relationships in the data, improving the accuracy and stability of prediction. Through the later fusion model, the prediction results of the regression and classification models are comprehensively utilized to improve the overall performance of the warning system. The MLP and Transformer models are respectively suitable for different data characteristics and task requirements, providing more flexible choices. The later fusion model can reduce the bias and error of a single model, improve the stability and accuracy of prediction, and finally achieve multi-parameter fusion warning of hot dry rock fracturing-induced earthquakes through the fusion of regression prediction and hazard classification results.

[0174] The exemplary embodiments of the engineering rock mass fracturing-induced earthquake multi-information fusion early warning method and system proposed by the present invention have been described in detail above with reference to the preferred embodiments. However, those skilled in the art can understand that, without departing from the concept of the present invention, various modifications and variations can be made to the above specific embodiments, and various combinations of the technical features and structures proposed by the present invention can be made without exceeding the protection scope of the present invention. The protection scope of the present invention is determined by the appended claims.

Claims

1. A multi-information fusion early warning method for engineering rock mass fracturing induced seismicity, characterized in that: The steps include: S10: collecting original waveform data induced by the rock fracturing vibration process, and filtering the obtained original waveform data to obtain effective waveform data; S20: extracting characteristic coefficients and performing focal mechanism inversion on the effective waveform data to obtain waveform characteristic parameters, focal location, and spatiotemporal strong multi-dimensional early warning indicator data of focal mechanism parameters, and taking the original waveform data induced by fracturing and the spatiotemporal strong multi-dimensional early warning indicator data as data sets; S30: Input the data set into the TFT regression model for regression analysis, and perform preprocessing by normalization and filling missing values, train the model and evaluate the model performance to obtain predicted spatiotemporal strong multi-dimensional data; S40: extracting characteristic variables from the spatiotemporal strong multi-dimensional early warning indicator data set, and obtaining a target variable on whether an earthquake occurs by analyzing historical microseismic data; preprocessing the characteristic variables extracted from the data set and the target variable on whether an earthquake occurs, and then inputting them into a random forest classification model for classification analysis, performing hazard classification using a multi-decision tree integration method, and obtaining a predicted earthquake probability by training and evaluating the performance of the preprocessed data using the random forest classification model; S50: The predicted spatiotemporal strong multi-dimensional data and the earthquake probability obtained by hazard classification are used as input data, and a new data set is re-integrated and divided into a training set and a test set. The training set data is trained and evaluated using a multi-layer perception MLP and Transformer fusion model. The outputs of the multi-layer perception MLP and Transformer are weightedly fused using the test set data to obtain the multi-parameter fusion warning results of rock fracturing earthquakes.

2. The multi-information fusion early warning method for engineering rock mass fracturing and seismic induction according to claim 1 is characterized in that: In step S10, the acquired original waveform data is filtered to obtain effective waveform data, which specifically includes the following steps: S11: Decompose the original waveform data to determine the wavelet basis function and the number of decomposition layers; S12: performing threshold processing on the decomposed signal to remove noise components; S13: Reconstruct the signal after noise removal to obtain filtered effective waveform data.

3. The multi-information fusion early warning method for engineering rock mass fracturing and seismic induction according to claim 1 is characterized in that: In step S20, characteristic coefficients are extracted and focal mechanism inversion is performed on the effective waveform data to obtain temporal and spatial strong multi-dimensional early warning indicator data of waveform characteristic parameters, focal location and focal mechanism parameters, including: S21: Perform feature recognition on the induced waveform and extract waveform feature coefficients, which include ring count, energy release rate, frequency band energy distribution, fractal characteristics, seismic time and initial amplitude; S22: Inversion of focal mechanism parameters, which include rupture location index and rupture intensity index. The rupture location index includes spatial concentration factor, spatial distribution intensity coefficient, error space form, tension-shear attribute, rupture surface occurrence, and extension azimuth. The rupture intensity index includes earthquake moment magnitude, b value, stress drop, and corner frequency. S23: Establishing a data set according to the waveform characteristic coefficient data and focal mechanism parameter data.

4. The multi-information fusion early warning method for engineering rock mass fracturing and seismic induction according to claim 1 is characterized in that: The step S30 specifically includes the following steps: S31: constructing a regression data set, combining waveform characteristic coefficients and focal mechanism parameters in the effective waveform data to form a regression data set; S32: construct a TFT regression model, and input the regression data set into the TFT regression model for regression prediction analysis; S33: Performance evaluation and result prediction, using the mean absolute scaled error (MASE) to quantitatively evaluate the prediction results of the model; output multi-parameter prediction data and prediction intervals through the trained model to obtain future earthquake activity trends.

5. The engineering rock mass fracturing induced seismic multi-information fusion early warning method according to claim 4 is characterized in that: The TFT regression model includes: A data processing and feature selection module configured to select the most significant features from the input using a variable selection network VSN and encode a context vector of static covariates through a static covariate encoder SCE; The information flow modeling module is configured to use a gated residual network (GRN) to ensure effective information flow, a temporal self-attention layer (TSL) to learn the long-term dependencies of temporal data, and a position-oriented feed-forward layer (PFL) to perform additional nonlinear processing on the output of the self-attention layer; A decoding and prediction module configured to enhance temporal features using a static enrichment layer SEL in a temporal fusion decoder TFD; The training optimization module is configured to use mean square error (MSE) as the loss function during training, the Adam optimizer for parameter optimization, and the mini-batch stochastic gradient descent algorithm to iteratively update the model parameters.

6. The engineering rock mass fracturing induced seismic multi-information fusion early warning method according to claim 1, characterized in that: In S40, the characteristic variables extracted from the data set and the target variable regarding whether an earthquake occurs are preprocessed and then input into a random forest classification model for classification analysis, which specifically includes the following steps: S41: Dataset division and data preprocessing: The feature variables extracted from the spatiotemporal strong multi-dimensional early warning indicator data set are combined with the target variable regarding whether an earthquake has occurred to form a classification data set; S42: Construct a random forest model: Generate multiple subsamples from the classification data set through bootstrap sampling, train a decision tree for each subsample, reduce the risk of overfitting by integrating multiple trees, and finally determine the final category through majority voting; S43: Model training and evaluation: Use the cross entropy loss function to optimize the model and calculate the classification effect of each tree during training; The performance of the model is evaluated through evaluation indicators, and the classification performance of the model is analyzed through confusion matrix; the evaluation indicators include: accuracy, recall rate and comprehensive accuracy; S44: Use the trained random forest model to classify and predict future earthquake data and output the probability of an earthquake occurring at each time point.

7. The engineering rock mass fracturing induced seismic multi-information fusion early warning method according to claim 6, characterized in that: The accuracy rate A, recall rate R and comprehensive accuracy rate F1 are calculated as follows: Wherein, TP is the hypothesized true positive: the actual situation is earthquake-causing and the model recognition result is earthquake-causing, that is, the earthquake-causing phenomenon is correctly identified; FP is the false positive: the actual situation is not earthquake-causing but the model recognition result is earthquake-causing, that is, the non-seismic phenomenon is incorrectly identified; TN is the true negative: the actual situation is not earthquake-causing and the model recognition result is not earthquake-causing, that is, the non-seismic phenomenon is correctly identified; FN is the false negative: the actual situation is earthquake-causing but the model recognition result is not earthquake-causing, that is, the earthquake-causing phenomenon is incorrectly identified.

8. The multi-information fusion early warning method for engineering rock mass fracturing and seismic induction according to claim 1 is characterized in that: In step S50, the training set data is trained and evaluated using a multi-layer perception MLP and Transformer fusion model, and the specific classification steps are as follows: S51: Constructing a dataset: Integrate the output data of the TFT regression model and the random forest classification model, i.e. the predicted spatiotemporal strong multi-dimensional data and the earthquake probability, into a new dataset; S52: Build an MLP and Transformer fusion model: Train the MLP model and the Transformer model separately, and automatically select the MLP or Transformer model according to the task characteristics of regression or classification, determine the best hyperparameter combination, and further tune the hyperparameters, calculate the errors of the training set and the validation set, determine whether the model is overfitting or underfitting, and evaluate the model performance; S53: Output the multi-parameter fusion warning results of rock fracturing earthquake: The predicted spatiotemporal strong multi-dimensional data output by the MLP and Transformer models and the classified predicted earthquake probability are weightedly fused to finally generate the multi-dimensional parameter fusion warning results of microseismic monitoring related to the waveform difference characteristics, source occurrence time, location and intensity of rock fracturing earthquake.

9. The engineering rock mass fracturing induced seismic multi-information fusion early warning method according to claim 8, characterized in that: Building an MLP and Transformer fusion model includes the following steps: S521: Model training: When training MLP, the mean square error (MSE) is used as the loss function, the Adam optimizer is used for parameter optimization, and the model parameters are iteratively updated through the mini-batch stochastic gradient descent algorithm; When training Transformer, we use the cross entropy loss function and the Adam optimizer for optimization, and use the multi-head attention mechanism to capture the complex relationship between features. S522: Model selection and hyperparameter tuning: Automatically select MLP or Transformer model according to the task characteristics of regression or classification, determine the best hyperparameter combination through cross-validation, and further tune the hyperparameters to optimize the performance of the model; S523: Model evaluation: During training, the errors of the training set and the validation set are calculated through MSE and cross entropy to determine whether the model is overfitting or underfitting. For regression tasks, the root mean square error RMSE is used to evaluate the model performance, and for classification tasks, the accuracy, precision, and recall are used for evaluation.

10. An engineering rock mass fracturing induced seismic multi-information fusion early warning system, characterized in that: An acquisition module is configured to acquire original waveform data induced by the rock fracturing vibration process, and filter the acquired original waveform data to obtain effective waveform data; A processing module is configured to extract characteristic coefficients and perform focal mechanism inversion on the effective waveform data to obtain waveform characteristic parameters, focal location, and spatial-temporal strong multi-dimensional early warning indicator data of focal mechanism parameters, and use the original waveform data induced by fracturing and the spatial-temporal strong multi-dimensional early warning indicator data as data sets; A regression model module is configured to input the data set into the TFT regression model for regression analysis, and perform preprocessing by normalization and filling missing values, train the model and evaluate the model performance to obtain predicted spatiotemporal strong multi-dimensional data; The classification model module is configured to extract characteristic variables from a spatiotemporal strong multi-dimensional early warning indicator data set, and obtain a target variable on whether an earthquake occurs by analyzing historical microseismic data; pre-process the characteristic variables extracted from the data set and the target variable on whether an earthquake occurs, and then input them into a random forest classification model for classification analysis, and perform hazard classification using a multi-decision tree integration method, and obtain a predicted earthquake probability by training and evaluating the performance of the pre-processed data using the random forest classification model; The weighted fusion module is configured to use the predicted spatiotemporal strong multi-dimensional data and the seismic probability obtained by hazard classification as input data, reintegrate to establish a new data set, and divide it into a training set and a test set. The training set data is trained and evaluated using the multi-layer perception MLP and Transformer fusion model, and the output of the multi-layer perception MLP and Transformer is weightedly fused using the test set data to obtain the multi-parameter fusion warning result of rock fracturing seismic.

Images