Mine dust concentration prediction and control method based on GNN and Transform self-supervised learning

Abstract

The invention relates to a mine dust concentration prediction and control method based on GNN and Transform self-supervised learning, and belongs to the field of coal mine disaster early warning. The method comprises the following steps: acquiring multi-modal mine environment parameters and historical dust concentration data, and carrying out standardization and characteristic engineering processing on the data; partially masking the processed data by adopting an MAE strategy, and learning the internal structure of the data by reconstructing the masked part; generating synthetic data similar to an actual mine environment by using the GAN so as to expand a training set; the spatial features extracted by the GNN module and the time sequence features extracted by the Transform module are fused, and comprehensive feature representation is generated; predicting a dust concentration value of each position of the mine based on the comprehensive characteristics, and generating a ventilation control parameter suggestion; and dynamically adjusting the mine ventilation system according to the prediction result to realize dust concentration control.

Description

Technical Field

[0001] The present invention belongs to the technical field of coal mine disaster warning, and relates to a method for predicting and controlling mine dust concentration based on self-supervised learning of GNN and Transformer. Background Art

[0002] In the field of dust concentration prediction and control in complex environments such as mine roadways, the existing technical system mainly focuses on two types of methods: traditional experience-driven and data-driven, but both face significant technical bottlenecks. The specific background is as follows:

[0003] 1) Limitations of traditional methods: The governance of mine dust has long relied on preset parameter configurations and empirical formulas (such as ventilation volume calculation formulas, dust settlement empirical models), which affect the optimization and control effect of the ventilation and dust removal system. Its core defects are: (1) Insufficient dynamic adaptability: Fixed parameters cannot respond to the dynamic changes of roadway working conditions (such as the movement of mining equipment, air flow disorder), resulting in a large prediction error of dust concentration; (2) Excessive mechanism simplification: Empirical models usually ignore the non-linear characteristics of dust diffusion (such as particle size distribution, turbulent flow effect), and it is difficult to accurately describe the dust migration law under the complex roadway topology.

[0004] 2) Introduction and improvement of deep learning technology: To improve the prediction accuracy, recent research has turned to data-driven methods. Typical technical routes include: (1) Spatiotemporal coupling modeling: Adopting a CNN-LSTM hybrid architecture, using a convolutional neural network (CNN) to extract local spatial features of sensor data (such as dust concentration field distribution), and capturing time-dependent relationships through a long short-term memory network (LSTM) or a gated recurrent unit (GRU) to improve the short-term prediction accuracy; (2) Multi-parameter fusion: Some studies attempt to integrate multi-source heterogeneous data such as wind speed, humidity, and operating status of ventilation equipment to construct a multi-input-multi-output (MIMO) prediction model.

[0005] Although deep learning has significantly improved the prediction performance, there are still the following key problems in complex mine scenarios:

[0006] (1) Defects in spatial topology modeling: CNN is only applicable to regular grid data and is difficult to represent irregular topological structures such as roadway branches and cross-section mutations; ignoring the path dependence of dust diffusion (such as air flow direction, roadway connectivity) results in incomplete extraction of spatial features.

[0007] (2) Limitations in temporal modeling: The capture ability of LSTM / GRU for long-term dependencies (>1 hour) is limited, and it is difficult to predict the lag response of dust concentration; the physical constraints of dust diffusion (such as mass conservation, momentum transfer) are not explicitly modeled, resulting in prediction results that violate physical laws.

[0008] (3) Weak generalization ability: Deep learning models rely on a large amount of labeled data for training. However, the cost of collecting and labeling mine environment data is relatively high. There are significant differences in the environmental conditions and air flow distributions of different mines, resulting in insufficient adaptability of the trained models in new environments and difficulty in stably predicting the changes in dust concentration under different mine conditions.

[0009] (4) Insufficient dynamic regulation: Some studies have attempted to adjust the air volume parameters through optimization algorithms (such as particle swarm optimization or genetic algorithms). However, these methods usually rely on fixed rules for optimization and are difficult to achieve dynamic regulation in complex and changeable environments.

[0010] In summary, the existing technologies are difficult to simultaneously model the spatial topological relationship of mine roadways and the temporal variation of dust concentration. At the same time, the prediction performance decreases under limited data, and the generalization ability is limited, unable to meet the requirements of accurate prediction and dynamic optimization in complex mine environments. Summary of the Invention

[0011] In view of this, the purpose of the present invention is to provide a method for predicting and controlling mine dust concentration based on self-supervised learning of GNN and Transformer, which is used to improve the accuracy of dust concentration prediction and the intelligent level of control strategies in complex mine environments.

[0012] To achieve the above purpose, the present invention provides the following technical solutions:

[0013] Solution 1: A method for predicting and controlling mine dust concentration based on self-supervised learning of GNN and Transformer, specifically including the following steps:

[0014] S1: Obtain multi-modal mine environment parameters, including air volume ratio, roadway cross-sectional area, environmental humidity, temperature, and historical dust concentration data, and perform standardization and feature engineering processing on the data;

[0015] S2: Self-supervised learning pre-training: Use the masked autoencoder (MAE) strategy to partially mask the processed data, and learn the internal structure of the data by reconstructing the masked part;

[0016] S3: Data augmentation: Use GAN to generate synthetic data similar to the actual mine environment to expand the training set; where GAN represents a generative adversarial network;

[0017] S4: Model training: Integrate the spatial features extracted by the graph neural network (GNN) module and the temporal features extracted by the Transformer module to generate a comprehensive feature representation;

[0018] S5: Predict the dust concentration: Based on the comprehensive features, predict the dust concentration values at various positions in the mine and generate ventilation control parameter suggestions;

[0019] S6: Dynamically adjust the mine ventilation system according to the prediction results, including wind speed and wind direction parameters, to achieve dust concentration control.

[0020] Furthermore, in step S2, the calculation formula of the loss function of the MAE is as follows:

[0021]

[0022] where L MAE represents the loss of the masked autoencoder, X pred,i represents the masked part predicted by the model, X ture,i is the real data, and N is the number of training samples.

[0023] Furthermore, in step S3, the GAN includes a generator and a discriminator; the function of the generator is to generate synthetic data, and the specific formula is as follows:

[0024] G(z) = σ(W g z + b g )

[0025] where G(z) represents the output of the generator, z represents the random noise vector, W g represents the weight matrix of the generator, b g represents the bias term of the generator, and σ(·) represents the activation function, usually ReLU;

[0026] The function of the discriminator is to discriminate whether the data is real or synthetic, and the specific formula is as follows:

[0027] D(x) = sigmoid(W d x + b d )

[0028] where D(x) represents the output result of the discriminator, x represents the input of the discriminator, W d represents the weight matrix of the discriminator, b d represents the bias term of the discriminator, and sigmoid(·) represents the Sigmoid activation function;

[0029] Finally, the calculation formula of the GAN loss function is as follows:

[0030]

[0031] where L GAN represents the GAN loss, Preal represents the real data distribution, and Pz represents the random noise distribution.

[0032] Further, in step S4, the GNN module constructs a topological structure diagram of the mine roadway, takes the sensor positions as nodes and the roadway connection relationships as edges, and uses a graph convolutional network (GCN) and a graph attention network (GAT) to extract the spatial features of the dust concentration; where GCN represents a graph convolutional network and GAT represents a graph attention network.

[0033] Further, in step S4, the formula for the GCN to extract spatial features is as follows:

[0034]

[0035] Where, is the feature representation of node v at the l-th layer, N(v) is the set of neighbors of v, u is the sensor position directly connected to node v, and u ∈ N(v) represents the neighbor node u of node v. W (l) and b (l) are training parameters, and σ is the activation function ReLU;

[0036] The GAT adaptively adjusts the influence of neighbor nodes on the target node through an attention mechanism, enabling the model to pay more attention to neighbor nodes that contribute more to the change in dust concentration. The attention weight α vu is calculated as follows:

[0037]

[0038] Where LeakyReLU is an improved version of ReLU and is also an activation function used to enhance numerical stability and avoid gradient disappearance; α vu represents the attention weight of neighbor node u on target node v, k is the neighbor node index, α is the attention parameter vector, is the feature of neighbor node u at the l-th layer, and || represents the vector concatenation operation.

[0039] Further, in step S4, the Transformer module processes time series data, adopts a multi-head self-attention mechanism to capture the long-term dependence relationship of the dust concentration; the calculation formula is as follows:

[0040]

[0041] Where Attention(Q, K, V) represents the result of self-attention calculation, and Q, K, and V represent the query vector, key vector, and value vector respectively; d k is the feature dimension during attention calculation;

[0042] And position encoding is adopted to capture time information, enabling the Transformer to perceive the time position relationship in the time series; the calculation formula is as follows:

[0043]

[0044] Among them, PE (pos,2i) is the value of the position encoding at the 2i-th dimension at the pos-th time step, where i is half of the dimension index of the current position encoding, pos represents the time step index, and d model is the model dimension.

[0045] Furthermore, in step S4, when extracting spatial features, a convolutional neural network (CNN) can also be used to perform grid processing on the mine environment to extract local spatial features.

[0046] Furthermore, in step S4, when extracting temporal features, a recurrent neural network (RNN) or a long short-term memory network (LSTM) can also be used for temporal modeling.

[0047] Furthermore, in step S2, contrastive learning can also be adopted for self-supervised learning.

[0048] Solution 2: A mine dust concentration prediction and control system based on self-supervised learning of GNN and Transformer, characterized in that the system includes an input layer, a self-supervised learning module, a GAN module, a GNN module, a Transformer module, a fusion layer, a hidden layer, and an output layer;

[0049] Input layer: used to receive multi-modal mine environment parameters and historical dust concentration data, the parameters include air volume ratio, roadway cross-sectional area, environmental humidity, temperature, and perform standardization and feature engineering processing on the data;

[0050] Self-supervised learning (SSL) module: adopts the masked autoencoder (MAE) strategy to learn the internal structure by reconstructing the masked input data;

[0051] Generative adversarial network (GAN) module: generates synthetic data similar to the actual mine environment to enhance the robustness of the model;

[0052] GNN module: constructs a mine roadway topological structure diagram, takes the sensor positions as nodes and the roadway connection relationships as edges, and extracts spatial features through a graph convolutional network (GCN) and a graph attention network (GAT) to capture the dust diffusion pattern;

[0053] Transformer module: adopts a multi-head self-attention mechanism to process time series data, and combines position encoding to capture the long-term dependence relationship of dust concentration;

[0054] Fusion layer: fuses the spatial features extracted by the GNN module and the temporal features extracted by the Transformer module to form a comprehensive feature representation;

[0055] Hidden layer: Based on the fusion layer, a multi-layer perceptron (MLP) is constructed for deep feature extraction and non-linear mapping. This layer consists of three layers of fully connected neurons, containing 64, 32, and 16 neurons respectively. Each layer uses the ReLU activation function to enhance the non-linear expression ability, and combines Dropout during training for regularization to prevent overfitting. The final hidden layer further optimizes and integrates the spatial features extracted by the GNN and the temporal features extracted by the Transformer, and transmits them to the output layer.

[0056] Output layer: Outputs the predicted dust concentration value and the corresponding ventilation control suggestions.

[0057] The beneficial effects of the present invention are as follows: By combining the methods of graph neural network (GNN), Transformer, and self-supervised learning (SSL), the present invention improves the accuracy of dust concentration prediction, enhances the model's understanding ability of the mine topological structure, and improves the adaptability and generalization ability of the model under limited data conditions, thereby optimizing the intelligent level of the ventilation and dust removal system. The specific advantages of the present invention are as follows:

[0058] 1) The topological structure of the mine roadway is constructed by the GNN, and multi-parameters such as roadway cross-sectional area, wind speed and direction, humidity, and dust concentration are input and modeled as graph data to capture spatial features. Combining the Transformer for time series modeling, the self-attention mechanism is used to extract the long-term change trend of dust concentration, enhance the adaptability of the model to the dynamic mine environment, and achieve more accurate dust concentration prediction, breaking through the problem of low prediction accuracy of traditional methods under complex working conditions. Compared with traditional methods, the prediction accuracy is improved, effectively guiding mine ventilation and dust removal, reducing prediction errors, and improving air quality.

[0059] 2) Self-supervised learning (SSL) improves the generalization ability of the model. Adopting the masked autoencoder (MAE) strategy, during the training process, some input data are randomly masked, and the model learns the internal structure of the data by reconstructing the missing part, enabling the model to still learn environmental features under limited or partially missing data conditions, reducing the dependence on large-scale labeled data. Compared with traditional supervised learning methods, the generalization ability in different mine environments is improved, making the model applicable to various working conditions.

[0060] 3) Generate realistic synthetic data through GAN to simulate the dust diffusion characteristics under different mine environments, improve the model robustness in the case of scarce data, expand the training data set by generating synthetic data similar to the real mine environment, reduce the error caused by data imbalance, alleviate the problem of insufficient actual mine data, enable the model to maintain stable performance in various environments, improve the prediction accuracy, and avoid overfitting caused by data imbalance. Compared with the method without data augmentation, the model prediction accuracy is improved in extreme environments, and the system stability is enhanced.

[0061] 4) Form an intelligent mine dust control system integrating data collection, feature modeling, self-supervised pre-training, GAN data augmentation, GNN-Transformer prediction, and dynamic air volume optimization, enhance the intelligent level of mine ventilation and dust removal, improve the dust removal efficiency, reduce energy consumption, and provide guarantee for the safe operation of mines.

[0062] Other advantages, objectives, and features of the present invention will be described to some extent in the subsequent specification, and to some extent, will be obvious to those skilled in the art based on the study of the following text, or can be taught from the practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the following specification. Brief Description of the Drawings

[0063] In order to make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be described in detail preferably with reference to the accompanying drawings, where:

[0064] Figure 1 is a mine dust concentration prediction model based on the GNN+Transformer structure;

[0065] Figure 2 is a flow chart of a mine dust concentration prediction and control method based on self-supervised learning of GNN and Transformer. Detailed Embodiments

[0066] The following specific examples illustrate the embodiments of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the diagrams provided in the following embodiments only illustrate the basic concept of the present invention in a schematic manner. Without conflict, the following embodiments and the features in the embodiments can be combined with each other.

[0067] Among them, the attached drawings are only for illustrative purposes, showing only schematic diagrams rather than physical diagrams, and should not be construed as a limitation to the present invention; for better illustration of the embodiments of the present invention, some components in the attached drawings will be omitted, enlarged or reduced, which does not represent the size of the actual product; for those skilled in the art, it is understandable that some well-known structures and their descriptions in the attached drawings may be omitted.

[0068] Embodiment 1:

[0069] This embodiment provides a mine dust concentration prediction model structure based on self-supervised learning of GNN and Transformer, as Figure 1 shown. This model combines graph neural network (GNN), Transformer, and self-supervised learning (SSL) for mine dust concentration prediction and control, aiming to improve the accuracy of dust concentration prediction and the intelligent level of control strategies in complex mine environments.

[0070] Model structure:

[0071] Input layer: Receives multi-modal mine environmental parameters, including air volume ratio, roadway cross-sectional area, environmental humidity, temperature, etc. These parameters are processed through standardization and feature engineering to ensure the stability and effectiveness of the model input.

[0072] GNN module: Constructs a topological structure diagram of the mine roadway, treats the sensor positions as nodes, and the roadway connection relationships as edges. Utilizes graph convolutional network (GCN) and graph attention network (GAT) to extract spatial features and capture the diffusion patterns of dust in complex roadway structures. The formula for spatial feature extraction by graph convolutional network (GCN) is as follows:

[0073]

[0074] Where is the feature representation of node v at the l-th layer, N(v) is the set of neighbors of v, W (l) and b (l) are training parameters, σ is the activation function ReLU. The attention weight α vu is as follows. Through the attention mechanism, GNN focuses on neighbor nodes that have a greater impact on dust diffusion:

[0075]

[0076] Transformer module: Processes time series data, adopts the multi-head self-attention mechanism, and captures the long-term dependence relationship of dust concentration. The formula is as follows:

[0077]

[0078] Among them, the input X is mapped to three different subspaces to obtain the query vector (Q), key vector (K), and value vector (V). d k is the feature dimension during attention calculation. Softmax normalization is used to make the sum of attention weights equal to 1. And positional encoding is adopted to capture time information, enabling the Transformer to perceive the temporal positional relationship in the time series. The formula is as follows:

[0079]

[0080] Among them, PE (pos,2i) is the value of the positional encoding at the dimension with index 2i at the pos-th time step. i is half of the dimension index where the current positional encoding acts, pos represents the time step index, and d model is the model dimension.

[0081] Fusion layer: The spatial features extracted by the GNN are fused with the temporal features extracted by the Transformer to form a comprehensive feature representation.

[0082] Self-supervised learning module: The masked autoencoder (MAE) strategy is adopted to partially mask the input data. The model learns the internal structure of the data by reconstructing the masked part, enhancing the generalization ability of the model. The calculation formula of the masked autoencoder loss function is as follows:

[0083]

[0084] Among them, X pred,i represents the masked part predicted by the model, X rure,i is the real data, and N is the number of training samples.

[0085] Generative adversarial network module: The GAN is used to generate synthetic data similar to the actual mine environment, enrich the training dataset, and improve the robustness of the model in the case of scarce data. The generative adversarial network consists of a generator and a discriminator. The function of the generator is to generate synthetic data. The specific formula is as follows:

[0086] G(z) = σ(W g z + b g ) (6)

[0087] The function of the discriminator is to discriminate whether the data is real or synthetic. The specific formula is as follows:

[0088] D(x) = sigmoid(W d x + b d ) (7)

[0089] The calculation formula of the final GAN loss function is as follows:

[0090]

[0091] Among them, Preal represents the real data distribution, and Pz represents the random noise distribution.

[0092] Hidden layer: Based on the fusion layer, a multi-layer perceptron (MLP) is constructed for deep feature extraction and non-linear mapping; this layer consists of three layers of fully connected neurons, containing 64, 32, and 16 neurons respectively; each layer uses the ReLU activation function to enhance the non-linear expression ability, and combines Dropout during the training process for regularization to prevent overfitting; the final hidden layer further optimizes and integrates the spatial features extracted by the GNN and the temporal features extracted by the Transformer, and transmits them to the output layer.

[0093] Output layer: Predicts the dust concentration values at different positions in the mine and provides corresponding ventilation control suggestions.

[0094] Please refer to Figure 2 , this embodiment provides a method for predicting and controlling mine dust concentration based on self-supervised learning of GNN and Transformer. The working process is as follows: First, preprocess the data, collect mine environmental parameters and historical dust concentration data, and perform cleaning, standardization, and feature engineering on the data. Use self-supervised learning to pre-train the model and learn the internal structure of the data. Perform data augmentation, use GAN to generate synthetic data, enrich the training data set, and improve the generalization ability of the model. Then, perform model training, combine the actual data with the synthetic data, perform supervised learning training on the fusion model, and optimize the model parameters. Predict the dust concentration, input real-time environmental parameters, the model extracts spatial and temporal features through GNN and Transformer, and outputs the predicted dust concentration value. Finally, according to the predicted dust concentration, adjust the ventilation parameters such as wind speed and wind direction to achieve effective control of the dust concentration.

[0095] In this embodiment, by combining graph neural network (GNN), Transformer, self-supervised learning (SSL), and generative adversarial network (GAN), this patent realizes high-precision prediction, intelligent optimization control, and data augmentation of mine roadway dust concentration.

[0096] 1) GNN combines with Transformer for feature extraction. The mine topology structure is modeled by GNN to capture the spatial features of dust diffusion, and the multi-head self-attention mechanism of Transformer is used to model the temporal relationship. Compared with traditional methods, the prediction accuracy is improved by about 10%-15%, effectively guiding mine ventilation and dust removal, reducing prediction errors, and improving air quality.

[0097] 2) Self-supervised learning (SSL) enhances the generalization ability of the model. By adopting the masked autoencoder (MAE) strategy, the model can still learn environmental features in the case of limited or partially missing data, reducing the dependence on manual annotation. Compared with traditional supervised learning methods, the prediction error is reduced by about 12% under different mine working conditions, improving the adaptability of the model.

[0098] 3) GAN is used for data augmentation to improve the robustness of the model in the case of scarce data. By generating synthetic data similar to the real mine environment, the training data set is expanded, reducing the error caused by data imbalance. Compared with the method without data augmentation, the model prediction accuracy is increased by about 9% in extreme environments, enhancing the system stability.

[0099] 4) Reinforcement learning combined with self-supervised learning optimizes the dust control parameters, dynamically adjusts the radial-axial air volume ratio, the extracted air volume, and the distance of the dust control device, enabling the ventilation system to perform real-time optimization and adjustment according to the dust concentration prediction results. Compared with the traditional fixed parameter strategy, the energy consumption can be reduced by about 15%, the dust control efficiency can be increased by more than 13%, enhancing the intelligent level of mine ventilation control and ensuring mine safety.

[0100] 5) Combining cross-validation and self-supervised learning improves the stability of the model in complex mine environments, optimizes the training process, and maintains high stability and low error. Compared with traditional methods, the generalization performance of the model in various mine environments is improved by more than 10%, making it more flexible in dealing with complex environments.

[0101] Example 2: A convolutional neural network (CNN) can be used to grid the mine environment and extract local spatial features, but it may not be able to fully capture the information of complex topological structures.

[0102] Example 3: A recurrent neural network (RNN) or a long short-term memory network (LSTM) can be used for time series modeling, but it may not be as good as Transformer in capturing long-term dependence relationships.

[0103] Example 4: Other self-supervised learning methods such as contrastive learning can be adopted to improve the generalization ability of the model.

[0104] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions, and they should all be covered within the scope of the claims of the present invention.

Claims

1. A method for predicting and controlling mine dust concentration based on self-supervised learning of GNN and Transformer, characterized in that, The method specifically includes the following steps: S1: Obtain multi-modal mine environment parameters and historical dust concentration data, and perform standardization and feature engineering processing on the data; S2: Self-supervised learning pre-training: Use the MAE strategy to partially mask the processed data, and learn the internal structure of the data by reconstructing the masked part; where, MAE represents the Masked Autoencoder; S3: Data augmentation: Use GAN to generate synthetic data similar to the actual mine environment to expand the training set; where, GAN represents the Generative Adversarial Network; S4: Model training: Integrate the spatial features extracted by the GNN module and the temporal features extracted by the Transformer module to generate a comprehensive feature representation; where, GNN represents the Graph Neural Network; S5: Predict the dust concentration: Based on the comprehensive features, predict the dust concentration values at various positions in the mine and generate ventilation control parameter suggestions; S6: Dynamically adjust the mine ventilation system according to the prediction results to achieve dust concentration control.

2. The method for predicting and controlling the mine dust concentration according to claim 1, wherein In step S2, the calculation formula of the loss function of the MAE is as follows: Among them, L MAE represents the masked autoencoder loss, X pred,i represents the masked part predicted by the model, and X ture,i is the true data, and N is the number of training samples.

3. The method for predicting and controlling the concentration of mine dust according to claim 1, characterized in that In step S3, the GAN includes a generator and a discriminator; the function of the generator is to generate synthetic data, and the specific formula is as follows: G(z) = σ(W g z + b g ) Among them, G(z) represents the output of the generator, z represents the random noise vector, W g represents the weight matrix of the generator, b g represents the bias term of the generator, and σ(·) represents the activation function; The function of the discriminator is to discriminate whether the data is real or synthetic, and the specific formula is as follows: D(x) = sigmoid(W d x + b d ) Among them, D(x) represents the output result of the discriminator, x represents the input of the discriminator, W d represents the weight matrix of the discriminator, b d represents the bias term of the discriminator, and sigmoid(·) represents the Sigmoid activation function; Finally, the calculation formula of the GAN loss function is as follows: Among them, L GAN represents the GAN loss, Preal represents the real data distribution, and Pz represents the random noise distribution.

4. The method for predicting and controlling the mine dust concentration according to claim 1, wherein, In step S4, the GNN module constructs a topological structure diagram of the mine roadway, takes the sensor positions as nodes and the roadway connection relationships as edges, and uses GCN and GAT to extract the spatial features of the dust concentration; where, GCN represents the Graph Convolutional Network, and GAT represents the Graph Attention Network.

5. The method for predicting and controlling the mine dust concentration according to claim 4, wherein In step S4, the formula for the GCN to extract spatial features is as follows: Among them, is the feature representation of node v at the l-th layer, N(v) is the set of neighbors of v, u is the position of the sensor directly connected to node v, u ∈ N(v) represents the neighbor node u of node v, W (l) and b (l) are training parameters, and σ is the activation function ReLU; GAT adaptively adjusts the influence of neighbor nodes on the target node through the attention mechanism, enabling the model to pay more attention to neighbor nodes that contribute more to the change in dust concentration. The attention weight α vu is calculated as follows: Among them, LeakyReLU is an improved version of ReLU and is also an activation function; α vu represents the attention weight of neighbor node u to target node v, k is the neighbor node index, α is the attention parameter vector, is the feature of neighbor node u at the l-th layer, and || represents the vector concatenation operation.

6. The method for predicting and controlling the concentration of mine dust according to claim 1, characterized in that, In step S4, the Transformer module processes time series data, adopts the multi-head self-attention mechanism to capture the long-term dependence relationship of the dust concentration; the calculation formula is as follows: Among them, Attention(Q, K, V) represents the result of self-attention calculation, and Q, K, and V represent the query vector, key vector, and value vector respectively; d k is the feature dimension during attention calculation; And uses position encoding to capture time information, so that the Transformer can perceive the time position relationship in the time series; the calculation formula is as follows: Among them, PE (pos,2i) is the value of the dimension with index 2i at the pos-th time step of the position encoding, where i is half of the index of the dimension on which the current position encoding acts, pos represents the time step index, and d model is the model dimension.

7. The method for predicting and controlling the mine dust concentration according to claim 1, wherein In step S4, when extracting spatial features, a convolutional neural network can also be used to perform grid processing on the mine environment to extract local spatial features.

8. The method for predicting and controlling the mine dust concentration according to claim 1, characterized in that In step S4, when extracting temporal features, a recurrent neural network or a long short-term memory network can also be used for temporal modeling.

9. The method for predicting and controlling the mine dust concentration according to claim 1, wherein In step S2, contrastive learning methods can also be used for self-supervised learning.

10. A mine dust concentration prediction and control system based on self-supervised learning of GNN and Transformer for implementing the method according to any one of claims 1 to 9, characterized in that, The system includes an input layer, a self-supervised learning module, a GAN module, a GNN module, a Transformer module, a fusion layer, a hidden layer, and an output layer; Input layer: Used to receive multi-modal mine environment parameters and historical dust concentration data, and perform standardization and feature engineering processing on the data; Self-supervised learning module: Adopt the MAE strategy to learn the internal structure by reconstructing the masked input data; GAN module: Generate synthetic data similar to the actual mine environment; GNN module: Construct a topological structure diagram of the mine roadway, take the sensor positions as nodes and the roadway connection relationships as edges, and extract spatial features through GCN and GAT to capture the dust diffusion pattern; Transformer module: It uses the multi-head self-attention mechanism to process time series data and combines positional encoding to capture the long-term dependencies of dust concentration; Fusion layer: It fuses the spatial features extracted by the GNN module and the temporal features extracted by the Transformer module to form a comprehensive feature representation; Hidden layer: A multi-layer perceptron is constructed on the basis of the fusion layer for deep feature extraction and non-linear mapping; This layer consists of three layers of fully connected neurons, containing 64, 32, and 16 neurons respectively; Each layer uses the ReLU activation function to enhance the non-linear expression ability, and Dropout is combined during training for regularization to prevent overfitting; The final hidden layer further optimizes and integrates the spatial features extracted by the GNN and the temporal features extracted by the Transformer, and transmits them to the output layer; Output layer: It outputs the predicted dust concentration value and the corresponding ventilation control suggestions.

Images