Rock mass fracture mode automatic identification method and system based on heterogeneous integrated model

Abstract

The invention provides a heterogeneous integrated model-based rock mass fracture mode automatic identification method and system. The method comprises the following steps of: acquiring waveform data of a rock mass fracture event on a construction site and preprocessing the waveform data; based on the preprocessed rock mass fracture event waveform data and the seismic source mechanism inversion model, obtaining sample events of different rock mass fracture modes; automatically extracting feature data of the sample event based on the sample event, the AWT model and the Python function model; constructing a heterogeneous integration model by fusing a machine learning model based on a Stacking heterogeneous integration technology; training the heterogeneous integration model based on the sample event feature data; and obtaining to-be-identified event feature data of a to-be-identified rock mass fracture event, and inputting the to-be-identified event feature data into the trained heterogeneous integrated model to obtain an identification result of the rock mass fracture mode. According to the method, automatic identification of the internal fracture mode of the rock mass in the deep complex engineering environment is realized, and the method has important practical significance for improving the identification accuracy of the internal fracture mode of the rock mass.

Description

Technical Field

[0001] This invention belongs to the field of deep engineering rock mass fracture and damage detection technology, specifically involving an automatic identification method and system for rock mass fracture modes based on a heterogeneous integrated model. Background Technology

[0002] Rock mass fracturing and instability, a key scientific problem in rock mechanics, has always been a major research focus for scholars both domestically and internationally. Different fracturing modes in rock masses are closely related to their instability mechanisms. Shear fracturing, through progressive shear slip, dominates the systematic instability of structural planes; tensile fracturing, through sudden tensile collapse, triggers localized brittle failure; and mixed fracturing induces catastrophic instability through shear-tensile coupling. In deep underground mineral resource production, under the combined effects of high static stress and mining disturbance stress, numerous fracturing modes of different types are easily induced within the rock mass. These fracturings mostly occur within the rock mass, and due to the opacity of the rock mass, the fracturing modes cannot be directly determined visually. Therefore, accurately and automatically identifying the fracturing modes induced within the rock mass has become one of the key research topics in the field of deep engineering rock masses.

[0003] To effectively identify rock fracture modes, engineers have conducted in-depth research and technological development. Chinese invention patent CN112183643 A discloses a "Method and Device for Identifying Hard Rock Tensile-Shear Fracture Based on Acoustic Emission." This technical solution obtains three-dimensional planar acoustic signatures (time-frequency-amplitude) of rock tensile and shear fracture signals through indoor experiments. Then, HOG features are extracted from the acoustic signatures as input data, and finally, an IVM model is used to determine the fracture mode of the signal to be identified. However, in deep mining operations, the characteristics of rock fracture signals collected on-site are more diverse and complex, differing somewhat from those obtained through indoor experiments. Furthermore, experimental results show that the integrated model effectively combines the advantages of multiple independent models, achieving higher classification accuracy than a single machine learning model.

[0004] Microseismic systems can acquire seismic wave signals generated by rock mass fractures in real time, making them an effective technology for monitoring rock mass fracture information. Inversion models based on moment tensors can calculate and determine rock mass fracture modes using waveform information from fracture events. However, the calculation process involves related theories from different fields such as seismology and elastic dynamics, specifically covering a series of mathematical models including far-field displacement at any point in the rock mass, P-wave far-field displacement based on the excitation matrix, Green's function, waveform attenuation, low-frequency displacement amplitude, and determination of the P-wave initial motion direction. Tests show that determining rock mass fracture modes using moment tensor-based inversion models is complex, computationally intensive, and time-consuming. Furthermore, the microseismic waveforms of rock mass fractures acquired in the field contain a large amount of noise. Due to the complexity of the noise components, some noise shares common frequency bands with the useful signals of the rock mass fracture microseismic events, making complete removal impossible. Therefore, by extracting useful feature information from the collected microseismic waveforms of rock mass fracture events as data support, and combining technologies such as databases, machine learning, and heterogeneous integration, a method for accurately, conveniently, and efficiently identifying rock mass fracture modes is proposed. This method is of great significance for revealing the mechanism of rock mass fracture instability and disaster and for the stability control of engineering rock masses. Summary of the Invention

[0005] To address the problems existing in the prior art, this invention provides an automatic identification method and system for rock mass fracture modes based on a heterogeneous integrated model. The aim is to use the collected rock mass fracture microseismic data as the basic data, and to further extract useful feature data from the preprocessed waveform data of rock mass fracture events with residual noise using AWT combined with Python functions. Then, a heterogeneous integrated model is established using Stacking technology to call the feature data to achieve accurate identification of the fracture mode of the rock mass fracture event to be identified.

[0006] To achieve the above objectives, the present invention provides the following solution: An automatic identification method for rock mass fracture modes based on a heterogeneous integrated model, the method comprising: Collect waveform data of rock mass fracture events at the construction site and perform preprocessing; Based on the preprocessed rock mass fracture event waveform data and the source mechanism inversion model, sample events of different rock mass fracture modes are obtained; Based on sample events, the AWT model, and the Python function model, feature data of sample events is obtained; A heterogeneous ensemble model is constructed by fusing machine learning models with Stacking heterogeneous ensemble technology. Training heterogeneous ensemble models based on sample event feature data; The feature data of the rock mass fracture event to be identified is obtained, and the feature data is input into the trained heterogeneous ensemble model to obtain the identification result of the rock mass fracture mode.

[0007] Preferably, the method for obtaining sample events of different rock mass fracture modes based on preprocessed rock mass fracture event waveform data and focal mechanism inversion models includes: Establish a source mechanism inversion model based on moment tensor, interpret rock mass fracture event information to determine fracture modes, and select the same number of fracture events with tensile, shear and mixed fracture modes as sample events, with no less than 500 sample events for each fracture mode; by DC The value of % determines the fracture mode of rock mass fracture events, when DC When % ≥ 60%, it indicates that the event is a shear fracture mode; when DC When % ≤ 40%, it indicates that the event is a tension rupture mode; when 40% < DC When % < 60%, it indicates that the event is a mixed rupture mode; the specific expression for DC% is: ; in, M DC , M CLVD and M ISO These are the specific mathematical expressions for the double couple part, the isotropic part, and the compensated linear vector dipole part, respectively.

[0008] Preferred methods for obtaining sample event feature data based on sample events, AWT models, and Python function models include: The AWT model is used to decompose the waveform of the sample event to obtain the useful detail coefficients of the sample event waveform; The Python function is used to extract the mean, standard deviation, energy, peak-to-peak value, and dominant frequency position information of the waveform of the sample event as the feature data of the sample event, and a database is built to classify and store the feature data of the sample event according to the fracture mode.

[0009] Preferably, methods for obtaining useful detail coefficients of sample event waveforms by decomposing the waveforms of sample events using the AWT model include: An AWT model is constructed. Based on the signal-to-noise ratio maximization criterion, the optimal number of decomposition levels is selected by comparing the SNR values ​​of the reconstructed signals at different decomposition levels and finding that the decomposition level with the maximum SNR is the optimal number of decomposition levels. Set the number of layers in the AWT model to the optimal number of decomposition layers, decompose the waveform of the sample event and perform noise reduction processing to obtain the useful detail coefficients of the sample event waveform.

[0010] Preferred methods for constructing heterogeneous ensemble models by fusing machine learning models based on Stacking heterogeneous ensemble technology include: The machine learning model to be integrated is selected from a set of pre-defined machine learning models using three indicators: SEN, TPR, and kappa coefficient. The Stacking heterogeneous ensemble technique is used to fuse the machine learning models to be integrated to build a heterogeneous ensemble model. The heterogeneous ensemble model includes a base model layer and a meta-model layer. The base model layer includes at least 3 different machine learning models, and the meta-model layer includes 1 machine learning model.

[0011] Preferably, the constructed heterogeneous integration model includes: The machine learning models in the base model layer are: ET, RF, and XGB; The machine learning model in the meta-model layer is: L.

[0012] Preferably, methods for training heterogeneous ensemble models based on sample event feature data include: New feature matrix establishment: During training, the base model layer extracts effective feature data from the sample event feature data and establishes a new feature matrix; Meta-model layer training: The meta-model layer is trained based on the newly established feature matrix.

[0013] Preferably, the method for obtaining the feature data of the rock mass fracture event to be identified, and inputting the feature data of the event to be identified into a trained heterogeneous ensemble model to obtain the identification result of the rock mass fracture mode includes: Real-time acquisition of microseismic waveform data of rock mass fracture events to be identified at the construction site, followed by noise reduction preprocessing; The AWT model is used to decompose the microseismic waveform data of the rock mass fracture event to be identified, and the useful detail coefficients of the waveform of the rock mass fracture event to be identified are obtained. Extract the mean, standard deviation, energy, peak-to-peak value, and dominant frequency position information of the waveform of the rock mass fracture event to be identified, and obtain the characteristic data of the event to be identified; The feature data of the event to be identified is input into the trained heterogeneous ensemble model to obtain the identification results of the rock mass fracture mode.

[0014] The present invention also provides an automatic identification system for rock mass fracture modes based on a heterogeneous integrated model. The system is used to implement the aforementioned method and includes: a data acquisition and processing module, a sample event acquisition module, a feature data acquisition module, an integrated module construction module, a model training module, and an identification module. The data acquisition and processing module is used to acquire waveform data of rock mass fracture events at the construction site and perform preprocessing. The sample event acquisition module is used to acquire sample events of different rock mass fracture modes based on the preprocessed rock mass fracture event waveform data and the source mechanism inversion model. The feature data acquisition module is used to acquire feature data of sample events based on sample events, AWT models, and Python function models. The integration module is used to build heterogeneous integration models by fusing machine learning models based on Stacking heterogeneous integration technology. The model training module is used to train heterogeneous ensemble models based on sample event feature data; The identification module is used to acquire the feature data of the rock mass fracture event to be identified, input the feature data of the event to be identified into the trained heterogeneous ensemble model, and obtain the identification result of the rock mass fracture mode.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) This invention uses microseismic waveform data of rock mass fracture monitored in the field as basic data, and uses AWT combined with Python functions to extract useful detail coefficient energy, dominant frequency position information and other features from the basic data as feature data. At the data level, unlike the existing methods of obtaining sample data through indoor tests, this invention further extracts useful signal feature information from waveform data with mixed noise, laying a data foundation for the automatic and effective identification of rock mass fracture modes.

[0016] (2) This invention selects high-performance models from various machine learning algorithms through testing, and uses the Stacking technology framework to effectively fuse these high-performance models, thus establishing a high-performance heterogeneous integrated model. At the algorithm level, the heterogeneous integrated model effectively combines the advantages of different types of high-performance machine learning models, and can accurately and automatically identify rock mass fracture modes.

[0017] (3) The principle of automatic identification of rock mass fracture mode in this invention is simple. It is different from the binary classification method based on indoor tensile and shear test to obtain sample data, and realizes the automatic identification of rock mass tensile, shear and mixed failure modes. Attached Figure Description

[0018] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a schematic diagram of the automatic identification method for rock mass fracture modes based on a heterogeneous integrated model according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the waveform feature extraction process for rock fracture events according to an embodiment of the present invention; Figure 3This is a schematic diagram of a typical rock fracture event waveform according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the data processing flow for identifying rock fracture patterns using a heterogeneous integrated model constructed for an embodiment of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0022] Example 1 like Figure 1 As shown, this invention provides an automatic identification method for rock mass fracture modes based on a heterogeneous integrated model, comprising: Collect waveform data of rock mass fracture events at the construction site and perform preprocessing; Based on the preprocessed rock mass fracture event waveform data and the source mechanism inversion model, sample events of different rock mass fracture modes are obtained; Based on sample events, the AWT model, and the Python function model, feature data of sample events is obtained. A heterogeneous ensemble model is constructed by fusing machine learning models with Stacking heterogeneous ensemble technology. Training heterogeneous ensemble models based on sample event feature data; The feature data of the rock mass fracture event to be identified is obtained, and the feature data is input into the trained heterogeneous ensemble model to obtain the identification result of the rock mass fracture mode.

[0023] Furthermore, the specific implementation process of this invention is as follows: The deployed microseismic monitoring system was used to collect waveform data of rock mass fracture events at the construction site, and the built-in denoising model of the microseismic monitoring system was used to perform denoising preprocessing on the collected rock mass fracture event waveform data.

[0024] Furthermore, based on preprocessed rock mass fracture event waveform data and focal mechanism inversion models, methods for obtaining sample events of different rock mass fracture modes include: A source mechanism inversion model based on moment tensor is established. The preprocessed rock mass fracture event waveform data is input into the source mechanism inversion model to interpret the rock mass fracture event information to determine the fracture mode. The same number of fracture events with tensile, shear and mixed fracture modes are selected as sample events, and there are no less than 500 sample events for each fracture mode. by DC The value of % determines the fracture mode of rock mass fracture events, when DC When % ≥ 60%, it indicates that the event is a shear fracture mode; when DC When % ≤ 40%, it indicates that the event is a tension rupture mode; when 40% < DC When % < 60%, it indicates that the event is a mixed rupture mode; the specific expression for DC% is: ; in, M DC , M CLVD and M ISO These are the two-couple parts ( DC ), isotropic part ( ISO ) and the compensated linear vector dipole part ( CLVD The specific mathematical expression of ).

[0025] Furthermore, such as Figure 2 As shown, methods for obtaining sample event feature data based on sample events, the AWT model (Adaptive Wavelet Thresholding Model), and Python function models include: The AWT model is used to decompose the waveform of the sample event to obtain the useful detail coefficients of the sample event waveform; The Python function is used to extract the mean, standard deviation, energy, peak-to-peak value, and dominant frequency position information of the waveform of the sample event as the feature data of the sample event, and a database is built to classify and store the feature data of the sample event according to the fracture mode.

[0026] Specifically, methods for obtaining useful detail coefficients from the waveform of sample events by decomposing the waveform using the AWT model include: Sample events include: Event 1: A1=[ a (1,1) , a (1,2) ,..., a (1,n) Event 2: A2 = [ a (2,1) , a (2,2) ,..., a (2,n)],...,event i A i =[ a (i,1) , a (i,,2) ,..., a (i,,n) ].in, a (i,,n) Indicates the first i In the data sequence of the microseismic event, the first... n Amplitude values.

[0027] An AWT model is established. Based on the signal-to-noise ratio (SNR) maximization criterion, the optimal number of decomposition levels is selected by comparing the SNR values ​​of the reconstructed signals under different decomposition levels (L=3-8) and selecting the decomposition level corresponding to the maximum SNR. Set the number of layers in the AWT model to the optimal number of decomposition layers, decompose the waveform of the sample event and perform noise reduction processing to obtain the useful detail coefficients of the sample event waveform.

[0028] Figure 3 The waveforms represent typical rock mass fracturing events. As shown in Table 1, the optimal number of decomposition layers for AWT decomposition of rock mass fracturing event waveforms is 7.

[0029] Table 1. SNR values ​​of AWT denoised reconstructed signals under different decomposition levels The mean, std, sum, find_peaks, and argmax functions of the Python programming language were used to calculate the mean (M), standard deviation (S), energy (E), peak-to-peak value (V), and dominant frequency position (C) of the sample event waveforms at each layer of the AWT model. These parameters were used as features of the sample events to form a dataset (divided proportionally into training and test sets). The extracted detail coefficient features of each layer for typical rock mass fracture events are shown in Table 2. In Table 2, Cd1-Cd7 represent detail information, and Ca7 represents approximate information obtained from the decomposition to the 7th layer.

[0030] Table 2. Details of each layer of typical rock mass fracture events extracted. Three forms are created using database technology: a tensile failure form, a shear failure form, and a mixed failure form, which store the feature matrix data of the corresponding failure mode events.

[0031] Furthermore, methods for constructing heterogeneous ensemble models based on Stacking heterogeneous ensemble technology and integrating machine learning models include: The machine learning model to be integrated is selected from a set of pre-set machine learning models using three indicators: SEN, TPR, and kappa coefficient. The Stacking heterogeneous ensemble technique is used to fuse the machine learning models to be integrated to build a heterogeneous ensemble model. The heterogeneous ensemble model includes a base model layer and a meta-model layer. The base model layer includes at least 3 different machine learning models, and the meta-model layer includes 1 machine learning model.

[0032] Specifically, multiple machine learning models are established and trained and tested using sample event feature data according to a certain strategy. The machine learning model with good performance in identifying event rupture patterns is selected. The selected different machine learning models are effectively integrated through Stacking heterogeneous integration technology to establish a new heterogeneous integrated model.

[0033] In this preferred embodiment, 10 machine learning models were established: Logistic regression model (L), K-Nearest Neighbor (KNN), Random Forest (RF), ExtraTrees (ET), Naive Bayes (NB), Decision Tree (DT), XGBoost (XGB), Multi-layer Perception (MLP), SupportVector Machine (SVM), and GradientBoosting. These models were trained and tested using event feature data stored in the database. The performance of these 10 machine learning models in identifying event rupture patterns was evaluated using Sensitivity (SEN), True Positive Rate (TPR), and kappa coefficient as indicators. The four models with the highest SEN and other three indicators were selected as the best-performing models. Table 3 shows the SEN, TPR, and kappa coefficient values ​​for different machine learning models.

[0034] Table 3. SEN, TPR, and kappa coefficients for different machine learning models In this preferred embodiment, the principle for selecting base models and meta-models is as follows: using models with excellent prediction performance and different types (heterogeneity) as base learners to improve the diversity and generalization ability of the ensemble model and ensure the independence and effectiveness of the meta dataset; using simple and stable models as meta-learners to balance the output of each base model, avoid the risk of overfitting, integrate the advantages of multiple models to achieve complementary effects, and improve the prediction accuracy and robustness of the model.

[0035] The final heterogeneous integration model includes: The machine learning models in the base model layer are: ET, RF, and XGB; The machine learning model in the meta-model layer is: L.

[0036] In this preferred embodiment, to verify the predictive effect of the heterogeneous integrated model constructed in this invention, a prediction experiment was conducted under the same operating environment and training data as the aforementioned 10 machine learning models. As shown in Table 4, the SEN, TPR, and kappa coefficients of the heterogeneous integrated model constructed in this invention are 0.946, 0.941, and 0.928, respectively, which are 7.2%, 6.3%, and 5.2% higher than those of the RF prediction results. This demonstrates that the heterogeneous integrated model constructed in this invention effectively integrates the advantages of different types of models, achieving high accuracy in identifying rock mass fracture modes, and its identification accuracy is superior to other single models.

[0037] Table 4: Evaluation of the predictive performance of the constructed ensemble model In this preferred embodiment, to further verify the predictive performance of the heterogeneous ensemble model constructed in this invention, under the premise of identical operating environment and training data, machine learning models with good predictive performance (L, RF, ET, XGB, MLP, SVM, and GB) were selected as the base models, and multiple ensemble models were established for predictive performance evaluation. Table 5 shows the three best-performing models among all ensemble models. Comparative analysis shows that the SEN, TPR, and kappa coefficients of the heterogeneous ensemble model constructed in this invention are improved by 2.5%, 1.4%, and 1.3% respectively compared to model A; by 2.7%, 1.2%, and 0.5% respectively compared to model B; and by 4.9%, 5.2%, and 10.7% respectively compared to model C. This proves that the heterogeneous ensemble model constructed in this invention has high accuracy in identifying rock mass fracture modes and its identification accuracy is superior to other models.

[0038] Table 5: Evaluation index values ​​of the three best-performing ensemble models Furthermore, such as Figure 4 As shown, the data processing flow of the heterogeneous integration model constructed in this invention includes: (1) Data partitioning: using K The cross-validation algorithm divides the training set into... K (2) Base model training and prediction: RF model (corresponding to) Figure 4 Model 1) According to the cross-validation method, firstly, the first... KThe first subset of the dataset is used as the prediction set, and the remaining subset is used as the training set for training and prediction. Features of the prediction set data are then extracted. Next, the first subset is used as the training set. K- One subset of the dataset is used as the prediction set, and the remaining subset is used as the training set for training and prediction, extracting features from the prediction set data; this process continues until the features of the first subset of the dataset are extracted, ending the loop. The method for extracting features is the same as that of the RF model. Then, the ET model (corresponding to...) is used sequentially. Figure 4 Model 2), XGB model (corresponding to) Figure 4 The model in m (3) Feature combination: Combine the new features extracted from each base model (corresponding to the dataset features). Figure 4 New features 1, new features 2, new features m (4) Metamodel training and prediction: Use the L model to call the new training set data to train the model, and predict the test set data, and finally output the prediction results.

[0039] Furthermore, methods for training heterogeneous ensemble models based on sample event feature data include: New feature matrix establishment: During training, the base model layer extracts effective feature data from the sample event feature data, simplifies the data to establish a new feature matrix, avoids model overfitting, and improves model running efficiency; Meta-model layer training: The meta-model layer is trained based on the newly established feature matrix.

[0040] Furthermore, methods for obtaining the feature data of the rock mass fracture event to be identified, and inputting the feature data into a trained heterogeneous ensemble model to obtain the identification result of the rock mass fracture mode include: Real-time acquisition of microseismic waveform data of rock mass fracture events to be identified at the construction site, followed by noise reduction preprocessing; The AWT model is used to decompose the microseismic waveform data of the rock mass fracture event to be identified, and the useful detail coefficients of the waveform of the rock mass fracture event to be identified are obtained. Extract the mean, standard deviation, energy, peak-to-peak value, and dominant frequency position information of the waveform of the rock mass fracture event to be identified, and obtain the characteristic data of the event to be identified; The feature data of the event to be identified is input into the trained heterogeneous ensemble model to obtain the identification results of the rock mass fracture mode.

[0041] The waveform data of the rock mass fracture event to be identified is input into AWT for feature extraction. Then, the extracted feature data is input into the trained heterogeneous ensemble model for automatic fracture mode identification, and the identification results are output.

[0042] Preferably, a cyclic mode is used to sequentially identify the rock mass fracture events to be identified. First, the waveform feature data of the first rock mass fracture event to be identified is extracted sequentially using a cyclic method. The feature matrix is ​​then called through a heterogeneous integrated model to identify and output the event fracture mode. Then, the waveform feature data of the second rock mass fracture event to be identified is extracted sequentially using a cyclic method to identify the fracture mode. This process is repeated to identify the fracture mode of the collected and preprocessed rock mass fracture events in real time.

[0043] In summary, the rock mass fracture mode automatic identification method based on the heterogeneous integrated model of the present invention differs from existing methods that obtain sample data through indoor experiments in terms of data, and further extracts useful signal feature information from waveform data with mixed noise; in terms of algorithm optimization, it integrates multiple excellent models, which effectively improves the accuracy of automatic identification of fracture modes inside deep rock masses in mines. The implementation of the present invention is of great significance for revealing the mechanism of rock mass fracture and instability disaster in complex underground construction environments and for rock mass stability control.

[0044] Example 2 Based on the same inventive concept, the present invention also provides an automatic rock mass fracture mode identification system based on a heterogeneous integrated model, used to implement the method described in the foregoing embodiments. The system includes: a data acquisition and processing module, a sample event acquisition module, a feature data acquisition module, an integrated module construction module, a model training module, and an identification module. The data acquisition and processing module is used to acquire waveform data of rock mass fracture events at the construction site and perform preprocessing. The sample event acquisition module is used to acquire sample events of different rock mass fracture modes based on the preprocessed rock mass fracture event waveform data and the source mechanism inversion model. The feature data acquisition module is used to acquire feature data of sample events based on sample events, AWT models, and Python function models. The integration module is used to build heterogeneous integration models by fusing machine learning models based on Stacking heterogeneous integration technology. The model training module is used to train heterogeneous ensemble models based on sample event feature data; The identification module is used to acquire the feature data of the rock mass fracture event to be identified, input the feature data of the event to be identified into the trained heterogeneous ensemble model, and obtain the identification result of the rock mass fracture mode.

[0045] Furthermore, the sample event acquisition module acquires sample events for different rock mass fracture modes based on the preprocessed rock mass fracture event waveform data and the source mechanism inversion model. This includes: establishing a source mechanism inversion model based on moment tensor, interpreting rock mass fracture event information to determine the fracture mode, and selecting the same number of fracture events of tensile, shear and mixed fracture modes as sample events, with no less than 500 sample events for each fracture mode. by DC The value of % determines the fracture mode of rock mass fracture events, when DC When % ≥ 60%, it indicates that the event is a shear fracture mode; when DC When % ≤ 40%, it indicates that the event is a tension rupture mode; when 40% < DC When % < 60%, it indicates that the event is a mixed rupture mode; the specific expression for DC% is: ; in, M DC , M CLVD and M ISO These are the specific mathematical expressions for the double couple part, the isotropic part, and the compensated linear vector dipole part, respectively.

[0046] Furthermore, the feature data acquisition module acquires sample event feature data based on sample events, the AWT model, and the Python function model, including: The AWT model is used to decompose the waveform of the sample event to obtain the useful detail coefficients of the sample event waveform; The Python function is used to extract the mean, standard deviation, energy, peak-to-peak value, and dominant frequency position information of the waveform of the sample event as the feature data of the sample event, and a database is built to classify and store the feature data of the sample event according to the fracture mode.

[0047] Furthermore, methods for obtaining useful detail coefficients from the waveform of sample events by decomposing the waveform using the AWT model include: An AWT model is constructed. Based on the signal-to-noise ratio maximization criterion, the optimal number of decomposition levels is selected by comparing the SNR values ​​of the reconstructed signals at different decomposition levels and finding that the decomposition level with the maximum SNR is the optimal number of decomposition levels. Set the number of layers in the AWT model to the optimal number of decomposition layers, decompose the waveform of the sample event and perform noise reduction processing to obtain the useful detail coefficients of the sample event waveform.

[0048] Furthermore, the integration module builds a heterogeneous integration model by fusing machine learning models based on Stacking heterogeneous integration technology, including: The machine learning model to be integrated is selected from a set of pre-defined machine learning models using three indicators: SEN, TPR, and kappa coefficient. The Stacking heterogeneous ensemble technique is used to fuse the machine learning models to be integrated to build a heterogeneous ensemble model. The heterogeneous ensemble model includes a base model layer and a meta-model layer. The base model layer includes at least 3 different machine learning models, and the meta-model layer includes 1 machine learning model.

[0049] Furthermore, the constructed heterogeneous integration model includes: The machine learning models in the base model layer are: ET, RF, and XGB; The machine learning model in the meta-model layer is: L.

[0050] Furthermore, the model training module trains the heterogeneous ensemble model based on sample event feature data, including: New feature matrix establishment: During training, the base model layer extracts effective feature data from the sample event feature data and establishes a new feature matrix; Meta-model layer training: The meta-model layer is trained based on the newly established feature matrix.

[0051] Furthermore, the identification module acquires the feature data of the rock mass fracture event to be identified, inputs the feature data into the trained heterogeneous ensemble model, and obtains the identification results of the rock mass fracture mode, including: Real-time acquisition of microseismic waveform data of rock mass fracture events to be identified at the construction site, followed by noise reduction preprocessing; The AWT model is used to decompose the microseismic waveform data of the rock mass fracture event to be identified, and the useful detail coefficients of the waveform of the rock mass fracture event to be identified are obtained. Extract the mean, standard deviation, energy, peak-to-peak value, and dominant frequency position information of the waveform of the rock mass fracture event to be identified, and obtain the characteristic data of the event to be identified; The feature data of the event to be identified is input into the trained heterogeneous ensemble model to obtain the identification results of the rock mass fracture mode.

[0052] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for automatic identification of rock mass fracture modes based on a heterogeneous integrated model, characterized in that, The method includes: Collect waveform data of rock mass fracture events at the construction site and perform preprocessing; Based on the preprocessed rock mass fracture event waveform data and the source mechanism inversion model, sample events of different rock mass fracture modes are obtained; Based on sample events, the AWT model, and the Python function model, feature data of sample events is obtained; A heterogeneous ensemble model is constructed by fusing machine learning models with Stacking heterogeneous ensemble technology. Training heterogeneous ensemble models based on sample event feature data; The feature data of the rock mass fracture event to be identified is obtained, and the feature data is input into the trained heterogeneous ensemble model to obtain the identification result of the rock mass fracture mode.

2. The method according to claim 1, characterized in that, Methods for obtaining sample events of different rock mass fracture modes based on preprocessed rock mass fracture event waveform data and focal mechanism inversion models include: Establish a source mechanism inversion model based on moment tensor, interpret rock mass fracture event information to determine fracture modes, and select the same number of fracture events with tensile, shear and mixed fracture modes as sample events, with no less than 500 sample events for each fracture mode; by DC The value of % determines the fracture mode of rock mass fracture events, when DC When % ≥ 60%, it indicates that the event is a shear fracture mode; when DC When % ≤ 40%, it indicates that the event is a tension rupture mode; when 40% < DC When % < 60%, it indicates that the event is a mixed rupture mode; the specific expression for DC% is: ； in, M DC , M CLVD and M ISO These are the specific mathematical expressions for the double couple part, the isotropic part, and the compensated linear vector dipole part, respectively.

3. The method according to claim 1, characterized in that, Methods for obtaining feature data of sample events based on sample events, AWT models, and Python function models include: The AWT model is used to decompose the waveform of the sample event to obtain the useful detail coefficients of the sample event waveform; The Python function is used to extract the mean, standard deviation, energy, peak-to-peak value, and dominant frequency position information of the waveform of the sample event as the feature data of the sample event, and a database is built to classify and store the feature data of the sample event according to the fracture mode.

4. The method according to claim 3, characterized in that, Methods for obtaining useful detail coefficients from the waveform of a sample event by decomposing it using the AWT model include: An AWT model is constructed. Based on the signal-to-noise ratio maximization criterion, the optimal number of decomposition levels is selected by comparing the SNR values ​​of the reconstructed signals at different decomposition levels and choosing the decomposition level corresponding to the maximum SNR. Set the number of layers in the AWT model to the optimal number of decomposition layers, decompose the waveform of the sample event and perform noise reduction processing to obtain the useful detail coefficients of the sample event waveform.

5. The method according to claim 1, characterized in that, Methods for constructing heterogeneous ensemble models based on stacking heterogeneous ensemble technology and integrating machine learning models include: The machine learning model to be integrated is selected from a set of pre-defined machine learning models using three indicators: SEN, TPR, and kappa coefficient. The Stacking heterogeneous ensemble technique is used to fuse the machine learning models to be integrated to build a heterogeneous ensemble model. The heterogeneous ensemble model includes a base model layer and a meta-model layer. The base model layer includes at least 3 different machine learning models, and the meta-model layer includes 1 machine learning model.

6. The method according to claim 5, characterized in that, The constructed heterogeneous integration model includes: The machine learning models in the base model layer are: ET, RF, and XGB; The machine learning model in the meta-model layer is: L.

7. The method according to claim 6, characterized in that, Methods for training heterogeneous ensemble models based on sample event feature data include: New feature matrix establishment: During training, the base model layer extracts effective feature data from the sample event feature data and establishes a new feature matrix; Meta-model layer training: The meta-model layer is trained based on the newly established feature matrix.

8. The method according to claim 1, characterized in that, Methods for obtaining the feature data of the rock mass fracture event to be identified, and inputting the feature data into a trained heterogeneous ensemble model to obtain the identification results of the rock mass fracture mode include: Real-time acquisition of microseismic waveform data of rock mass fracture events to be identified at the construction site, followed by noise reduction preprocessing; The AWT model is used to decompose the microseismic waveform data of the rock mass fracture event to be identified, and the useful detail coefficients of the waveform of the rock mass fracture event to be identified are obtained. Extract the mean, standard deviation, energy, peak-to-peak value, and dominant frequency position information of the waveform of the rock mass fracture event to be identified, and obtain the characteristic data of the event to be identified; The feature data of the event to be identified is input into the trained heterogeneous ensemble model to obtain the identification results of the rock mass fracture mode.

9. An automatic rock mass fracture mode identification system based on a heterogeneous integrated model, the system being used to implement the method described in any one of claims 1-8, characterized in that, The system includes: a data acquisition and processing module, a sample event acquisition module, a feature data acquisition module, an integration module construction module, a model training module, and a recognition module; The data acquisition and processing module is used to acquire waveform data of rock mass fracture events at the construction site and perform preprocessing. The sample event acquisition module is used to acquire sample events of different rock mass fracture modes based on the preprocessed rock mass fracture event waveform data and the source mechanism inversion model. The feature data acquisition module is used to acquire feature data of sample events based on sample events, AWT models, and Python function models. The integration module is used to build heterogeneous integration models by fusing machine learning models based on Stacking heterogeneous integration technology. The model training module is used to train heterogeneous ensemble models based on sample event feature data; The identification module is used to acquire the feature data of the rock mass fracture event to be identified, input the feature data of the event to be identified into the trained heterogeneous ensemble model, and obtain the identification result of the rock mass fracture mode.

Images