A method and system for intelligently predicting air volume requirements in mines

Abstract

The present invention discloses a method and system for intelligently predicting mine air volume demand, comprising: S1: deploying multi-source sensors to collect mine air volume demand status data; S2: constructing an intelligent flow evolution model based on the mine air volume demand status data, and obtaining optimal optimization parameters through the intelligent flow evolution model; S3: constructing a dynamic graph network based on the mine air flow environment, combining the dynamic graph network with the optimal optimization parameters to model the spatiotemporal evolution of mine air flow, and capturing the optimal mutation state data in the mine air volume demand data according to the modeling results; S4: introducing a dynamic convolution kernel deformation mechanism based on a deep spatiotemporal convolutional network to optimize the ontology network, using the optimal mutation state data as the input of the deep spatiotemporal convolutional network, and outputting optimal prediction strategy data, thereby significantly improving the model's feature extraction efficiency and prediction accuracy for different mutation types such as gas outburst and fan failure.

Description

Technical Field

[0001] The present invention relates to the technical field of mine air volume demand prediction, and in particular to a mine air volume demand intelligent prediction method and system. Background Art

[0002] As mineral resource development progresses towards deeper and more complex locations, mine ventilation systems are facing increasingly severe challenges. Traditional ventilation technologies are no longer able to meet the safety and efficiency requirements of modern mines. Statistics show that the geological complexity faced by deep mining is becoming increasingly prominent. Increased rock stress leads to abnormal gas emissions that are 3-5 times higher than those in shallow mines, and the geothermal gradient rises by 3-4°C per 100 meters, directly increasing the difficulty of mine ventilation.

[0003] At present, traditional mine air volume forecasting methods usually use fixed analysis models, which are difficult to cope with dynamic fluctuations in air volume demand. The risks of gas accumulation and dust explosion have increased significantly. In addition, ventilation equipment has been in a high-flow, low-efficiency operation state for a long time, and energy consumption accounts for 40%-60% of the total energy consumption of the mine. Relying on manual experience and control makes it difficult to achieve refined management. Therefore, an intelligent prediction method for mine air volume demand is proposed here. Summary of the Invention

[0004] In order to overcome the above-mentioned defects of the prior art and to achieve the above-mentioned objectives, the present invention proposes the following technical solutions:

[0005] A method for intelligently predicting required air volume in a mine, comprising:

[0006] S1: deploy multi-source sensors to collect mine air volume status data;

[0007] S2: Build an intelligent flow evolution model based on the mine air volume state data, and obtain the best optimization parameters through the intelligent flow evolution model;

[0008] S3: Based on the mine airflow environment, a dynamic graph network is constructed. The dynamic graph network is combined with the optimal optimization parameters to model the spatiotemporal evolution of the mine airflow. The optimal mutation state data in the mine air volume demand data is captured based on the modeling results.

[0009] S4: Based on the deep spatiotemporal convolutional network, a dynamic convolution kernel deformation mechanism is introduced to optimize the ontology network, the optimal mutation state data is used as the input of the deep spatiotemporal convolutional network, and the optimal prediction strategy data is output.

[0010] The mine air volume requirement status data includes wind speed data, gas concentration data and temperature data.

[0011] The process of constructing the intelligent flow pattern evolution model is as follows:

[0012] Based on neural differential equations as the basic logical framework of intelligent fluid evolution model;

[0013] The state vector of the mine airflow at time t is obtained based on the mine air volume state data. The dynamic change process of the airflow state vector is regarded as a continuous state based on the neural differential equation and the optimal optimization parameters are introduced to obtain the intelligent flow state evolution model. The model is expressed as follows:

[0014]

[0015] Where f is a function parameterized by a neural network, is the best optimization parameter, represents the state vector at time t.

[0016] The best optimization parameters The acquisition process is:

[0017] Use a multi-layer perceptron as the neural network structure of function f and preset an initial parameter , and adjust the initial parameters;

[0018] Input the mine air volume state data at the initial moment, use the numerical solution method to solve the neural differential equation to obtain the predicted value of the mine air volume state vector at different times t, compare these predicted values ​​with the actual monitoring data, and calculate the loss function;

[0019] Through meta-gradient learning, a small perturbation is made to the initial parameters to obtain the adjusted parameters. The adjusted parameters are then used to calculate the loss of the model on the current task. The losses under different adjusted parameters are compared to obtain the best optimized parameters. .

[0020] The dynamic graph network construction process is as follows:

[0021] Build a dynamic graph network based on the mine airflow environment. The dynamic graph network includes edges, graph nodes, and node attributes.

[0022] The mine tunnel nodes are regarded as nodes of the graph, the connections between the tunnels are regarded as edges, and an edge weight is defined based on the optimal optimization parameters. Dynamically update edge weights and node attributes to build a dynamic graph network.

[0023] The process of modeling the spatiotemporal evolution of mine airflow is as follows:

[0024] The processing process of the intelligent flow evolution model is applied on each node of the dynamic graph network to realize the spatiotemporal evolution modeling of the mine airflow. The formula is expressed as:

[0025]

[0026] in, For the The neural differential function of nodes, is the best optimization parameter, represents the set of edge weights connected to the i-th node, represents the edge weight of the i-th node, represents the rate of change of the wind flow state vector of the i-th node with time t.

[0027] The optimal mutation state data is obtained by directly extracting the corresponding propagation characteristics of the output of the mine wind flow spatiotemporal evolution modeling in the spatiotemporal dimension.

[0028] The process of optimizing the ontology network by introducing the dynamic convolution kernel deformation mechanism based on the deep spatiotemporal convolutional network is as follows:

[0029] The optimal mutation state data is used as the input data of the deep spatiotemporal convolutional network;

[0030] A dynamic convolution kernel deformation mechanism is introduced into the traditional deep spatiotemporal convolutional network. The traditional deep spatiotemporal convolutional network receives the optimal mutation state data through the input layer, optimizes it through the dynamic convolution kernel deformation mechanism, and outputs the optimal prediction strategy data through the fully connected layer.

[0031] Construct a deformation network, which consists of two convolutional layers and convolution kernel deformation parameters. The first convolutional layer performs preliminary feature extraction on the input data, and the second convolutional layer further performs deep feature extraction to obtain deep mutation state features.

[0032] Based on the deep mutation state features, the attention mechanism is used to calculate the attention weights of different sub-mutation features, and different sub-mutation features are encoded separately through an independent feature encoding sub-network;

[0033] The encoded low-dimensional mutation feature vector is input into a deformation parameter generator to generate the convolution kernel deformation parameters, and the convolution kernel is optimized based on the deformation network control to realize the dynamic convolution kernel deformation mechanism.

[0034] An intelligent prediction system for mine air volume requirement, comprising:

[0035] Status acquisition module: deploys multi-source sensors to collect mine air volume status data;

[0036] Evolution optimization module: Build an intelligent flow evolution model based on the mine air volume state data, and obtain the best optimization parameters through the intelligent flow evolution model;

[0037] Mutation capture module: This module builds a dynamic graph network based on the mine airflow environment, combines the dynamic graph network with the optimal optimization parameters to model the spatiotemporal evolution of the mine airflow, and captures the optimal mutation state data in the mine air volume data based on the modeling results.

[0038] Prediction output module: Based on the deep spatiotemporal convolutional network, a dynamic convolution kernel deformation mechanism is introduced to optimize the ontology network, the optimal mutation state data is used as the input of the deep spatiotemporal convolutional network, and the optimal prediction strategy data is output.

[0039] The present invention has the following beneficial effects:

[0040] In the present invention, the problem of insufficient reliability of a single data source is solved by fusing heterogeneous data such as wind speed, gas concentration, and temperature through a multi-source sensor network and trust function theory, thereby improving data credibility. In addition, neural differential equations and dynamic graph networks are introduced to construct a continuous spatiotemporal evolution model. Neural differential equations can dynamically model the rate of change of wind flow states, accurately capturing the nonlinear effects of factors such as changes in fan speed and changes in tunnel structure on wind flow states. The dynamic graph network updates the tunnel network topology in real time, clearly presenting the propagation path and interaction of wind flow in the tunnel network. The combination of the two breaks through the limitations of traditional static models, better adapts to nonlinear dynamic scenarios such as tunnel penetration and equipment start-up and shutdown during mining, and accurately simulates changes in wind flow states. Dynamic modeling of wind flow state change rates and real-time updates of tunnel network topologies are achieved, breaking through the bottleneck of the adaptability of traditional static models to nonlinear dynamic scenarios.

[0041] Secondly, by embedding a dynamic convolution kernel deformation mechanism in the deep spatiotemporal convolutional network and using the deformation network to adaptively generate convolution kernel parameters, dynamic focusing on the spatiotemporal dimensions of mutation features is achieved (such as focusing on short-term change trends in the time dimension and locking the core area of ​​mutation in the spatial dimension), breaking through the limitations of traditional fixed convolution kernels and significantly improving the model's feature extraction efficiency and prediction accuracy for different mutation types such as gas outburst and fan failure. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 This is a method step diagram of a method and system for intelligently predicting mine air volume requirements proposed by the present invention.

[0043] Figure 2 This is a system block diagram of an intelligent prediction method and system for mine air volume demand proposed by the present invention. DETAILED DESCRIPTION

[0044] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0045] Example 1: Figure 1 As shown, the present invention proposes an intelligent prediction method for mine air volume requirement, comprising:

[0046] S1: deploy multi-source sensors to collect mine air volume status data;

[0047] Wind speed sensors, gas concentration sensors, and temperature sensors are deployed in the mine environment. The sensors collect data in real time at a frequency of once per second and convert it into electrical data for collection;

[0048] Specifically, for the wind speed sensor, a device based on the ultrasonic time difference method is selected to capture subtle changes in wind speed and obtain wind speed data v. The gas concentration sensor uses the composite principle of catalytic combustion and infrared absorption, with a detection accuracy of ±0.01%. It monitors the gas concentration in real time to obtain gas concentration data c, and promptly warns of dangerous situations such as gas leaks. The temperature sensor uses a platinum resistance type with an accuracy of ±0.1°C. It monitors the ambient temperature changes that affect air density and wind flow characteristics to obtain temperature data. ;

[0049] The collected data is converted into electrical data and normalized. The original data point after electrical data conversion is assumed to be , then the normalization process is expressed as:

[0050]

[0051] in, is the maximum value of the data points, is the minimum value of the data point, The normalized data point is the mine air volume required status data.

[0052] S2: Build an intelligent flow evolution model based on the mine air volume state data, and obtain the best optimization parameters through the intelligent flow evolution model;

[0053] The construction process of the intelligent flow evolution model is as follows:

[0054] Based on neural differential equations as the basic logical framework of intelligent fluid evolution model;

[0055] Based on the mine air volume state data, the state vector of the mine air flow at time t is obtained, which is expressed as ,in represents the wind speed state at time t, represents the wind direction at time t, Indicates the gas concentration state at time t;

[0056] Based on the neural differential equation, the dynamic change process of the wind flow state vector is regarded as a continuous state and an optimal optimization parameter is introduced. The intelligent flow state evolution model is expressed as follows:

[0057]

[0058] Where f is a function parameterized by a neural network, is the best optimization parameter, represents the state vector at time t;

[0059] Specifically, f can capture the comprehensive impact of various factors on the change of wind flow state, for example, in simulating the wind speed in a certain tunnel When it changes, f can include modeling of the impact on the fan. The fan is a key device that provides power in the mine ventilation system. Changes in parameters such as fan speed and power will directly affect the wind speed. f will incorporate these effects through nonlinear mapping. At the same time, Determines the specific form and parameter values ​​of the function f, by adjusting , which can make f better fit the actual wind flow dynamic changes;

[0060] Best optimization parameters The acquisition process is:

[0061] Use a multi-layer perceptron (MLP) as the neural network structure of f and preset an initial parameter , for the initial parameters Make adjustments;

[0062] Enter the initial time Mine air volume demand status data , then, use numerical solution methods to solve the neural differential equation Get the predicted value of the mine air volume state vector at different times t , these predicted values Compared with actual monitoring data Make comparisons and calculate the loss function;

[0063] The loss function is calculated by the mean square error, and its expression is:

[0064]

[0065] Where N is the number of samples, represents the norm of a vector;

[0066] Specifically, this loss function measures the average difference between the predicted value and the true value;

[0067] Obtain the best optimization parameters based on the value of the loss function through meta-gradient learning , specifically:

[0068] Meta-gradient learning makes small adjustments to the parameters in different directions and magnitudes, and then observes how the model's loss changes on the task after the adjustments. By trying different adjustments multiple times, the meta-learner can calculate a meta-gradient that reflects which parameter adjustment direction is most effective in reducing the loss on different tasks.

[0069] For the initial parameters , first initialize the parameters Make a small perturbation , get the adjusted parameters, and then use the adjusted parameters to calculate the loss of the model on the current task , by comparing the loss under different adjusted parameters , the meta-learner can find an optimal parameter adjustment direction and amplitude to make the overall loss on multiple tasks decrease the fastest. This optimal parameter is the best optimization parameter , according to this optimal optimization parameter Adjusting the parameters of the neural network can adapt the model's ability to predict dynamic changes in mine airflow more quickly and accurately;

[0070] S3: Based on the mine airflow environment, a dynamic graph network is constructed. The dynamic graph network is combined with the optimal optimization parameters to model the spatiotemporal evolution of the mine airflow. The optimal mutation state data in the mine air volume demand data is captured based on the modeling results.

[0071] Build a dynamic graph network based on the mine wind flow environment:

[0072] Dynamic graph networks include edges, graph nodes, and node attributes;

[0073] Consider the mine tunnel nodes as nodes A of the graph, the connections between tunnels as edges B, and define an edge weight Reflects the degree of wind flow influence between different nodes and is based on the best optimization parameters For dynamic updates, the formula is:

[0074]

[0075] Where g is the update function, The best parameters obtained for the neural differential equation, for The air volume required by the mine at the moment, for The air volume required by the mine at the moment;

[0076] Specifically, and It includes various parameters such as wind speed, wind direction, gas concentration, etc., which comprehensively reflects the wind flow conditions in the mine at different times. It is the optimal parameter obtained by the neural differential equation (intelligent flow evolution model), which plays a key role in updating the function g and determines the mapping method of the function g to the wind flow state parameters. For example, if The wind speed in a certain lane suddenly increases at the moment, and the update function g will be combined with as well as The wind flow state of the relevant lane at all times recalculates the weight of the edge between the two nodes , thus reflecting the change in the degree of influence of the wind speed change on the airflow in the adjacent lanes;

[0077] The node attribute is set to a(t), which includes the wind flow state data at different nodes, updated with time and edge information, and is updated through edge weights. Update the node attributes to obtain the updated node attributes. The node attribute update formula is:

[0078]

[0079] in, For the The neighbor set of a node, is the best optimization parameter, The set of edge weights connected to the i-th node;

[0080] Specifically, the best optimization parameters In the process of updating node attributes, the current node's own attribute a(t), the edge weight set connected to the i-th node are comprehensively considered. And the current mine air volume demand status For example, when the wind flow state of a node's neighboring nodes changes, the edge weight The best optimization parameters Based on these changes, the current node's own properties and the overall mine air volume demand status, the next moment will be calculated. The properties of this node , thereby achieving real-time update of node attributes and accurately reflecting the dynamic changes of wind flow status at the node;

[0081] By continuously updating node attributes, dynamic graph networks can be constructed;

[0082] Specifically, by repeatedly updating node attributes, the attributes of each node can reflect changes in the surrounding airflow state and its own dynamic situation in real time. The continuous updating and mutual connection of many node attributes constitute a dynamic graph network. In the dynamic graph network, nodes represent key locations in the mine tunnels, and edges represent the connection relationships between tunnels. The dynamic changes in node attributes and edge weights enable the entire dynamic graph network to reflect changes in the airflow state of the mine ventilation system in real time.

[0083] The process of modeling the spatiotemporal evolution of mine airflow is as follows:

[0084] The processing process of the intelligent flow evolution model is applied on each node of the dynamic graph network to realize the spatiotemporal evolution modeling of the mine airflow. The formula is expressed as:

[0085]

[0086] in, For the The neural differential function of nodes, is the best optimization parameter, represents the set of edge weights connected to the i-th node, represents the edge weight of the i-th node, represents the rate of change of the wind flow state vector of node i over time t;

[0087] Specifically, in this way, while considering the dynamic changes of the node's own wind flow, the influence of adjacent nodes transmitted through edges is also considered;

[0088] When the mine air volume data suddenly changes (such as a sudden surge of gas in a certain area causing a sudden change in gas concentration), the state vector at the node will change, at this time, the neural differential function in the intelligent fluid evolution model It will quickly adjust the calculation of the node's own wind flow dynamic changes according to the changes in the node's own state vector;

[0089] At the same time, the edge weights in the dynamic graph network It will also be recalculated based on the best optimization parameters. The changes in the state vectors of adjacent nodes will be transmitted through edge weights, further affecting other connected nodes, thereby propagating this mutation information throughout the dynamic graph network.

[0090] The process of obtaining optimal mutation state data is as follows:

[0091] First, through the dynamic change items of the node itself, , which comprehensively considers the wind flow status of the current node , time t and optimal optimization parameters , characterizes the dynamic changes of the node's own wind flow. For example, if the gas concentration at a certain node increases, the function will calculate the impact of the gas concentration on the change of the node's wind flow state based on these factors;

[0092] Adjacent node influence items: Reflect adjacent nodes through edge weights The impact on the current node i, for example, when the wind speed of the adjacent node changes, it will affect the wind flow state calculation of the current node through the edge weight transmission;

[0093] Adding the above two parts together, we get , that is, the rate of change of the wind flow state vector of the i-th node with time t, which reflects the dynamic change of the wind flow state of the node;

[0094] According to the output of the mine airflow spatiotemporal evolution model (i.e., the rate of change of the airflow state vector of the i-th node captured by the mine airflow spatiotemporal evolution model over time t), ), directly extract The corresponding propagation characteristics in the time and space dimensions, such as the rate of change of wind flow state parameters at each node in different time periods, the boundaries of the mutation impact range, etc., these key features constitute the optimal mutation state data ;

[0095] Specifically, the optimized intelligent flow evolution model and dynamic graph network work together to continuously track the propagation of mutation information in the spatiotemporal dimensions. The model records the path of the mutation from the starting node to the adjacent nodes in the tunnel network, such as determining which tunnel the gas concentration mutation starts from and which connected tunnels it diffuses through in sequence. By calculating the changes in the state vectors of each node at different time points, the speed of mutation propagation is determined, such as the distance the gas concentration diffuses in the tunnel per second. The optimal mutation state data captured in this way takes into account the dynamic changes in the node's own airflow, as well as the influence of adjacent nodes transmitted through edges, thereby realizing the spatiotemporal evolution modeling of the entire mine airflow system.

[0096] S4: Based on the deep spatiotemporal convolutional network, a dynamic convolution kernel deformation mechanism is introduced to optimize the ontology network, the optimal mutation state data is used as the input of the deep spatiotemporal convolutional network, and the optimal prediction strategy data is output;

[0097] The optimal mutation state data is used as the input data of the deep spatiotemporal convolutional network and expressed as , where T represents the time dimension, reflecting the mutation state feature sequence collected over a period of time; H is the spatial dimension, which can represent the feature distribution related to the spatial position of the mine tunnel; C is the channel dimension, corresponding to different types of mutation feature information, such as gas concentration changes, wind speed changes, etc.

[0098] A dynamic convolution kernel deformation mechanism is introduced into the traditional deep spatiotemporal convolutional network. The traditional deep spatiotemporal convolutional network receives the optimal mutation state data through the input layer, optimizes it through the dynamic convolution kernel deformation mechanism, and finally outputs the optimal prediction strategy data through the fully connected layer;

[0099] The dynamic convolution kernel deformation mechanism is achieved by constructing a deformation network g to optimize the convolution kernel;

[0100] The construction process of the deformation network g is:

[0101] The deformation network g consists of two convolutional layers and convolution kernel deformation parameters;

[0102] The first convolution layer performs preliminary feature extraction on the input data X to mine potential local feature patterns in the data. The second convolution layer further performs deep feature extraction to obtain deep mutation state features.

[0103] Specifically, the design of a two-layer convolutional layer and a convolution kernel deformation parameter structure gradually extracts data features and adjusts the convolution kernel according to specific rules, which can better adapt to the sudden changes in the data. The two-layer convolutional layer is set up to gradually deepen the feature mining process, and the convolution kernel deformation parameter gives the convolution kernel the ability to adapt to changes, thereby enhancing the network's processing effect on different sudden changes.

[0104] The process of obtaining the convolution kernel deformation parameters is as follows:

[0105] The attention mechanism is used to calculate the attention weights of different sub-mutation features (gas concentration, wind speed, temperature) for deep mutation state features. 、 、 ;

[0106] Different sub-mutation features are encoded separately through an independent feature encoding sub-network. Let the encoding sub-network function be Encode, then:

[0107]

[0108] in, , is the low-dimensional mutation feature vector obtained by converting the change trend of the sub-mutation feature in the spatiotemporal dimension;

[0109] Specifically, the attention mechanism enables the network to focus on the more important parts of different sub-mutation features. For example, in some scenarios, a sudden change in gas concentration may be crucial to ventilation safety. Therefore, the attention mechanism assigns a higher weight to the gas concentration feature, allowing the network to focus more on analyzing the gas concentration feature in subsequent processing.

[0110] The encoded low-dimensional mutation feature vector is input into a deformation parameter generator (obtained through neural network training) to generate the convolution kernel deformation parameters , the formula is: ,in, Represents the deformation parameter generator;

[0111] Specifically, the deformation parameter generator generates convolution kernel deformation parameters that can adjust the shape of the convolution kernel based on the input low-dimensional mutation feature vector. The convolution kernel deformation parameters will determine how the convolution kernel better captures the mutation characteristics in the data in subsequent convolution operations. For example, when the gas concentration increases rapidly, the generated deformation parameters will cause the convolution kernel to pay more attention to the rate of concentration change in the corresponding area in the time dimension and focus more on the core area of ​​increased concentration in the spatial dimension.

[0112] Based on two convolutional layers and convolution kernel deformation parameters, the deformation network is constructed;

[0113] Based on the deformation network g, the convolution kernel is optimized to realize the dynamic convolution kernel deformation mechanism, such as:

[0114] When a low-dimensional mutation feature vector change is detected (such as a rapid increase in gas concentration and a decrease in wind speed), the convolution kernel deformation parameter of the deformation network g is This will make the convolution kernel in the traditional deep spatiotemporal convolutional network focus on the concentration change rate in the time dimension of the corresponding area, and focus on the core area of ​​increased concentration in the spatial dimension to realize the dynamic convolution kernel deformation mechanism;

[0115] Based on the deformable network g, the convolution kernel in the traditional deep spatiotemporal convolutional network is adjusted and the convolution operation is performed on the input data X to obtain the best prediction strategy data. ;

[0116] Specifically, during the convolution process, due to the convolution kernel deformation parameter It is generated based on the input data X and can better adapt to the local characteristics of the mutation features in the data. For example, in the area where the gas concentration changes suddenly, the convolution kernel will change according to the convolution kernel deformation parameters. Adjust the shape and parameters to better fit the concentration variation characteristics of the area, and the best prediction strategy data The expression form accurately presents the control parameters of various equipment in the mine ventilation system and the predicted values ​​of airflow state parameters with specific numerical values;

[0117] For example, in a mining operation area of ​​a certain mine, the gas concentration in the mining area suddenly increases. The current gas concentration is 0.7%, the wind speed is 1.3 m / s, the fan speed is 1600 rpm, and the ventilation valve opening is 45 degrees. As the mining work progresses, it is expected that the gas concentration in the area will rise to 0.9% in 1 hour. The data from the best prediction strategy shows that in order to ensure that the gas concentration is within a safe range (below 1%), the fan speed needs to be increased to 1900 rpm and the ventilation valve opening should be increased to 55 degrees. These specific values ​​can directly guide operators to predict and control the ventilation equipment to ensure mine ventilation safety.

[0118] Example 2: Figure 2As shown, the present invention proposes an intelligent prediction system for mine air volume demand, comprising:

[0119] Status acquisition module: deploys multi-source sensors to collect mine air volume status data;

[0120] Evolution optimization module: Build an intelligent flow evolution model based on the mine air volume state data, and obtain the best optimization parameters through the intelligent flow evolution model;

[0121] Mutation capture module: This module builds a dynamic graph network based on the mine airflow environment, combines the dynamic graph network with the optimal optimization parameters to model the spatiotemporal evolution of the mine airflow, and captures the optimal mutation state data in the mine air volume data based on the modeling results.

[0122] Prediction output module: Based on the deep spatiotemporal convolutional network, a dynamic convolution kernel deformation mechanism is introduced to optimize the ontology network, the optimal mutation state data is used as the input of the deep spatiotemporal convolutional network, and the optimal prediction strategy data is output.

[0123] In the application, several formulas involved are calculated by taking their numerical values ​​after removing the dimensions, and the formulas are established by collecting a large amount of data and performing software simulation to obtain a formula for the most recent real situation. Some coefficients or weights in the formulas are set by technical personnel in this field according to actual conditions, so they will not be elaborated here.

[0124] The above embodiments can be implemented in whole or in part by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented in whole or in part in the form of a computer program product. Those skilled in the art will appreciate that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution.

[0125] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A method for intelligently predicting air volume required in a mine, characterized in that: include: S1: deploy multi-source sensors to collect mine air volume status data; S2: Build an intelligent flow evolution model based on the mine air volume state data, and obtain the best optimization parameters through the intelligent flow evolution model; The process of constructing the intelligent flow pattern evolution model is as follows: Based on neural differential equations as the basic logical framework of intelligent fluid evolution model; The state vector of the mine airflow at time t is obtained based on the mine air volume state data. The dynamic change process of the airflow state vector is regarded as a continuous state based on the neural differential equation and the optimal optimization parameters are introduced to obtain the intelligent flow state evolution model. The model is expressed as follows: ； Where f is a function parameterized by a neural network, is the best optimization parameter, represents the state vector at time t; S3: Based on the mine airflow environment, a dynamic graph network is constructed. The dynamic graph network is combined with the optimal optimization parameters to model the spatiotemporal evolution of the mine airflow. The optimal mutation state data in the mine air volume demand data is captured based on the modeling results. The process of modeling the spatiotemporal evolution of mine airflow is as follows: The processing process of the intelligent flow evolution model is applied on each node of the dynamic graph network to realize the spatiotemporal evolution modeling of the mine airflow. The formula is expressed as: ； in, For the The neural differential function of nodes, is the best optimization parameter, represents the set of edge weights connected to the i-th node, represents the edge weight of the i-th node, represents the rate of change of the wind flow state vector of the i-th node with time t; S4: Based on the deep spatiotemporal convolutional network, a dynamic convolution kernel deformation mechanism is introduced to optimize the ontology network, the optimal mutation state data is used as the input of the deep spatiotemporal convolutional network, and the optimal prediction strategy data is output.

2. The method for intelligently predicting air volume required in a mine according to claim 1, characterized in that: The mine air volume requirement status data includes wind speed data, gas concentration data and temperature data.

3. The method for intelligently predicting air volume required in a mine according to claim 1, characterized in that: The best optimization parameters The acquisition process is: Use a multi-layer perceptron as the neural network structure of function f and preset an initial parameter , and adjust the initial parameters; Input the mine air volume state data at the initial moment, use the numerical solution method to solve the neural differential equation to obtain the predicted value of the mine air volume state vector at different times t, compare these predicted values ​​with the actual monitoring data, and calculate the loss function; Through meta-gradient learning, a small perturbation is made to the initial parameters to obtain the adjusted parameters. The adjusted parameters are then used to calculate the loss of the model on the current task. The losses under different adjusted parameters are compared to obtain the best optimized parameters. .

4. The method for intelligently predicting air volume required in a mine according to claim 1, characterized in that: The dynamic graph network construction process is as follows: Build a dynamic graph network based on the mine airflow environment. The dynamic graph network includes edges, graph nodes, and node attributes. The mine tunnel nodes are regarded as nodes of the graph, the connections between the tunnels are regarded as edges, and an edge weight is defined based on the optimal optimization parameters. Dynamically update edge weights and node attributes to build a dynamic graph network.

5. The method for intelligently predicting air volume required in a mine according to claim 1, characterized in that: The optimal mutation state data is obtained by directly extracting the corresponding propagation characteristics of the output of the mine wind flow spatiotemporal evolution modeling in the spatiotemporal dimension.

6. The method for intelligently predicting air volume required in a mine according to claim 1, characterized in that: The process of optimizing the ontology network by introducing the dynamic convolution kernel deformation mechanism based on the deep spatiotemporal convolutional network is as follows: The optimal mutation state data is used as the input data of the deep spatiotemporal convolutional network; A dynamic convolution kernel deformation mechanism is introduced into the traditional deep spatiotemporal convolutional network. The traditional deep spatiotemporal convolutional network receives the optimal mutation state data through the input layer, optimizes it through the dynamic convolution kernel deformation mechanism, and outputs the optimal prediction strategy data through the fully connected layer.

7. The method for intelligently predicting air volume required in a mine according to claim 6, characterized in that: The implementation process of the dynamic convolution kernel deformation mechanism is as follows: Construct a deformation network, which consists of two convolutional layers and convolution kernel deformation parameters. The first convolutional layer performs preliminary feature extraction on the input data, and the second convolutional layer further performs deep feature extraction to obtain deep mutation state features. Based on the deep mutation state features, the attention mechanism is used to calculate the attention weights of different sub-mutation features, and different sub-mutation features are encoded separately through an independent feature encoding sub-network; The encoded low-dimensional mutation feature vector is input into a deformation parameter generator to generate the convolution kernel deformation parameters, and the convolution kernel is optimized based on the deformation network control to realize the dynamic convolution kernel deformation mechanism.

8. An intelligent prediction system for mine air volume demand, using the method according to any one of claims 1 to 7, characterized in that: include: Status acquisition module: deploys multi-source sensors to collect mine air volume status data; Evolution optimization module: Build an intelligent flow evolution model based on the mine air volume state data, and obtain the best optimization parameters through the intelligent flow evolution model; Mutation capture module: This module builds a dynamic graph network based on the mine airflow environment, combines the dynamic graph network with the optimal optimization parameters to model the spatiotemporal evolution of the mine airflow, and captures the optimal mutation state data in the mine air volume data based on the modeling results. Prediction output module: Based on the deep spatiotemporal convolutional network, a dynamic convolution kernel deformation mechanism is introduced to optimize the ontology network, the optimal mutation state data is used as the input of the deep spatiotemporal convolutional network, and the optimal prediction strategy data is output.

Images