Coal rock instability precursor identification method and system based on fracture topological parameters

Abstract

The invention discloses a coal rock instability precursor identification method and system based on fracture topological parameters, and the method comprises the steps: obtaining sound wave data through micro-seismic monitoring, and extracting coal rock fracture core parameters in combination with seismic source positioning and seismic source mechanism inversion; building a fracture network topology model based on a graph theory method by considering the geometric dimension and direction characteristics of the fracture, and obtaining fracture space-time topology evolution parameters to build a multivariable time sequence topology characteristic data set; introducing a multivariable time sequence graph model MTAD-GAT, dynamically capturing a structural dependency relationship among topological parameters by using a graph attention mechanism, mining collaborative anomaly characteristics of the structural dependency relationship, and capturing an evolution trend of the structural dependency relationship in combination with a time sequence modeling module; and fusing the prediction error and the reconstruction error to form an abnormal score as an instability precursor index, extracting covariant features and evolution trends of a plurality of key topological indexes in the precursor stage, and assisting in distinguishing an instability critical state, thereby realizing unsupervised automatic identification and intelligent early warning of the topological precursor.

Description

Technical Field

[0001] The present invention relates to the field of coal rock instability early warning, and in particular to a coal rock instability precursor identification method and system based on fracture topology parameters. Background Art

[0002] As coal mining continues to deepen, the geological structure of the mines becomes increasingly complex, and the coal and rock masses are subjected to long-term nonlinear loading environments characterized by high stress and multi-source disturbances. This complex stress evolution process can easily trigger unstable expansion and sudden rupture of internal cracks in the coal and rock masses, leading to typical dynamic disasters such as rock bursts, coal and gas outbursts, tunnel instability, and roof collapse. Before instability and rupture, coal and rock masses often undergo a series of microstructural evolution and energy release processes, exhibiting certain precursory characteristics. Therefore, identifying these hidden precursory signals before a disaster occurs and promptly determining whether the coal and rock masses have entered the stage of instability evolution has become a key issue in the current field of coal mine disaster prevention and control.

[0003] Microseismic monitoring, as a highly sensitive, non-destructive detection method, can capture in real time the elastic wave signals released by microfracture activity in coal and rock masses during loading. By comprehensively analyzing the spatiotemporal distribution, frequency characteristics, and energy evolution of microseismic events, the precursory characteristics of coal and rock fractures can be effectively identified, thereby achieving early perception and accurate early warning of dynamic disasters. Patent CN114895352A proposes a method and device for predicting rock instability based on microseismic monitoring. It constructs a logarithmic potential energy ratio based on source parameters, volumetric potential, and energy release, and establishes a fuzzy early warning time model based on the cumulative apparent volume to achieve a timely early warning of rock instability. Patent CN117932371A discloses a full-time and spatiotemporal prediction method for rock burst based on the "time-space-intensity" of microseismic data. It uses an LSTM network to predict the occurrence time of rock burst based on the precursory pattern sequence of microseismic data, and uses microseismic spatiotemporal characteristic indicators to achieve spatial identification of dangerous areas through multi-scale spatial cloud map fusion. Patent CN117150911A proposes a method and system for predicting coal rock instability and fracture based on graph neural networks. This method divides microseismic signals into time and space to construct a network graph. Through graph embedding and neural networks, it predicts future topological evolution and identifies important nodes as fracture precursors, enabling accurate early warning. These methods can, to a certain extent, reflect the energy release and disturbance response of the rock mass during the fracture process. Combined with models such as neural networks, they can identify precursor patterns and predict time and space. However, they still have significant limitations: they primarily focus on analyzing the spatiotemporal distribution, frequency characteristics, and energy parameters of the microseismic signals themselves, starting from the "signal level." They fail to deeply explore the evolutionary laws of the fracture structure reflected behind the microseismic events, making it difficult to fully reflect the complex fracture behavior during the fracture evolution stage. Summary of the Invention

[0004] In response to the problems and needs raised above, this solution proposes a method and system for identifying coal and rock instability precursors based on fracture topology parameters. By adopting the following technical features, it can achieve the above technical objectives and bring about many other technical effects.

[0005] One object of the present invention is to provide a method for identifying coal rock instability precursors based on fracture topology parameters, comprising the following steps:

[0006] S10: The original acoustic wave data is collected by the microseismic monitoring system, and combined with the earthquake source location calculation and focal mechanism inversion, the core physical parameters directly related to the microfracture source in the coal and rock fracture process are obtained;

[0007] S20: Four types of topological relationships are defined by comprehensively considering the geometric size and orientation characteristics of the fractures. The fracture penetration index (BCI) is quantitatively calculated. The fracture network is abstracted into a graph model based on graph theory. A topological model that describes the spatiotemporal evolution of fractures is constructed. Multidimensional time series topological parameters are extracted to obtain a multivariate topological time series dataset reflecting the evolution of coal and rock fractures.

[0008] S30: Divide the constructed multivariate topological time series dataset into a training set and a test set, and perform data normalization and outlier processing on each dimension of the topological structure parameter time series;

[0009] S40: Based on the preprocessed training data, a multivariate time series graph model MTAD-GAT is constructed to form a feature-oriented and time-oriented dual-graph attention mechanism to model the structural dependency and temporal evolution trend between topological parameters respectively. Multi-head attention is introduced in the graph attention layer, and different attention scores are calculated through multiple independent attention heads. The local and global features extracted by the graph attention layer and the one-dimensional convolutional layer are fused and input into the gated recurrent unit (GRU) to deeply capture the dynamic evolution pattern and long-term dependency of topological features. During the training process, the model is iteratively optimized by minimizing the loss function and combining backpropagation and gradient update mechanisms.

[0010] S50: Apply the trained model to the test set data, and use the gated recurrent unit (GRU) output to pass it into the prediction module for predicting the observation value at the next moment and the reconstruction module for characterizing the overall temporal distribution characteristics; fuse the errors of the two to construct an anomaly score as an instability precursor index; determine the rupture precursor interval through the temporal changes of the anomaly score, and then identify the high-contribution topological parameters and their change characteristics that have changed significantly within the precursor interval.

[0011] In addition, the coal rock instability precursor identification method and system based on fracture topology parameters according to the present invention may also have the following technical features:

[0012] In one example of the present invention, in step S10, raw acoustic wave data is collected and combined with earthquake source location calculation and focal mechanism inversion to obtain core parameters of the coal and rock fracture source directly related to the fracture, specifically including the following steps:

[0013] S11: The original acoustic wave data collected by the microseismic monitoring system is post-processed and the first wave arrival time and amplitude information are extracted using the red delay information criterion method;

[0014] S12: The spatial positioning of the acoustic emission source is calculated using the simplex positioning robust algorithm through the first wave arrival time data of sensors at different positions;

[0015] S13: Based on the constraints of the tension-shear rupture source model and combined with the first-wave amplitude data of sensors at different locations, the six components of the source rupture moment tensor are solved;

[0016] S14: Through the calculated moment tensor components, the core physical parameters of the coal rock fracture source are quantitatively inverted, where the core physical parameters include positioning coordinates, spatial orientation and fracture volume.

[0017] In one example of the present invention, in step S20, four types of topological relationships are defined by comprehensively considering the geometric size and directional characteristics of the fractures, the fracture penetration index (BCI) is quantitatively calculated, the fracture network is abstracted into a graph model based on graph theory, a topological model is constructed to describe the spatiotemporal evolution characteristics of the fractures, and multidimensional time series topological parameters are extracted to obtain a multivariate topological time series dataset reflecting the evolution of coal and rock fractures, including the following steps:

[0018] S21: Based on geometric relationships, using the fracture location coordinates, volume radius, and spatial orientation, construct the normal vector and parametric equation of the intersection of any two fracture planes, and combine them with the equation of the sphere where the fracture is located. By analyzing the spatial position relationship of the intersection point, the four topological relationships between the fractures are determined: intersection, connection, embedding, or separation. At the same time, based on the distance between the fracture sphere radius and the sphere center, the diameter of the intersection circle is calculated to characterize the geometric connection characteristics of the fracture.

[0019] S22: For any fractures i and j in the fracture network, after determining whether there is a connection relationship based on the spatial topological relationship between the fractures, the fracture penetration index is calculated as follows:

[0020]

[0021] Where, Z(V i ) and Z(V j ) are the normalized volume values ​​of the i-th crack and the j-th crack, d ij is the Euclidean distance between the i-th crack and the j-th crack;

[0022] Assuming that there are n through-crack nodes around crack i, the total fracture penetration index BCI is i The calculation formula is as follows:

[0023]

[0024] S23: After determining the connection relationship between the fractures, a fracture network topology model is constructed based on the geometric relationship between the fracture center and the intersection center;

[0025] S24: Extract the multidimensional topological attribute parameters of the fracture network and, combined with the time series information of microseismic events, construct a multivariate topological time series feature dataset that characterizes the time-varying evolution characteristics of the fracture network structure. The multidimensional topological attribute parameters include: connection type, network density, clustering coefficient, cumulative openness, and permeability index.

[0026] In one example of the present invention, in step S23, the process of constructing the fracture network topology model includes the following steps:

[0027] Graph theory is used to abstract the fracture network into a graph structure, where fractures are nodes and divided into I-nodes, T-nodes, Y-nodes, X-nodes, and D-nodes according to the number of connecting edges. The degree of a node is defined as the total number of connecting edges.

[0028] The geometric information of the crack is assigned to the edges connecting the nodes, and the edges are divided into unidirectional edges and bidirectional edges according to the connectivity index (BCI). Among them, unidirectional edges represent lower connectivity and weak interaction, while bidirectional edges represent higher connectivity and strong interaction, which can form stable crack channels. The crack network topology model includes:

[0029] In one example of the present invention, in step S30, dividing the multivariate topological time series data set and performing data preprocessing includes the following steps:

[0030] S31: Divide the constructed multivariate topological time series dataset into a training set and a test set. The training set selects data from the stable evolution stage from the early to middle stage of coal rock fracture evolution, and the test set selects data from the late stage before unstable fracture.

[0031] S32: Normalize the time series of topological structure parameters of each dimension in the training set and the test set. For the topological parameters x1, x2, ..., x n-1 , x n , using the minimum-maximum normalization method, its expression is:

[0032]

[0033] Where, is the minimum value of the topological parameter, is the maximum value of the topological parameter, y i ∈[0,1];

[0034] S33: The spectral residual method is used to identify and remove abnormal points in the time series of topological structure parameters of each dimension in the training set.

[0035] In one example of the present invention, in step S40, the MTAD-GAT model includes a 1-D convolutional layer, which is used to extract high-level features of each timestamp from the input multivariate time series data, specifically including:

[0036] Let the multivariate time series input tensor be X∈R T×N×D , where T is the maximum timestamp length, N is the number of variables, and D is the input feature dimension of each variable. The feature extraction formula through the 1-D convolution layer is as follows:

[0037] X conv =Conv1D(X)∈R T×N×D′

[0038] Where, X conv is the output tensor of the 1-D convolution layer, and D′ is the local feature dimension extracted after the 1-D convolution layer.

[0039] In one example of the present invention, in step S40, the dual-graph attention mechanism includes the following steps:

[0040] S41: Linear transformation: For each attention head m=1,2,.....,M, the output of each node after the 1-D convolution layer

[0041] Feature X conv ∈R T×N×D′ Perform linear mapping, the expression is:

[0042]

[0043] Where: h i ∈R D′ is the output feature of the i-th node after the convolution layer, W (m) ∈R F×D′ is a learnable linear mapping matrix, F is the output dimension of each attention head, is the feature after mapping;

[0044] S42: Calculate the attention coefficient: For two nodes i and j, the attention mechanism is used to calculate their similarity. The expression of the attention coefficient is:

[0045]

[0046] Where: is the non-normalized attention score between nodes i and j in the mth attention head, LeakyReLU represents the nonlinear activation function, α (m) ∈R 2F is a learnable multi-head attention vector, Represents vector concatenation operation;

[0047] S43: Normalized attention weight: Use Softmax to normalize the attention score, and its expression is:

[0048]

[0049] Where, is the normalized attention score between nodes i and j in the mth attention head, Softmax represents the activation function, and N(i) is the set of neighboring nodes of node i;

[0050] S44: Aggregate neighbor features: By weighting the features of neighbor nodes through attention, the expression for updating the node is:

[0051]

[0052] Where, is the aggregate feature output of node i in the mth attention head, and σ is a nonlinear activation function;

[0053] S45: Multi-head output aggregation: The output features of each head are spliced ​​together by dimension. The expression is:

[0054]

[0055] Where z i is the final representation of node i after all heads are fused;

[0056] Among them, the final output of the feature-oriented graph attention layer is The final output of the time-oriented graph attention layer is The output of the feature-oriented graph attention layer and the output of the time-oriented graph attention layer are arranged into a unified tensor and concatenated with the output of the 1-D convolutional layer to obtain the total representation of the concatenated output, which is then input into the GRU layer for dynamic temporal modeling:

[0057]

[0058] Where M·F is the dimension of the final output of each GAT layer, and D″ is the total feature dimension after splicing.

[0059] In one example of the present invention, step S50 includes the following steps:

[0060] S51: Joint optimization strategy to construct a multi-objective loss function: Combine the prediction module and the reconstruction module, and introduce a joint loss function in the training process to optimize the two sub-networks simultaneously. The formula is as follows:

[0061] Loss=Loss for +Loss rec

[0062] Where, Loss for is the loss function of the prediction model, Loss rec is the loss function of the reconstruction model;

[0063] S52: Construct an anomaly scoring mechanism: The model generates prediction errors and reconstruction probabilities separately, and then fuses them to construct a comprehensive anomaly scoring function. A peak-to-threshold algorithm is used to select the anomaly threshold on the validation set. If the anomaly score for a time step is greater than the anomaly threshold, the time step is marked as "abnormal," otherwise it is marked as "normal." The "abnormal" score is used as an instability precursor index to identify the rupture precursor interval.

[0064] The calculation formula of the comprehensive anomaly scoring function is as follows:

[0065]

[0066] Where, is the predicted value With the actual value x i The square error between (1-p i ) is the probability of encountering an outlier value of feature i according to the reconstruction model, and γ is the hyperparameter selected in the validation set to balance the introduction of the two modules;

[0067] S53: Count and analyze the topological evolution trend in the precursory interval: Count the characteristic quantities of each topological parameter in the precursory interval and compare them with the corresponding parameters in the normal interval; use parameter difference analysis and t-test method to determine whether there are significant changes in the topological parameters in the precursory stage, so as to identify the high-contribution topological parameters and their change characteristics that play a key role in the rupture evolution process.

[0068] In one example of the present invention, the prediction module is configured to access three fully connected layers with a hidden dimension of d2 after the temporal feature representation output by the gated recurrent unit GRU, for performing nonlinear prediction of the data of the next time step; wherein the prediction loss function adopts the root mean square error (RMSE) and is defined as follows:

[0069]

[0070] Where: x n For the input sequence x=(x0,x1,.....x n-1)’s next time step true value, x n,i is the true value of the i-th feature, is the predicted value, K is the total number of features;

[0071] The reconstruction module is configured to use a variational autoencoder to learn the probability distribution of the entire input sequence in the latent space, reconstruct the input and calculate the reconstruction error; wherein the reconstruction loss function is expressed as:

[0072]

[0073] Where, the first term is the reconstruction error, which measures the closeness between the reconstructed value and the original input, q φ (z|x) is the encoder, is the decoder; the second term is the KL divergence, which regularizes the potential space distribution, the latent variable Represents the implicit structure of the data.

[0074] Another object of the present invention is to provide a coal rock instability precursor identification system based on fracture topology parameters, comprising:

[0075] A source parameter acquisition module is configured to acquire raw acoustic wave data through a microseismic monitoring system, and combine source location calculation and focal mechanism inversion to obtain core physical parameters directly related to the microfracture source during coal and rock fracture;

[0076] The topological parameter acquisition module is configured to comprehensively consider the geometric size and orientation characteristics of the fracture to define four types of topological relationships, quantitatively calculate the fracture penetration index (BCI), and abstract the fracture network into a graph model based on graph theory. It is configured to construct a topological model that describes the spatiotemporal evolution characteristics of the fracture, extract multidimensional time series topological parameters, and obtain a multivariate topological time series dataset reflecting the evolution of coal and rock fractures.

[0077] A data set processing and partitioning module is configured to partition the constructed multivariate topological time series data set into a training set and a test set, and perform data normalization and outlier processing on each dimension of the topological structure parameter time series;

[0078] The precursor recognition model construction module is configured to construct a multivariate time series graph model (MTAD-GAT) based on preprocessed training data. This model forms a feature-oriented and time-oriented dual-graph attention mechanism to model the structural dependencies and temporal evolution trends between topological parameters, respectively. Multi-head attention is introduced into the graph attention layer, with different attention scores calculated using multiple independent attention heads. The local and global features extracted by the graph attention layer and the one-dimensional convolutional layer are fused and then input into a gated recurrent unit (GRU) to deeply capture the dynamic evolution patterns and long-term dependencies of topological features. During training, the model is iteratively optimized by minimizing the loss function and combining backpropagation and gradient update mechanisms.

[0079] The instability precursor identification module is configured to pass the output of the gated recurrent unit (GRU) into a prediction module for predicting the observation value at the next moment and a reconstruction module for characterizing the overall temporal distribution characteristics; the errors of the two are fused to construct an anomaly score as an instability precursor index; the rupture precursor interval is determined by the temporal changes of the anomaly score, and then the high-contribution topological parameters and their change characteristics that have changed significantly within the precursor interval are identified.

[0080] Compared with the prior art, the present invention has the following beneficial effects:

[0081] The present invention uses microseismic monitoring data to perform focal mechanism inversion, which can accurately capture the source characteristics of cracks, dynamically track the occurrence, expansion and penetration process of cracks, and then construct a crack topology structure that is closer to the actual evolution law. It provides a new path for achieving more physically based coal rock instability warning, and significantly improves the scientificity and reliability of precursor identification.

[0082] The present invention constructs a fracture network topology analysis model based on graph theory. By extracting multi-dimensional topological parameters, it characterizes the evolution law and local abnormal characteristics of the fracture network, realizes the structural identification and quantitative characterization of the precursors of instability and rupture, breaks through the limitations of traditional signal feature-based early warning, and provides a more accurate and explanatory intelligent early warning method for coal and rock dynamic disasters.

[0083] The present invention introduces the multivariate time series graph model MTAD-GAT, which adopts an unsupervised learning mechanism. It does not require preset labels or manually defined feature thresholds, and can automatically mine potential abnormal patterns from historical crack topology evolution data, significantly improving the model's adaptability and generalization capabilities. Secondly, MTAD-GAT has multivariate recognition capabilities and can simultaneously process the joint evolution characteristics of multiple topological parameters, avoiding the problems of misjudgment and missed judgment that may be caused by a single indicator, and improving the comprehensiveness and robustness of precursor identification. At the same time, multi-head attention is introduced to improve the model, enabling it to capture different structural and time-dependent features in parallel from multiple subspaces, effectively enhancing the stability and expressiveness of feature representation, and further improving the perception of complex crack evolution patterns and the accuracy of anomaly detection.

[0084] The present invention fully integrates graphical modeling and multivariate time series graphical model, realizes a new path of precursor identification and disaster warning based on the evolution of fracture structure, breaks through the traditional warning method based on signal analysis, and establishes an intelligent identification framework for the evolution mechanism of fracture structure. It provides more scientific and efficient technical support for the early warning and active prevention and control of coal and rock dynamic disasters, and has significant engineering practical value and broad prospects for promotion and application.

[0085] Hereinafter, the best embodiment of the present invention will be described in more detail with reference to the accompanying drawings so that the features and advantages of the present invention can be easily understood. BRIEF DESCRIPTION OF THE DRAWINGS

[0086] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings of the embodiments of the present invention. The drawings are only used to illustrate some embodiments of the present invention, but not to limit all embodiments of the present invention thereto.

[0087] Figure 1 Flowchart of a method for identifying coal rock instability precursors based on fracture topology parameters according to an embodiment of the present invention;

[0088] Figure 2 Graph showing four types of spatial topological relationships between cracks according to an embodiment of the present invention;

[0089] Figure 3 The fracture network according to an embodiment of the present invention is generalized into a graph structure;

[0090] Figure 4 MTAD-GAT is a principle structure diagram of a multivariate time series graph model according to an embodiment of the present invention;

[0091] Figure 5 It is a feature-oriented multi-head graph attention GAT layer according to an embodiment of the present invention. DETAILED DESCRIPTION

[0092] In order to make the purpose, technical solution and advantages of the technical solution of the present invention clearer, the technical solution of the embodiment of the present invention will be clearly and completely described below in conjunction with the drawings of specific embodiments of the present invention. The same figure marks in the drawings represent the same parts. It should be noted that the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the described embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0093] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning understood by persons of ordinary skill in the field to which the invention belongs. The words "first", "second" and similar terms used in the patent application specification and claims of the present invention do not indicate any order, quantity or importance, but are only used to distinguish different components. Similarly, words such as "a" or "an" do not necessarily indicate a quantity limitation. Words such as "include" or "comprising" mean that the elements or objects preceding the word include the elements or objects listed after the word and their equivalents, without excluding other elements or objects. Words such as "connected" or "connected" are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. "Up", "down", "left", "right" and the like are only used to indicate relative positional relationships. When the absolute position of the object being described changes, the relative positional relationship may also change accordingly.

[0094] To make up for this shortcoming, monitoring and early warning methods based on fracture structures have gradually become a research hotspot. By constructing a fracture topology network, analyzing the topological changes of the network structure in the time-series evolution, and deeply exploring the correlation and collaborative abnormal characteristics between fractures, potential fracture precursors are identified from the structural level, which significantly improves the physical interpretability and reliability of the early warning. At the same time, the introduction of neural networks for time-series modeling of fracture topology parameters can effectively capture the nonlinear relationship between topological features and their dynamic trends over time, thereby improving the sensitivity and accuracy of precursor identification. Therefore, the precursor identification method based on fracture network structure and neural network can more comprehensively and accurately reveal the evolution process of fractures inside coal and rock masses, provide strong technical support for mine safety management, and greatly improve the efficiency and accuracy of precursor identification.

[0095] According to the first aspect of the present invention, a method for identifying coal rock instability precursors based on fracture topology parameters is provided. Figure 1 、 Figure 4 As shown, the following steps are included:

[0096] S10: The original acoustic wave data is collected by the microseismic monitoring system, and combined with the source location calculation and focal mechanism inversion, the core physical parameters directly related to the microfracture source in the coal and rock fracture process are obtained;

[0097] S20: Four types of topological relationships are defined by comprehensively considering the geometric size and direction characteristics of the cracks, and the fracture penetration index (BCI) is quantitatively calculated. The fracture network is abstracted into a graph model based on graph theory, and a topological model is constructed to characterize the spatiotemporal evolution characteristics of the cracks. Multidimensional time series topological parameters such as connection type and network density are extracted to obtain a multivariate topological time series dataset reflecting the evolution of coal and rock fractures, providing structured input data for precursor identification.

[0098] S30: Divide the constructed multivariate topological time series dataset into a training set and a test set, and perform data normalization and outlier processing on each dimension of the topological structure parameter time series; data normalization is applied to both the training set and the test set, while outlier cleaning is only applied to the training set.

[0099] S40: Based on the preprocessed training data, a multivariate time series graph model MTAD-GAT is constructed to form a feature-oriented and time-oriented dual-graph attention mechanism to model the structural dependency and temporal evolution trend between topological parameters respectively. Multi-head attention is introduced in the graph attention layer, and different attention scores are calculated through multiple independent attention heads to obtain a more stable and expressive feature representation, thereby enhancing the model's ability to model local structures and temporal patterns. The local and global features extracted by the graph attention layer and the one-dimensional convolutional layer are fused and input into the gated recurrent unit (GRU) to deeply capture the dynamic evolution pattern and long-term dependency of topological features. During the training process, the model is iteratively optimized by minimizing the loss function and combining backpropagation and gradient update mechanisms.

[0100] S50: The trained model is applied to the test set data, and the output of the gated recurrent unit (GRU) is passed to the prediction module for predicting the observation value at the next moment and the reconstruction module for characterizing the overall temporal distribution characteristics. The errors of the two are fused to construct an anomaly score as an instability precursor index. The rupture precursor interval is determined by the temporal changes of the anomaly score, and then the high-contribution topological parameters and their change characteristics that have changed significantly within the precursor interval are identified.

[0101] It is understandable that if Figure 1 and Figure 4 As shown, the multivariate time series graph model MTAD-GAT includes: a one-dimensional convolutional layer, a graph attention layer, a gated recurrent unit GRU, a prediction module and a reconstruction module, wherein the graph attention layer includes a parallel feature-oriented graph attention GAT layer and a time-oriented graph attention GAT layer, the front end of the graph attention layer is coupled with the one-dimensional convolutional layer, and the back end of the graph attention layer is coupled with the gated recurrent unit GRU through splicing, and the gated recurrent unit GRU is coupled with the parallel prediction module and reconstruction module respectively.

[0102] This identification method uses microseismic monitoring data to perform source mechanism inversion, which can accurately capture the source characteristics of the cracks, dynamically track the occurrence, expansion and penetration process of the cracks, and then construct a crack topology structure that is closer to the actual evolution law. It provides a new path for achieving more physically based coal rock instability warning, and significantly improves the scientific nature and reliability of precursor identification.

[0103] This identification method constructs a fracture network topology analysis model based on graph theory. By extracting multi-dimensional topological parameters, it characterizes the evolution law and local abnormal characteristics of the fracture network, realizes the structural identification and quantitative characterization of the precursors of instability and rupture, breaks through the limitations of traditional signal feature-based early warning, and provides a more accurate and explanatory intelligent early warning method for coal and rock dynamic disasters.

[0104] This identification method introduces the multivariate time series graph model MTAD-GAT, which uses an unsupervised learning mechanism. It does not require preset labels or manually defined feature thresholds, and can automatically mine potential abnormal patterns from historical crack topology evolution data, significantly improving the model's adaptability and generalization capabilities. Secondly, MTAD-GAT has multivariate recognition capabilities and can simultaneously process the joint evolution characteristics of multiple topological parameters, avoiding the misjudgment and omission problems that may be caused by a single indicator, and improving the comprehensiveness and robustness of precursor identification. At the same time, multi-head attention is introduced to improve the model, enabling it to capture different structural and time-dependent features in parallel from multiple subspaces, effectively enhancing the stability and expressiveness of feature representation, and further improving the perception of complex crack evolution patterns and the accuracy of anomaly detection.

[0105] This identification method fully integrates graphical modeling and multivariate time series graphical model, realizing a new path for precursor identification and disaster warning based on the evolution of fracture structure. It breaks through the traditional warning method based on signal analysis and establishes an intelligent identification framework for the evolution mechanism of fracture structure. It provides more scientific and efficient technical support for the early warning and active prevention and control of coal and rock dynamic disasters, and has significant engineering practical value and broad prospects for promotion and application.

[0106] In one example of the present invention, in step S10, as Figure 2 As shown in the figure, the original acoustic wave data is collected and combined with the earthquake source location calculation and focal mechanism inversion to obtain the core parameters of the coal and rock fracture source directly related to the fracture. The specific steps include the following:

[0107] S11: The original acoustic wave data collected by the microseismic monitoring system is post-processed and the first wave arrival time and amplitude information are extracted using the red delay information criterion method;

[0108] S12: The spatial positioning of the acoustic emission source is calculated using the simplex positioning robust algorithm through the first wave arrival time data of sensors at different positions;

[0109] S13: Based on the constraints of the tension-shear rupture source model and combined with the first-wave amplitude data of sensors at different locations, the six components of the source rupture moment tensor are solved;

[0110] S14: Through the calculated moment tensor components, the core physical parameters of the coal rock fracture source are quantitatively inverted, where the core physical parameters include positioning coordinates, spatial orientation and fracture volume.

[0111] In one example of the present invention, in step S20, four types of topological relationships are defined by comprehensively considering the geometric size and directional characteristics of the fractures, the fracture penetration index (BCI) is quantitatively calculated, the fracture network is abstracted into a graph model based on graph theory, a topological model is constructed to describe the spatiotemporal evolution characteristics of the fractures, and multidimensional time series topological parameters are extracted to obtain a multivariate topological time series dataset reflecting the evolution of coal and rock fractures, including the following steps:

[0112] S21: Based on geometric relationships, using the fracture location coordinates, volume radius, and spatial orientation, construct the normal vector and parametric equation of the intersection of any two fracture planes, and combine them with the equation of the sphere where the fracture is located. By analyzing the spatial position relationship of the intersection point, the four topological relationships between the fractures are determined: intersection, connection, embedding, or separation. At the same time, based on the distance between the fracture sphere radius and the sphere center, the diameter of the intersection circle is calculated to characterize the geometric connection characteristics of the fracture.

[0113] S22: For any fractures i and j in the fracture network, after determining whether there is a connection relationship based on the spatial topological relationship between the fractures, the fracture penetration index is calculated as follows:

[0114]

[0115] Where, Z(V i ) and Z(V j ) are the normalized volume values ​​of the i-th crack and the j-th crack, d ij is the Euclidean distance between the i-th crack and the j-th crack;

[0116] Assuming that there are n through-crack nodes around crack i, the total fracture penetration index BCI is i The calculation formula is as follows:

[0117]

[0118] S23: After determining the connection relationship between the fractures, a fracture network topology model is constructed based on the geometric relationship between the fracture center and the intersection center;

[0119] S24: Extract the multidimensional topological attribute parameters of the fracture network and, combined with the time series information of microseismic events, construct a multivariate topological time series feature dataset that characterizes the time-varying evolution characteristics of the fracture network structure. The multidimensional topological attribute parameters include: connection type, network density, clustering coefficient, cumulative openness, and permeability index.

[0120] In one example of the present invention, in step S23, as Figure 3 As shown in Figure 2, the process of constructing the fracture network topology model includes the following steps:

[0121] Graph theory is used to abstract the fracture network into a graph structure, where fractures are nodes and divided into I-nodes, T-nodes, Y-nodes, X-nodes, and D-nodes according to the number of connected edges. The degree of a node is defined as the total number of connected edges.

[0122] The geometric information of the crack is assigned to the edges connecting the nodes, and the edges are divided into unidirectional edges and bidirectional edges according to the connectivity index (BCI). Among them, unidirectional edges represent lower connectivity and weak interaction, while bidirectional edges represent higher connectivity and strong interaction, which can form stable crack channels.

[0123] In one example of the present invention, Figure 5 As shown, in step S24, multi-dimensional topological attribute parameters such as connection type, network density, clustering coefficient, cumulative openness, and connectivity index of the fracture network are extracted, specifically:

[0124] ① Connection type: including node type, edge type, and edge directionality. The five node types, 15 edge types (II, IT, IY, IX, ID, TT, TY, TX, TD, YY, YX, YD, XX, XD, DD), and the number and proportion of unidirectional and bidirectional edges in the fracture network model are analyzed along with the stage-by-stage change trend of coal and rock deformation.

[0125] ② Network density: The ratio of the number of edges in a fracture network to the maximum number of edges it can accommodate. It is used to describe the edge density between nodes in the network and indicates the density of fracture distribution.

[0126]

[0127] Where N e is the total number of edges in the fracture network, and N is the total number of nodes;

[0128] ③ Clustering coefficient: The local clustering coefficient C(v) is the ratio of the actual number of edges between a node's neighboring nodes to the total number of possible edges between the node's neighboring nodes, reflecting the compactness of the local network structure around a node. The global clustering coefficient C is the average of the local clustering coefficients of all nodes, measuring the degree of clustering of all nodes in the graph and reflecting the compactness of the global structure of the fracture network:

[0129]

[0130] Where, E v is the actual number of edges between node v’s neighbor nodes, k v is the degree of node v, and N is the total number of nodes in the graph.

[0131] ④ Cumulative aperture: The sum of the apertures of all nodes in the fracture network changes with the stage-by-stage deformation of coal and rock.

[0132] ⑤ Connectivity index: The spatial distribution of the connectivity index of all nodes in the fracture network changes with the stage-by-stage deformation of coal and rock.

[0133] In one example of the present invention, in step S30, dividing the multivariate topological time series data set and performing data preprocessing includes the following steps:

[0134] S31: Divide the constructed multivariate topological time series dataset into a training set and a test set. The training set selects data from the stable evolution stage from the early to middle stage of coal rock fracture evolution, and the test set selects data from the late stage before unstable fracture.

[0135] S32: To eliminate the dimensional inconsistency problem between different topological parameters, the time series of each dimension of topological structure parameters in the training set and the test set are normalized. For the topological parameters x1, x2, ..., x n-1 , x n , using the minimum-maximum normalization method, its expression is:

[0136]

[0137] Where, is the minimum value of the topological parameter, is the maximum value of the topological parameter, y i ∈[0,1];

[0138] S33: A spectral residual (SR) method is used to identify and remove outliers in the time series of topological parameters in each dimension of the training set. This method extracts the frequency domain features of the time series through Fourier transform, calculates the spectral residual of its amplitude spectrum, and then reconstructs the abnormal enhancement signal through inverse transform. Finally, a threshold is set based on the amplitude distribution of the reconstructed signal to identify and remove potential outliers.

[0139] In one example of the present invention, in step S40, the MTAD-GAT model includes a 1-D convolutional layer (with a kernel size of 7), which is used to extract high-level features of each timestamp from the input multivariate time series data, specifically including:

[0140] Let the multivariate time series input tensor be X∈R T×N×D , where T is the maximum timestamp length, N is the number of variables, and D is the input feature dimension of each variable. The feature extraction formula through the 1-D convolution layer is as follows:

[0141] X conv =Conv1D(X)∈RT×N×D′

[0142] Where, X conv is the output tensor of the 1-D convolution layer, and D′ is the local feature dimension extracted after the 1-D convolution layer.

[0143] In one example of the present invention, in step S40, the dual-graph attention mechanism includes the following steps:

[0144] S41: Linear transformation: For each attention head m=1,2,.....,M, the output of each node after the 1-D convolution layer

[0145] Feature X conv ∈R T×N×D′ Perform linear mapping, the expression is:

[0146]

[0147] Where: h i ∈R D′ is the output feature of the i-th node after the convolution layer, W (m) ∈R F×D′ is a learnable linear mapping matrix, F is the output dimension of each attention head, is the feature after mapping;

[0148] S42: Calculate the attention coefficient: For two nodes i and j, the attention mechanism is used to calculate their similarity. The expression of the attention coefficient is:

[0149]

[0150] Where: is the non-normalized attention score between nodes i and j in the mth attention head, LeakyReLU represents the nonlinear activation function, α (m) ∈R 2F is a learnable multi-head attention vector, Represents vector concatenation operation;

[0151] S43: Normalized attention weight: Use Softmax to normalize the attention score, and its expression is:

[0152]

[0153] Where, is the normalized attention score between nodes i and j in the mth attention head, Softmax represents the activation function, and N(i) is the set of neighboring nodes of node i;

[0154] S44: Aggregate neighbor features: By weighting the features of neighbor nodes through attention, the expression for updating the node is:

[0155]

[0156] Where, is the aggregate feature output of node i in the mth attention head, and σ is a nonlinear activation function;

[0157] S45: Multi-head output aggregation: The output features of each head are spliced ​​together by dimension. The expression is:

[0158]

[0159] Where z i is the final representation of node i after all heads are fused;

[0160] Among them, the final output of the feature-oriented graph attention layer is The final output of the time-oriented graph attention layer is The output of the feature-oriented graph attention layer and the output of the time-oriented graph attention layer are arranged into a unified tensor and concatenated with the output of the 1-D convolutional layer to obtain the total representation of the concatenated output, which is then input into the GRU layer for dynamic temporal modeling:

[0161]

[0162] Where M·F is the dimension of the final output of each GAT layer, and D″ is the total feature dimension after splicing.

[0163] In one example of the present invention, in step S40, a multivariate time series graph model MTAD-GAT is introduced. The gated recurrent unit in the model is a recurrent neural network that can effectively model time series data. Its structure is suitable for capturing the long-term dependencies of time series data. Specifically,

[0164] After feature extraction by the 1-D convolution and GAT layers, the obtained high-dimensional features are input to the GRU layer. The output dimension of this layer is set to d1, which means that the GRU layer will further compress and learn the temporal features of the input data, thereby capturing the complex patterns and dynamic evolution laws in multiple time series.

[0165] In one example of the present invention, step S50 includes the following steps:

[0166] S51: Joint optimization strategy to construct a multi-objective loss function: Combine the prediction module and the reconstruction module, and introduce a joint loss function in the training process to optimize the two sub-networks simultaneously. The formula is as follows:

[0167] Loss=Lossfor +Loss rec

[0168] Where, Loss for is the loss function of the prediction model, Loss rec is the loss function of the reconstruction model;

[0169] S52: Construct an anomaly scoring mechanism: The model generates prediction errors and reconstruction probabilities separately, and then fuses them to construct a comprehensive anomaly scoring function. The peak over threshold (POT) algorithm is used to select the anomaly threshold on the validation set. If the anomaly score of a time step is greater than the anomaly threshold, the time step is marked as "abnormal," otherwise it is marked as "normal." The "abnormal" score is used as an instability precursor index to identify the rupture precursor interval.

[0170] The calculation formula of the comprehensive anomaly scoring function is as follows:

[0171]

[0172] Where, is the predicted value With the actual value x i The square error between (1-p i ) is the probability of encountering an outlier value of feature i according to the reconstruction model, and γ is the hyperparameter selected in the validation set to balance the introduction of the two modules;

[0173] S53: Statistically analyze the topological evolution trend within the precursory interval: Statistically calculate the mean, variance, slope and other characteristic quantities of each topological parameter within the precursory interval, and compare them with the corresponding parameters in the normal interval; use parameter difference analysis and t-test methods to determine whether there are significant changes in the topological parameters during the precursory stage, so as to identify the high-contribution topological parameters and their change characteristics that play a key role in the rupture evolution process; if multiple parameters show synchronous or coordinated change trends within the same precursory interval, such as synchronous increases or synchronous fluctuation enhancements, it may indicate that the overall fracture network structure has entered an unstable evolutionary state, providing important topological precursor information for the impending rupture.

[0174] In one example of the present invention, the prediction module is configured to access three fully connected layers with a hidden dimension of d2 after the temporal feature representation output by the gated recurrent unit GRU, for performing nonlinear prediction of the data of the next time step; wherein the prediction loss function adopts the root mean square error (RMSE) and is defined as follows:

[0175]

[0176] Where: x n For the input sequence x=(x0,x1,.....x n-1)’s next time step true value, x n,i is the true value of the i-th feature, is the predicted value, K is the total number of features;

[0177] The reconstruction module is configured to use a variational autoencoder (VAE) to learn the probability distribution of the entire input sequence in the latent space, reconstruct the input and calculate the reconstruction error; wherein the reconstruction loss function is expressed as:

[0178]

[0179] Where the first term is the reconstruction error (negative log likelihood), which measures how close the reconstructed value is to the original input, and q φ (z|x) is the encoder, is the decoder; the second term is the KL divergence, which regularizes the potential space distribution, the latent variable Represents the implicit structure of the data.

[0180] Among them, the VAE model includes the encoder q φ (z|x) maps the input sequence to a latent variable distribution; the decoder Reconstruct the input sequence based on the latent variables.

[0181] According to a second aspect of the present invention, a coal rock instability precursor identification system based on fracture topology parameters comprises:

[0182] A source parameter acquisition module is configured to acquire raw acoustic wave data through a microseismic monitoring system, and combine source location calculation and focal mechanism inversion to obtain core physical parameters directly related to the microfracture source during coal and rock fracture;

[0183] The topological parameter acquisition module is configured to define four types of topological relationships by comprehensively considering the geometric size and directional characteristics of the cracks, quantitatively calculate the fracture penetration index (BCI), and abstract the crack network into a graph model based on graph theory. It is configured to construct a topological model that describes the spatiotemporal evolution characteristics of the cracks, extract multidimensional time series topological parameters such as connection type and network density, and obtain a multivariate topological time series data set reflecting the evolution of coal and rock fractures, providing structured input data for precursor identification.

[0184] The dataset processing and partitioning module is configured to divide the constructed multivariate topological time series dataset into a training set and a test set, and perform data normalization and outlier processing on each dimensional topological structure parameter time series; data normalization is applied to both the training set and the test set, while outlier cleaning is only applied to the training set.

[0185] The precursor recognition model construction module is configured to construct a multivariate time series graph model (MTAD-GAT) based on preprocessed training data, forming a feature-oriented and time-oriented dual-graph attention mechanism to model the structural dependencies and temporal evolution trends between topological parameters, respectively. Multi-head attention is introduced in the graph attention layer, and different attention scores are calculated through multiple independent attention heads to obtain a more stable and expressive feature representation, thereby enhancing the model's ability to model local structures and temporal patterns. The local and global features extracted by the graph attention layer and the one-dimensional convolutional layer are fused and input into the gated recurrent unit (GRU) to deeply capture the dynamic evolution patterns and long-term dependencies of topological features. During the training process, the model is iteratively optimized by minimizing the loss function and combining backpropagation and gradient update mechanisms.

[0186] The instability precursor identification module is configured to apply the trained model to the test set data, and use the output of the gated recurrent unit (GRU) to pass it into the prediction module for predicting the observation value at the next moment and the reconstruction module for characterizing the overall time series distribution characteristics; the errors of the two are fused to construct an anomaly score as the instability precursor index; the rupture precursor interval is determined by the temporal changes of the anomaly score, and then the high-contribution topological parameters and their change characteristics that have changed significantly within the precursor interval are identified.

[0187] This identification system uses microseismic monitoring data to perform source mechanism inversion, which can accurately capture the source characteristics of cracks, dynamically track the occurrence, expansion and penetration process of cracks, and then construct a crack topology structure that is closer to the actual evolution law. It provides a new path for achieving more physically based coal rock instability warning, and significantly improves the scientific nature and reliability of precursor identification.

[0188] This identification system constructs a fracture network topology analysis model based on graph theory. By extracting multi-dimensional topological parameters, it characterizes the evolution law and local abnormal characteristics of the fracture network, realizes the structural identification and quantitative characterization of the precursors of instability and rupture, breaks through the limitations of traditional signal feature-based early warning, and provides a more accurate and explanatory intelligent early warning method for coal and rock dynamic disasters.

[0189] The recognition system introduces the multivariate time series graph model MTAD-GAT, which adopts an unsupervised learning mechanism. It does not require preset labels or manually defined feature thresholds, and can automatically mine potential abnormal patterns from historical crack topology evolution data, significantly improving the model's adaptability and generalization capabilities. Secondly, MTAD-GAT has multivariate recognition capabilities and can simultaneously process the joint evolution characteristics of multiple topological parameters, avoiding the misjudgment and omission problems that may be caused by a single indicator, and improving the comprehensiveness and robustness of precursor identification. At the same time, multi-head attention is introduced to improve the model, enabling it to capture different structural and time-dependent features in parallel from multiple subspaces, effectively enhancing the stability and expressiveness of feature representation, and further improving the perception of complex crack evolution patterns and the accuracy of anomaly detection.

[0190] This identification system fully integrates graphical modeling and multivariate time series graphical models, realizing a new path for precursor identification and disaster warning based on the evolution of fracture structures. It breaks through the traditional warning method based on signal analysis and establishes an intelligent identification framework for the evolution mechanism of fracture structures. It provides more scientific and efficient technical support for the early warning and active prevention and control of coal and rock dynamic disasters, and has significant engineering practical value and broad prospects for promotion and application.

[0191] The above describes in detail the exemplary implementation of the method and system for identifying coal rock instability precursors based on fracture topology parameters proposed in the present invention with reference to the preferred embodiments. However, those skilled in the art will understand that, without departing from the concept of the present invention, various modifications and variations can be made to the above-mentioned specific embodiments, and various technical features and structures proposed in the present invention can be combined in various ways without exceeding the scope of protection of the present invention, which is determined by the appended claims.

Claims

1. A method for identifying coal rock instability precursors based on fracture topology parameters, characterized in that: The steps include: S10: The original acoustic wave data is collected by the microseismic monitoring system, and combined with the earthquake source location calculation and focal mechanism inversion, the core physical parameters directly related to the microfracture source in the coal and rock fracture process are obtained; S20: Four types of topological relationships are defined by comprehensively considering the geometric size and orientation characteristics of the fractures. The fracture penetration index (BCI) is quantitatively calculated. The fracture network is abstracted into a graph model based on graph theory. A topological model that describes the spatiotemporal evolution of fractures is constructed. Multidimensional time series topological parameters are extracted to obtain a multivariate topological time series dataset reflecting the evolution of coal and rock fractures. S30: Divide the constructed multivariate topological time series dataset into a training set and a test set, and perform data normalization and outlier processing on each dimension of the topological structure parameter time series; S40: Based on the preprocessed training data, a multivariate time series graph model MTAD-GAT is constructed to form a feature-oriented and time-oriented dual-graph attention mechanism to model the structural dependency and temporal evolution trend between topological parameters respectively. Multi-head attention is introduced in the graph attention layer to calculate different attention scores through multiple independent attention heads. The local and global features extracted by the graph attention layer and the one-dimensional convolutional layer are fused and input into the gated recurrent unit (GRU) to deeply capture the dynamic evolution pattern and long-term dependency of topological features. During the training process, the model is iteratively optimized by minimizing the loss function and combining backpropagation and gradient update mechanisms; S50: Apply the trained model to the test set data, and use the gated recurrent unit (GRU) output to pass it into the prediction module for predicting the observation value at the next moment and the reconstruction module for characterizing the overall temporal distribution characteristics; fuse the errors of the two to construct an anomaly score as an instability precursor index; determine the rupture precursor interval through the temporal changes of the anomaly score, and then identify the high-contribution topological parameters and their change characteristics that have changed significantly within the precursor interval.

2. The method for identifying coal rock instability precursors based on fracture topology parameters according to claim 1, characterized in that: In step S10, raw acoustic wave data is collected and combined with earthquake source location calculation and focal mechanism inversion to obtain core parameters of the coal and rock fracture source directly related to the fracture, which specifically includes the following steps: S11: The original acoustic wave data collected by the microseismic monitoring system is post-processed and the first wave arrival time and amplitude information are extracted using the red delay information criterion method; S12: The spatial positioning of the acoustic emission source is calculated using the simplex positioning robust algorithm through the first wave arrival time data of sensors at different positions; S13: Based on the constraints of the tension-shear rupture source model and combined with the first-wave amplitude data of sensors at different locations, the six components of the source rupture moment tensor are solved; S14: Through the calculated moment tensor components, the core physical parameters of the coal rock fracture source are quantitatively inverted, where the core physical parameters include positioning coordinates, spatial orientation and fracture volume.

3. The method for identifying coal rock instability precursors based on fracture topology parameters according to claim 1, characterized in that: In step S20, four types of topological relationships are defined by comprehensively considering the geometric size and directional characteristics of the fractures, the fracture penetration index (BCI) is quantitatively calculated, the fracture network is abstracted into a graph model based on graph theory, a topological model is constructed to describe the spatiotemporal evolution characteristics of the fractures, and multidimensional time series topological parameters are extracted to obtain a multivariate topological time series dataset reflecting the evolution of coal and rock fractures. The process includes the following steps: S21: Based on geometric relationships, using the fracture location coordinates, volume radius, and spatial orientation, the normal vector and parametric equation of the intersection line of any two fracture planes are constructed and combined with the equation of the sphere where the fracture is located. By analyzing the spatial position relationship of the intersection point, four types of topological relationships between the fractures are determined: intersection, connection, embedding, or separation. At the same time, based on the distance between the fracture sphere radius and the sphere center, the diameter of the intersection circle is calculated to characterize the geometric connection characteristics of the fracture. S22: For any fractures i and j in the fracture network, after determining whether there is a connection relationship based on the spatial topological relationship between the fractures, the fracture penetration index is calculated as follows: Where, Z(V i ) and Z(V j ) are the normalized volume values ​​of the i-th crack and the j-th crack, d ij is the Euclidean distance between the i-th crack and the j-th crack; Assuming that there are n through-crack nodes around crack i, the total fracture penetration index BCI is i The calculation formula is as follows: S23: After determining the connection relationship between the fractures, a fracture network topology model is constructed based on the geometric relationship between the fracture center and the intersection center; S24: Extract the multidimensional topological attribute parameters of the fracture network and, combined with the time series information of microseismic events, construct a multivariate topological time series feature dataset that characterizes the time-varying evolution characteristics of the fracture network structure. The multidimensional topological attribute parameters include: connection type, network density, clustering coefficient, cumulative openness, and permeability index.

4. The method for identifying coal rock instability precursors based on fracture topology parameters according to claim 1, characterized in that: In step S23, the process of constructing the fracture network topology model includes the following steps: Graph theory is used to abstract the fracture network into a graph structure, where fractures are nodes and divided into I-nodes, T-nodes, Y-nodes, X-nodes, and D-nodes according to the number of connecting edges. The degree of a node is defined as the total number of connecting edges. The geometric information of the crack is assigned to the edges connecting the nodes, and the edges are divided into unidirectional edges and bidirectional edges according to the connectivity index (BCI). Among them, unidirectional edges represent lower connectivity and weak interaction, while bidirectional edges represent higher connectivity and strong interaction, which can form stable crack channels.

5. The method for identifying coal rock instability precursors based on fracture topology parameters according to claim 1, characterized in that: In step S30, the multivariate topological time series data set is divided and data preprocessing is performed, including the following steps: S31: Divide the constructed multivariate topological time series dataset into a training set and a test set. The training set selects data from the stable evolution stage from the early to middle stage of coal rock fracture evolution, and the test set selects data from the late stage before unstable fracture. S32: Normalize the time series of topological structure parameters of each dimension in the training set and the test set. For the topological parameters x1, x2, ..., x n-1 , x n , using the minimum-maximum normalization method, its expression is: Where, is the minimum value of the topological parameter, is the maximum value of the topological parameter, y i ∈[0,1]; S33: The spectral residual method is used to identify and remove abnormal points in the time series of topological structure parameters of each dimension in the training set.

6. The method for identifying coal rock instability precursors based on fracture topology parameters according to claim 1, characterized in that: In step S40, the MTAD-GAT model includes a 1-D convolutional layer, which is used to extract high-level features of each timestamp from the input multivariate time series data, specifically including: Let the multivariate time series input tensor be X∈R T×N×D , where T is the maximum timestamp length, N is the number of variables, and D is the input feature dimension of each variable. The feature extraction formula through the 1-D convolution layer is as follows: X conv =Conv1D(X)∈R T×N×D′ Where, X conv is the output tensor of the 1-D convolution layer, and D′ is the local feature dimension extracted after the 1-D convolution layer.

7. The method for identifying coal rock instability precursors based on fracture topology parameters according to claim 1, characterized in that: In step S40, the dual-graph attention mechanism includes the following steps: S41: Linear transformation: For each attention head m=1,2,.....,M, the output feature X of each node after the 1-D convolution layer conv ∈R T×N×D′ Perform linear mapping, the expression is: Where: h i ∈R D′ is the output feature of the i-th node after the convolution layer, W (m) ∈R F×D′ is a learnable linear mapping matrix, F is the output dimension of each attention head, is the feature after mapping; S42: Calculate the attention coefficient: For two nodes i and j, the attention mechanism is used to calculate their similarity. The expression of the attention coefficient is: Where: is the non-normalized attention score between nodes i and j in the mth attention head, LeakyReLU represents the nonlinear activation function, α (m) ∈ R 2F is a learnable multi-head attention vector, Represents vector concatenation operation; S43: Normalized attention weight: Use Softmax to normalize the attention score, and its expression is: Where, is the normalized attention score between nodes i and j in the mth attention head, Softmax represents the activation function, and N(i) is the set of neighboring nodes of node i; S44: Aggregate neighbor features: By weighting the features of neighbor nodes through attention, the expression for updating the node is: Where, is the aggregate feature output of node i in the mth attention head, and σ is a nonlinear activation function; S45: Multi-head output aggregation: The output features of each head are spliced ​​together by dimension. The expression is: Where z i is the final representation of node i after all heads are fused; Among them, the final output of the feature-oriented graph attention layer is The final output of the time-oriented graph attention layer is The output of the feature-oriented graph attention layer and the output of the time-oriented graph attention layer are arranged into a unified tensor and concatenated with the output of the 1-D convolutional layer to obtain the total representation of the concatenated output, which is then input into the GRU layer for dynamic temporal modeling: Where M·F is the dimension of the final output of each GAT layer, and D″ is the total feature dimension after splicing.

8. The method for identifying coal rock instability precursors based on fracture topology parameters according to claim 1, characterized in that: The step S50 includes the following steps: S51: Joint optimization strategy to construct a multi-objective loss function: Combine the prediction module and the reconstruction module, and introduce a joint loss function in the training process to optimize the two sub-networks simultaneously. The formula is as follows: Loss=Loss for +Loss rec Where, Loss for is the loss function of the prediction model, Loss rec is the loss function of the reconstruction model; S52: Construct an anomaly scoring mechanism: The model generates prediction errors and reconstruction probabilities separately, and then fuses them to construct a comprehensive anomaly scoring function. The peak-to-threshold algorithm is used to select the anomaly threshold on the validation set. If the anomaly score of a time step is greater than the anomaly threshold, the time step is marked as "abnormal", otherwise it is marked as "normal". The "abnormal" score is used as an instability precursor index to identify the rupture precursor interval. The calculation formula of the comprehensive anomaly scoring function is as follows: Where, is the predicted value The square error between the actual value xi, (1-p i ) is the probability of encountering an outlier value of feature i according to the reconstruction model, and γ is the hyperparameter selected in the validation set to balance the introduction of the two modules; S53: Count and analyze the topological evolution trend in the precursory interval: Count the characteristic quantities of each topological parameter in the precursory interval and compare them with the corresponding parameters in the normal interval; use parameter difference analysis and t-test method to determine whether there are significant changes in the topological parameters in the precursory stage, so as to identify the high-contribution topological parameters and their change characteristics that play a key role in the rupture evolution process.

9. The method for identifying coal rock instability precursors based on fracture topology parameters according to claim 1, characterized in that: The prediction module is configured to access three fully connected layers with a hidden dimension of d2 after the temporal feature representation output by the gated recurrent unit (GRU) to perform nonlinear prediction of the data of the next time step; wherein the prediction loss function is defined as the root mean square error as follows: Where: x n For the input sequence x=(x0,x1,.....x n-1 )’s next time step true value, x n,i is the true value of the i-th feature, is the predicted value, K is the total number of features; The reconstruction module is configured to use a variational autoencoder to learn the probability distribution of the entire input sequence in the latent space, reconstruct the input and calculate the reconstruction error; wherein the reconstruction loss function is expressed as: Where, the first term is the reconstruction error, which measures the closeness between the reconstructed value and the original input, q φ (z|x) is the encoder, is the decoder; the second term is the KL divergence, which regularizes the potential space distribution, the latent variable Represents the implicit structure of the data.

10. A coal rock instability precursor identification system based on fracture topology parameters, characterized in that: include: A source parameter acquisition module is configured to acquire raw acoustic wave data through a microseismic monitoring system, and combine source location calculation and focal mechanism inversion to obtain core physical parameters directly related to the microfracture source during coal and rock fracture; The topological parameter acquisition module is configured to comprehensively consider the geometric size and orientation characteristics of the fracture to define four types of topological relationships, quantitatively calculate the fracture penetration index (BCI), and abstract the fracture network into a graph model based on graph theory. It is configured to construct a topological model that describes the spatiotemporal evolution characteristics of the fracture, extract multidimensional time series topological parameters, and obtain a multivariate topological time series dataset reflecting the evolution of coal and rock fractures. A data set processing and partitioning module is configured to partition the constructed multivariate topological time series data set into a training set and a test set, and perform data normalization and outlier processing on each dimension of the topological structure parameter time series; The precursor recognition model construction module is configured to construct a multivariate time series graph model MTAD-GAT based on preprocessed training data, forming a feature-oriented and time-oriented dual-graph attention mechanism to model the structural dependencies and temporal evolution trends between topological parameters, respectively. Multi-head attention is introduced in the graph attention layer, and different attention scores are calculated through multiple independent attention heads. The local and global features extracted by the graph attention layer and the one-dimensional convolutional layer are fused and input into the gated recurrent unit (GRU) to deeply capture the dynamic evolution patterns and long-term dependencies of topological features. During the training process, the model is iteratively optimized by minimizing the loss function and combining backpropagation and gradient update mechanisms; The instability precursor identification module is configured to apply the trained model to the test set data, and use the output of the gated recurrent unit (GRU) to pass it into the prediction module for predicting the observation value at the next moment and the reconstruction module for characterizing the overall temporal distribution characteristics; the errors of the two are combined to construct an anomaly score as the instability precursor index; the rupture precursor interval is determined by the temporal changes of the anomaly score, and then the high-contribution topological parameters and their change characteristics that have changed significantly within the precursor interval are identified.

Images