Method, medium and system for establishing dynamic distribution model of gas flow field in mining space

Abstract

The invention provides an excavation space gas flow field dynamic distribution model establishment method, medium and system, and belongs to the technical field of excavation space gas flow fields, and the method comprises the steps: firstly, regularly collecting air data, including wind speed, wind direction, temperature, humidity, air pressure, gas concentration and the like, of a plurality of point locations in an excavation space; for each acquisition moment, an air dense matrix is constructed by using an interpolation method, and the air dense matrixes are combined into an air time sequence matrix along a time sequence. Carrying out normalization processing on the matrix, and carrying out multivariate mixed decomposition calculation to obtain a stable and fluctuating time sequence matrix; then, Fourier transform is carried out on the two matrixes respectively, and stable and fluctuating frequency domain and time domain features are obtained; based on the features, a multilayer tensor network structure is constructed, and a tensor decomposition method is used to extract a low-dimensional feature tensor. And finally, inputting the feature tensor into a pre-designed neural network model, and training to obtain a model of the mining space gas flow field for effectively describing the dynamic distribution features of the gas in the mining space.

Description

Technical Field

[0001] The present invention belongs to the technical field of gas flow field in mining space, and in particular, relates to a method, medium and system for establishing a dynamic distribution model of gas flow field in mining space. Background Art

[0002] At present, coal mining operations have been widely used in underground mining, surface mining, open-pit mining and other methods. In these mining operations, gas is a safety hazard that must be focused on prevention and control. Gas is a naturally occurring flammable gas in coal seams or coal-bearing strata. Its main component is methane, which is highly flammable and explosive. During the mining process, if gas cannot be effectively controlled, it is very easy to cause serious gas explosion accidents, resulting in huge casualties and property losses.

[0003] In order to achieve safe gas control, mining companies generally establish gas monitoring and early warning systems. By arranging various types of sensor equipment in the mining space, they continuously monitor gas concentration, wind speed and direction, temperature and humidity, and use a special gas early warning model to analyze the monitoring data, detect gas anomalies in time, and remind mining workers to take emergency measures. This gas early warning mechanism based on real-time monitoring and model analysis has improved the safety of mining operations to a certain extent.

[0004] However, the existing gas early warning models have some limitations: these models usually only focus on the changing trend of gas concentration, lack effective description of the dynamic distribution characteristics of gas in the mining space, and are difficult to accurately reflect the gas flow field characteristics in complex working environments. Therefore, it is urgent to develop a new gas flow field dynamic distribution model that can more accurately describe and predict the gas distribution characteristics in complex mining environments, and provide more reliable technical support for gas safety management and control. Summary of the invention

[0005] In view of this, the present invention provides a method, medium and system for establishing a dynamic distribution model of the gas flow field in the mining space, which can solve the technical problem that the prior art usually only focuses on the changing trend of the gas concentration and lacks an effective description of the dynamic distribution characteristics of the gas in the mining space.

[0006] The present invention is achieved in that:

[0007] A first aspect of the present invention provides a method for establishing a dynamic distribution model of a gas flow field in a mining space, comprising the following steps:

[0008] S10, collecting air data at multiple collection points in the mining space at preset time intervals, including wind speed, wind direction, temperature, humidity, air pressure and gas concentration;

[0009] S20, at each collection moment, using the air data of each collection point to establish an air matrix, and using an interpolation method to interpolate and supplement the space between the collection points, so as to convert the air matrix into an air dense matrix;

[0010] S30, merging the air dense matrix at each acquisition moment into an air time series matrix according to the time series, and normalizing each element of the air time series matrix to obtain an air normalized time series matrix;

[0011] S40, decomposing the air normalized time series matrix by using multivariate mixed decomposition calculation to obtain a stable time series matrix and a fluctuating time series matrix;

[0012] S50, performing fast Fourier transform on the stable time series matrix and the fluctuating time series matrix respectively, and retaining the original time domain information, to obtain a stable frequency domain feature matrix, a stable time domain feature matrix, a fluctuating frequency domain feature matrix, and a fluctuating time domain feature matrix;

[0013] S60, constructing a multi-layer tensor network structure, taking a stable frequency domain feature matrix, a stable time domain feature matrix, a fluctuation frequency domain feature matrix, and a fluctuation time domain feature matrix as input tensors; using a tensor decomposition method to reduce the dimension and extract features of the input tensor, and obtaining a feature tensor represented by a low dimension;

[0014] S70, inputting the feature tensor into a pre-designed embedded multi-branch pyramid neural network for training to obtain a dynamic distribution model of the gas flow field in the mining space.

[0015] Specifically, the step S10 specifically includes: setting a number of collection points in the mining space, deploying wind speed sensors, wind direction sensors, temperature sensors, humidity sensors, pressure sensors and gas concentration sensors, and setting a predetermined collection time interval, for example, collecting once every 10 minutes. At each collection moment, these sensors will automatically record the wind speed, wind direction, temperature, humidity, air pressure and gas concentration and other air data monitored at the current moment. After a certain period of continuous collection, the air data of multiple collection points in the mining space at different times can be obtained, laying the foundation for the establishment of the gas flow field model in the subsequent steps.

[0016] The step S20 specifically includes: firstly, constructing an air matrix using the air data of each collection point at each collection moment, and then interpolating and supplementing the air matrix using the bilinear interpolation method to obtain a denser air matrix. This can more comprehensively reflect the distribution of each measurement parameter in the mining space at that moment.

[0017] Wherein, the step S30 specifically includes: stacking the air density matrix at each acquisition moment in time series to form a three-dimensional air time series matrix. In order to eliminate the dimension and value range differences between different parameters, each element of the time series matrix needs to be normalized to obtain an air normalized time series matrix.

[0018] The step S40 specifically includes: decomposing the normalized time series matrix using a multivariate hybrid decomposition algorithm to decompose it into two relatively independent parts, one is a stable time series matrix reflecting the stable change trend, and the other is a fluctuating time series matrix describing the fluctuating change characteristics. This decomposition is conducive to more comprehensive description of the dynamic distribution characteristics of the gas flow field.

[0019] The step S50 specifically includes: performing fast Fourier transform on the stable time series matrix and the fluctuating time series matrix respectively to obtain their characteristic matrices in the frequency domain, while retaining the original time domain information. This can reflect the frequency domain characteristics of the flow field changes and retain the time domain characteristics, which helps to more comprehensively characterize the dynamic characteristics of the gas flow field.

[0020] Wherein, the step S60 specifically includes: constructing a multi-layer tensor network structure, splicing the above-extracted stable frequency domain feature matrix, stable time domain feature matrix, fluctuation frequency domain feature matrix and fluctuation time domain feature matrix into a 4th-order input tensor, and then using the Tucker decomposition algorithm to reduce the dimension and extract features of the tensor to obtain a lower-dimensional feature tensor. This feature tensor will be used as the input of the subsequent neural network model for learning and predicting the dynamic distribution of the gas flow field.

[0021] Wherein, the step S70 specifically includes: designing an embedded multi-branch pyramid neural network structure, the network structure includes: an embedded equation group module, an equation group error term branch network, an adaptive resolution module, a model branch network of four resolutions, and a fusion summary network. Wherein, the embedded equation group module calculates the basic distribution characteristics according to the physical characteristics of the gas flow field and the air data, the equation group error term branch network is used to compensate for the calculation error of the module, the adaptive resolution module can dynamically adjust the network resolution according to the input data, the model branch networks of four resolutions respectively generate gas flow field distribution results of different resolutions, and the fusion summary network integrates these results to output the final multi-resolution gas flow field dynamic distribution model. This network structure can make full use of the physical characteristics and multi-scale spatiotemporal information of the gas flow field, and greatly improve the expression ability of the model.

[0022] On the basis of the above technical solution, the method for establishing a dynamic distribution model of gas flow field in mining space of the present invention can also be improved as follows:

[0023] Among them, the embedded multi-branch pyramid neural network includes an embedded set of equations module, an equations error term branch network, an adaptive resolution module, a first resolution model branch network, a second resolution model branch network, a third resolution model branch network, a fourth resolution model branch network and a fusion summary network.

[0024] Furthermore, the embedded equations module is used to calculate the basic distribution characteristics of the gas flow field according to the physical properties of the gas flow field and air data, including the gas diffusion equation, the gas velocity field equation, the gas pressure field equation and the gas distribution equation.

[0025] Among them, the gas diffusion equation takes into account the concentration gradient and convection effect, and is used to describe the diffusion process of gas in the mining space;

[0026] Among them, the gas velocity field equation is based on the Navier-Stokes equation, which is used to describe the movement characteristics of gas in the mining space;

[0027] Among them, the gas pressure field equation is based on the state equation and the principle of mass conservation, and is used to describe the distribution and changes of gas pressure in the mining space;

[0028] Among them, the gas distribution equation takes into account source and sink terms and boundary conditions, and is used to describe the overall distribution characteristics of gas in the mining space.

[0029] Specifically, the embedded equations include:

[0030] a) Gas diffusion equation:

[0031]

[0032] Where c is the gas concentration (mol / m 3 ), t is time (s), u is velocity vector (m / s), D is diffusion coefficient (m 2 / s), Q s is the source term (mol / (m 3 ·s)), Q r is the sink term (mol / (m 3 ·s)).

[0033] Parameter acquisition method:

[0034] c is obtained by measuring the gas concentration sensor;

[0035] u is obtained by measuring the wind speed sensor;

[0036] D is obtained by fitting the experimental data, D = D0·(T / T0) 1.75(p0 / p), where D0 is the diffusion coefficient under standard conditions, T is temperature, T0 is standard temperature, p is pressure, and p0 is standard pressure;

[0037] Q s and Q r It is obtained through experimental measurement or numerical simulation.

[0038] b) Gas flow rate field equation:

[0039]

[0040] Where ρ is the gas density (kg / m 3 ), u is the velocity vector (m / s), p is the pressure (Pa), μ is the dynamic viscosity (Pa·s), g is the gravitational acceleration vector (m / s 2 ), F is the external force (such as ventilation) (N / m 3 ).

[0041] Parameter acquisition method:

[0042] ρ is calculated by the gas state equation, ρ = pM / (RT), where M is the molar mass, R is the gas constant, and T is the temperature;

[0043] u is obtained by measuring the wind speed sensor;

[0044] p is obtained by measuring the pressure sensor;

[0045] μ is obtained by experimental measurement or table lookup;

[0046] g is a known constant;

[0047] F is obtained through experimental measurement or numerical simulation.

[0048] c) Gas pressure field equation:

[0049]

[0050] Where p is pressure (Pa), u is velocity vector (m / s), γ is specific heat ratio (dimensionless), k is permeability (m 2 ), φ is the porosity (dimensionless), μ is the dynamic viscosity (Pa·s), Q p is the pressure source term (Pa / s).

[0051] Parameter acquisition method:

[0052] p is obtained by measuring the pressure sensor;

[0053] u is obtained by measuring the wind speed sensor;

[0054] γ is obtained by experimental measurement or table lookup;

[0055] k and φ were determined experimentally using rock samples;

[0056] μ is obtained by experimental measurement or table lookup;

[0057] Q p It is obtained through experimental measurement or numerical simulation.

[0058] d) Gas distribution equation:

[0059]

[0060] Where c is the gas concentration (mol / m 3 ), t is time (s), u is velocity vector (m / s), d is diffusion coefficient (m 2 / s), Q s is the source term (mol / (m 3 ·s)), Q r is the sink term (mol / (m 3 ·s)), λ is the gas attenuation coefficient (1 / s), K ij is the turbulent diffusion tensor (m 2 / s).

[0061] Parameter acquisition method:

[0062] c,u,D,Q s and Q r The method of obtaining is the same as the gas diffusion equation;

[0063] λ is obtained through experimental measurement, λ=ln2 / t 1 / 2 , where t 1 / 2 is the half-life of gas concentration;

[0064] K ij Calculated by turbulence model, Among them C μ is an empirical constant, k is the turbulent kinetic energy, and ε is the turbulent dissipation rate.

[0065] These equations constitute a complete physical model that describes the dynamic distribution of the gas flow field in the mining space, covering the diffusion, flow, pressure changes and overall distribution characteristics of the gas.

[0066] Furthermore, the equation group error term branch network is used to compensate for the calculation error of the embedded equation group module, the input is the output of the embedded equation group module and actual measurement data, the output is the error compensation term, and the structure is a multi-layer perceptron network.

[0067] Furthermore, the adaptive resolution module is used to dynamically adjust the network resolution according to the characteristics of the input data. Specifically, the most suitable resolution is determined by analyzing the spatial and temporal variation characteristics of the input data. The input is a feature tensor, and the output is a resolution selection signal. The structure is an adaptive network obtained by replacing the convolution module of the ResNet network with a dilated convolution and deleting the fully connected layer.

[0068] Further, the first resolution model branch network is used to generate a gas flow field distribution of a first resolution, the input is a feature tensor and an output of an embedded equation group module, the output is a gas flow field distribution of a first resolution, and the structure is a U-Net network;

[0069] The second resolution model branch network is used to generate a second resolution gas flow field distribution. The input is the feature tensor and the down-sampling result of the first resolution model output. The output is the second resolution gas flow field distribution. The structure is a simplified U-Net network followed by a GFSR module. The GFSR module increases the second resolution to twice the first resolution.

[0070] The third resolution model branch network is used to generate a gas flow field distribution at a third resolution. The input is the feature tensor and the down-sampling result of the second resolution model output. The output is the gas flow field distribution at a third resolution. The structure is a ResNet-18 network followed by two connected GFSR modules, which increases the third resolution to 4 times the first resolution.

[0071] The fourth resolution model branch network is used to generate a fourth resolution gas flow field distribution. The input is the feature tensor and the downsampling result of the third resolution model output. The output is the fourth resolution gas flow field distribution. The structure is a DenseNet-121 network followed by three connected GFSR modules, which increases the fourth resolution to 8 times the first resolution.

[0072] Furthermore, the fusion summary network is used to integrate the gas flow field distribution results of different resolutions. The input is the gas flow field distribution results of four different resolutions, and the output is the final multi-resolution gas flow field dynamic distribution model. The structure is an attention mechanism fusion network.

[0073] Furthermore, the pyramid structure of the embedded multi-branch pyramid neural network specifically reduces the spatial resolution from the first resolution to the fourth resolution to expand the receptive field, and then gradually improves the spatial resolution through the GFSR algorithm while maintaining information transmission and fusion between different resolutions.

[0074] Among them, each higher-resolution network receives the output from the low-resolution network, performs super-resolution processing and feature fusion through GFSR, thereby achieving effective integration of multi-scale information.

[0075] The first resolution is specifically the original grid resolution of the mining space;

[0076] The second resolution is specifically twice the first resolution;

[0077] The third resolution is specifically 4 times the first resolution;

[0078] The fourth resolution is specifically 8 times the first resolution;

[0079] This improved structural design can not only capture the changing characteristics of the gas flow field at different spatial scales, but also reconstruct high-resolution detail information through the GFSR algorithm, greatly improving the model's ability to express gas distribution in complex mining spaces. The physical constraint loss function of the GFSR algorithm ensures that the physical characteristics of the gas flow field are maintained during the super-resolution process, so that the model output has both high resolution and conforms to the actual gas flow law.

[0080] Among them, GFSR stands for Gas Flow Super-Resolution module, and its structure includes:

[0081] a) Generator Network:

[0082] The generator network uses residual dense block (RDB) as the basic unit, and the specific structure is as follows:

[0083]

[0084] Among them, F RDB (x) is the output of RDB, x is the input, is the residual learning function, defined as:

[0085]

[0086] In the formula, the generator network contains 5 convolutional layers, W1~W5 are the convolutional layer weights of the generator network, and σ is the activation function (such as ReLU).

[0087] b) Discriminator network:

[0088] The discriminator adopts the PatchGAN structure and outputs an NxN matrix, where each element represents the authenticity score of the corresponding patch:

[0089] D(x)=σ′(W′ n ·f n-1 (W′ n-1 ·...f1(W′1·x′)));

[0090] Where D(x) is the discriminator output, which represents the authenticity score of the input x′, x′ is the input of the discriminator network, and W′ i is the convolutional layer weight of the discriminator network, f is the feature extraction function, σ′ is the sigmoid activation function, n: represents the number of convolutional layers in the discriminator network, the default value is 3; that is, the output of the discriminator network is an N×N matrix, each element represents the authenticity score of the corresponding image block, N defaults to 70,

[0091] c) Loss function:

[0092] The total loss function consists of three parts:

[0093] L total =αL percep +βL adv +γL phys ;

[0094] Among them, L percep is the perceptual loss, L adv To combat the loss, L phys is the physical constraint loss, and α, β, and γ are weight coefficients.

[0095] Among them, the GFSR module performs the following steps:

[0096] Step 1: Input low-resolution gas flow field data into the generator network.

[0097] Step 2: The generator outputs high-resolution gas flow field data.

[0098] Step 3: The discriminator evaluates the generated high-resolution data and the real high-resolution data respectively.

[0099] Step 4: Calculate the total loss function.

[0100] Step 5: Update the parameters of the generator and discriminator through back-propagation.

[0101] Step 6: Repeat steps 1-5 until the model converges.

[0102] Among them, the training data set of the GFSR module is:

[0103] The training data set consists of low-resolution and high-resolution gas flow field data pairs. The acquisition method is as follows:

[0104] Step 1: Use high-precision sensors to densely sample in the mining space to obtain high-resolution gas flow field data.

[0105] Step 2: Downsample the high-resolution data to obtain the corresponding low-resolution data.

[0106] Step 3: Pair high-resolution data with low-resolution data to form training samples.

[0107] Among them, the training process of the GFSR module is:

[0108] Step 1: Initialize the generator and discriminator network parameters.

[0109] Step 2: Randomly extract a batch of samples from the training dataset.

[0110] Step 3: Input the low-resolution sample into the generator to obtain the generated high-resolution result.

[0111] Step 4: Calculate the loss function.

[0112] Step 5: Update the network parameters through back-propagation.

[0113] Step 6: Repeat steps 2-5 until the model converges or the preset number of iterations is reached.

[0114] A second aspect of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores program instructions, and when the program instructions are run in a computer, they are used to execute the above-mentioned method for establishing a dynamic distribution model of a gas flow field in a mining space.

[0115] A third aspect of the present invention provides a system for establishing a dynamic distribution model of a gas flow field in a mining space, which includes the above-mentioned computer-readable storage medium.

[0116] Compared with the prior art, the method, medium and system for establishing a dynamic distribution model of gas flow field in mining space provided by the present invention have the following beneficial effects:

[0117] First, this method makes full use of multi-source air monitoring data, including wind speed, wind direction, temperature, humidity, air pressure and gas concentration, to comprehensively characterize the dynamic change characteristics of the gas flow field in the mining space. It can not only reflect the changing trend of gas concentration, but also describe the flow distribution of gas in space.

[0118] Secondly, this method uses mathematical analysis methods such as multivariate mixed decomposition and fast Fourier transform to decompose the original complex spatiotemporal gas flow field data into two relatively independent parts: stable changes and fluctuating changes. It can not only extract the long-term trend characteristics of the flow field, but also capture the short-term fluctuation characteristics, greatly improving the model's ability to express the gas distribution law in complex mining environments.

[0119] Furthermore, this method designs an embedded multi-branch pyramid neural network structure, which integrates two modeling ideas: physical model and data-driven model. On the one hand, the network contains an equation group module based on the physical laws of the gas flow field, which can use prior knowledge to calculate the basic distribution characteristics of gas such as diffusion, flow, and pressure; on the other hand, the network also has the ability of adaptive resolution adjustment and multi-resolution feature fusion, which can extract representative spatiotemporal features from the original monitoring data and realize accurate modeling of complex gas flow fields through end-to-end deep learning. This fusion innovation not only ensures the physical rationality of the model output, but also greatly improves the ability to express the actual gas distribution law.

[0120] Finally, the GFSR super-resolution module in this method can reconstruct high-resolution detail information based on low-resolution gas flow field data, and achieve a refined description of the complex mining environment under the premise of ensuring physical constraints. This is not only conducive to better analyzing the changing characteristics of gas at different scales, but also provides more reliable data support for subsequent gas safety warnings.

[0121] To sum up, the method for establishing the dynamic distribution model of the gas flow field in the mining space proposed in the present invention has made innovative breakthroughs in mathematical analysis, deep learning modeling and physical constraint fusion. Compared with the existing technology, it has significantly improved the ability to express the gas distribution law in complex mining environments, and solved the technical problem that the existing technology usually only focuses on the changing trend of gas concentration and lacks an effective description of the dynamic distribution characteristics of gas in the mining space. BRIEF DESCRIPTION OF THE DRAWINGS

[0122] Figure 1 A flow chart of the method provided by the present invention;

[0123] Figure 2 It is a multi-resolution gas distribution comparison map;

[0124] Figure 3 It is a forecast diagram of dynamic changes in gas concentration;

[0125] Figure 4 It is a multi-parameter gas flow field analysis diagram;

[0126] Figure 5 It is a safety risk assessment heat map;

[0127] Figure 6 It is a 3D scatter plot of gas concentration, temperature and humidity. DETAILED DESCRIPTION

[0128] In order to make the purpose, technical solution and advantages of the embodiments of the present invention more clear, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0129] like Figure 1 FIG. 1 is a flow chart of a method for establishing a dynamic distribution model of a gas flow field in a mining space provided by the first aspect of the present invention. The method comprises the following steps:

[0130] S10, collecting air data at multiple collection points in the mining space at preset time intervals, including wind speed, wind direction, temperature, humidity, air pressure and gas concentration;

[0131] S20, at each collection moment, using the air data of each collection point to establish an air matrix, and using the interpolation method to interpolate and supplement the space between the collection points, so as to convert the air matrix into an air dense matrix;

[0132] S30, merging the air dense matrix at each acquisition moment into an air time series matrix according to the time series, and normalizing each element of the air time series matrix to obtain an air normalized time series matrix;

[0133] S40, decomposing the normalized time series matrix of air by using multivariate mixed decomposition calculation to obtain a stable time series matrix and a fluctuating time series matrix;

[0134] S50, performing fast Fourier transform on the stable time series matrix and the fluctuating time series matrix respectively, and retaining the original time domain information, to obtain a stable frequency domain feature matrix, a stable time domain feature matrix, a fluctuating frequency domain feature matrix, and a fluctuating time domain feature matrix;

[0135] S60, constructing a multi-layer tensor network structure, taking a stable frequency domain feature matrix, a stable time domain feature matrix, a fluctuation frequency domain feature matrix, and a fluctuation time domain feature matrix as input tensors; using a tensor decomposition method to reduce the dimension and extract features of the input tensor, and obtaining a feature tensor represented by a low dimension;

[0136] S70. Input the feature tensor into a pre-designed embedded multi-branch pyramid neural network for training to obtain a dynamic distribution model of the gas flow field in the mining space.

[0137] The specific implementation methods of the above steps are described in detail below:

[0138] The specific implementation method of step S10 is to use multiple sensor devices to perform regular collection at preset collection points in the mining space. First, several collection points are set in the mining space, and the positions of these collection points need to fully consider the uniformity of spatial distribution to ensure comprehensive and accurate collection of the gas flow field characteristics of the entire mining space. Secondly, wind speed sensors, wind direction sensors, temperature sensors, humidity sensors, pressure sensors and gas concentration sensors are deployed at each collection point. These sensors can monitor the air flow characteristics and gas concentration at the collection point in real time. Third, a predetermined collection time interval is set, such as collecting once every 10 minutes. This time interval needs to be determined according to the speed of change of the gas flow field at the actual mining site to ensure that the collection frequency can fully reflect the dynamic changes of the gas flow field. At each collection moment, these sensors will automatically record the data monitored by each at the current moment, including wind speed, wind direction, temperature, humidity, air pressure and gas concentration. After a certain period of continuous collection, the air data of multiple collection points in the mining space at different times can be obtained. These data lay the foundation for the establishment of the gas flow field model in the subsequent steps.

[0139] The specific implementation of step S20 is to first process and integrate the data at each collection time. For a certain collection time t, an air matrix A(t) is constructed using the air data of each collection point at that time. The number of rows in this air matrix is ​​equal to the total number of collection points, and the number of columns is 6, corresponding to the six parameters of wind speed, wind direction, temperature, humidity, air pressure and gas concentration. In other words, the element a of the matrix A(t) is i,j (t) represents the value of the jth parameter measured at the i-th acquisition point at time t.

[0140] Since the distribution of collection points is often not dense enough, it is difficult to completely cover the entire mining space. Therefore, it is necessary to use interpolation methods to fill and supplement the space between collection points. Commonly used interpolation methods include Kriging interpolation, spline interpolation, etc. The most common bilinear interpolation method is used here. Specifically, the air matrix A(t) is interpolated and supplemented to obtain a denser air matrix B(t). The number of rows and columns of this dense matrix has greatly increased, which can more comprehensively reflect the distribution of each measurement parameter in the mining space at that moment.

[0141] The specific implementation of step S30 is to stack the air dense matrix B(t) at each acquisition moment in time sequence to form a three-dimensional air time series matrix C. The first dimension of this time series matrix represents time, and the second and third dimensions correspond to spatial coordinates. In other words, the element c of the matrix C i,j,k It represents a parameter value measured at the position with coordinates (j, k) in the mining space at time i.

[0142] Since the dimensions and numerical ranges of different parameters vary greatly, in order to eliminate the impact of such differences on subsequent analysis, each element of the time series matrix C needs to be normalized. The specific normalization method can be minimum-maximum normalization or Z-score normalization. The matrix obtained after normalization is recorded as D, which is called the air-normalized time series matrix.

[0143] The specific implementation method of step S40 is to decompose the normalized time series matrix D using a multilinear decomposition algorithm. This algorithm can decompose the three-dimensional matrix D into the product of two two-dimensional matrices and a diagonal matrix. Among them, the two two-dimensional matrices represent the eigenvectors of the time dimension and the space dimension respectively, and the diagonal matrix contains the singular values ​​of these eigenvectors. Through this decomposition, the original complex time series matrix D can be decomposed into two relatively independent parts: one is a stable time series matrix E that reflects the stable change trend, and the other is a fluctuating time series matrix F that describes the fluctuating change characteristics.

[0144] The specific implementation method of step S50 is to perform Fast Fourier Transform (FFT) on the stable time series matrix E and the fluctuating time series matrix F respectively. FFT is an efficient discrete Fourier transform algorithm that can convert time domain signals into frequency domain representations. By performing FFT transformation on E and F, their characteristic matrices in the frequency domain can be obtained, which are denoted as G and H respectively. At the same time, in order to retain the original time domain information, the time domain characteristic matrices of E and F need to be retained, which are denoted as I and J.

[0145] In summary, the purpose of steps S40 and S50 is to extract two types of features from the gas flow field time series data: one type is the feature that reflects the long-term stable change trend, and the other type is the feature that describes the short-term fluctuation change. By retaining both the time domain and frequency domain features, the dynamic distribution characteristics of the gas flow field can be more comprehensively characterized.

[0146] The specific implementation of step S60 is to construct a multi-layer tensor network structure, taking the feature matrix extracted above as input. This network structure includes the following modules:

[0147] 1. Feature input module: concatenate the stable frequency domain feature matrix G, the stable time domain feature matrix I, the fluctuating frequency domain feature matrix H and the fluctuating time domain feature matrix J into a 4th-order input tensor.

[0148] 2. Tensor decomposition module: Tucker decomposition algorithm is used to reduce the dimension and extract features of the input tensor. Tucker decomposition can decompose a high-dimensional tensor into the product of a core tensor and several low-dimensional feature matrices. Through this decomposition, a lower-dimensional feature tensor can be obtained, which contains the main features of the input tensor.

[0149] 3. Supervised learning module: The extracted feature tensor is input into a pre-designed neural network model for supervised learning training. The specific structure of this neural network model will be introduced in detail in step S70.

[0150] Through this multi-layer tensor network structure, representative features can be effectively extracted from the original spatiotemporal gas flow field data, providing important input for the subsequent construction of the gas flow field model.

[0151] The specific implementation of step S70 is to design an embedded multi-branch pyramid neural network structure to realize the training and prediction of the gas flow field dynamic distribution model. This network structure includes the following key modules:

[0152] 1. Embedded equation module: This module calculates the basic distribution characteristics of the gas flow field based on the physical characteristics of the gas flow field and air data. Specifically, it includes the gas diffusion equation, gas velocity field equation, gas pressure field equation and gas distribution equation. These physical equations can describe the diffusion, flow, pressure change and overall distribution characteristics of gas in the mining space.

[0153] 2. Equation group error term branch network: The function of this branch network is to compensate for the calculation error of the embedded equation group module. It takes the output of the embedded equation group module and the actual measurement data as input and outputs an error compensation term. The network is implemented using a multi-layer perceptron (MLP) structure.

[0154] 3. Adaptive resolution module: This module can dynamically adjust the resolution of the network according to the characteristics of the input data. It analyzes the spatial and temporal variation characteristics of the input data and selects the most appropriate resolution as the output. This module uses an improved ResNet network structure, replaces the original convolution module with a dilated convolution, and deletes the fully connected layer.

[0155] 4. Multi-resolution model branch network: Four different resolution model branch networks are designed here. The first resolution model uses the U-Net network structure, the second resolution model uses the GFSR module based on the simplified version of U-Net, the third resolution model uses the ResNet-18 network plus two cascaded GFSR modules, and the fourth resolution model uses DenseNet-121 plus three cascaded GFSR modules. These networks can generate gas flow field distribution results of different resolutions.

[0156] 5. Fusion summary network: The function of this network is to integrate the gas flow field distribution results of different resolutions and output the final multi-resolution gas flow field dynamic distribution model. It uses the attention mechanism to fuse the network structure.

[0157] This embedded multi-branch pyramid neural network structure can make full use of the physical characteristics and multi-scale spatiotemporal information of the gas flow field, and effectively capture the dynamic change law of the gas flow field in complex mining space. The GFSR module can reconstruct high-resolution gas flow field details based on low-resolution input, further improving the expression ability of the model. The entire network structure embodies the integration and innovation from physical model to data-driven model.

[0158] A second aspect of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores program instructions, and when the program instructions are run in a computer, they are used to execute the above-mentioned method for establishing a dynamic distribution model of a gas flow field in a mining space.

[0159] A third aspect of the present invention provides a system for establishing a dynamic distribution model of a gas flow field in a mining space, which includes the above-mentioned computer-readable storage medium.

[0160] In order to better understand and implement the present invention, a specific embodiment 1 of the method of the present invention is provided below. The steps of this embodiment 1 are described in detail as follows: First, in step S10, various types of sensor equipment need to be deployed at multiple collection points in the mining space to regularly collect air data. Specifically, each collection point should be equipped with a wind speed sensor, a wind direction sensor, a temperature sensor, a humidity sensor, a pressure sensor, and a gas concentration sensor. These sensors can measure and record parameters such as wind speed u, wind direction θ, temperature T, humidity H, air pressure p, and gas concentration c at the current moment. Set a predetermined collection time interval Δt, such as collecting once every 10 minutes. After a certain period of continuous collection, the air data of multiple collection points in the mining space at different times can be obtained. These data can be expressed as:

[0161]

[0162] Among them, x i and i represents the spatial coordinates of the i-th acquisition point, t j Indicates the jth acquisition moment. Through these air data, the foundation can be laid for the establishment of the gas flow field model in the subsequent steps.

[0163] In step S20, the data at each acquisition time needs to be processed and integrated. j, using the air data of each collection point at that moment to construct an air matrix A(t j ):

[0164]

[0165] Where n represents the total number of collection points. Since the distribution of collection points is often not dense enough, interpolation methods are needed to fill and supplement the space between collection points. Common interpolation methods include Kriging interpolation and spline interpolation. The most common bilinear interpolation method is used here. Specifically, for the air matrix A(t j ) is interpolated and supplemented to obtain a denser air matrix B(t j ):

[0166]

[0167] Here, m represents the number of rows and columns of the dense matrix, which is much larger than the number of acquisition points n. This can more comprehensively reflect the distribution of each measurement parameter in the mining space at that moment.

[0168] In step S30, the air density matrix B(t j ) are stacked in time series to form a three-dimensional air time series matrix C:

[0169]

[0170] Where k represents the total number of collected moments. Since the dimensions and numerical ranges of different parameters vary greatly, in order to eliminate the impact of such differences on subsequent analysis, each element of the timing matrix C needs to be normalized. Commonly used normalization methods include minimum-maximum normalization and Z-score normalization. Here, Z-score normalization is used:

[0171]

[0172] in, Represents the normalized element value, x i,j,l Represents the original element value, μ i,j,l and σ i,j,l Respectively represent the mean and standard deviation of the element. The matrix obtained after normalization is denoted as D, which is called the air normalized time series matrix.

[0173] In step S40, the normalized time series matrix D needs to be decomposed using a multilinear decomposition algorithm. This algorithm can decompose the three-dimensional matrix D into the product of two two-dimensional matrices and a diagonal matrix:

[0174] D=G×1U×2V T ;

[0175] Among them, G is a diagonal matrix containing the singular values ​​of D; U and V are two orthogonal matrices, representing the eigenvectors of the time dimension and the space dimension respectively. Through this decomposition, the original complex time series matrix D can be decomposed into two relatively independent parts:

[0176] E = G × 1U;

[0177] F=G×2V T ;

[0178] Among them, E is a stable time series matrix reflecting the stable change trend, and F is a fluctuating time series matrix describing the fluctuating change characteristics. This decomposition is conducive to a more comprehensive description of the dynamic distribution characteristics of the gas flow field.

[0179] In step S50, it is necessary to perform Fast Fourier Transform (FFT) on the stable time series matrix E and the fluctuating time series matrix F respectively. FFT is an efficient discrete Fourier transform algorithm that can convert time domain signals into frequency domain representation. Specifically, by performing FFT transformation on E and F, their characteristic matrices in the frequency domain can be obtained:

[0180]

[0181]

[0182] in, represents the FFT transform operator. At the same time, in order to preserve the original time domain information, the time domain feature matrices of E and F also need to be preserved:

[0183] I = E;

[0184] J = F;

[0185] By retaining both time domain and frequency domain features, the dynamic distribution characteristics of the gas flow field can be more comprehensively characterized.

[0186] In step S60, a multi-layer tensor network structure needs to be constructed, taking the above-extracted feature matrices G, I, H and J as input. This network structure includes the following modules:

[0187] 1. Feature input module: concatenate the four feature matrices into a 4th-order input tensor X:

[0188]

[0189] 2. Tensor decomposition module: Tucker decomposition algorithm is used to reduce the dimension and extract features of the input tensor X. Tucker decomposition can decompose a high-dimensional tensor X into a core tensor And the product form of several low-dimensional feature matrices U, V, W:

[0190]

[0191] Through this decomposition, a lower-dimensional feature tensor can be obtained It contains the main features of the input tensor X.

[0192] 3.Supervised learning module: extract the feature tensor The input is then sent to a pre-designed neural network model for supervised learning training. The specific structure of this neural network model will be described in detail in step S70.

[0193] In summary, through this multi-layer tensor network structure, representative features can be effectively extracted from the original spatiotemporal gas flow field data, providing important input for the subsequent construction of the gas flow field model.

[0194] Finally, in step S70, an embedded multi-branch pyramid neural network structure needs to be designed to implement the training and prediction of the gas flow field dynamic distribution model. This network structure includes the following key modules:

[0195] 1. Embedded equation module: This module calculates the basic distribution characteristics of the gas flow field based on the physical characteristics of the gas flow field and air data. Specifically, it includes the following physical equations:

[0196] Gas diffusion equation:

[0197]

[0198] Gas velocity field equation:

[0199]

[0200] Gas pressure field equation:

[0201]

[0202] Gas distribution equation:

[0203]

[0204] These physical equations can describe the diffusion, flow, pressure change and overall distribution characteristics of gas in the mining space.

[0205] 2. Equation group error term branch network: The function of this branch network is to compensate for the calculation error of the embedded equation group module. It takes the output of the embedded equation group module and the actual measurement data as input and outputs an error compensation term. The network is implemented using a multi-layer perceptron (MLP) structure.

[0206] 3. Adaptive resolution module: This module can dynamically adjust the resolution of the network according to the characteristics of the input data. It analyzes the spatial and temporal variation characteristics of the input data and selects the most appropriate resolution as the output. This module uses an improved ResNet network structure, replaces the original convolution module with a dilated convolution, and deletes the fully connected layer.

[0207] 4. Multi-resolution model branch network: Four different resolution model branch networks are designed here. The first resolution model uses the U-Net network structure, the second resolution model uses the GFSR module based on the simplified version of U-Net, the third resolution model uses the ResNet-18 network plus two cascaded GFSR modules, and the fourth resolution model uses DenseNet-121 plus three cascaded GFSR modules. These networks can generate gas flow field distribution results of different resolutions.

[0208] 5. Fusion summary network: The function of this network is to integrate the gas flow field distribution results of different resolutions and output the final multi-resolution gas flow field dynamic distribution model. It adopts the attention mechanism to fuse the network structure.

[0209] This embedded multi-branch pyramid neural network structure can make full use of the physical characteristics and multi-scale spatiotemporal information of the gas flow field, and effectively capture the dynamic changes of the gas flow field in complex mining space. The GFSR module can reconstruct high-resolution gas flow field details based on low-resolution input, further improving the model's expressiveness.

[0210] In general, through the specific implementation of the above 7 steps, this embodiment 1 can construct a high-precision, high-resolution gas flow field dynamic distribution model, providing important decision-making support for mining control and safe production.

[0211] Specifically, the principle of the present invention is to make full use of multi-source air monitoring data, and accurately capture the dynamic change characteristics of the gas flow field in the mining environment through a modeling method that combines mathematical analysis and deep learning. Specifically, the method mainly includes the following key technical links:

[0212] 1. Data collection and preprocessing

[0213] First, multiple monitoring points are set up in the mining space, and sensors such as wind speed, wind direction, temperature, humidity, pressure, and gas concentration are deployed to regularly collect various air parameter data. These data lay the foundation for subsequent flow field feature analysis and model establishment. In order to eliminate the differences in dimensions and numerical ranges between different parameters, the collected data needs to be normalized.

[0214] 2. Flow field characteristics analysis

[0215] For each acquisition moment, a spatial matrix is ​​constructed using the air data of each monitoring point at that moment, and then the matrix is ​​supplemented by interpolation to obtain a denser air matrix. The dense matrices of all moments are stacked in time series to form a three-dimensional space-time gas flow field data matrix. Next, the multivariate hybrid decomposition algorithm is used to decompose the matrix into a matrix reflecting the stable change trend and a matrix describing the fluctuation change characteristics. At the same time, the two matrices are fast Fourier transformed to extract their characteristics in the frequency domain and time domain. These characteristics characterize the dynamic change law of the gas flow field at different time scales, laying the foundation for subsequent deep learning modeling.

[0216] 3. Deep Learning Modeling

[0217] Based on the spatiotemporal features extracted above, an embedded multi-branch pyramid neural network is designed. The network contains four key modules:

[0218] (1) Embedded Equation Module: This module constructs a set of partial differential equations describing the basic distribution law of gas based on the physical characteristics of the gas flow field, such as gas diffusion, flow, pressure change, etc. These equations can calculate the preliminary distribution characteristics of the gas flow field using the measured air parameter data.

[0219] (2) Equation group error term branch network: The function of this network is to compensate for the calculation error of the embedded equation group module and further improve the accuracy of the model. It takes the output of the equation group module and the actual measurement data as input and obtains an error compensation term through end-to-end learning.

[0220] (3) Adaptive resolution module: This module can dynamically adjust the resolution of the neural network according to the temporal and spatial variation characteristics of the input data. This is conducive to capturing the characteristics of the gas flow field at different scales.

[0221] (4) Multi-resolution model branch network: This network contains four sub-models with different resolutions, which can generate gas flow field distribution prediction results with different spatial resolutions. These results are integrated through an attention mechanism fusion network to output the final multi-resolution gas flow field dynamic distribution model.

[0222] In addition, a GFSR super-resolution module is embedded in the multi-resolution model branch network. Based on the idea of ​​generative adversarial networks, this module can reconstruct high-resolution gas distribution details from low-resolution flow field data, providing support for the refined description of complex mining environments.

[0223] In summary, the method for establishing a dynamic distribution model of gas flow field in mining space proposed in this invention fully combines the two modeling ideas of mathematical analysis and deep learning. On the one hand, it uses multi-source air monitoring data and mathematical analysis methods such as spatiotemporal feature extraction to comprehensively characterize the dynamic change characteristics of the gas flow field; on the other hand, it designs an embedded multi-branch neural network architecture that integrates physical models and data-driven models, which can effectively learn and express the actual distribution law of gas in complex mining environments.

[0224] In order to further better understand and implement the present invention, the following is an example 2 of a specific application scenario of the present invention: A coal mining enterprise conducts dynamic distribution monitoring and modeling of gas flow field in a mining operation area, aiming to improve gas safety management and control capabilities. The mining area covers an area of ​​about 2 square kilometers, with complex and diverse terrain, including different mining methods such as underground caverns and surface open pits. In order to fully grasp the spatiotemporal variation characteristics of the gas flow field in the area, the enterprise has taken the following specific measures:

[0225] 1. Multi-point air data collection

[0226] In the mining area, the company has set up 50 collection points, evenly distributed underground and on the surface. Each collection point is equipped with monitoring equipment such as wind speed sensors, wind direction sensors, temperature sensors, humidity sensors, pressure sensors and gas concentration sensors. These sensors can record the parameter data they monitor in real time, including wind speed u, wind direction θ, temperature T, humidity H, air pressure p and gas concentration c. Figure 6 The figure shows a 3D scatter plot of gas concentration, temperature and humidity, which can reveal the potential relationship between these three parameters and help understand the impact of environmental factors on gas concentration.

[0227] In order to fully reflect the dynamic characteristics of the gas flow field, the company set a 10-minute collection time interval Δt, that is, data collection and recording were performed at each collection point every 10 minutes. After three months of continuous monitoring, the mining area collected a total of 12,960 time points (3 months, 30 days, 24 hours * 6 intervals) of air parameter data.

[0228] 2. Air data preprocessing

[0229] For the original monitoring data obtained, preprocessing is first required to eliminate the dimension and value range differences between different parameters. The specific steps are as follows:

[0230] (1) Constructing the air matrix: For a certain acquisition time t j , using the monitoring data of each collection point at that moment to construct an air matrix A(t j ), the matrix size is 50×6:

[0231]

[0232] (2) Interpolation supplement: Since the distribution of the acquisition points is not dense enough, it is necessary to use bilinear interpolation to supplement the air matrix to obtain a denser air matrix B(t j ), the size is 200×200.

[0233] (3) Time series matrix construction: The air density matrix B(t j ) are stacked in time series to form a 3D gas flow field time series matrix C with a size of 12,960×200×200.

[0234] (4) Data normalization: Perform Z-score normalization on each element of the time series matrix C to eliminate the dimensional differences between different parameters and obtain the normalized air time series matrix D.

[0235] Through the above preprocessing steps, the company obtained a gas flow field dynamic data containing 12,960 time points and 200×200 spatial resolution, laying the foundation for subsequent feature analysis and model building.

[0236] 3. Gas flow field feature extraction

[0237] For the normalized time series matrix D that has been preprocessed, the enterprise uses a multivariate hybrid decomposition algorithm to decompose it, with the aim of extracting the stable change characteristics and fluctuating change characteristics in the gas flow field. The specific steps are as follows:

[0238] (1) Multivariate mixed decomposition: Decompose the 3D matrix D into two 2D matrices E and F and a diagonal matrix G, i.e. D = G × 1E × 2F T Among them, E reflects the stable time characteristics of the gas flow field, F reflects the stable spatial characteristics of the gas flow field, and G contains the singular values ​​of these characteristics.

[0239] (2) Fast Fourier Transform: Perform Fast Fourier Transform (FFT) on E and F respectively to obtain their characteristic matrices G and H in the frequency domain. At the same time, the characteristic matrices i and J of E and F in the time domain are also retained.

[0240] Through the above feature extraction steps, the company obtained four feature matrices: stable frequency domain feature G, stable time domain feature i, fluctuating frequency domain feature H and fluctuating time domain feature J. These features comprehensively describe the dynamic change law of the gas flow field in the mining area in time and space dimensions, including long-term trends and short-term fluctuation characteristics, and provide important input for subsequent neural network modeling.

[0241] 4. Neural network model training

[0242] Based on the dynamic characteristics of the gas flow field extracted above, the company designed an embedded multi-branch pyramid neural network structure to train and predict the gas flow field distribution in the mining area. The network mainly consists of the following modules:

[0243] (1) Embedded equations module: This module constructs the following partial differential equations describing the basic distribution characteristics of gas based on the physical laws of the gas flow field:

[0244] Gas diffusion equation:

[0245]

[0246] Gas flow field equation:

[0247]

[0248] Gas pressure field equation:

[0249]

[0250] Gas distribution equation:

[0251]

[0252] This module uses the collected measured data, such as u, θ, T, H, p and c, to calculate these physical equations and obtain the preliminary distribution characteristics of the gas flow field.

[0253] (2) Equation group error term branch network: This branch network takes the output of the embedded equation group module and the actual measurement data as input, and outputs a term for compensating the calculation error of the equation group through end-to-end learning.

[0254] (3) Adaptive resolution module: This module can automatically determine the optimal resolution of the neural network based on the temporal and spatial variation characteristics of the input data. The specific approach is to use an improved ResNet network structure, replace the original convolution module with a dilated convolution, and delete the fully connected layer.

[0255] (4) Multi-resolution model branch network: Four sub-model networks with different resolutions are designed here:

[0256] The first resolution model: uses the U-Net network structure to output 200×200 gas flow field distribution;

[0257] Second resolution model: Based on the simplified version of U-Net, the GFSR module is connected to output the flow field distribution of 400×400;

[0258] The third resolution model uses a ResNet-18 network plus two cascaded GFSR modules to output a flow field distribution of 800×800;

[0259] Fourth resolution model: Use DenseNet-121 plus three cascaded GFSR modules to output 1600×1600 flow field distribution.

[0260] (5) Fusion summary network: The function of this network is to integrate the gas flow field distribution results of different resolutions and output the final multi-resolution gas flow field dynamic model.

[0261] During the training process, the company first initialized the network parameters of each module, then randomly extracted training samples from the previously extracted feature matrices G, I, H, and J, and input them into the neural network for end-to-end supervised learning. Through repeated iterative optimization, a deep learning model was finally obtained that can accurately predict the dynamic distribution of the gas flow field in the mining area.

[0262] 5. Model application effect

[0263] The company applied the trained neural network model to the actual production in the mining area and achieved remarkable results:

[0264] (1) Refined description of gas distribution: The model can output gas flow field distribution results with multiple resolutions, ranging from 200×200 to 1600×1600, to meet the needs of refined description in different application scenarios. For example, for underground cavern areas, a higher resolution (such as 800×800) model output can be used to fully reflect the distribution characteristics of gas in complex terrain; for surface open-air areas, a lower resolution (such as 400×400) model can be used to take into account both computational efficiency and accuracy requirements. Figure 2 Shown is a multi-resolution gas distribution comparison chart, which shows the gas distribution at different resolutions from 200x200 to 1600x1600.

[0265] (2) Prediction of gas dynamic changes: The model can not only describe the gas distribution at a certain moment, but also predict the dynamic changes of the gas flow field within a certain period of time in the future (such as 1 hour later) based on historical data. This provides mining workers with more timely safety warnings and buys valuable response time for adjusting operation plans and taking emergency measures. Figure 3 The following is a forecast of the dynamic change of gas concentration, showing the actual and predicted values ​​of gas concentration within 24 hours, as well as the confidence interval of the prediction.

[0266] (3) Decision support: The output of the model includes not only the gas concentration distribution, but also multiple important parameters such as wind speed and wind direction. This helps mining managers to comprehensively analyze the formation mechanism of the gas flow field, scientifically formulate ventilation system transformation plans, and improve the pertinence and effectiveness of gas control. Figure 4 The figure shows a multi-parameter gas flow field analysis diagram, which also shows the changes of three parameters: gas concentration, wind speed and wind direction over time.

[0267] (4) Safety risk reduction: Through accurate prediction and analysis of the dynamic distribution of gas flow fields, enterprises can discover possible high-gas concentration areas in advance and take targeted monitoring and prevention measures, effectively reducing the risk of gas-induced accidents. In the three-month application practice, no gas accidents occurred in the mining area, and the safety production level was significantly improved. Figure 5 As shown in the figure, it is a safety risk assessment heat map, which shows the distribution of safety risks in the mining area and marks the high-risk areas; at the same time, Figure 5 Too Figure 2 Detail picture of.

[0268] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A method for establishing a dynamic distribution model of gas flow field in a mining space, characterized in that: The following steps are involved: S10, collecting air data at multiple collection points in the mining space at preset time intervals, including wind speed, wind direction, temperature, humidity, air pressure and gas concentration; S20, at each collection moment, using the air data of each collection point to establish an air matrix, and using an interpolation method to interpolate and supplement the space between the collection points, so as to convert the air matrix into an air dense matrix; S30, merging the air dense matrix at each acquisition moment into an air time series matrix according to the time series, and normalizing each element of the air time series matrix to obtain an air normalized time series matrix; S40, decomposing the air normalized time series matrix by using multivariate mixed decomposition calculation to obtain a stable time series matrix and a fluctuating time series matrix; S50, performing fast Fourier transform on the stable time series matrix and the fluctuating time series matrix respectively, and retaining the original time domain information, to obtain a stable frequency domain feature matrix, a stable time domain feature matrix, a fluctuating frequency domain feature matrix, and a fluctuating time domain feature matrix; S60, constructing a multi-layer tensor network structure, and taking the stable frequency domain feature matrix, the stable time domain feature matrix, the fluctuation frequency domain feature matrix, and the fluctuation time domain feature matrix as input tensors; Use tensor decomposition to reduce the dimension and extract features of the input tensor to obtain a low-dimensional feature tensor; S70, inputting the feature tensor into a pre-designed embedded multi-branch pyramid neural network for training to obtain a dynamic distribution model of the gas flow field in the mining space.

2. The method for establishing a dynamic distribution model of gas flow field in a mining space according to claim 1, characterized in that: The embedded multi-branch pyramid neural network includes an embedded set of equations module, an equations error term branch network, an adaptive resolution module, a first resolution model branch network, a second resolution model branch network, a third resolution model branch network, a fourth resolution model branch network and a fusion summary network.

3. The method for establishing a dynamic distribution model of gas flow field in a mining space according to claim 2, characterized in that: The embedded equation group module is used to calculate the basic distribution characteristics of the gas flow field according to the physical characteristics of the gas flow field and air data, including the gas diffusion equation, the gas velocity field equation, the gas pressure field equation and the gas distribution equation.

4. The method for establishing a dynamic distribution model of gas flow field in a mining space according to claim 3, characterized in that: The equation group error term branch network is used to compensate for the calculation error of the embedded equation group module, the input is the output of the embedded equation group module and actual measurement data, the output is the error compensation term, and the structure is a multi-layer perceptron network.

5. The method for establishing a dynamic distribution model of gas flow field in a mining space according to claim 4, characterized in that: The adaptive resolution module is used to dynamically adjust the network resolution according to the characteristics of the input data. Specifically, the most suitable resolution is determined by analyzing the spatial and temporal variation characteristics of the input data. The input is a feature tensor, and the output is a resolution selection signal. The structure is an adaptive network obtained by replacing the convolution module of the ResNet network with a dilated convolution and deleting the fully connected layer.

6. The method for establishing a dynamic distribution model of gas flow field in a mining space according to claim 5, characterized in that: The first resolution model branch network is used to generate a first resolution gas flow field distribution, the input is a feature tensor and an output of an embedded equation group module, the output is a first resolution gas flow field distribution, and the structure is a U-Net network; The second resolution model branch network is used to generate a second resolution gas flow field distribution. The input is the feature tensor and the down-sampling result of the first resolution model output. The output is the second resolution gas flow field distribution. The structure is a simplified U-Net network followed by a GFSR module. The GFSR module increases the second resolution to twice the first resolution. The third resolution model branch network is used to generate a gas flow field distribution at a third resolution. The input is the feature tensor and the down-sampling result of the second resolution model output. The output is the gas flow field distribution at a third resolution. The structure is a ResNet-18 network followed by two connected GFSR modules, which increases the third resolution to 4 times the first resolution. The fourth resolution model branch network is used to generate a fourth resolution gas flow field distribution. The input is the feature tensor and the downsampling result of the third resolution model output. The output is the fourth resolution gas flow field distribution. The structure is a DenseNet-121 network followed by three connected GFSR modules, which increases the fourth resolution to 8 times the first resolution.

7. The method for establishing a dynamic distribution model of gas flow field in a mining space according to claim 6, characterized in that: The fusion summary network is used to integrate the gas flow field distribution results of different resolutions. The input is the gas flow field distribution results of four resolutions. The output is the final multi-resolution gas flow field dynamic distribution model. The structure is an attention mechanism fusion network.

8. The method for establishing a dynamic distribution model of gas flow field in a mining space according to claim 7, characterized in that: The pyramid structure of the embedded multi-branch pyramid neural network specifically reduces the spatial resolution from the first resolution to the fourth resolution to expand the receptive field, and then gradually improves the spatial resolution through the GFSR algorithm while maintaining information transmission and fusion between different resolutions.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores program instructions, and when the program instructions are run in a computer, they are used to execute the method for establishing a dynamic distribution model of a gas flow field in a mining space as described in any one of claims 1-8.

10. A system for establishing a dynamic distribution model of gas flow field in mining space, characterized in that: A computer-readable storage medium comprising the computer-readable storage medium of claim 9.

Images