Mine pollutant migration multi-process coupling prediction method and device based on deep learning

Abstract

The application relates to a mine pollutant migration multi-process coupling prediction method and device based on deep learning, which comprises the following steps: acquiring first features of real-time data of each node of a target mine; calculating first attention coefficients based on the first features of the nodes and edge features of connected nodes; updating the first features by using first attention weights standardized by the first attention coefficients to obtain second features of the nodes; calculating second attention coefficients of the connected nodes based on environmental variables and the second features of the target mine, updating the second features by using second attention weights standardized by the second attention coefficients to obtain third features of the nodes; calculating third attention coefficients of the connected nodes based on oxidation-reduction potentials and the third features of the nodes, updating the third features by using third attention weights standardized by the third attention coefficients to obtain fourth features of the nodes, and mapping the fourth features into predicted pollutant concentrations. The method can improve prediction accuracy and comprehensively perform risk assessment.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of environmental science and artificial intelligence, in particular to a mine pollutant migration multi-process coupling prediction method and device based on deep learning. BACKGROUND

[0002] When production activities are carried out in the environment, the production activities carried out may produce pollutants, and the produced pollutants may migrate from one place to another, thereby causing the migration and diffusion of the pollutants. For example, in the process of mining, the migration of pollutants such as fissure water, ammonia nitrogen and heavy metals in groundwater is influenced by complex geological structure, hydrological conditions and meteorological factors. Therefore, it is necessary to analyze the migration and diffusion process of the pollutants to predict the concentration distribution of the pollutants at a future time, so that measures can be taken in time according to the concentration distribution to prevent irreversible damage to the environment.

[0003] At present, the models for predicting the concentration of pollutants are mostly based on a single process or static parameters, and it is difficult to capture the nonlinear dynamic characteristics of multi-process coupling, resulting in low prediction accuracy. For example, the traditional pollutant migration model usually relies on static geological parameters and assumed initial conditions to predict the concentration, and the artificial intelligence algorithm model mostly makes local concentration prediction based on single-point data of a sensor, and cannot generate a complete spatial concentration field distribution, and the local concentration is difficult to comprehensively evaluate the pollution risk. SUMMARY

[0004] Therefore, it is necessary to provide a mine pollutant migration multi-process coupling prediction method and device based on deep learning which can improve the prediction accuracy and comprehensively evaluate the risk.

[0005] A mine pollutant migration multi-process coupling prediction method based on deep learning, the method comprising:

[0006] S1, acquiring first features of real-time data of each node in a target mine; the real-time data comprising pollutant data;

[0007] Preferably, the target mine refers to a region where pollution exists.

[0008] Preferably, the pollutant data includes but is not limited to pollutant concentration and pollutant type.

[0009] Preferably, the real-time data of each node can be collected by a real-time data collection device.

[0010] S2, calculating first attention coefficients of connected nodes based on the first features of the nodes and edge features between the connected nodes;

[0011] Preferably, the first attention coefficient is a coefficient used to represent the importance or weight of the connection between connected nodes in the graph attention network.

[0012] Preferably, the edge feature is used to describe the interaction relationship between two connected nodes.

[0013] S3, updating the first feature using the first attention weight standardized by the first attention coefficient to obtain a second feature of the node;

[0014] S4, calculating a second attention coefficient of the connected node based on the environmental variable of the target mine and the second feature, and updating the second feature based on the second attention weight standardized by the second attention coefficient to obtain a third feature of the node; the environmental variable includes water flow rate and pollutant diffusion coefficient;

[0015] Preferably, the second attention coefficient is a parameter used to dynamically adjust the weight of the pollutant transmission path.

[0016] Preferably, the pollutant diffusion coefficient is a parameter used to describe the ability of the pollutant to diffuse between two connected nodes.

[0017] S5, calculating a third attention coefficient of the connected node based on the oxidation-reduction potential of the node and the third feature, and updating the third feature using the third attention weight standardized by the third attention coefficient to obtain a fourth feature of the node, and mapping the fourth feature to the predicted pollutant concentration.

[0018] Preferably, the third attention coefficient is used to model the oxidation-reduction reaction in the chemical form transformation process of heavy metals in each node.

[0019] Preferably, the oxidation-reduction potential is used to reflect the ability of the pollutant to undergo oxidation-reduction reaction at the node.

[0020] In one of the embodiments, the way of obtaining the first feature in step S1 includes:

[0021] Obtaining the geological parameters, pollutant parameters and meteorological parameters of each node in the target mine, and converting the geological parameters, pollutant parameters and meteorological parameters into geological vectors, pollutant vectors and meteorological vectors in vector form;

[0022] Splicing the geological vectors, pollutant vectors and meteorological vectors to obtain the first features of each node;

[0023] The expression for splicing the first features of each node is:

[0024] ;

[0025] in, represents the first feature of node i; represents the geological vector of node i, p i represents the pollutant vector of node i, m i represents the meteorological vector of node i, || represents a vector concatenation operation; , d input represents the total dimension, represents the dimension of the geological vector, represents the dimension of the pollutant vector, Indicates the dimension of the weather vector.

[0026] Preferably, geological parameters include the internal structure and surface features of the target mine. Pollutant parameters include parameters used to describe the characteristics of the pollutants themselves. Meteorological parameters include various physical quantities that describe the atmospheric conditions of the target mine. In one embodiment, the calculation formula of the first attention coefficient in step S2 is:

[0027] ;

[0028] in, represents the first attention coefficient between the connected nodes i and j, is a nonlinear activation function; for The transposed matrix of the first learnable parameter vector of dimension ; is the first learnable weight matrix, F is the input dimension, is the output dimension; is the first edge feature transformation matrix; is the partition coefficient calculation function, Represents the edge features between connected nodes i and j; represents the first feature of node i, Represents the first feature of node j.

[0029] In one embodiment, the calculation formula of the first attention weight in step S3 is:

[0030] ;

[0031] The calculation formula of the second feature of the node in step S3 is:

[0032] ;

[0033] in, is the first attention weight between the connected nodes i and j; softmax is the activation function; denotes a first attention coefficient between connected node i and node j, denotes a first attention coefficient between connected node i and node k; is a set of nodes connected to node i; is an activation function, denotes a first feature of node j, is a first learnable weight matrix; is a second feature of node i.

[0034] In one of the embodiments, the formula for calculating the second attention coefficient in step S4 is:

[0035] ;

[0036] wherein, denotes a second attention coefficient between connected node i and node j; LeakyReLU is a nonlinear activation function; is a transpose matrix of a second learnable parameter vector of dimension D; is a second learnable weight matrix, is an input dimension, is an output dimension; is a second edge feature transformation matrix; is a flow velocity vector from node i to node j; is a diffusion coefficient from node i to node j; denotes a second feature of node i, denotes a second feature of node j.

[0037] In one of the embodiments, the environmental variables further include rainfall and mining activity intensity of the target mine, and when the rainfall reaches a first threshold and / or the mining activity intensity reaches a second threshold, the formula for calculating the second attention coefficient in step S4 is:

[0038] ;

[0039] wherein, denotes a second attention coefficient between connected node i and node j; LeakyReLU is a nonlinear activation function; is a transpose matrix of a second learnable parameter vector of dimension D; is a second learnable weight matrix, F is an input dimension, is an output dimension; is a second edge feature transformation matrix; W R and W M are learnable parameters; R tM is the rainfall amount for time t t is the mining activity intensity for time t; is the water flow velocity vector from node i to node j; is the diffusion coefficient from node i to node j.

[0040] Preferably, the mining activity intensity represents the mining frequency of the target mine.

[0041] In one embodiment, the calculation formula of the second attention weight in step S4 is:

[0042] ;

[0043] The calculation formula of the third feature of the node in step S4 is:

[0044] ;

[0045] wherein, is the second attention weight between the connected nodes i and j, softmax is an activation function, denotes the second attention coefficient between the connected nodes i and j, denotes the second attention coefficient between the connected nodes i and k, is the set of nodes connected to node i; is an activation function, denotes the second feature of node j, is the third feature of node i.

[0046] In one embodiment, the calculation formula of the third attention coefficient in step S5 is:

[0047] ;

[0048] The calculation formula of is:

[0049] ;

[0050] wherein, denotes the third attention coefficient between the connected nodes i and j; LeakyReLU is a nonlinear activation function; is the transpose matrix of the third learnable parameter vector of dimension ; is the third learnable weight matrix, F is the input dimension, is the output dimension; is the diffusion coefficient from node i to node j; || denotes the vector concatenation operation, denotes the oxidation-reduction potential of node i and the third feature the result of splicing, the redox potential of node j a third feature the result of splicing.

[0051] In one of the embodiments, the calculation formula of the third attention weight in step S5 is:

[0052]

[0053] The calculation formula of the fourth feature in step S5 is:

[0054]

[0055] The mapping formula of mapping the fourth feature to the predicted pollutant concentration in step S5 is:

[0056]

[0057] wherein, is the third attention weight between the connected node i and node j, and softmax is an activation function, denotes the third attention coefficient between the connected node i and node j, denotes the third attention coefficient between the connected node i and node k, is the set of nodes connected to node i; is an activation function, the redox potential of node i a third feature the result of splicing, is the fourth feature of node i, c i is the predicted pollutant concentration of node i, w out is the output weight matrix, b out is the bias vector.

[0058] A mine pollutant migration multi-process coupling prediction device based on deep learning, the device comprising:

[0059] A first feature acquisition module is configured to acquire a first feature composed of real-time data of each node in a target mine; the real-time data includes pollutant data.

[0060] A coefficient calculation module is configured to calculate a first attention coefficient of a connected node based on a first feature of the node and an edge feature between the connected nodes.

[0061] A second feature acquisition module is configured to update the first feature using a first attention weight standardized by the first attention coefficient to obtain a second feature of the node. ​​​

[0062] a third feature acquisition module configured to calculate a second attention coefficient of the connected node based on the environmental variable of the target mine and the second feature, and update the second feature based on a second attention weight standardized based on the second attention coefficient, to obtain a third feature of the node; the environmental variable includes a water flow rate and a pollutant diffusion coefficient;

[0063] a concentration prediction module configured to calculate a third attention coefficient of the connected node based on the oxidation-reduction potential of the node and the third feature, and update the third feature based on a third attention weight standardized based on the third attention coefficient, to obtain a fourth feature of the node, and map the fourth feature to a predicted pollutant concentration.

[0064] The above-mentioned deep learning-based mine pollutant migration multi-process coupling prediction method and device, through S1, the first feature of the real-time data of each node in the target mine is acquired; the real-time data includes pollutant data, S2, the first attention coefficient of the connected node is calculated based on the first feature of the node and the edge feature between the connected nodes, S3, the second feature of the node is obtained by updating the first feature based on the first attention weight standardized based on the first attention coefficient, S4, the second attention coefficient of the connected node is calculated based on the environmental variable of the target mine and the second feature, and the second feature is updated based on the second attention weight standardized based on the second attention coefficient, to obtain the third feature of the node; the environmental variable includes a water flow rate and a pollutant diffusion coefficient, S5, the third attention coefficient of the connected node is calculated based on the oxidation-reduction potential of the node and the third feature, and the third feature is updated based on the third attention weight standardized based on the third attention coefficient, to obtain the fourth feature of the node, and the fourth feature is mapped to a predicted pollutant concentration, on the one hand, the predicted pollutant concentration of each node in the target mine can be obtained, so as to comprehensively evaluate the risk of the target mine according to the predicted pollutant concentration of each node, on the other hand, since the real-time data of the node is used, the dynamic update of the edge feature can be realized, the whole prediction process can be self-adapted to the dynamic change of the environment in the target mine, the deviation caused by the static assumption can be reduced, and the prediction accuracy can be improved. In addition, by layer-by-layer fusion of the first feature, the edge feature, the environmental variable, the second feature, the oxidation-reduction potential and the third feature of each node, multi-process coupling modeling and migration prediction under the synergistic action of multiple mechanisms can be realized. BRIEF DESCRIPTION OF DRAWINGS

[0065] Figure 1 It is an application environment diagram of the deep learning-based mine pollutant migration multi-process coupling prediction method in one embodiment;

[0066] Figure 2 It is a flowchart of the deep learning-based mine pollutant migration multi-process coupling prediction method in one embodiment;

[0067] Figure 3 This is a structural block diagram of a multi-process coupled prediction device for mine pollutant migration based on deep learning in one embodiment;

[0068] Figure 4 FIG. 1 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0069] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0070] The multi-process coupling prediction method for mine pollutant migration based on deep learning provided in the embodiment of the present application can be applied to Figure 1 In the application environment shown. The terminal 102 interacts with the server 104 via a wired / wireless channel. The data storage system can store data that the server 104 needs to process. S1. Obtain a first feature consisting of real-time data of each node in the target mine; the real-time data includes pollutant data; S2. Calculate the first attention coefficient of the connected node based on the first feature of the node and the edge feature between the connected nodes; S3. Update the first feature using the first attention weight normalized by the first attention coefficient to obtain the second feature of the node; S4. Calculate the second attention coefficient of the connected node based on the environmental variables and the second feature of the target mine, and update the second feature based on the second attention weight normalized by the second attention coefficient to obtain the third feature of the node; the environmental variables include water flow rate and pollutant diffusion coefficient; S5. Calculate the third attention coefficient of the connected node based on the redox potential of the node and the third feature, and update the third feature using the third attention weight normalized by the third attention coefficient to obtain the fourth feature of the node, and map the fourth feature to the predicted pollutant concentration. The terminal 102 can be, but is not limited to, various personal computers, laptops, smartphones, tablets, IoT devices, etc. The server 104 may be a single server, a server cluster composed of multiple servers, or a cloud computing center composed of multiple servers.

[0071] In one embodiment, Figure 2 As shown in the figure, a multi-process coupling prediction method for mine pollutant migration based on deep learning is provided. Figure 1 Taking the server 104 in the example as an example, the following steps are included:

[0072] S1. Acquire a first feature of real-time data of each node in the target mine; the real-time data includes pollutant data.

[0073] The target mine refers to a mine with pollution. Specifically, during the mining of ores, chemical agents are used for mineral extraction. If the chemical substances are not properly managed during the extraction process, they will leak into the surrounding environment, causing soil and water pollution. The mine includes but is not limited to metal mines and rare earth mines. The metal mine includes but is not limited to iron mines and gold mines.

[0074] The distribution of nodes in the entire target mine can be uniform or non-uniform. The node refers to the location of the deployed real-time data acquisition device in the target mine and the location of the important but undeployed real-time data acquisition device in the target mine. Specifically, in the target mine, the node includes the location of the deployed real-time data acquisition device, and the location of the undeployed real-time data acquisition device in the deep fissure, remote slope, and high monitoring cost. Further, the real-time data acquisition device includes but is not limited to sensors and data recorders. Further, the real-time data of the node is acquired by the deployed real-time data acquisition device, and the real-time data of the node without deploying the real-time data acquisition device is determined according to the geological prior knowledge of the target mine, which includes but is not limited to historical fissure distribution and permeability coefficient gradient. The historical fissure distribution refers to the formation, development, and spatial distribution of cracks or faults in a region of rock in geological history. The permeability coefficient gradient refers to the trend or rate of change of the permeability coefficient with position, i.e., in a particular direction, how the permeability coefficient changes with distance. The permeability coefficient is a parameter that describes the ability of a fluid to pass through a porous medium.

[0075] The first feature is a feature vector obtained from the real-time data of the node. Further, when the real-time data is one kind, the first feature is the feature vector of the real-time data, and when the real-time data is multiple kinds, the first feature is the feature vector obtained by splicing various real-time data.

[0076] The pollutant data can also be referred to as pollutant parameters. The pollutant data includes but is not limited to the concentration of pollutants at the location of the node and the type of pollutants. In some embodiments, pollutants can be transferred from one node to another node through water flow, soil, etc. Therefore, the real-time data of the node is constantly changing, so the first feature of the node is also constantly changing.

[0077] In some embodiments, the process of obtaining the first feature of the real-time data of each node in the target mine includes: obtaining the initial data of each node in the target mine in real time, performing data preprocessing on the initial data to obtain real-time data, and obtaining the first feature of the real-time data. Data preprocessing includes removing outliers in the initial data and standardizing the initial data. Data preprocessing can ensure input consistency.

[0078] In the pollution concentration prediction of each node, the real-time data of each node in the target mine is input into the trained graph attention network to predict the pollution concentration through each layer of the graph attention network. When the real-time data is input into the graph attention network, the input real-time data is accompanied by node information, which is used to determine which node the input real-time data belongs to. The graph attention network belongs to a kind of deep learning. Specifically, the graph attention network is a subclass of graph neural networks, which is specially used to process graph structure data.

[0079] S2, based on the first feature of the node and the edge feature between the connected nodes, the first attention coefficient of the connected nodes is calculated.

[0080] Among them, the connected nodes refer to nodes with physical association or chemical association. The physical association includes but is not limited to groundwater flow direction, fracture connectivity. The groundwater flow direction refers to the flow direction of groundwater in the underground porous medium, and the fracture connectivity refers to the mutual connection degree between the fractures in the rock, which reflects the openness and continuity of the fracture network. The chemical association includes but is not limited to the migration path of the pollutant. The migration path of the pollutant refers to the route of the pollutant from the source to other locations. For example, in the target mine, the groundwater from node A can flow to node B, so node A and node B are connected nodes; there is a main fracture near node C, which is connected with the main fracture zone intersection (node D) and a deep hidden fracture (node E), which can form a fracture network channel, so node C, node D and node E are connected nodes; the diffusion path of the pollutant leaked from node F is often detected by nodes G and H, and the concentration change has correlation, so nodes F, G and H are connected nodes.

[0081] The edge feature is used to describe the interaction relationship between two connected nodes. For example, node A and node B are connected nodes, and the edge feature between node A and node B can be the adsorption behavior strength when the pollutant migrates from node A to node B. The edge feature is calculated in real time according to the first feature of the connected nodes, and the first feature of the connected nodes is real-time changing, so the edge feature will also change constantly.

[0082] The first attention coefficient is a coefficient used to represent the importance or weight of the connection between the connected nodes in the graph attention network. The calculation formula of the first attention coefficient can be , , which represents the first attention coefficient between the connected nodes i and j, and LeakyReLU is a nonlinear activation function; is the transpose matrix of the first learnable parameter vector of dimension d; is the allocation coefficient calculation function, ​represents an edge feature between connected node i and node j; represents a first feature of node i, represents a first feature of node j.

[0083] S3, updating the first feature using the first attention weight standardized by the first attention coefficient to obtain a second feature of the node.

[0084] The first attention weight is obtained by standardizing the first attention coefficient. The standardization methods include but are not limited to using Softmax function for standardization, normalization processing, threshold-based truncation processing, sparse processing, and local standardization processing.

[0085] The first attention weight, the second attention weight, and the third attention weight are attention weights between two connected nodes, mainly used to measure the connection strength or importance between two connected nodes.

[0086] The updating method of the first feature of the node includes but is not limited to weighted summation updating, high-order aggregation strategy updating, jump connection updating, and self-attention mechanism updating. The formula for updating the first feature by the high-order aggregation strategy is:

[0087] ;

[0088] The calculation formula for updating the first feature by the jump connection is:

[0089] ;

[0090] The calculation formula for updating the first feature by the self-attention mechanism is:

[0091] ;

[0092] wherein, is the second feature of node i, h i is the first feature of node i, W is a weight matrix, is the first attention weight between connected node i and node j, is the gating vector between connected node i and node j, N i is a set of nodes connected to node i, h j is the first feature of node j, is the self-attention weight of node i, is an activation function.

[0093] S4, calculating the second attention coefficient of the connected node based on the environmental variable and the second feature of the target mine, and updating the second feature based on the second attention weight standardized by the second attention coefficient to obtain a third feature of the node.

[0094] wherein the environmental variables include water flow velocity and contaminant diffusion coefficient. The water flow velocity is the water flow velocity between two connected nodes in the target mine. The contaminant diffusion coefficient is a parameter used to describe the ability of the contaminant to diffuse between two connected nodes. The environmental variables can be obtained in advance or in real time.

[0095] The second attention coefficient is a parameter used to dynamically adjust the weight of the contaminant transmission path. The calculation formula of the second attention coefficient can be wherein, denotes the second attention coefficient between the connected node i and node j; LeakyReLU is a nonlinear activation function; is the transpose matrix of the second learnable parameter vector of dimension d; is the water flow velocity vector from node i to node j; is the diffusion coefficient from node i to node j; denotes the second feature of node i, denotes the second feature of node j.

[0096] The second attention weight is obtained by standardizing the second attention coefficient. The standardization methods include but are not limited to using the Softmax function for standardization, normalization processing, threshold-based truncation processing, sparse processing, and local standardization processing. The second attention coefficient can be obtained by concatenating the second feature and the environmental variable feature.

[0097] The updating method of the second feature of the node includes but is not limited to updating by weighted summation, updating by high-order aggregation strategy, updating by jump connection, and updating by self-attention mechanism.

[0098] S5, based on the oxidation-reduction potential and the third feature of the node, calculating the third attention coefficient of the connected node, and updating the third feature using the third attention weight standardized by the third attention coefficient to obtain the fourth feature of the node, and mapping the fourth feature to the predicted contaminant concentration.

[0099] wherein the oxidation-reduction potential is used to reflect the ability of the contaminant to undergo oxidation-reduction reaction at the node. By controlling the "form / toxicity of the contaminant in the node changes" in the graph attention network through the oxidation-reduction potential, the chemical form of the heavy metal under different environmental conditions and its change process can be simulated and predicted. The oxidation-reduction potential of the node can be measured by a sensor.

[0100] The third attention coefficient is used to model the redox reaction in the chemical speciation transformation of heavy metals of each node. Specifically, by introducing the redox potential as part of the node features, the interaction weight between the connected nodes is dynamically adjusted, that is, the third attention coefficient is dynamically adjusted. The third attention coefficient can be obtained by concatenating the redox potential and the third feature. The calculation formula of the third attention coefficient can be

[0101] wherein, represents the third attention coefficient between the connected node i and node j; LeakyReLU is a nonlinear activation function; is the transpose matrix of the third learnable parameter vector of dimension d; is the pollutant diffusion coefficient from node i to node j; || represents the vector concatenation operation, represents the redox potential of node i and the third feature concatenated, represents the redox potential of node j and the third feature concatenated.

[0102] The third attention weight is obtained by standardizing the third attention coefficient. The standardization methods include but are not limited to using the Softmax function for standardization, normalization processing, threshold-based truncation processing, sparse processing, and local standardization processing.

[0103] The updating method of the third feature of the node includes but is not limited to weighted summation updating, high-order aggregation strategy updating, jump connection updating, and self-attention mechanism updating.

[0104] The way of mapping the fourth feature to the predicted pollutant concentration includes but is not limited to through a linear regression model and through a multilayer perception. Specifically, the process of mapping the fourth feature through a linear regression model is that the fitting relationship between the training features and the pollutant concentration is obtained by fitting through the training features and the pollutant concentration, and the predicted pollutant concentration is obtained based on the fitting relationship and the actual fourth feature. The process of mapping the fourth feature through a multilayer perception is that the fourth feature is taken as the data of the input layer of the multilayer perception, and the predicted pollutant concentration value is output through the learning and conversion of the hidden layer of the multilayer perception.

[0105] The predicted pollutant concentration obtained by mapping is the pollutant concentration of each node in the target mine at a future time. By predicting the pollutant concentration of each node at a future time, the migration of the pollutant in the target mine can be analyzed, so that the target mine can be managed according to the predicted pollutant concentration. Specifically, a pollutant concentration alarm threshold is determined, and a warning signal is triggered when the predicted pollutant concentration of any node reaches the alarm threshold. The alarm threshold includes a low-risk alarm threshold, a medium-risk alarm threshold, and a high-risk alarm threshold. Specifically, the node whose predicted pollutant concentration reaches the high-risk alarm threshold is automatically pushed to the management system to prompt the management personnel to manage the node whose predicted pollutant concentration reaches the high-risk alarm threshold.

[0106] In some embodiments, the pollutant concentration of each node in the managed target mine is collected to obtain collected pollutant concentration; the collected pollutant concentration of each node is analyzed by difference with the predicted pollutant concentration, and if the difference between the collected pollutant concentration and the predicted pollutant concentration is greater than a preset difference, it is determined that the management measure for the target mine is better.

[0107] In some embodiments, the predicted pollutant concentration of each node is pushed to a mobile terminal to realize real-time monitoring and remote decision-making through the mobile terminal.

[0108] Since the nodes are distributed throughout the target mine, when the predicted pollutant concentration of each node is obtained, the concentration field distribution of the target mine can be determined based on the predicted pollutant concentration of each node. Specifically, according to the spatial coordinates and the predicted pollutant concentration of each node, a continuous three-dimensional pollutant concentration field is generated by an interpolation algorithm, so that multi-process coupling prediction of the migration of the pollutant in the target mine can be realized.

[0109] The above deep learning-based mine pollutant migration multi-process coupling prediction method comprises the following steps: S1, obtaining the first features of each node in the target mine by real-time data; the real-time data comprises pollutant data, S2, calculating the first attention coefficient of the connected nodes based on the first features of the nodes and the edge features between the connected nodes, S3, updating the first features using the first attention weight after standardization of the first attention coefficient to obtain the second features of the nodes, S4, calculating the second attention coefficient of the connected nodes based on the environmental variables of the target mine and the second features, and updating the second features using the second attention weight after standardization of the second attention coefficient to obtain the third features of the nodes; the environmental variables comprise water flow rate and pollutant diffusion coefficient, S5, calculating the third attention coefficient of the connected nodes based on the oxidation-reduction potential of the nodes and the third features, and updating the third features using the third attention weight after standardization of the third attention coefficient to obtain the fourth features of the nodes; and mapping the fourth features to the predicted pollutant concentration. On the one hand, the predicted pollutant concentration of each node in the target mine can be obtained, so that the target mine can be comprehensively risk evaluated according to the predicted pollutant concentration of each node; on the other hand, since the real-time data of the nodes are used, the edge features can be dynamically updated, the whole prediction process can be self-adapted to the dynamic changes of the environment in the target mine, the deviation caused by static assumptions can be reduced, and the prediction accuracy can be improved. In addition, by layer-by-layer fusing the first features of each node, the edge features, the environmental variables, the second features, the oxidation-reduction potential and the third features, multi-process coupling modeling and migration prediction under the synergistic action of multiple mechanisms can be realized.

[0110] In one embodiment, the first feature in step S1 is obtained in the following manner:

[0111] Obtain the geological parameters, pollutant parameters and meteorological parameters of each node in the target mine, and convert the geological parameters, pollutant parameters and meteorological parameters into geological vectors, pollutant vectors and meteorological vectors in vector form.

[0112] Splice the geological vectors, pollutant vectors and meteorological vectors to obtain the first features of each node.

[0113] The expression for splicing the first features of each node is:

[0114] ;

[0115] wherein, Xi represents the first feature of node i; Gi represents the geological vector of node i, p i Xi represents the pollutant vector of node i, m i Xi represents the meteorological vector of node i, || represents a vector splicing operation; , d input Xi represents the total dimension, dimensions of geology vectors, dimensions of contaminant vectors, dimensions of meteorology vectors.

[0116] The geology parameters, the contaminant parameters and the meteorology parameters are real-time data of the nodes. The geology parameters include internal structure and surface features of the target mine. For example, the geology parameters include types and properties of rocks in the target mine, a proportion of pore space in the total volume in the rocks or soil, an ability to affect fluid passing, a degree of difficulty of fluid moving in porous media, a distribution of cracks in the rocks and a degree of connection between the cracks. The contaminant parameters include parameters for describing characteristics of the contaminant itself. For example, a concentration of the contaminant in the target mine, types of the contaminant. The meteorology parameters include various physical quantities for describing atmospheric conditions of the target mine. For example, temperature, humidity, wind speed and direction, precipitation, sunshine hours of the target mine.

[0117] The geology parameters are converted into geology vectors, the contaminant parameters are converted into contaminant vectors, and the meteorology parameters are converted into meteorology vectors. The way of converting the geology parameters, the contaminant parameters and the meteorology parameters into the geology vectors, the contaminant vectors and the meteorology vectors in the form of vectors is mainly determined according to types of the geology parameters, the contaminant parameters and the meteorology parameters. Specifically, if the geology parameters, the contaminant parameters and the meteorology parameters are numerical data, the numerical data can be directly taken as vectors; if the geology parameters, the contaminant parameters and the meteorology parameters are categorical data, the geology parameters, the contaminant parameters and the meteorology parameters can be converted into the geology vectors, the contaminant vectors and the meteorology vectors through one-hot encoding. One-hot encoding is a method of converting categorical data into a format that can be provided to a machine learning algorithm for processing. One-hot encoding realizes the conversion by representing each classification value as a binary vector, in which only one element is 1 and the rest are 0.

[0118] Further, the numerical geology parameters, the numerical contaminant parameters and the numerical meteorology parameters are standardized, and the data after the standardization is determined as the geology vectors, the contaminant vectors and the meteorology vectors. Further, types of the geology parameters include but are not limited to rock types and surface cover types; types of the contaminant parameters include but are not limited to contaminant types; and types of the meteorology parameters include but are not limited to sunny days and rainy days.

[0119] Further, the process of converting the geological parameters, the pollutant parameters and the meteorological parameters into geological vectors, pollutant vectors and meteorological vectors by one-hot encoding is that: presetting one-hot encodings of each geological type, each pollutant type and each meteorological type, querying a matching geological type from each geological type according to a target geological type of the obtained geological parameters, querying a matching pollutant type from each pollutant type according to a target pollutant type of the obtained pollutant parameters, querying a matching meteorological type from each meteorological type according to a target meteorological type of the obtained meteorological parameters, taking the one-hot encoding of the matching geological type as the geological vector, taking the one-hot encoding of the matching pollutant type as the pollutant vector, and taking the one-hot encoding of the matching meteorological type as the meteorological vector.

[0120] Total dimension d input is the sum of the dimensions of the geological vectors, the pollutant vectors and the meteorological vectors converted from the geological parameters, the pollutant parameters and the meteorological parameters input into the graph attention network, which can also be referred to as the sum of the lengths of the vectors. For example, the geological vector has 5 numerical values, the pollutant vector has 3 numerical values, and the meteorological vector has 2 numerical values, so the total dimension d input is 5+3+2=10.

[0121] Concatenation is the process of connecting two or more vectors together to form a new vector. Concatenation does not change the content of the geological vector, the pollutant vector and the meteorological vector, but combines the geological vector, the pollutant vector and the meteorological vector to facilitate subsequent data processing. Specifically, the geological vector is [0.6, 75], d 地质 =2, the pollutant vector is [1, 0], d 污染物 =2, and the meteorological vector is [20, 80, 10], d 气象 =3, then the first feature after concatenation is [0.6, 75, 1, 0, 20, 80, 10], d input =d 地质 +d 污染物 +d 气象 =7.

[0122] In this embodiment, by converting the geological parameters, the pollutant parameters and the meteorological parameters into the geological vectors, the pollutant vectors and the meteorological vectors in the form of vectors, different types of geological parameters, pollutant parameters and meteorological parameters can be converted into a unified vector form, ensuring that the geological parameters, the pollutant parameters and the meteorological parameters can be processed and analyzed. By concatenating the geological vectors, the pollutant vectors and the meteorological vectors, a comprehensive first feature can be created to provide more comprehensive information for the graph attention network.

[0123] In one embodiment, the calculation formula of the first attention coefficient in step S2 is:

[0124] ;

[0125] wherein, represents the first attention coefficient between the connected node i and node j, and LeakyReLU is a nonlinear activation function; is the transpose matrix of the first learnable parameter vector of dimension d; is the first learnable weight matrix, F is the input dimension, is the output dimension; is the first edge feature transformation matrix; is the allocation coefficient calculation function, represents the edge feature between the connected node i and node j; represents the first feature of node i, represents the first feature of node j.

[0126] The edge feature is used to describe the attribute of the interaction relationship between two connected nodes. For example, the edge feature is used to describe the adsorption behavior strength of the pollutant when migrating from the connected node A to the node B. High edge feature indicates that the pollutant is easy to be adsorbed in the soil, not easy to migrate with the groundwater or the pollutant is not easy to desorb into the water body, and the migration risk is low; low edge feature indicates that the pollutant is not easy to be adsorbed in the soil, easy to migrate with the groundwater or the pollutant is easy to desorb into the water body, and the migration risk is high.

[0127] Specifically, after the specific values of the first learnable weight matrix , the first feature of node i , the first feature of node j , the first edge feature transformation matrix , the edge feature between the connected node i and node j are determined, the specific values are directly substituted into the calculation formula of the first attention coefficient for calculation, so that the first attention coefficient can be obtained.

[0128] In this embodiment, by considering the edge feature between the two connected nodes and the first features of the two connected nodes when calculating the first attention coefficient, the complex interaction mechanism between the nodes in the actual environment is met, the relationship between the nodes can be more comprehensively described, and the first attention coefficient can be more accurately obtained.

[0129] In one embodiment, the calculation formula of the first attention weight in step S3 is:

[0130] ;

[0131] The calculation formula of the second feature of the node in step S3 is:

[0132] ;

[0133] wherein, is the first attention weight between the connected node i and node j; softmax is an activation function; denotes the first attention coefficient between the connected node i and node j, denotes the first attention coefficient between the connected node i and node k; is the set of nodes connected to node i; is an activation function, denotes the first feature of node j, is the first learnable weight matrix; is the second feature of node i.

[0134] The Softmax function is used to normalize the first attention coefficient to obtain the first attention weight. The second feature is obtained by weighted summation of the first features of the connected nodes. The second feature is an enhanced representation based on the aggregation of the adjacent node information in the graph structure, which reflects the local context information of the node in the graph structure. For example, the second feature of node A is obtained by weighted summation of the first features of the nodes connected to node A.

[0135] Contaminant migration is a process of "distribution-transportation-reaction" in series. When calculating the first attention coefficient, the edge feature between the two connected nodes and the first features of the two connected nodes are considered, and the first feature is updated using the first attention weight after normalization of the first attention coefficient to obtain the second feature of the node. The distribution process in the series coupling process. The hierarchical design can make the graph attention network learn the physically interpretable decision logic, and the hierarchical design can significantly improve the accuracy and robustness through the physical interpretation of the coupling of the phased feature transformation and the attention mechanism.

[0136] Specifically, after determining the specific value of the first attention coefficient between the connected nodes , the specific value is directly substituted into the calculation formula of the first attention weight for calculation, so that the first attention weight can be obtained.

[0137] Specifically, after determining the specific value of the first attention weight between the connected nodes , the first feature of the connected nodes , the first learnable weight matrix , the specific value can be directly substituted into the calculation formula of the second feature for calculation, so that the second feature can be obtained.

[0138] In this embodiment, the first attention coefficient is normalized by using the softmax function to obtain the first attention weight, and the second feature of each node is obtained by weighted summing the first features of the connected nodes. In this way, the second feature of the node can integrate the first features of the connected nodes, strengthen the features of the connected nodes, suppress the influence of the first features of the unconnected nodes, and avoid information overload or noise interference.

[0139] In one embodiment, the calculation formula of the second attention coefficient in step S4 is:

[0140] ;

[0141] in, Represents the second attention coefficient between the connected nodes i and j; LeakyReLU is a nonlinear activation function; for The transposed matrix of the second learnable parameter vector of dimension ; is the second learnable weight matrix, F is the input dimension, is the output dimension; is the second edge feature transformation matrix; is the water flow velocity vector from node i to node j; is the pollutant diffusion coefficient from node i to node j; represents the second feature of node i, Represents the second feature of node j.

[0142] Specifically, after determining the specific values ​​of the second learnable weight matrix, the second feature of node i, the second feature of node j, the second edge feature transformation matrix, the pollutant diffusion coefficient, and the water flow rate vector, the specific values ​​are directly substituted into the calculation formula of the second attention coefficient for calculation, so that the second attention coefficient can be obtained.

[0143] In some embodiments, the contaminant diffusion coefficient It can be determined based on the temperature and / or medium type between nodes i and j. Specifically, the initial diffusion coefficient of each medium is preset, and based on the target medium between the connected nodes, a matching medium that matches the target medium is searched from each medium, and the initial diffusion coefficient of the matching medium and the current temperature between the connected nodes are obtained; the initial diffusion coefficient is corrected by the current temperature to obtain the pollutant diffusion coefficient between the connected nodes. The higher the temperature, the more intense the molecular motion and the faster the diffusion speed, so the pollutant diffusion coefficient It will increase with increasing temperature. Different media have different resistances to the diffusion of pollutants. Generally speaking, the resistance of air medium is greater than that of soil medium, and the resistance of soil medium is greater than that of water medium.

[0144] In this embodiment, by considering the water flow rate and the pollutant diffusion coefficient between the connected nodes when calculating the second attention coefficient, the migration process of the pollutant can be more accurately modeled, and the simulated migration process of the pollutant is more consistent with the actual migration process.

[0145] In one embodiment, the environmental variables further include rainfall and mining activity intensity of the target mine, and the calculation formula of the second attention coefficient in step S4 is:

[0146] ;

[0147] wherein, represents the second attention coefficient between the connected node i and node j; LeakyReLU is a nonlinear activation function; is the transpose matrix of the second learnable parameter vector of dimension D; is the second learnable weight matrix, F is the input dimension, is the output dimension; is the second edge feature transformation matrix; W R and W M are learnable parameters; R t is the rainfall at time t, M t is the mining activity intensity at time t; is the water flow rate vector from node i to node j; is the pollutant diffusion coefficient from node i to node j.

[0148] The learnable parameters W R and W M respectively control the influence strength of rainfall and mining activity on the second attention weight. The rainfall R t and the mining activity intensity M t are normalized values. When the rainfall R t increases, the W R R t term enhances, and the second attention coefficient of the edge related to the water flow significantly improves; similarly, the mining activity intensity M t dynamically adjusts the second attention weight between the connected nodes through W M M t , reflecting the influence of human activities on the migration of pollutants. The mining activity intensity represents the mining frequency of the target mine.

[0149] Specifically, after the specific values of the second attention coefficient between the connected nodes, the second learnable weight matrix, the second feature of the node i, the second feature of the node j, the second edge feature transformation matrix, the pollutant diffusion coefficient, the vector of the water flow rate, the learnable parameter, the rainfall, and the intensity of the mining activity are determined, the specific values are directly substituted into the calculation formula of the second attention weight for calculation, so that the second attention weight can be obtained.

[0150] In some embodiments, the pollutant diffusion coefficient may be determined according to the temperature and / or medium type between the node i and the node j.

[0151] In this embodiment, the second attention coefficient is calculated in combination with the rainfall and the intensity of the mining activity when the rainfall of the target mine reaches the rainfall threshold and / or the intensity of the mining activity of the target mine reaches the second threshold, so that the data that has no significant influence on the prediction result can be filtered out, thereby avoiding the graph attention network from becoming complex or overfitting due to processing too much meaningless data.

[0152] In one embodiment, the calculation formula of the second attention weight in step S4 is:

[0153] ;

[0154] The calculation formula of the third feature of the node in step S4 is:

[0155] ;

[0156] wherein, is the second attention weight between the connected nodes i and j, and softmax is an activation function, denotes the second attention coefficient between the connected nodes i and j, denotes the second attention coefficient between the connected nodes i and k, is a set of nodes connected to the node i; is an activation function, denotes the second feature of the node j, is the third feature of the node i.

[0157] Specifically, after the specific values of the second attention coefficient between the connected nodes are determined, the specific values are directly substituted into the calculation formula of the second attention weight for calculation, so that the second attention weight can be obtained.

[0158] Specifically, after the specific values of the second attention weight between the connected nodes, the second features of the connected nodes, and the second learnable weight matrix are determined, the specific values can be directly substituted into the calculation formula of the third feature for calculation, so that the third feature can be obtained.

[0159] The Softmax function is used to normalize the second attention coefficient to obtain the second attention weight. The third feature is obtained by weighted summing the second features of connected nodes. For example, the third feature of node A is obtained by weighted summing the second features of the nodes connected to node A.

[0160] Pollutant migration is a process of "distribution-transmission-reaction" coupled in series. When calculating the second attention coefficient, the water flow rate between two connected nodes, the pollutant diffusion coefficient, and the second characteristics of the two connected nodes are considered. The second attention weight, normalized by the second attention coefficient, is used to update the second characteristic. The third characteristic of the node is the transmission process.

[0161] In this embodiment, the second attention coefficient is normalized by using the softmax function to obtain the second attention weight, and the third feature of each node is obtained by weighted summing the second features of the connected nodes. In this way, the third feature of the node can integrate the second features of the connected nodes, strengthen the features of the connected nodes, suppress the influence of the second features of the unconnected nodes, and avoid information overload or noise interference.

[0162] In one embodiment, the calculation formula of the third attention coefficient in step S5 is:

[0163] ;

[0164] The calculation formula is:

[0165] ;

[0166] in, Represents the third attention coefficient between the connected nodes i and j; LeakyReLU is a nonlinear activation function; for The transposed matrix of the third learnable parameter vector of dimension ; is the third learnable weight matrix, F is the input dimension, is the output dimension; is the pollutant diffusion coefficient from node i to node j; || represents the vector concatenation operation, represents the redox potential of node i With the third feature The result of splicing, represents the redox potential of node j With the third feature The result of splicing.

[0167] Specifically, after the third feature of the node and the redox potential are determined, the redox potential of the node and the splicing result of the third feature are obtained, the splicing result of the redox potential and the third feature is substituted into the calculation formula of the third attention coefficient for calculation, and thus the third attention coefficient can be obtained.

[0168] In this embodiment, by considering the redox potential of each node when calculating the third attention coefficient, the chemical form and its change process of the pollutant in each node under different environmental conditions can be simulated, and thus more fine modeling and prediction of the migration and transformation behavior of the pollutant can be realized.

[0169] In one embodiment, the calculation formula of the third attention weight in step S5 is:

[0170]

[0171] The calculation formula of the fourth feature in step S5 is:

[0172]

[0173] wherein, is the third attention weight between the connected node i and node j, and softmax is an activation function, denotes the third attention coefficient between the connected node i and node j, denotes the third attention coefficient between the connected node i and node k, is the set of nodes connected to node i; is an activation function, denotes the redox potential of node i and the splicing result of the third feature , is the fourth feature of node i.

[0174] The softmax function is used for standardizing the third attention coefficient to obtain the third attention weight. The fourth feature is obtained by weighted summation of the redox potential and the splicing result of the third feature of the connected nodes. For example, the fourth feature of node A is obtained by weighted summation of the redox potential and the splicing result of the third feature of the nodes connected to node A.

[0175] The migration of the pollutant is a process of "distribution-transportation-reaction" in series. When calculating the third attention coefficient, the redox potential and the splicing result of the third feature of the connected nodes are considered, and the third attention weight of the third attention coefficient after standardization is used to update the third feature, and thus the fourth feature of the node is obtained, which is the reaction process.

[0176] Specifically, after the third attention coefficient between the connected nodes is determined​​ After the specific value of the third attention weight is determined, the specific value is directly substituted into the calculation formula of the third attention weight for calculation, so that the third attention weight can be obtained.

[0177] Specifically, after the specific value of the third attention weight between the connected nodes, the spliced result of the third feature and the redox potential of the connected nodes, and the third learnable weight matrix are determined, the specific value can be directly substituted into the calculation formula of the fourth feature for calculation, so that the fourth feature can be obtained.

[0178] In this embodiment, the third attention weight is obtained by using the softmax function to standardize the third attention coefficient, and the fourth feature is obtained by weighted sum of the spliced result of the redox potential and the third feature of the connected nodes. In this way, the fourth feature of the node can integrate the third feature and the redox reaction of the connected nodes, suppress the influence of the features of the unconnected nodes, and avoid information overload or noise interference.

[0179] In one embodiment, the mapping formula for mapping the fourth feature to the predicted pollutant concentration in step S5 is:

[0180] ;

[0181] wherein c i is the predicted pollutant concentration of node i, is the fourth feature of node i, w out is the output weight matrix, and b out is the bias vector. Specifically, the output weight matrix w out and the bias vector b out are known data. After the specific value of the fourth feature of the node is obtained, the output weight matrix w out , the bias vector b out , and the specific value of the fourth feature can be substituted into the mapping formula for calculation, so that the predicted pollutant concentration of each node can be obtained.

[0182] The process of calculating the predicted pollutant concentration by the formula is carried out in the fully connected layer of the graph attention network. The fully connected layer carries out linear regression prediction by .

[0183] In this embodiment, the pollutant concentration is linearly regressed and predicted by using , which can simplify the prediction process and quickly obtain the predicted pollutant concentration of each node.

[0184] It should be understood that, although the various steps in the flowcharts involved in the various embodiments described above are displayed in sequence according to the instructions of the arrows, these steps are not necessarily executed in sequence in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order restriction on the execution of these steps, and these steps can be executed in other orders. Moreover, at least a portion of the steps in the flowcharts involved in the various embodiments described above can include multiple steps or multiple stages, and these steps or stages are not necessarily executed and completed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a portion of steps or stages in other steps.

[0185] Based on the same inventive concept, the embodiments of the present application also provide a device for multi-process coupled prediction of mine pollutant migration based on deep learning, which is used to implement the multi-process coupled prediction method for mine pollutant migration based on deep learning. The implementation solution provided by this device is similar to the implementation solution described in the above method. Therefore, the specific limitations of one or more embodiments of the multi-process coupled prediction device for mine pollutant migration based on deep learning provided below can be referred to the limitations of the multi-process coupled prediction method for mine pollutant migration based on deep learning above, and will not be repeated here.

[0186] In one embodiment, Figure 3 As shown, a multi-process coupled prediction device for mine pollutant migration based on deep learning is provided, including:

[0187] A first feature acquisition module 302 is configured to acquire a first feature consisting of real-time data of each node in the target mine; the real-time data includes pollutant data;

[0188] A coefficient calculation module 304 is configured to calculate a first attention coefficient of the connected node based on the first feature of the node and the edge feature between the connected nodes;

[0189] A second feature acquisition module 306 is configured to update the first feature using the first attention weight normalized by the first attention coefficient to obtain a second feature of the node;

[0190] a third feature acquisition module 308 for calculating a second attention coefficient of the connected node based on the environmental variables of the target mine and the second feature, and updating the second feature based on a second attention weight normalized by the second attention coefficient to obtain a third feature of the node; the environmental variables include water flow rate and pollutant diffusion coefficient;

[0191] The concentration prediction module 310 is configured to calculate a third attention coefficient of the connected node based on the redox potential of the node and the third feature, update the third feature by using the third attention weight after the third attention coefficient is normalized to obtain a fourth feature of the node, and map the fourth feature to a predicted pollutant concentration.

[0192] The modules in the deep learning-based mine pollutant migration multi-process coupling prediction device can be implemented by software, hardware, or a combination thereof. The modules can be embedded in or independent of a processor in a computer device in hardware form, or stored in a memory in the computer device in software form, so that the processor can call and execute the operations of the modules.

[0193] In one embodiment, a computer device, which can be a server, is provided. An internal structure diagram of the computer device can be as shown in FIG. 8. Figure 4 The computer device includes a processor, a memory, and a network interface connected by a system bus. The processor of the computer device is configured to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for running the operating system and the computer program in the non-volatile storage medium. The database of the computer device is configured to store real-time data, a first feature, an edge feature, a first attention coefficient, a first attention weight, a second feature, an environmental variable, a second attention coefficient, a second attention weight, a third feature, a redox potential, a third attention coefficient, a third attention weight, a fourth feature, and a predicted pollutant concentration. The network interface of the computer device is configured to communicate with a terminal outside through a network connection. The computer program is executed by the processor to implement a deep learning-based mine pollutant migration multi-process coupling prediction method.

[0194] Those skilled in the art can understand that Figure 4 The structure shown in FIG. 8 is only a block diagram of part of the structure related to the scheme of the present application, and does not limit the computer device to which the scheme of the present application is applied. Specifically, the computer device can include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0195] In one embodiment, a computer device is also provided, which includes a memory and a processor. The memory stores a computer program. The processor executes the computer program to implement the steps in the above method embodiments.

[0196] In an embodiment, a computer readable storage medium is provided, having stored thereon a computer program which, when executed by a processor, implements the steps of any of the above method embodiments.

[0197] In an embodiment, a computer program product is provided, comprising a computer program which, when executed by a processor, implements the steps of any of the above method embodiments.

[0198] It should be noted that the user information (including but not limited to user equipment information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in the present application are all information and data authorized by the user or authorized by all parties.

[0199] A person of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiments can be completed by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer readable storage medium. When the computer program is executed, it can include the processes of the above-mentioned embodiments. Any reference to a memory, database or other medium used in the embodiments provided by the present application can include at least one of a non-volatile and volatile memory. The non-volatile memory can include a read-only memory (ROM), a magnetic tape, a floppy disk, a flash memory, an optical storage, a high-density embedded non-volatile memory, a resistive memory (ReRAM), a magnetoresistive random access memory (MRAM), a ferroelectric memory (FRAM), a phase change memory (PCM), a graphene memory, etc. The volatile memory can include a random access memory (RAM) or an external cache memory, etc. As an illustration but not limitation, the RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc. The database involved in the embodiments provided by the present application can include at least one of a relational database and a non-relational database. The non-relational database can include a distributed database based on a block chain, etc., without being limited thereto. The processor involved in the embodiments provided by the present application can be a general-purpose processor, a central processing unit, a graphics processing unit, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., without being limited thereto.

[0200] Any combination of the technical features in the above embodiments can be made. For the sake of brevity, the foregoing description has not described all possible combinations of the technical features in the above embodiments, however, as long as the combination of the technical features does not contradict, it should be considered within the scope of the present disclosure.

[0201] The above embodiments only express several implementation manners of the present application, and the description is relatively specific and detailed, but it should not be understood as a limitation on the patent scope of the present application. It should be pointed out that, for ordinary skilled persons in the art, several modifications and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the appended claims.

Claims

1. A multi-process coupling prediction method for mine pollutant migration based on deep learning, characterized by: The method comprises: S1. Obtaining a first feature consisting of real-time data of each node in a target mine; the real-time data includes pollutant data; S2. Calculating a first attention coefficient of the connected node based on the first feature of the node and the edge feature between the connected nodes; S3. Update the first feature using the first attention weight normalized by the first attention coefficient to obtain a second feature of the node; S4. Calculating a second attention coefficient of the connected node based on the environmental variables of the target mine and the second feature, and updating the second feature based on a second attention weight normalized by the second attention coefficient to obtain a third feature of the node; the environmental variables include water flow rate and pollutant diffusion coefficient; S5. Calculate a third attention coefficient of the connected node based on the redox potential of the node and the third feature, and update the third feature using a third attention weight normalized by the third attention coefficient to obtain a fourth feature of the node, and map the fourth feature to a predicted pollutant concentration. The method of obtaining the first feature in step S1 includes: Obtaining geological parameters, pollutant parameters, and meteorological parameters of each node in the target mine, and converting the geological parameters, pollutant parameters, and meteorological parameters into vector-type geological vectors, pollutant vectors, and meteorological vectors; splicing the geological vector, the pollutant vector, and the meteorological vector to obtain a first feature of each node; The expression for concatenating the first features of the nodes is: ； in, Representation node i The first characteristic of Representation node i The geological vector, p i Representation node i The pollutant vector, m i Representation node i The meteorological vector, || represents a vector splicing operation; , d input represents the total dimension, represents the dimension of the geological vector, represents the dimension of the pollutant vector, Represents the dimension of the weather vector.

2. The method according to claim 1, characterized in that The calculation formula of the first attention coefficient in step S2 is: ； in, Indicates connected nodes i and nodes j The first attention coefficient between, LeakyReLU is a nonlinear activation function; for The transposed matrix of the first learnable parameter vector of dimension ; is the first learnable weight matrix, F is the input dimension, is the output dimension; is the first edge feature transformation matrix; is the partition coefficient calculation function, Indicates connected nodes i and nodes j Edge features between Representation node i The first characteristic, Representation node j The first feature.

3. The method according to claim 2, characterized in that The calculation formula of the first attention weight in step S3 is: ； The calculation formula of the second feature of the node in step S3 is: ； in, For connected nodes i and nodes j The first attention weight between softmax is the activation function; Indicates connected nodes i and nodes j The first attention coefficient between Indicates connected nodes i and nodes k The first attention coefficient between For nodes i A collection of connected nodes; is the activation function, Representation node j The first characteristic, is the first learnable weight matrix; For nodes i The second characteristic.

4. The method according to claim 1, wherein The calculation formula of the second attention coefficient in step S4 is: ； in, Indicates connected nodes i and nodes j The second attention coefficient between the two; LeakyReLU is a nonlinear activation function; for The transposed matrix of the second learnable parameter vector of dimension ; is the second learnable weight matrix, F is the input dimension, is the output dimension; is the second edge feature transformation matrix; For nodes i To Node j The water velocity vector; For nodes i To Node j The diffusion coefficient of Representation node i The second characteristic, Representation node j The second characteristic.

5. The method according to claim 1, wherein The environmental variables also include the rainfall and mining activity intensity of the target mine. When the rainfall reaches a first threshold and / or the mining activity intensity reaches a second threshold, the calculation formula of the second attention coefficient in step S4 is: ； in, Indicates connected nodes i and nodes j The second attention coefficient between the two; LeakyReLU is a nonlinear activation function; for The transposed matrix of the second learnable parameter vector of dimension ; is the second learnable weight matrix, F is the input dimension, is the output dimension; is the second edge feature transformation matrix; W R and W M is a learnable parameter; R t For time t The amount of rainfall, M t For time t the intensity of mining activities; For nodes i To Node j The water velocity vector; For nodes i To Node j The diffusion coefficient of .

6. The method according to claim 4 or 5, characterized in that The calculation formula of the second attention weight in step S4 is: ； The calculation formula of the third feature of the node in step S4 is: ； in, For connected nodes i and nodes j The second attention weight between softmax is the activation function, Indicates connected nodes i and nodes j The second attention coefficient of time, Indicates connected nodes i and nodes k The second attention coefficient of time, For nodes i A collection of connected nodes; is the activation function, Representation node j The second characteristic, For nodes i The third characteristic.

7. The method according to claim 1, characterized in that The calculation formula of the third attention coefficient in step S5 is: ； The calculation formula is: ； in, Indicates connected nodes i and nodes j The third attention coefficient between the two; LeakyReLU is a nonlinear activation function; for The transposed matrix of the third learnable parameter vector of dimension ; is the third learnable weight matrix, F is the input dimension, is the output dimension; For nodes i To Node j The diffusion coefficient of || represents the vector splicing operation, Representation node i Redox potential With the third feature The result of splicing, Representation node j Redox potential With the third feature The result of splicing.

8. The method according to claim 7, characterized in that The calculation formula of the third attention weight in step S5 is: ； The calculation formula of the fourth feature in step S5 is: ； The mapping formula for mapping the fourth feature to the predicted pollutant concentration in step S5 is: ； in, For connected nodes i and nodes j The third attention weight between softmax is the activation function, Indicates connected nodes i and nodes j The third attention coefficient between Indicates connected nodes i and nodes k The third attention coefficient between For nodes i A collection of connected nodes; is the activation function, Representation node i Redox potential With the third feature The result of splicing, For nodes i The fourth characteristic of c i For nodes i The predicted pollutant concentration, w out is the output weight matrix, b out is the bias vector.

9. A multi-process coupling prediction device for mine pollutant migration based on deep learning, characterized in that: The device comprises: A first feature acquisition module is used to acquire a first feature composed of real-time data of each node in the target mine; the real-time data includes pollutant data; A coefficient calculation module, configured to calculate a first attention coefficient of the connected node based on the first feature of the node and the edge feature between the connected nodes; A second feature acquisition module is configured to update the first feature using the first attention weight normalized by the first attention coefficient to obtain a second feature of the node; a third feature acquisition module, configured to calculate a second attention coefficient of the connected node based on the environmental variables of the target mine and the second feature, and update the second feature based on a second attention weight normalized by the second attention coefficient to obtain a third feature of the node; the environmental variables include water flow rate and pollutant diffusion coefficient; a concentration prediction module, configured to calculate a third attention coefficient of the connected node based on the redox potential of the node and the third feature, update the third feature using a third attention weight normalized by the third attention coefficient, obtain a fourth feature of the node, and map the fourth feature into a predicted pollutant concentration; Methods for obtaining the first feature include: Obtaining geological parameters, pollutant parameters, and meteorological parameters of each node in the target mine, and converting the geological parameters, pollutant parameters, and meteorological parameters into vector-type geological vectors, pollutant vectors, and meteorological vectors; splicing the geological vector, the pollutant vector, and the meteorological vector to obtain a first feature of each node; The expression for concatenating the first features of the nodes is: ； in, Representation node i The first characteristic of Representation node i The geological vector, p i Representation node i The pollutant vector, m i Representation node i The meteorological vector, || represents a vector splicing operation; , d input represents the total dimension, represents the dimension of the geological vector, represents the dimension of the pollutant vector, Represents the dimension of the weather vector.

Images