Space structure instability rock burst prediction method and system based on graph neural network

Abstract

The invention discloses a spatial structure instability rock burst prediction method based on a graph neural network. A plurality of multi-source monitoring sensors are arranged at a plurality of key monitoring nodes; collecting multi-source monitoring data to obtain original monitoring data; preprocessing the original monitoring data to obtain processed multi-source monitoring data; regarding each key monitoring node as a node in a graph, and constructing an edge based on a spatial distance or geological association to form a graph structure; wherein the monitoring data of each key monitoring node is used as a node embedding vector; constructing and training a GCN network, and obtaining a GNN prediction model after training; inputting real-time multi-source monitoring data into the graph structure to obtain a node embedding vector; inputting the node embedding vector into a GNN prediction model for risk prediction, and outputting the risk probability at the current moment; and when the risk probability of any node exceeds the limit, the early warning module is controlled to scan and alarm. According to the system and the method, high-precision and low-cost prediction of the rock burst can be realized, and the mine safety is improved.

Description

Technical Field

[0001] This invention belongs to the field of intelligent prediction technology, specifically a method and system for predicting spatial structural instability and rockburst based on graph neural networks, which is applicable to safety monitoring of underground engineering such as mines and tunnels. Background Technology

[0002] Rockbursts are a major geological hazard in mining operations, primarily caused by sudden changes in rock mass stress, which can lead to equipment damage and casualties. Traditional prediction methods rely on physical models or statistical learning, such as threshold alarms based on stress monitoring or time series analysis. However, these methods have limitations: firstly, they struggle to capture the complex dependencies of spatial structures; secondly, they are highly susceptible to environmental interference; and thirdly, they are computationally complex and lack real-time performance. In recent years, machine learning methods such as neural networks have been introduced, but traditional neural networks ignore spatial topology, resulting in insufficient prediction accuracy. Graph Neural Networks (GNNs) can directly process graph-structured data and effectively model the relationships between nodes, providing a new approach to this type of problem. However, existing GNN applications are mostly concentrated in social networks or bioinformatics, and a systematic solution is lacking in the field of mining safety. Therefore, there is an urgent need for a low-cost, high-precision method and system for predicting spatial structural instability rockbursts based on graph neural networks to improve the intelligence level of mine safety monitoring. Summary of the Invention

[0003] To address the problems existing in the prior art, this invention provides a method and system for predicting spatial structural instability rockbursts based on graph neural networks. This system and method can achieve high-precision and low-cost prediction of rockbursts, thereby improving mine safety.

[0004] To achieve the above objectives, this invention provides a method for predicting spatial structural instability and rockburst based on graph neural networks, comprising the following steps: Step 1: Monitoring point deployment and data collection; S11: Deploy multiple multi-source monitoring sensors at multiple key monitoring nodes in the mine space; S12: Collect multi-source monitoring data using multi-source monitoring sensors to obtain raw monitoring data; multi-source monitoring data includes stress data, displacement data, and geological parameters; Step 2: Data preprocessing; The raw monitoring data is preprocessed to obtain processed multi-source monitoring data; Step 3: Graph structure modeling; Each key monitoring node is treated as a node in the graph, and edges are constructed based on spatial distance or geological correlation to form a graph structure; where the monitoring data of each key monitoring node is used as a node embedding vector. Step 4: GNN Model Construction and Training S41: Construct a 2-layer GCN network structure with a hidden layer dimension of 64 and using the ReLU activation function; The graph convolutional layer uses formula (1) to propagate features layer by layer; The output layer uses the Sigmoid function to output the risk probability value of each node. ; (1); In the formula, For the first The node feature matrix of the layer; It is the adjacency matrix of the graph; yes The degree matrix; For the first The trainable weight matrix of the layer; For activation functions; The output after propagation and transformation through the graph convolutional layer is the first... Layer node characteristics; S42: Use historical monitoring data from the past 6 months and corresponding rockburst event labels to train the GCN network, and obtain the GNN prediction model after training; Step 5: Real-time forecasting and risk assessment; S51: Input real-time multi-source monitoring data into a graph structure to obtain node embedding vectors; then input the node embedding vectors into a GNN prediction model for risk prediction, and output the risk probability at the current time. ; S52: Setting Risk Thresholds When the risk probability of any node At that time, the control and early warning module will scan and trigger an alarm.

[0005] As a preferred option, in step two, the preprocessing consists of normalization and denoising. Normalization eliminates dimensional differences between different features, while denoising filters out outliers and interference signals, thereby improving data quality.

[0006] As a preferred option, in step S42 of step four, during the training process, the cross-entropy loss function in formula (2) is used. The prediction error is measured and the parameters are updated through backpropagation using the Adam optimizer until the model converges.

[0007] (2); In the formula, The total number of samples; This is a real label; To predict probabilities.

[0008] As a preferred embodiment, in step five, S52, the risk threshold... The value is 0.7.

[0009] As a preferred option, in step five, S52, the alarm action is an audible and visual alarm, and at the same time, a network alarm reminder message is pushed to the designated communication terminal. Simultaneously, the emergency control system is linked to execute predetermined protective measures.

[0010] As a preferred option, in step three, the edge weights in the graph structure are calculated based on the Euclidean distance between nodes, and the edge weights are obtained according to formula (3). , The weight of ); (3); In the formula, For nodes Between distance, It is a scale parameter; As a preferred option, in step S41 of step four, the risk probability is obtained according to formula (4). ; (4); In the formula, It is a linear combination of the output layer values ​​of the GNN.

[0011] The present invention also provides a spatial structure instability rockburst prediction system based on graph neural networks, which is used to implement a spatial structure instability rockburst prediction method based on graph neural networks, including a data acquisition module, a data processing terminal and an early warning module; The data acquisition module is a multi-source monitoring sensor, which is deployed at key monitoring nodes to collect multi-source monitoring data; The data processing terminal includes a data preprocessing module, a graph structure module, a GNN prediction module, and a decision module. The data preprocessing module is used to preprocess multi-source monitoring data. The graph structure module is used to output node embedding vectors based on the multi-source monitoring data. The GNN prediction module is used to predict risk probabilities based on the node embedding vectors and output risk probability values. The decision module is used to compare the risk probability values ​​with a set risk threshold, and when the risk probability value is greater than the set risk threshold, it sends an early warning signal to the early warning module. The early warning module is used to execute the corresponding alarm action after receiving an early warning signal.

[0012] As a preferred embodiment, the early warning module integrates an audible and visual alarm and a communication module.

[0013] This invention provides a method and system for predicting rockburst risks based on spatial structural instability using graph neural networks, addressing the shortcomings of traditional prediction methods in spatial modeling. Traditional methods, such as stress threshold alarms or time series analysis, often struggle to capture the dependencies within the complex spatial structure of mines, are susceptible to environmental interference, and suffer from computational complexity and poor real-time performance. This invention abstracts the mine's spatial region into a graph structure and utilizes the inherent spatial perception capabilities of graph neural networks (GNNs) to achieve high-precision, real-time prediction of rockburst risks. This method not only improves prediction accuracy but also offers advantages such as simple implementation, strong anti-interference capabilities, and low cost, making it suitable for large-scale mine monitoring networks.

[0014] In the data acquisition step, this invention uses a sensor network (such as stress gauges, displacement gauges, and acoustic emission sensors) deployed in key areas of the mine to collect multidimensional monitoring data in real time, including stress data, displacement data, geological parameters (such as lithology and hardness), and time-series data. Data preprocessing includes normalization and denoising operations to ensure input quality. For example, each monitoring point can collect d-dimensional features, forming a feature matrix X ∈ R^(N×d), where N is the number of monitoring points. This multi-source dataset provides a rich input foundation for subsequent graph construction.

[0015] In the graph structure construction step, this invention divides the mine spatial area into multiple nodes, each corresponding to a physical monitoring point. Node features are composed of monitoring data. Edge sets are constructed based on spatial distance or geological correlation, forming a graph structure G=(V, E), where V is the node set and E is the edge set. Edge construction is typically based on Euclidean distance; if the distance is less than a threshold R, an edge is added. This weighted graph structure effectively captures the spatial proximity and geological similarity between monitoring points, providing topological input for the GNN. The graph construction process transforms the physical monitoring network into a topological graph that can be processed by the GNN, intuitively displaying spatial correlations.

[0016] In the GNN training step, this invention employs architectures such as Graph Convolutional Networks (GCN) for model training. Graph convolutional layers learn spatial dependency patterns by aggregating features of the node itself and its neighbors. The model is trained using historical monitoring data and corresponding rockburst event labels, employing cross-entropy loss as the loss function. Backpropagation and an optimizer (such as Adam) minimize the loss, enabling the model to identify hazardous patterns. Graph convolutional operations, through multi-layer stacking, progressively capture spatial features from local to global perspectives.

[0017] The method of this invention is computationally efficient, can be embedded in edge devices, and can be deployed at low cost. Through spatial modeling using GNNs, the reliability of rockburst prediction is significantly improved. This system and method can achieve high-precision, low-cost prediction of rockbursts, thereby enhancing mine safety. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the overall system architecture and workflow; Figure 2 A diagram illustrating the overall structure of the prediction system and the collaborative workflow between its modules; Figure 3 The actual physical monitoring network of the mine is transformed into the topological graph structure required for a graph neural network. Figure 4 The diagram illustrates the convolution operation. Detailed Implementation

[0019] The invention will now be further described with reference to the accompanying drawings.

[0020] like Figures 1 to 4 As shown, this invention provides a method for predicting spatial structural instability and rockburst based on graph neural networks, comprising the following steps: Step 1: Monitoring point deployment and data collection; S11: Multiple multi-source monitoring sensors are deployed at multiple key monitoring nodes in the mine space (working face and roadway, etc.); Preferably, the multi-source monitoring sensor includes a stress sensor, an acoustic emission monitoring system, and a displacement sensor; S12: Collect multi-source monitoring data using multi-source monitoring sensors to obtain raw monitoring data; the multi-source monitoring data includes stress data. Displacement data Geological parameters such as lithology and hardness; Step 2: Data preprocessing; The raw monitoring data is preprocessed to obtain processed multi-source monitoring data; Step 3: Graph structure modeling; Each key monitoring node is treated as a node in the graph, and edges are constructed based on spatial distance or geological association to form a graph structure G = (V, E), where V is a node and E is an edge; the graph structure is used to convert multi-source monitoring data into graph data, where the monitoring data of each key monitoring node is used as a node embedding vector; Step 4: GNN Model Construction and Training S41: Construct a 2-layer GCN network structure with a hidden layer dimension of 64 and use the ReLU activation function; The graph convolutional layer uses formula (1) to propagate features layer by layer to effectively learn spatial instability modes; The output layer uses the Sigmoid function to output the risk probability value of each node. ; (1); In the formula, For the first The node feature matrix of the layer; It is the adjacency matrix of the graph. ; yes The degree matrix, ; For the first The trainable weight matrix of the layer; For activation functions; The output after propagation and transformation through the graph convolutional layer is the first... Layer node characteristics; ; S42: Use historical monitoring data from the past 6 months and corresponding rockburst event labels to train the GCN network, and obtain the GNN prediction model after training; Step 5: Real-time forecasting and risk assessment; S51: Input real-time multi-source monitoring data into a graph structure to obtain node embedding vectors; then input the node embedding vectors into a GNN prediction model for risk prediction, and output the risk probability at the current time. ; S52: Setting Risk Thresholds When the risk probability of any node At that time, the control and early warning module will scan and trigger an alarm.

[0021] As a preferred option, in step two, the preprocessing consists of normalization and denoising. Normalization eliminates dimensional differences between different features, while denoising filters out outliers and interference signals, thereby improving data quality.

[0022] As a preferred option, in step S42 of step four, during the training process, the cross-entropy loss function in formula (2) is used. The prediction error is measured and the parameters are updated through backpropagation using the Adam optimizer until the model converges.

[0023] (2); In the formula, The total number of samples; The labels are true (1 indicates that a rockburst occurred, and 0 indicates that it did not occur). To predict probabilities.

[0024] As a preferred embodiment, in step five, S52, the risk threshold... The value is 0.7.

[0025] As a preferred option, in step five, S52, the alarm action is an audible and visual alarm, and at the same time, a network alarm reminder message is pushed to the designated communication terminal. Simultaneously, the emergency control system is linked to execute predetermined protective measures.

[0026] As a preferred option, in step three, the edge weights in the graph structure are calculated based on the Euclidean distance between nodes, and the edge weights are obtained according to formula (3). , The weight of ); (3); In the formula, For nodes Between distance, It is a scaling parameter used to control the rate of weight decay, and is usually taken as the average distance; As a preferred option, in step S41 of step four, the risk probability is obtained according to formula (4). ; (4); In the formula, It is a linear combination of the output layer values ​​of the GNN.

[0027] The present invention also provides a spatial structure instability rockburst prediction system based on graph neural networks, which is used to implement a spatial structure instability rockburst prediction method based on graph neural networks, including a data acquisition module, a data processing terminal and an early warning module; The data acquisition module is a multi-source monitoring sensor, which is deployed at key monitoring nodes to collect multi-source monitoring data; The data processing terminal includes a data preprocessing module, a graph structure module, a GNN prediction module, and a decision module. The data preprocessing module is used to preprocess multi-source monitoring data. The graph structure module is used to output node embedding vectors based on the multi-source monitoring data. The GNN prediction module is used to predict risk probabilities based on the node embedding vectors and output risk probability values. The decision module is used to compare the risk probability values ​​with a set risk threshold, and when the risk probability value is greater than the set risk threshold, it sends an early warning signal to the early warning module. The early warning module is used to execute the corresponding alarm action after receiving an early warning signal.

[0028] As a preferred embodiment, the early warning module integrates an audible and visual alarm and a communication module, which simultaneously sends early warning information to the receiving terminal of the monitoring center when an alarm is triggered.

[0029] Example: Taking a mine working face as an example, 100 monitoring points are set up, with a feature dimension d=5 (including stress, displacement, etc.). The graph construction parameter R = 50m. The GNN structure consists of a 2-layer GCN with 64 hidden layers. Training data covers 6 months of historical data, and testing shows a prediction accuracy of 95% and a false positive rate of less than 5%.

[0030] This invention provides a method and system for predicting rockburst risks based on spatial structural instability using graph neural networks, addressing the shortcomings of traditional prediction methods in spatial modeling. Traditional methods, such as stress threshold alarms or time series analysis, often struggle to capture the dependencies within the complex spatial structure of mines, are susceptible to environmental interference, and suffer from computational complexity and poor real-time performance. This invention abstracts the mine's spatial region into a graph structure and utilizes the inherent spatial perception capabilities of graph neural networks (GNNs) to achieve high-precision, real-time prediction of rockburst risks. This method not only improves prediction accuracy but also offers advantages such as simple implementation, strong anti-interference capabilities, and low cost, making it suitable for large-scale mine monitoring networks.

[0031] In the data acquisition step, this invention uses a sensor network (such as stress gauges, displacement gauges, and acoustic emission sensors) deployed in key areas of the mine to collect multidimensional monitoring data in real time, including stress data, displacement data, geological parameters (such as lithology and hardness), and time-series data. Data preprocessing includes normalization and denoising operations to ensure input quality. For example, each monitoring point can collect d-dimensional features, forming a feature matrix X ∈ R^(N×d), where N is the number of monitoring points. This multi-source dataset provides a rich input foundation for subsequent graph construction.

[0032] In the graph structure construction step, this invention divides the mine spatial area into multiple nodes, each corresponding to a physical monitoring point. Node features are composed of monitoring data. Edge sets are constructed based on spatial distance or geological correlation, forming a graph structure G=(V, E), where V is the node set and E is the edge set. Edge construction is typically based on Euclidean distance; if the distance is less than a threshold R, an edge is added. This weighted graph structure effectively captures the spatial proximity and geological similarity between monitoring points, providing topological input for the GNN. The graph construction process transforms the physical monitoring network into a topological graph that can be processed by the GNN, intuitively displaying spatial correlations.

[0033] In the GNN training step, this invention employs architectures such as Graph Convolutional Networks (GCN) for model training. Graph convolutional layers learn spatial dependency patterns by aggregating features of the node itself and its neighbors. The model is trained using historical monitoring data and corresponding rockburst event labels, employing cross-entropy loss as the loss function. Backpropagation and an optimizer (such as Adam) minimize the loss, enabling the model to identify hazardous patterns. Graph convolutional operations, through multi-layer stacking, progressively capture spatial features from local to global perspectives.

[0034] The method of this invention is computationally efficient, can be embedded in edge devices, and can be deployed at low cost. Through spatial modeling using GNNs, the reliability of rockburst prediction is significantly improved. This system and method can achieve high-precision, low-cost prediction of rockbursts, thereby enhancing mine safety.

Claims

1. A method for predicting spatial structural instability and rockburst based on graph neural networks, characterized in that, Includes the following steps: Step 1: Monitoring point deployment and data collection; S11: Deploy multiple multi-source monitoring sensors at multiple key monitoring nodes in the mine space; S12: Collect multi-source monitoring data using multi-source monitoring sensors to obtain raw monitoring data; multi-source monitoring data includes stress data, displacement data, and geological parameters; Step 2: Data preprocessing; The raw monitoring data is preprocessed to obtain processed multi-source monitoring data; Step 3: Graph structure modeling; Each key monitoring node is treated as a node in the graph, and edges are constructed based on spatial distance or geological correlation to form a graph structure; where the monitoring data of each key monitoring node is used as a node embedding vector. Step 4: GNN Model Construction and Training S41: Construct a 2-layer GCN network structure with a hidden layer dimension of 64 and using the ReLU activation function; The graph convolutional layer uses formula (1) to propagate features layer by layer; The output layer uses the Sigmoid function to output the risk probability value of each node. ; （1）； In the formula, For the first The node feature matrix of the layer; It is the adjacency matrix of the graph; yes The degree matrix; For the first The trainable weight matrix of the layer; For activation functions; The output after propagation and transformation through the graph convolutional layer is the first... Layer node characteristics; S42: Use historical monitoring data from the past 6 months and corresponding rockburst event labels to train the GCN network, and obtain the GNN prediction model after training; Step 5: Real-time forecasting and risk assessment; S51: Input real-time multi-source monitoring data into a graph structure to obtain node embedding vectors; then input the node embedding vectors into a GNN prediction model for risk prediction, and output the risk probability at the current time. ; S52: Setting Risk Thresholds When the risk probability of any node At that time, the control and early warning module will scan and trigger an alarm.

2. The method for predicting spatial structural instability and rockburst based on graph neural networks according to claim 1 or 2, characterized in that, In step two, the preprocessing consists of normalization and denoising. Normalization eliminates the dimensional differences between different features, while denoising filters out outliers and interference signals, thereby improving data quality.

3. The method for predicting spatial structural instability and rockburst based on graph neural networks according to claim 2, characterized in that, In step S42 of step four, during the training process, the cross-entropy loss function in formula (2) is used. The prediction error is measured and the parameters are updated by backpropagation through the Adam optimizer until the model converges. （2）； In the formula, The total number of samples; This is a real label; To predict probabilities.

4. The method for predicting spatial structural instability and rockburst based on graph neural networks according to claim 3, characterized in that, In step five, S52, the risk threshold The value is 0.

7.

5. The method for predicting spatial structural instability and rockburst based on graph neural networks according to claim 1, characterized in that, In step S52 of step five, the alarm action is an audible and visual alarm, and at the same time, a network alarm reminder message is pushed to the designated communication terminal. Simultaneously, the emergency control system is activated to execute predetermined protective measures.

6. The method for predicting spatial structural instability and rockburst based on graph neural networks according to claim 1, characterized in that, In step three, the edge weights in the graph structure are calculated based on the Euclidean distance between nodes, and the edge weights are obtained according to formula (3). , The weight of ); （3）； In the formula, For nodes Between distance, It is a scale parameter.

7. The method for predicting spatial structural instability and rockburst based on graph neural networks according to claim 1, characterized in that, In step S41 of step four, the risk probability is obtained according to formula (4). ; （4）； In the formula, It is a linear combination of the output layer values ​​of the GNN.

8. A spatial structural instability rockburst prediction system based on graph neural networks, used to implement the spatial structural instability rockburst prediction method based on graph neural networks as described in any one of claims 1 to 7, characterized in that, It includes a data acquisition module, a data processing terminal, and an early warning module; The data acquisition module is a multi-source monitoring sensor, which is deployed at key monitoring nodes to collect multi-source monitoring data; The data processing terminal includes a data preprocessing module, a graph structure module, a GNN prediction module, and a decision module. The data preprocessing module is used to preprocess multi-source monitoring data. The graph structure module is used to output node embedding vectors based on the multi-source monitoring data. The GNN prediction module is used to predict risk probabilities based on the node embedding vectors and output risk probability values. The decision module is used to compare the risk probability value with a set risk threshold. When the risk probability value is greater than the set risk threshold, it sends an early warning signal to the early warning module. The early warning module is used to execute the corresponding alarm action after receiving an early warning signal.

9. A spatial structural instability rockburst prediction system based on graph neural networks according to claim 8, characterized in that, The early warning module integrates an audible and visual alarm and a communication module.

Images