A Dynamic Prediction Method for Mine Gas Outburst Based on Spatiotemporal Graph Networks
By constructing a gas fractal spatiotemporal topology map and a local polynomial prediction model, the problem of integrating spatial topological information with temporal evolution laws in gas outburst prediction is solved, improving prediction accuracy and adaptability, extending prediction time, and making it suitable for mine gas disaster prevention and control.
Patent Information
- Application Number
- CN202611351155.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-02
- Publication Date
- 2026-11-10
- Estimated Expiration
- 2046-09-02
AI Technical Summary
Existing gas outburst prediction technologies are insufficient in integrating mine spatial topology information with temporal evolution patterns, and are unable to effectively handle noise in high-frequency monitoring data, resulting in a poor balance between prediction accuracy and predictable time.
A spatiotemporal graph network-based approach is adopted. By constructing a gas fractal spatiotemporal topology graph and combining aerodynamic parameters and fractal dimension, the maximum Lyapunov exponent and correlation dimension are calculated to build a local polynomial prediction model for dynamic prediction of gas concentration.
It achieves a precise characterization of the cascade transmission effect of gas concentration, improves the adaptability and accuracy of the prediction model, extends the effective prediction time range, and provides sufficient decision-making time for mine gas disasters.
Smart Images

Figure CN122862895B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of mine gas disaster prevention and control technology, and in particular to a dynamic prediction method for mine gas outbursts based on spatiotemporal graph networks. Background Technology
[0002] Mine gas outbursts are one of the major hazards in coal mining. Accurately predicting the dynamic changes in gas outburst volume is crucial for ensuring safe coal mine production and preventing gas explosions. The gas outburst process is influenced by multiple factors, including geological structure, mining intensity, and ventilation system status, exhibiting complex nonlinear dynamic characteristics. Furthermore, gas concentrations in different regions show significant spatiotemporal correlations and evolutionary patterns, placing high demands on the adaptability and accuracy of gas prediction technologies.
[0003] Traditional gas outburst prediction technologies have evolved over a long period, resulting in several mature technical approaches. Statistical analysis-based methods, such as time series analysis and multiple regression analysis, offer simple computational processes and low engineering difficulty, providing relatively reliable prediction results under relatively stable production conditions. Physical model-based methods, starting from the fundamental mechanisms of gas migration, can essentially describe the diffusion and convection processes of gas in roadways and goafs, providing theoretical support for prediction. In recent years, with the development of artificial intelligence technology, deep learning methods such as recurrent neural networks and long short-term memory networks have been gradually introduced into the field of gas prediction. Their strong fitting ability to nonlinear time series further enriches the technical means of gas outburst prediction.
[0004] In actual mine production environments, the complex topology of ventilation networks has a significant impact on the spatial transmission of gas concentration. At the same time, the chaotic characteristics exhibited by the gas outburst process create a certain balance between prediction accuracy and predictable time.
[0005] There is still room for further exploration in fully integrating mine spatial topology information with temporal evolution patterns in some existing methods, and the adaptability to noise in high-frequency monitoring data also needs to be improved. Summary of the Invention
[0006] The purpose of this invention is to propose a dynamic prediction method for mine gas outburst based on spatiotemporal graph networks in order to solve the above-mentioned problems.
[0007] To achieve the above objectives, the present invention adopts the following technical solution: A dynamic prediction method for mine gas outbursts based on spatiotemporal graph networks includes: Acquire high-frequency multi-source monitoring data, a three-dimensional spatial model of the mine, and aerodynamic parameters; perform noise reduction processing on the high-frequency multi-source monitoring data of the mine; and construct a spatiotemporal data matrix for multi-source monitoring of the mine. The fractal dimension of the mine ventilation network is calculated based on the three-dimensional spatial model of the mine. The edge weight matrix is calculated by combining aerodynamic parameters. The multi-source monitoring spatiotemporal data matrix of the mine is mapped to the nodes to construct a gas fractal spatiotemporal topology map. For the time series in the spatiotemporal data matrix of multi-source monitoring in the mine, the delay time and embedding dimension are determined. In the gas fractal spatiotemporal topology graph, the state vectors of neighboring nodes are fused with the edge weight matrix to construct the spatiotemporal joint phase space matrix. In the spatiotemporal joint phase space matrix, the maximum Lyapunov exponent and correlation dimension are calculated, and the maximum predictable time scale is determined based on the maximum Lyapunov exponent. The prediction step size is set according to the maximum predictable time scale, the local neighborhood parameters are determined by combining the correlation dimension, the prediction center point is extracted in the spatiotemporal joint phase space matrix, and the local neighborhood set is searched according to the local neighborhood parameters to construct a local polynomial prediction model, solve the polynomial coefficient matrix, and deduce the dynamic prediction sequence of gas concentration.
[0008] Preferably, the noise reduction processing of the high-frequency multi-source monitoring data in the mine specifically includes: By selecting the embedding dimension and delay time, the noisy gas concentration time series is mapped to a high-dimensional phase space to construct a phase space point set; in the phase space, the nearest neighbors of the center point in terms of spatial distance are found to form a local neighborhood matrix, and the covariance matrix of the local neighborhood matrix is calculated to reflect the distribution variance of the data points in each dimension. Perform singular value decomposition on the covariance matrix, sort the singular values according to their size, and retain the eigenvectors corresponding to the singular values that are greater than the singular value threshold to form the projection matrix; After subtracting the local mean from the center point and its neighborhood points, the points are projected onto the signal subspace composed of dominant singular values using a projection matrix to obtain the denoised phase space points. The arithmetic mean of all projected values containing the same time point is then calculated to restore the one-dimensional time series.
[0009] Preferably, the calculation of the fractal dimension of the mine ventilation network using the box-counting dimension method based on the three-dimensional spatial model of the mine specifically includes: The three-dimensional space containing the mine ventilation network is divided into tiny cubic boxes, and the total number of boxes containing roadways or goaf pores is counted. Continuously reduce the side length of the tiny cube boxes and recount the total number of corresponding boxes; In a double logarithmic coordinate system, linear regression fitting is performed, and the slope of the fitted line is the fractal dimension, calculated using the following formula: ; in, For fractal dimension, Let be the side length of the tiny cube box. This represents the total number of boxes containing tunnels or goaf pores.
[0010] Preferably, the calculation of the edge weight matrix in conjunction with aerodynamic parameters specifically includes: The fractal equivalent length is obtained by fractal correction of the physical tunnel length using the fractal dimension. Combined with the aerodynamic friction coefficient, tunnel perimeter, and average cross-sectional area, the fractal-corrected ventilation resistance is calculated using the following formula: ; in, The ventilation resistance is after fractal correction. The coefficient of aerodynamic friction drag. The physical tunnel length, For fractal dimension, The topological dimension of a standard straight tunnel. The fractal amplification index, For the perimeter of the alley, This represents the average cross-sectional area of the tunnel. An exponential decay function containing convection and diffusion terms is constructed. Wind speed, gas diffusion coefficient, and fractal-corrected ventilation resistance are substituted into the exponential decay function to quantify the spatial decay ratio of gas concentration transfer between nodes, resulting in a boundary weight matrix normalized to the interval between zero and one.
[0011] Preferably, the process of determining the delay time includes: The numerical range of the time series is divided into equally spaced data boxes. The probability of a time series data point falling into each data box is calculated, as well as the probability of a time series data point falling into each data box after a delay time. Calculate the joint probability between the time series and the delayed time series, and construct a mutual information function to measure the degree of nonlinear correlation between the two time series; By continuously increasing the value of the delay time, we find the first local minimum point of the mutual information function as the delay time increases, and take the delay time corresponding to the first local minimum point as the optimal delay time.
[0012] Preferably, the process of determining the embedding dimension includes: Find the nearest neighbor of the state vector in the phase space and calculate the increase in distance between the state vector and the nearest neighbor after the phase space dimension is increased; The ratio of the increased component distance to the Euclidean distance in the original phase space is used to obtain the relative rate of change of distance; When the relative rate of change of distance is greater than the judgment threshold, the nearest neighbor is determined to be a false nearest neighbor caused by low-dimensional projection. Iterate through all points in the phase space to count the proportion of false nearest neighbors, gradually increase the embedding dimension, and when the proportion of false nearest neighbors drops to a set threshold for the first time, use the corresponding dimension as the optimal embedding dimension.
[0013] Preferably, the step of fusing the neighbor node state vectors with the edge weight matrix to construct the spatiotemporal joint phase space matrix specifically includes: For the target node, find the set of all neighboring nodes connected by directed edges in the gas fractal spatiotemporal topology graph; extract the phase space state vector of the neighboring nodes, perform spatial weighted average decay processing using the corresponding edge weights in the edge weight matrix, and aggregate to form a spatial perturbation term; By using the phase space vector of the target node itself as the dominant component and superimposing spatial perturbation terms, a spatiotemporal joint phase space matrix containing the cascade transmission effect of gas concentration in the local topological network is constructed.
[0014] Preferably, the process of calculating the maximum Lyapunov exponent is as follows: Find the nearest neighbor of the reference trajectory state point in the spatiotemporal joint phase space matrix, and apply a brief separation constraint greater than the average orbital period of the time series to avoid them being in the same continuous evolution segment. Calculate the Euclidean distance between the state point and its nearest neighbor after the discrete time step of evolution, and track the exponential divergence rate of adjacent trajectories over time. The arithmetic mean of the logarithmic distances of all reference points at the same evolutionary step size is calculated, and the evolution curve is plotted with evolution time as the x-axis and the mean logarithmic distance as the y-axis. The initial linearly rising region of the evolution curve is fitted with a straight line using the least squares method, and the slope of the fitted line is the maximum Lyapunov exponent.
[0015] Preferably, the process of calculating the correlation dimension is as follows: Given a spatial distance radius, the correlation integral is obtained by using the step function to count the number of point pairs in the spatiotemporal joint phase space matrix where the distance between any two state points is less than the spatial distance radius, and then dividing the number of point pairs that may exist in the phase space by the correlation integral. Calculate the corresponding correlation integrals by selecting a series of gradually increasing spatial distance radii, and plot a scatter plot of spatial distance radius versus correlation integral in a double logarithmic coordinate system; Find the scale-free region in the scatter plot that shows a power-law linear relationship, and use linear regression to fit the data points in the scale-free region. The slope of the fitted line is the correlation dimension.
[0016] Preferably, the process of constructing a local polynomial prediction model, solving for the polynomial coefficient matrix, and deriving the dynamic prediction sequence of gas concentration is as follows: The reciprocal of the largest Lyapunov exponent is used as the largest predictable time scale to set the prediction step size; the latest state vector of the target node is extracted as the prediction center point, and multiple historical state points with the smallest Euclidean distance are selected to form a local neighborhood set. Using the multidimensional Taylor expansion theorem, a local polynomial prediction model is constructed by performing a second-order truncated expansion near the prediction center point. The calculation formula is as follows: ; Construct a diagonal weight matrix with the distance decay function as the weights, and solve for the polynomial coefficient matrix using the weighted least squares method. The calculation formula is as follows: ; in, For a state point in a local neighborhood set, To evolve to a state one step ahead in the future, A vector of constant terms. It is a first-order linear coefficient matrix. It is a second-order nonlinear coefficient tensor. This is the error term; For inclusion , , The polynomial coefficient matrix, To design the matrix, It is a diagonal weight matrix. The response matrix; Substituting the state vector of the prediction center point into the local multinomial prediction model yields a single-step prediction value. The single-step prediction value is then used as the new prediction center point for multi-step iterative deduction. The first dimension component of the prediction state vector sequence is extracted to obtain the dynamic prediction sequence of gas concentration.
[0017] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: 1. This invention constructs a fractal spatiotemporal topology map of gas, fully explores the spatial topological characteristics of the mine ventilation network, uses the fractal dimension to correct the physical roadway length, and calculates the ventilation resistance after fractal correction by combining aerodynamic parameters, which can more accurately reflect the influence of complex roadway structure on gas migration; on this basis, the edge weight matrix quantifies the spatial attenuation ratio of gas concentration transmission between different monitoring nodes, and realizes an effective characterization of the cascade transmission effect of gas concentration.
[0018] 2. This invention reveals the nonlinear evolution characteristics of gas concentration time series by calculating the maximum Lyapunov exponent and correlation dimension, and rationally determines the maximum predictable time scale and local neighborhood parameters of the prediction model, providing a theoretical basis for parameter selection of the prediction model. It employs a local polynomial prediction model based on phase space reconstruction, combined with weighted least squares method to solve the polynomial coefficients, which can effectively capture the nonlinear evolution law of gas concentration and has good adaptability to noise in high-frequency monitoring data. Through multi-step iterative deduction, dynamic prediction of gas concentration is achieved, which to a certain extent extends the effective prediction time range and provides more sufficient decision-making time for early warning of mine gas disasters. Attached Figure Description
[0019] Further details, features, and advantages of this application are disclosed in the following description of exemplary embodiments in conjunction with the accompanying drawings, in which: Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation
[0020] Several embodiments of this application will now be described in more detail with reference to the accompanying drawings to enable those skilled in the art to implement this application. This application may be embodied in many different forms and for various purposes and should not be limited to the embodiments set forth herein. These embodiments are provided to make this application thorough and complete, and to fully convey the scope of this application to those skilled in the art. The embodiments described do not limit this application.
[0021] Unless otherwise defined, all terms used herein (including technical and scientific terms) shall have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. It will be further understood that terms such as those defined in commonly used dictionaries shall be interpreted as having a meaning consistent with their meaning in the relevant field and / or the context of this specification, and shall not be interpreted in an idealized or overly formal sense unless expressly defined herein.
[0022] Example 1
[0023] Its specific implementation method is combined with the appendix Figure 1 Please provide a detailed explanation.
[0024] Appendix Figure 1 The flowchart of a dynamic prediction method for mine gas outburst based on spatiotemporal graph networks provided in this embodiment of the invention illustrates the complete steps from acquiring high-frequency multi-source monitoring data of the mine to deriving the dynamic prediction sequence of gas concentration.
[0025] In this embodiment, it includes: Step 1: Acquisition of multi-source monitoring data in the mine: Acquire high-frequency multi-source monitoring data, a three-dimensional spatial model of the mine, and aerodynamic parameters. Perform phase space local projection nonlinear noise reduction processing on the high-frequency multi-source monitoring data to construct a spatiotemporal data matrix for multi-source monitoring in the mine. The core objective of this step is to acquire high-frequency, multi-dimensional time-series data that accurately reflects the nonlinear dynamics of gas outbursts from the complex physical environment of a mine, and to perform phase space noise reduction without disrupting the topology of the chaotic attractor, thus providing a high-quality data foundation for subsequent fractal network construction and chaotic inference.
[0026] Mine gas outburst is a complex nonlinear dynamic system that is influenced by multiple factors, including geological structure, coal seam occurrence conditions, ventilation system status, and mining activities.
[0027] Traditional low-frequency sampling data often misses the high-frequency, minute fluctuations in the gas outburst process, which are the specific manifestations of the "butterfly effect" (i.e., sensitive dependence on initial conditions) in chaotic dynamics.
[0028] This step begins by deploying a multi-source heterogeneous sensor network at key underground nodes in the mine (including but not limited to the upper corner of the coal face, return airway, intake airway, tunneling face, and near the sealed wall of the goaf). The types of sensors deployed include high-precision laser methane sensors, ultrasonic anemometers, high-resolution micro-differential pressure sensors, and temperature and humidity sensors.
[0029] In order to capture the chaotic evolution characteristics of gas concentration, the data acquisition frequency of the sensor must break the conventional minute-level limitation and be set to a high-frequency sampling mode (e.g., once per second or higher).
[0030] High-frequency sampling ensures that the acquired time series contains sufficiently rich dynamic information, thus enabling the complete unfolding of the fine structure of the chaotic attractor in subsequent phase space reconstruction.
[0031] Due to the vast space in the mine, network delays occur during data transmission from sensors in different locations, resulting in inconsistent timestamps.
[0032] To eliminate the impact of this time phase error on the synchronization of the spatiotemporal graph network node status, this step uses a precise time protocol to synchronize all downhole sensors at the nanosecond level, ensuring that multi-source data collected at the same physical moment are strictly aligned in the time dimension.
[0033] After acquiring the raw high-frequency multi-source monitoring data, an initial multidimensional time series matrix is formed. However, the harsh underground environment, including the start-up and shutdown of large electromechanical equipment, electromagnetic interference, and dust obstruction, inevitably introduces a large amount of broadband random noise into the raw data. If traditional linear filtering methods (such as moving average filtering, low-pass filtering, or Fourier transform filtering) are used, while filtering out noise, they will severely smooth out the inherent nonlinear abrupt changes in the gas concentration time series, destroying its inherent chaotic dynamic mechanism, leading to severe distortion of the subsequently calculated chaotic characteristic exponents.
[0034] To address this issue, this step employs a nonlinear noise reduction technique based on local phase space projection.
[0035] The core of this technology is that chaotic signals are confined to a low-dimensional deterministic attractor manifold in the reconstructed high-dimensional phase space, while random noise is uniformly distributed throughout the high-dimensional phase space.
[0036] By projecting the noisy signal onto the local subspace where the attractor manifold is located, noise can be effectively separated and eliminated, while the nonlinear dynamic characteristics of gas outburst can be fully preserved.
[0037] The specific operation process is as follows: Suppose that the time series of noisy gas concentration collected at a certain node is as follows: ,in , This represents the total length of the data.
[0038] Choose an appropriate embedding dimension and delay time One-dimensional time series Mapped to Construct a set of phase space points in 3D phase space. ,in .
[0039] In the reconstructed phase space, for each center point To find its spatial distance The nearest neighbors form a local neighborhood matrix. .
[0040] The dimension of this local neighborhood matrix is Then, the covariance matrix of this local neighborhood matrix is calculated. .
[0041] covariance matrix It reflects the variance of the distribution of data points in various dimensions within a local neighborhood.
[0042] For covariance matrix Perform singular value decomposition. Sort the singular values according to their magnitude. Larger singular values correspond to the dominant direction of the chaotic attractor manifold (i.e., the signal subspace), while smaller singular values correspond to the secondary direction of random noise (i.e., the noise subspace).
[0043] Set a singular value threshold, and retain the eigenvectors corresponding to singular values greater than the threshold to form a projection matrix. .
[0044] center point The points within the projection matrix are projected onto the points in the neighborhood of the projection matrix. In Zhang Cheng's signal subspace, the denoised phase space points are obtained. ; Right now: ; in, It is a local mean. Projection matrix The transpose of .
[0045] This local projection operation forcibly pulls noisy data points that have deviated from the attractor manifold back onto the deterministic manifold surface.
[0046] After completing the local projection of all phase space points, it is necessary to restore the denoised high-dimensional phase space point set into a one-dimensional time series. Since data from the same time point may appear in multiple different phase space vectors during phase space reconstruction, the arithmetic mean of all projected values containing the same time point is calculated during restoration to obtain the final denoised gas concentration time series. .
[0047] In addition to gas concentration data, multi-source data such as wind speed, air pressure, temperature, and humidity are also simultaneously subjected to the aforementioned phase space local projection noise reduction processing.
[0048] After processing, to address potential data loss issues (such as temporary sensor disconnection), this step employs a nonlinear interpolation method based on the neighborhood evolution of chaotic attractors. This involves finding the nearest neighbor trajectory of the preceding trajectory of the missing point in the phase space and using the subsequent evolution trend of the nearest neighbor trajectory to fill in the missing value, thereby ensuring the continuity of the time series in a dynamic sense.
[0049] The multi-source time series data after noise reduction and completion of all nodes are normalized to eliminate the magnitude difference between different physical dimensions, and a high-quality spatiotemporal data matrix of multi-source monitoring in the mine is constructed as input for subsequent steps.
[0050] Step 2, Gas fractal spatiotemporal topology construction steps: Based on the three-dimensional spatial model of the mine, the fractal dimension of the mine ventilation network is calculated using the box-counting dimension method. The edge weight matrix is calculated in combination with aerodynamic parameters. The multi-source monitoring spatiotemporal data matrix of the mine is mapped to the nodes to construct the gas fractal spatiotemporal topology. The core objective of this step is to abstract and map the complex physical ventilation network and gas transport channels of the mine into a mathematical graph structure model with fractal geometric characteristics, and to embed the multi-source time series data obtained in step one into this graph structure to form a gas fractal spatiotemporal topology map that can simultaneously express spatial cascade effects and temporal evolution laws.
[0051] The mine ventilation network is not a simple Euclidean geometric space; it is composed of main roadways, mining area roadways, working faces, and a complex network of ventilation channels in the goaf.
[0052] This network structure exhibits significant self-similarity at different scales; that is, the branching structure of local roadways and the overall mine ventilation network structure share a similar level of morphological complexity. This self-similarity is a typical characteristic of fractal geometry. Traditional graph network construction methods typically abstract roadways simply as nodes and edges, with edge weights considering only physical straight-line distances. This severely neglects the nonlinear hindering effects of roadway surface roughness, abrupt changes in cross-section, and the porous media characteristics of the goaf on gas diffusion and convection.
[0053] To accurately quantify this complex spatial topological relationship, this step introduces fractal dimension to correct the edge weights of the graph network, constructing a gas fractal spatiotemporal topological graph. ; Define a set of nodes The key intersections, air doors, regulating windows, coal mining faces, tunneling faces, and sensor deployment locations in the mine ventilation network are defined as nodes in the graph network. Each node Having fixed three-dimensional spatial coordinates and node type attributes; Define edge set If two nodes and If there is a direct physical tunnel connection or an obvious gas / fluid transport channel between the two points, then a directed edge is established between the two points. The direction of the edge is determined by the dominant airflow direction of the mine ventilation system. For areas with light or no wind where wind flow may reverse due to fluctuations, establish a two-way boundary.
[0054] Calculate the edge weight matrix that integrates fractal features and aerodynamic parameters. Edge weight Used to characterize nodes Gas concentration fluctuations at the location are transmitted to the node. The difficulty level and spatial attenuation ratio of the location.
[0055] To calculate the fractal dimension, this step uses the box-counting dimension method to quantify the three-dimensional ventilation network diagram of the mine.
[0056] The three-dimensional space encompassing the entire mine ventilation network is divided into sections with sides of length [missing information]. Tiny cube-shaped boxes; The total number of boxes containing pores in roadways or goaf areas is recorded as follows: ; Continuously reduce the side length of the box And recalculate the corresponding statistics. ; Fractal dimension of mine ventilation network Calculated using the following limit formula: ; In actual calculations, due to It is impossible for it to infinitely approach zero, therefore, by selecting a series of discrete... Values in a log-log coordinate system and Linear regression fitting is performed, and the slope of the fitted line is the fractal dimension of the region. .
[0057] fractal dimension The larger the value, the denser the tunnel network, the more complex the spatial structure, and the more tortuous the gas transport path in the area.
[0058] After obtaining the fractal dimension, edge weights are constructed by combining the ventilation resistance law from fluid mechanics. The computational model.
[0059] The movement of gas in roadways is mainly controlled by both airflow convection and concentration gradient diffusion.
[0060] Define nodes arrive The physical tunnel length is The average cross-sectional area of the tunnel is The perimeter of the alley is The aerodynamic friction drag coefficient is .
[0061] The traditional formula for calculating wind resistance is: ; To demonstrate the nonlinear influence of complex topology on gas transport, this step utilizes fractal dimension. For physical length Fractal correction is performed to obtain the fractal equivalent length. ,in This is the topological dimension of a standard straight tunnel (usually taken as 1). It is the fractal amplification index, used to adjust the nonlinear amplification effect of fractal features on length.
[0062] Calculate the fractal-corrected ventilation resistance based on the fractal equivalent length. ; The greater the ventilation resistance, the greater the loss of kinetic energy in the airflow carrying gas.
[0063] Considering the diffusion coefficient of the gas itself and wind speed between nodes .
[0064] Based on the above physical and geometric parameters, edge weights are defined. The mathematical expression for .
[0065] Edge weights are positively correlated with wind speed (higher wind speeds result in faster transmission), negatively correlated with the fractal-corrected ventilation resistance, and negatively correlated with the spatial distance between nodes. Quantification is achieved by constructing an exponential decay function that includes convection and diffusion terms. This normalizes the weight values to the (0,1) interval. The closer the weight value is to 1, the tighter the gas dynamic coupling between the two nodes; the closer the weight value is to 0, the more independent the gas states between the two nodes.
[0066] Define the node feature matrix The multi-source time series data (gas concentration, wind speed, air pressure, temperature and humidity, etc.) after phase space local projection noise reduction in step one are mapped to the corresponding nodes of the graph network according to the time step and node index.
[0067] At any time , graph network The state can be represented by a feature matrix. This means that the rows of the matrix correspond to the nodes in the network, and the columns correspond to the time interval of that node. Multidimensional monitoring data values.
[0068] Thus, the mine ventilation network in physical space is rigorously transformed into a spatial fractal topology (composed of nodes). ,side and fractal weights (Composition) and time-dynamic evolution characteristics (composed of feature matrix) The gas fractal spacetime topology diagram (consisting of).
[0069] This diagram not only preserves the physical connections between tunnels, but more importantly, through the deep integration of fractal dimension and aerodynamic parameters, it accurately describes the physical obstruction mechanism of nonlinear diffusion and convection of gas in complex confined spaces at the mathematical level, providing a mathematical framework that conforms to physical reality for subsequent chaotic dynamics deductions.
[0070] Step 3, Gas Outburst Phase Space Reconstruction and Deduction Steps: For the time series in the multi-source monitoring spatiotemporal data matrix of the mine, the mutual information method and the false nearest neighbor method are used to determine the delay time and embedding dimension. In the gas fractal spatiotemporal topology graph, the edge weight matrix is combined with the neighbor node state vector to construct the spatiotemporal joint phase space matrix. The core objective of this step is to expand the one-dimensional time series data in the gas fractal spatiotemporal topology map constructed in step two into a high-dimensional phase space through topologically equivalent mathematical mapping.
[0071] Since mine gas outburst is a nonlinear dynamic system affected by multivariate coupling, the one-dimensional time series (such as gas concentration) collected by a single sensor actually contains the historical traces of the evolution of all other related variables in the system (such as wind speed, air pressure, geological stress, etc.).
[0072] According to Takens' embedding theorem in nonlinear dynamics, by choosing appropriate delay time and embedding dimension, a high-dimensional phase space attractor structure that is completely equivalent to the dynamic characteristics of the original system can be recovered from a one-dimensional time series.
[0073] Furthermore, this step breaks away from the traditional reconstruction method of isolated nodes, introducing the spatial weight matrix in the graph network into the phase space reconstruction process to achieve spatiotemporal joint deduction.
[0074] For any target node in a fractal spacetime topology graph Extract the time series of gas concentration after noise reduction. ,in .
[0075] To perform phase space reconstruction, two key dynamic parameters must be calculated precisely: delay time. and embedding dimension .
[0076] Delay time Calculation: Traditional autocorrelation function methods can only extract linear correlations in time series and cannot handle highly nonlinear signals such as gas outbursts.
[0077] This step uses mutual information to determine the delay time. The mutual information method, based on the Shannon entropy concept in information theory, can comprehensively measure the degree of linear and nonlinear correlation between two time series.
[0078] Let time series and the process Delayed time series First, find the maximum and minimum values in the sequence, and divide the numerical range into several equally spaced smaller intervals (i.e., data "boxes").
[0079] Statistical series The data points in the middle fall into the first The probability of a box is denoted as . ; Statistical sequence The data points in the middle fall into the first The probability of a box is denoted as . Simultaneously, statistics Falling into A box and Falling into The joint probability of the boxes is denoted as . .
[0080] According to information theory, the mutual information function between these two sequences Defined as: ; Mutual Information Reflects the known Under the premise of being able to obtain information about Information content; when When the numbers are small, the two sequences are highly correlated. When the value is large, the reconstructed phase space trajectory will be squeezed near the diagonal and cannot be fully expanded; when When the size is too large, the two sequences become almost independent, and the reconstructed trajectories become chaotic and lose their original dynamic correlation.
[0081] Therefore, this step involves continuously increasing... Calculate the corresponding value. And search for mutual information functions. Follow The first local minimum point that appears due to the increase. This local minimum point corresponds to... The value ensures that the two sequences are sufficiently independent in a nonlinear sense while preserving the necessary dynamic connection, and is used as the target node. Optimal delay time .
[0082] Embedding dimension Calculation: After determining the delay time Next, the embedding dimension needs to be determined. ;if If it's too small, chaotic attractors in high-dimensional phase space will be squeezed and projected into low-dimensional space, causing trajectories that were originally far apart in space to overlap and intersect, producing "false nearest neighbors"; if... If the value is too large, it will not only increase the computational load exponentially, but also amplify the residual noise and destroy the geometry of the attractor.
[0083] This step uses the spurious nearest neighbor method to determine the optimal embedding dimension. Its core logic is: With the embedding dimension As the projection increases, the false neighboring points caused by the projection will gradually be separated.
[0084] for Any state vector in the phase space Find its nearest neighbor in the Euclidean distance sense. The distance between the two is denoted as .
[0085] When the dimension of phase space increases to When the dimension is increased, these two points will have an additional component in the new dimension, at which point the distance between them will become... .
[0086] Calculate the relative rate of change of distance: ; Set a judgment threshold (The value is usually between 10 and 15).
[0087] if This means that these two points are... Adjacency in 3D space is an artifact caused by low-dimensional projection, and the determination of adjacency is based on this. for The false nearest neighbor.
[0088] Iterate through all points in the phase space and count the proportion of false nearest neighbors.
[0089] Gradually increase The value is determined when the proportion of false nearest neighbors first drops to a very small set threshold (e.g., less than 5%) or no longer increases. When the increase leads to a significant decrease, at this time... The value is the optimal embedding dimension that can fully unfold the chaotic attractor. .
[0090] After obtaining a single node and Then, this step involves innovative operations of spatiotemporal joint simulation; Traditional phase space state vectors consist only of the node's own historical data, i.e. .
[0091] However, in mine ventilation networks, nodes The gas concentration is necessarily affected by its upstream or adjacent nodes. The impact of gas migration.
[0092] This step calls the edge weight matrix of the gas fractal spacetime topology graph constructed in step two. .
[0093] For the target node Find the set of all neighboring nodes that are connected to it by directed edges. Utilizing edge weights (represents a node) For nodes The influence weights are used to weight and fuse the phase space state vectors of neighboring nodes to construct a spatiotemporal joint phase space state vector. ; The specific construction method is as follows: [The node is...] Its own phase space vector As the dominant component, it is also subject to weighting. The decayed state vector of the neighboring node.
[0094] To maintain dimensional uniformity, a spatial weighted average is used to aggregate the influence of neighboring nodes, forming a spatial perturbation term, which is then superimposed on the phase space vector of the target node.
[0095] In this way, the reconstructed spatiotemporal joint phase space matrix not only represents the dynamic state of the target node at a specific moment in each row, but also implicitly includes the gas concentration cascade transmission effect of the entire local topological network under the fractal space structure. This lays a solid mathematical foundation for accurately capturing the spatiotemporal chaotic characteristics of gas outbursts.
[0096] Step 4, Calculation steps of chaotic dynamic characteristic index: In the spatiotemporal joint phase space matrix, calculate the maximum Lyapunov exponent and correlation dimension, and determine the maximum predictable time scale based on the maximum Lyapunov exponent; The core objective of this step is to extract characteristic indices that can quantitatively characterize the chaotic properties of the gas emission system within the spatiotemporal joint phase space constructed in step three, through rigorous mathematical calculations. These indices are key criteria for determining the predictability of the system and for determining the prediction timescale. This step mainly calculates two core indicators: the maximum Lyapunov exponent and the correlation dimension.
[0097] Lyapunov exponents are used to quantify the average divergence or convergence rate of adjacent trajectories in phase space over time.
[0098] If a dynamical system has at least one Lyapunov exponent greater than zero, it indicates that the system is extremely sensitive to initial conditions, i.e., the "butterfly effect" exists, which is the most prominent feature of chaotic systems.
[0099] To accurately compute LLE in a finite-length time series containing a small amount of residual noise, this step employs the small data method (Rosenstein algorithm); in the spatiotemporal joint phase space, let a state point on the reference trajectory be... .
[0100] Searching in phase space nearest neighbor To avoid the nearest neighbor and the reference point being on the same continuous evolution segment of the same time trajectory (which would invalidate the calculated divergence rate), a transient separation constraint must be imposed, i.e., requiring... ,in The average orbital period of the time series can be estimated by performing a Fast Fourier Transform (FFT) on the time series to obtain the dominant frequency.
[0101] Find the nearest neighbor that meets the conditions. Then, the evolution of these two initially very close points in phase space is calculated. The Euclidean distance after each discrete time step is denoted as . .
[0102] Its mathematical expression is: ; According to chaotic dynamics theory, if the system is chaotic, the distance between adjacent trajectories will diverge exponentially with time, i.e. ,in The maximum Lyapunov index, This represents the sampling time interval.
[0103] Taking the natural logarithm of both sides of the above equation, we get To eliminate errors caused by local fluctuations, all reference points in phase space are... Perform the above calculations, and apply the same evolutionary step size. Below Find the arithmetic mean, denoted as . .
[0104] In terms of evolution time The x-axis represents the average logarithmic distance. Plot the evolution curve with the vertical axis as the ordinate.
[0105] In the initial stage of the curve, there is usually a distinct linear ascending region. Using the least squares method to fit a straight line to this region, the slope of the fitted line is the maximum Lyapunov exponent of the target node under spatiotemporal coupling. .
[0106] If the calculation yields This mathematically proves that the gas outburst process of the node and its local topological region has chaotic characteristics.
[0107] At the same time, the reciprocal of the maximum Lyapunov index The maximum predictable timescale of the system (i.e., the Lyapunov time) is defined. Beyond this timescale, the prediction error will amplify exponentially, leading to prediction failure.
[0108] This metric provides a theoretical upper limit for setting the prediction step size in subsequent steps.
[0109] The correlation dimension is a fractal dimension that measures the geometric complexity of a chaotic attractor in phase space. It reflects the minimum number of independent variables required to describe the nonlinear dynamic system.
[0110] This step uses the GP algorithm for calculation; in the spatiotemporal joint phase space, a spatial distance radius is given. For any state point in phase space Statistical analysis of all other points in phase space arrive The distance is less than the radius The number of points. Add the statistical results of all points and divide by the total number of possible point pairs in the phase space to obtain the correlation integral. .
[0111] Related Integrals The physical meaning is: In phase space, the distance between any two state points is less than [a certain value]. The probability of this. Its mathematical formula includes a step function, which occurs when the distance is less than... When the step function is active, its value is 1; otherwise, it is 0.
[0112] According to fractal theory, for a chaotic attractor, when the radius... When varying within a certain range, the correlation integral With radius There exists a power-law relationship, that is... ,in This is the correlation dimension.
[0113] Taking the natural logarithm of both sides of this power-law relation, we get... In practice, a series of gradually increasing radii are selected. Value, calculate the corresponding correlation integral In a double logarithmic coordinate system and Draw a scatter plot, observe the scatter plot, and find an interval that shows a clear linear relationship. This interval is called the "scale-free region".
[0114] Linear regression is used to fit data points within the scale-free region; the slope of the fitted line is the correlation dimension. .
[0115] If the calculated correlation dimension If it is a non-integer (i.e., fractional dimension), it further confirms that the gas emission phase space attractor has fractal geometric characteristics.
[0116] Correlation dimension The size of the variable directly indicates the minimum dimension of the input variables required when building a nonlinear prediction model, thus avoiding redundancy or missing information in the model structure.
[0117] Through rigorous calculations in this step, not only were the spatiotemporal chaotic characteristics of the mine gas emission system quantitatively verified, but key dynamic parameters such as the maximum predictable time scale and system degrees of freedom were also obtained, providing mathematical constraints and parameter guidance for the final nonlinear dynamic prediction step.
[0118] Step 5, Nonlinear Dynamic Prediction of Gas Outburst: Set the prediction step size according to the maximum predictable time scale, determine the local neighborhood parameters by combining the correlation dimension, extract the prediction center point in the spatiotemporal joint phase space matrix and search the local neighborhood set according to the local neighborhood parameters, construct a local polynomial prediction model, solve the polynomial coefficient matrix, and obtain the dynamic prediction sequence of gas concentration through multi-step iterative deduction. The core objective of this step is to use the spatiotemporal joint phase space constructed in step three and the chaotic dynamic characteristic index extracted in step four to perform a high-precision dynamic prediction of the future evolution trajectory of mine gas outbursts through a purely nonlinear dynamic local approximation mathematical method, without relying on any black-box data-driven model (such as neural networks or deep learning).
[0119] Because chaotic systems are highly sensitive to initial conditions, global linear or nonlinear equations cannot accurately describe the complex folding and stretching processes of attractors throughout the phase space. Therefore, this step employs a nonlinear prediction mechanism based on local polynomial fitting of the phase space, combined with the spatial weights of a fractal network, to achieve multi-step dynamic prediction of gas emission concentration.
[0120] Before making a prediction, the effective time boundary for the prediction must first be established; based on the maximum Lyapunov exponent calculated in step four. The system's maximum predictable timescale (Lyapunov time) is defined as .
[0121] Any prediction that exceeds this timescale will have its error magnified exponentially and lose its physical meaning; Set prediction step size Must meet ,in This represents the data sampling time interval.
[0122] It is necessary to determine the number of nearest neighbors required for local polynomial fitting. To ensure that the polynomial coefficient solution of the system of equations has a unique solution and a certain degree of noise robustness, The value of must be strictly greater than the total number of coefficients to be determined in the polynomial model.
[0123] Let the embedding dimension of the spatiotemporal joint phase space be... If a second-order local polynomial is used for fitting, the total number of coefficients to be determined is: Therefore, the number of nearest neighbors is set. ,in This is a redundancy coefficient, typically ranging from 1.5 to 3.0, to ensure the numerical stability of the least squares solution process.
[0124] Let the current time be Extracting target nodes in the spatiotemporal joint phase space. The latest state vector is denoted as the prediction center point. .
[0125] This vector integrates the node's own historical delay coordinates with the spatially weighted states of adjacent nodes in the fractal network.
[0126] In the historical phase space trajectory (i.e., time) Calculate all historical state points from the set of all state points. With the predicted center point Euclidean distance between .
[0127] Arrange these distances in ascending order and select the ones with the smallest distances. These historical state points constitute the prediction center point. local neighborhood set ; this The nearest neighbor points are extremely close to the current state in phase space. According to the deterministic principle of chaotic dynamics, this... The next evolutionary direction of each point in its historical trajectory is highly similar to its current state. In the future direction of evolution.
[0128] In local neighborhood set Inside, it is assumed that the evolution of the phase space state follows a smooth nonlinear mapping function. .
[0129] Using the multidimensional Taylor expansion theorem, the nonlinear mapping function is expanded by second-order truncation near the prediction center point to construct a local polynomial prediction model.
[0130] For any state point in the neighborhood set Its state after evolving one step into the future. It can be represented as: ; in, A vector of constant terms; It is a first-order linear coefficient matrix (i.e., Jacobian matrix), which reflects the linear stretching and rotation of the local phase space; It is a second-order nonlinear coefficient tensor (i.e., Hessian matrix), which reflects the nonlinear folding and curvature of the local phase space; This is a combined term of higher-order truncation error and residual noise.
[0131] To solve for the set of unknown coefficients in the above model This step uses the weighted least squares method; The principle for weight allocation is: Distance from the predicted center point The closer the neighbors, the greater the reference value of their evolutionary trajectories for future predictions, and therefore they should be assigned higher weights. Define the distance weighting function. ,in for The minimum distance among the nearest neighbors. Construct a diagonal weight matrix from the weights of all neighbor points. .
[0132] Local neighborhood set All The state vectors of each point and their corresponding polynomial combination terms are arranged in rows to construct the design matrix. (its dimensions are) ). This The true state vectors of each point after one step of evolution in the historical trajectory are arranged in rows to construct the response matrix. (its dimensions are) Based on the principle of weighted least squares, the polynomial coefficient matrix is solved by minimizing the sum of squared weighted errors. .
[0133] Its matrix solution equation is: ; In the above solution process, to prevent the design matrix... Multicollinearity can lead to the failure of matrix inversion (i.e., the matrix is singular or ill-conditioned). Singular value decomposition (SVD) or Tikhonov regularization can be introduced to numerically stabilize the matrix inversion process.
[0134] Obtain the local polynomial coefficient matrix Then, the predicted center point state vector at the current time is... By substituting the values into the solved local polynomial model, the predicted phase space state vector of the target node at the next time step can be calculated. .
[0135] The above process achieves single-step prediction. In order to perform multi-step dynamic prediction, this step adopts an iterative inference strategy.
[0136] The single-step prediction obtained As the new prediction center point, the processes of "local neighborhood search", "weight matrix update", "polynomial coefficient solution" and "state substitution calculation" are repeated to obtain the predicted value in the second step. .
[0137] This process continues iteratively until the prediction step size reaches the maximum predictable time scale defined by the maximum Lyapunov exponent. .
[0138] After completing the multi-step prediction in phase space, it is necessary to restore the predicted trajectory in high-dimensional phase space to the gas concentration time series in physical space.
[0139] Due to the phase space state vector The first component corresponds to the gas concentration value of the target node at the current moment. Therefore, the predicted state vector sequence is directly extracted. The first dimension component in the matrix can be used to obtain the future value of the target node. A dynamic prediction sequence of gas concentration within a time step.
[0140] Combining the gas fractal spatiotemporal topology map from step two, the predicted sequences of all nodes are synchronously mapped back into the three-dimensional physical space model of the mine, forming a global dynamic picture of the spatiotemporal evolution of gas outburst. This picture not only provides specific values of future gas concentrations for each node, but also reveals the spatial cascading and spreading trend of gas anomalies in complex roadway topology through purely mathematical nonlinear dynamic deduction.
[0141] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.
[0142] It should be noted that, in this document, the use of relational terms such as "first" and "second" is merely for distinguishing one entity or operation from another, and does not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.
[0143] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.
[0144] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.
[0145] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.
Claims
1. A dynamic prediction method for mine gas outbursts based on spatiotemporal graph networks, characterized in that, include: Acquire high-frequency multi-source monitoring data, a three-dimensional spatial model of the mine, and aerodynamic parameters; perform noise reduction processing on the high-frequency multi-source monitoring data of the mine; and construct a spatiotemporal data matrix for multi-source monitoring of the mine. The fractal dimension of the mine ventilation network is calculated based on the three-dimensional spatial model of the mine. The edge weight matrix is calculated by combining aerodynamic parameters. The multi-source monitoring spatiotemporal data matrix of the mine is mapped to the nodes to construct a gas fractal spatiotemporal topology map. The calculation of the edge weight matrix in conjunction with aerodynamic parameters specifically includes: The fractal equivalent length is obtained by fractal correction of the physical tunnel length using the fractal dimension. Combined with the aerodynamic friction coefficient, tunnel perimeter, and average cross-sectional area, the fractal-corrected ventilation resistance is calculated using the following formula: ; in, The ventilation resistance is after fractal correction. The coefficient of aerodynamic friction drag. The physical tunnel length, For fractal dimension, The topological dimension of a standard straight tunnel. The fractal amplification index, For the perimeter of the alley, This represents the average cross-sectional area of the tunnel. An exponential decay function containing convection and diffusion terms is constructed. Wind speed, gas diffusion coefficient, and fractal-corrected ventilation resistance are substituted into the exponential decay function to quantify the spatial decay ratio of gas concentration transfer between nodes, and the edge weight matrix is obtained by normalizing it to the interval between zero and one. For the time series in the spatiotemporal data matrix of multi-source monitoring in the mine, the delay time and embedding dimension are determined. In the gas fractal spatiotemporal topology graph, the state vectors of neighboring nodes are fused with the edge weight matrix to construct the spatiotemporal joint phase space matrix. In the spatiotemporal joint phase space matrix, the maximum Lyapunov exponent and correlation dimension are calculated, and the maximum predictable time scale is determined based on the maximum Lyapunov exponent. The prediction step size is set according to the maximum predictable time scale, the local neighborhood parameters are determined by combining the correlation dimension, the prediction center point is extracted in the spatiotemporal joint phase space matrix, and the local neighborhood set is searched according to the local neighborhood parameters. The process of constructing a local polynomial prediction model, solving for the polynomial coefficient matrix, and deriving the dynamic prediction sequence of gas concentration is as follows: The reciprocal of the largest Lyapunov exponent is used as the largest predictable time scale to set the prediction step size; the latest state vector of the target node is extracted as the prediction center point, and multiple historical state points with the smallest Euclidean distance are selected to form a local neighborhood set. Using the multidimensional Taylor expansion theorem, a local polynomial prediction model is constructed by performing a second-order truncated expansion near the prediction center point. The calculation formula is as follows: ; Construct a diagonal weight matrix with the distance decay function as the weights, and solve for the polynomial coefficient matrix using the weighted least squares method. The calculation formula is as follows: ; in, For a state point in a local neighborhood set, To evolve to a state one step ahead in the future, A vector of constant terms. It is a first-order linear coefficient matrix. It is a second-order nonlinear coefficient tensor. This is the error term; For inclusion , , The polynomial coefficient matrix, To design the matrix, It is a diagonal weight matrix. The response matrix; Substituting the state vector of the prediction center point into the local multinomial prediction model yields a single-step prediction value. The single-step prediction value is then used as the new prediction center point for multi-step iterative deduction. The first dimension component of the prediction state vector sequence is extracted to obtain the dynamic prediction sequence of gas concentration.
2. The method for dynamic prediction of mine gas outburst based on spatiotemporal graph networks according to claim 1, characterized in that, Noise reduction processing for high-frequency, multi-source monitoring data in mines specifically includes: By selecting the embedding dimension and delay time, the noisy gas concentration time series is mapped to a high-dimensional phase space to construct a phase space point set; in the phase space, the nearest neighbors of the center point in terms of spatial distance are found to form a local neighborhood matrix, and the covariance matrix of the local neighborhood matrix is calculated to reflect the distribution variance of the data points in each dimension. Perform singular value decomposition on the covariance matrix, sort the singular values according to their size, and retain the eigenvectors corresponding to the singular values that are greater than the singular value threshold to form the projection matrix; After subtracting the local mean from the center point and its neighborhood points, the points are projected onto the signal subspace composed of dominant singular values using a projection matrix to obtain the denoised phase space points. The arithmetic mean of all projected values containing the same time point is then calculated to restore the one-dimensional time series.
3. The method for dynamic prediction of mine gas outbursts based on spatiotemporal graph networks according to claim 1, characterized in that, The fractal dimension of the mine ventilation network is calculated using the box-counting dimension method based on a three-dimensional spatial model of the mine. Specifically, this includes: The three-dimensional space containing the mine ventilation network is divided into tiny cubic boxes, and the total number of boxes containing roadways or goaf pores is counted. Continuously reduce the side length of the tiny cube boxes and recount the total number of corresponding boxes; In a double logarithmic coordinate system, linear regression fitting is performed, and the slope of the fitted line is the fractal dimension, calculated using the following formula: ; in, For fractal dimension, Let be the side length of the tiny cube box. This represents the total number of boxes containing tunnels or goaf pores.
4. The method for dynamic prediction of mine gas outburst based on spatiotemporal graph networks according to claim 1, characterized in that, The process of determining the delay time includes: The numerical range of the time series is divided into equally spaced data boxes. The probability of a time series data point falling into each data box is calculated, as well as the probability of a time series data point falling into each data box after a delay time. Calculate the joint probability between the time series and the delayed time series, and construct a mutual information function to measure the degree of nonlinear correlation between the two time series; By continuously increasing the value of the delay time, we find the first local minimum point of the mutual information function as the delay time increases, and take the delay time corresponding to the first local minimum point as the optimal delay time.
5. The method for dynamic prediction of mine gas outburst based on spatiotemporal graph networks according to claim 1, characterized in that, The process of determining the embedding dimension includes: Find the nearest neighbor of the state vector in the phase space and calculate the increase in distance between the state vector and the nearest neighbor after the phase space dimension is increased; The ratio of the increased component distance to the Euclidean distance in the original phase space is used to obtain the relative rate of change of distance; When the relative rate of change of distance is greater than the judgment threshold, the nearest neighbor is determined to be a false nearest neighbor caused by low-dimensional projection. Iterate through all points in the phase space to count the proportion of false nearest neighbors, gradually increase the embedding dimension, and when the proportion of false nearest neighbors drops to a set threshold for the first time, use the corresponding dimension as the optimal embedding dimension.
6. The method for dynamic prediction of mine gas outburst based on spatiotemporal graph networks according to claim 1, characterized in that, By combining the edge weight matrix with the neighbor node state vectors, a spatiotemporal joint phase space matrix is constructed, which specifically includes: For the target node, find the set of all neighboring nodes connected by directed edges in the gas fractal spatiotemporal topology graph; extract the phase space state vector of the neighboring nodes, perform spatial weighted average decay processing using the corresponding edge weights in the edge weight matrix, and aggregate to form a spatial perturbation term; By using the phase space vector of the target node itself as the dominant component and superimposing spatial perturbation terms, a spatiotemporal joint phase space matrix containing the cascade transmission effect of gas concentration in the local topological network is constructed.
7. The method for dynamic prediction of mine gas outburst based on spatiotemporal graph networks according to claim 1, characterized in that, The process of calculating the maximum Lyapunov exponent is as follows: Find the nearest neighbor of the reference trajectory state point in the spatiotemporal joint phase space matrix, and apply a brief separation constraint greater than the average orbital period of the time series to avoid them being in the same continuous evolution segment. Calculate the Euclidean distance between the state point and its nearest neighbor after the discrete time step of evolution, and track the exponential divergence rate of adjacent trajectories over time. The arithmetic mean of the logarithmic distances of all reference points at the same evolutionary step size is calculated, and the evolution curve is plotted with evolution time as the x-axis and the mean logarithmic distance as the y-axis. The initial linearly rising region of the evolution curve is fitted with a straight line using the least squares method, and the slope of the fitted line is the maximum Lyapunov exponent.
8. The method for dynamic prediction of mine gas outburst based on spatiotemporal graph networks according to claim 1, characterized in that, The process of calculating the correlation dimension is as follows: Given a spatial distance radius, the correlation integral is obtained by using the step function to count the number of point pairs in the spatiotemporal joint phase space matrix where the distance between any two state points is less than the spatial distance radius, and then dividing the number of point pairs that may exist in the phase space by the correlation integral. Calculate the corresponding correlation integrals by selecting a series of gradually increasing spatial distance radii, and plot a scatter plot of spatial distance radius versus correlation integral in a double logarithmic coordinate system; Find the scale-free region in the scatter plot that shows a power-law linear relationship, and use linear regression to fit the data points in the scale-free region. The slope of the fitted line is the correlation dimension.
Citation Information
Patent Citations
Multi-source sensing fusion method and system for state evaluation of contact network compensation device
CN121211363A
Gold mine resource potential prediction method fusing geological knowledge map and large model
CN121860802A
