A dual-network fusion intelligent identification and early warning method for coal mine roof gas disasters

Abstract

The application discloses a double-network fusion intelligent identification and early warning method for coal mine roof gas disasters, and steps are as follows: S1, collecting multi-source monitoring data related to gas and roof disasters of a coal mine working face, establishing an initial data set, and obtaining pretreatment data through abnormal value elimination, missing value interpolation and normalization processing; S2, integrating a graph attention network into a convolutional neural network, extracting time and local features of the monitoring data through the convolutional neural network, constructing a disaster and monitoring index correlation network by means of the graph attention network, learning disaster coupling influence and outputting double-disaster coupling features; S3, constructing a risk probability calculation model based on the coupling features, calculating single-disaster occurrence probability and a comprehensive risk index, and introducing a dynamic updating mechanism to realize dynamic risk assessment; and S4, constructing a grading standard according to the risk probability and the comprehensive risk index, realizing disaster grading early warning and outputting results. The method can improve disaster early warning accuracy and reliability and reduce false alarm and missed alarm rates.

Description

Technical Field

[0001] This invention relates to the field of coal mine safety monitoring and intelligent early warning technology, and in particular to a dual-network fusion intelligent identification and early warning method for coal mine roof gas disasters. Background Technology

[0002] As shallow coal resources dwindle and coal mining extends deeper, the geological environment and mining site structure become increasingly complex. Gas and roof collapse hazards, the two main types of disasters, exhibit a trend of overlapping and coupled development, significantly increasing accident risks and posing a serious threat to personnel safety and production stability. Therefore, systematically analyzing the correlations between multi-source monitoring indicators, the interaction mechanisms between gas and roof collapse hazards, and the characteristic information of the hazard incubation stage becomes a crucial foundation for effective prediction and early warning of coal mine disasters.

[0003] Currently, early warning methods for gas and roof collapse disasters mostly focus on single indicators or single disaster types, relying primarily on threshold judgment, trend analysis, or independent prediction methods based on models such as CNN and LSTM. These methods emphasize single time series characteristics or local data ranges, lacking a comprehensive consideration of the correlation between multiple indicators and the coupling effect between the two disasters. In actual complex working conditions, the triggering mechanisms of gas and roof collapse disasters exhibit significant interactivity and dynamism. For example, roof pressure may lead to abnormal gas outbursts, and gas re-rush and gas control can increase the risk of gas dynamic disasters. Relying solely on single-factor analysis is insufficient to comprehensively depict the disaster evolution process and accurately reflect the potential risk level.

[0004] Furthermore, existing methods still fall short in the collaborative utilization of multi-source information, failing to effectively integrate the correlation characteristics between different monitoring parameters such as gas concentration, roof pressure, and surrounding rock delamination with disaster types. This leads to problems such as insufficient information utilization and incomplete risk expression during risk identification, thus affecting the accuracy and reliability of early warning results. In complex environments, early warning models driven by single disaster types or single indicators are insufficient for identifying and intervening in the comprehensive risks of gas and roof. Summary of the Invention

[0005] To achieve the above objectives, this application adopts the following technical solution:

[0006] This application provides a dual-network fusion intelligent identification and early warning method for coal mine roof gas disasters, including:

[0007] Step S1: Data Acquisition and Preprocessing: Collect multi-source monitoring data related to gas and roof disasters in the coal mine working face, establish an initial dataset, identify outliers in the initial dataset, remove outliers, imput missing values, and normalize the data to form preprocessed data.

[0008] Step S2, Construction of Multi-Disaster Fusion Prediction Model: Integrate graph attention network into convolutional neural network model, and extract time variation features and local pattern features of multi-source monitoring data based on convolutional neural network. Then, construct the association network of gas disaster, roof disaster and corresponding monitoring indicators through graph attention network, and learn the dependency relationship between nodes and the coupling effect of disasters. Finally, output the dual-disaster coupling feature representation.

[0009] Step S3: Construction of the integrated early warning model: Based on the coupling features, a risk probability calculation model is constructed to calculate the probability of gas disaster occurrence, the probability of roof disaster occurrence, and the comprehensive risk index. A dynamic model update mechanism is introduced to realize dynamic risk assessment.

[0010] Step S4, Disaster Classification and Early Warning Output: Based on the single disaster occurrence probability and comprehensive risk index, a risk level classification standard is constructed to realize disaster risk classification and early warning and output the early warning results.

[0011] In some embodiments, the initial dataset in step S1 includes gas concentration data, carbon monoxide concentration data, fully mechanized mining resistance data, surrounding rock delamination data, and time series data of anchor cable stress.

[0012] The collected monitoring data is divided into training set, validation set, and test set;

[0013] Outlier identification and removal adopts the 3σ principle based on statistical distribution to identify abnormal jump points and outliers in each monitoring indicator;

[0014] Missing data imputation adopts the interpolation imputation method based on time series neighborhood; the min-max normalization method is used for normalization processing to uniformly map the data of each monitoring indicator to the interval [0, 1].

[0015] In some embodiments, step S2 includes step 2.1 temporal and local feature extraction based on convolutional neural networks and step 2.2 disaster coupling modeling based on graph attention networks; using the standardized data preprocessed in step S1 as input data, time series local feature extraction is realized based on one-dimensional convolutional neural networks, and multi-disaster coupling relationship modeling is realized based on graph attention networks, thus completing the deep fusion of monitoring features and disaster correlation.

[0016] In some embodiments, step 2.1, which involves temporal and local feature extraction based on a convolutional neural network, is specifically as follows:

[0017] The preprocessed multi-source monitoring data related to roof and gas disasters were constructed into a time window sequence, with each time window containing multiple monitoring indicators; a one-dimensional convolutional neural network was used to extract local features from the sequence data within each time window.

[0018] Let the input data be: Where T is the time window length and N is the number of monitoring indicators; after convolutional layers, a high-dimensional feature representation is obtained:

[0019] ;

[0020] in, These are the convolution kernel parameters; The bias is indicated by *; * indicates a convolution operation; d represents the output feature dimension. Temporal-spatial features extracted by a convolutional neural network.

[0021] In some embodiments, step 2.2, disaster coupling modeling based on graph attention network, is specifically as follows: the main disaster types in coal mines and their corresponding multi-source monitoring data are abstracted into nodes in a graph structure, and a disaster association graph is constructed: G=(V,E);

[0022] Among them, the node set m represents the total number of nodes; specifically, it includes gas disaster nodes, roof disaster nodes, and key monitoring indicator nodes such as gas concentration, fully mechanized mining resistance, surrounding rock delamination, and anchor cable stress.

[0023] The edge set E represents the relationships between nodes, predefined based on data correlation, focusing on characterizing the coupling influence path between gas and the roof, and outputting features using a convolutional neural network. Based on this, an initial feature vector is generated for each graph node through a linear mapping layer, forming a node feature matrix. ;

[0024] The disaster association graph G and the node feature matrix F are input into the graph attention network layer. The influence weights of neighboring nodes on the central node are adaptively learned through the self-attention mechanism to quantify the disaster coupling strength.

[0025] Attention coefficient between node i and its neighbor node j The calculation is as follows:

[0026] ;

[0027] ;;

[0028] in, For node features; For learnable weight matrix, For attention parameter vectors, This represents vector concatenation. Let i be the set of neighboring nodes. These are the normalized attention weights;

[0029] Then, the node features are updated as follows:

[0030] ;

[0031] A multi-head attention mechanism is used to concatenate and aggregate the outputs of K independent attention heads:

[0032] ;

[0033] Where K is the number of attention heads, For splicing operations;

[0034] Finally, the node features output by the graph attention network are... Flattening or global pooling yields the globally coupled feature vector:

[0035] The feature vector Z integrates the temporal variation characteristics of gas and roof multi-source monitoring data and the coupling relationship between the two hazards, and serves as the input for subsequent risk assessment and graded early warning modules.

[0036] In some embodiments, step S3 includes the following steps: step 3.1 single disaster risk probability calculation, step 3.2 comprehensive risk index calculation, and step 3.3 model dynamic update; based on the global coupled feature vector, the single disaster risk probability calculation, gas and roof disaster comprehensive risk index calculation are completed, and an online learning mechanism is introduced to realize the model dynamic update, ensuring the real-time performance and adaptability of the early warning model.

[0037] In some embodiments, step 3.1, the calculation of the single disaster risk probability, is as follows:

[0038] Based on the dual-hazard coupling feature vector Z output in step S2.2, a hazard risk probability calculation model is constructed to achieve comprehensive quantification of the single hazard risk of gas and roof and the coupling influence between the two hazards. A dynamic update mechanism is introduced to enable the early warning results to respond in real time to changes in the mine's safety status.

[0039] Suppose there are C=2 types of disasters. For the c-th type of disaster, what is the probability of the c-th type of disaster occurring? The calculation formula is:

[0040] ;

[0041] in, , For learnable weights, , For bias, It is a Sigmoid activation function, with an output range of [0,1]. This represents the probability of the c-th type of disaster occurring. The higher the value, the higher the risk of the disaster.

[0042] In some embodiments, step 3.2, the calculation of the comprehensive risk index, is as follows:

[0043] There are mutual induction and coupling enhancement effects between gas and roof disasters, requiring the construction of a coupling influence matrix. Quantify the intensity of the mutual influence between two disasters;

[0044] The coupling effect can be seen in the attention weights output by the graph attention layer. By aggregating the results, independent coupled learning modules can be designed:

[0045] ;

[0046] in, The node features are those corresponding to disaster i and disaster j; MLP stands for Multilayer Perceptron. This represents the coupling influence coefficient of disaster j on disaster i;

[0047] Taking into account the basic risks of gas and roof respectively, as well as the coupling effect between the two hazards, a comprehensive risk index is constructed. The formula for calculating the comprehensive risk index R is:

[0048] ;

[0049] Alternatively, a weighted summation method can be used:

[0050] ;

[0051] in, For adjustable weighting coefficients, satisfying Overall Risk Index The higher the value, the higher the overall risk of gas and roof. Let i be the probability of disaster i occurring. Let be the probability of disaster j occurring.

[0052] In some embodiments, step 3.3, dynamic model update, is specifically as follows:

[0053] To adapt to changes in mine operating conditions and drift in monitoring data distribution, an online learning and incremental update mechanism is introduced to achieve dynamic correction of the risk probability model; during the training phase, a weighted cross-entropy loss function is used, with a focus on high-risk samples.

[0054] ;

[0055] in, For the true label of the c-th type of disaster; These are class weights used to address the problem of imbalanced samples.

[0056] During the model deployment phase, new batches of real-time monitoring data are used. To update samples, sliding window retraining or online gradient update mechanisms are used to perform real-time parameter fine-tuning:

[0057] ;

[0058] in, These are the current model parameters. For learning rate, For new batches of monitoring data within the current time window, the model is updated periodically or triggered to adapt to new data distributions and disaster evolution patterns.

[0059] In some embodiments, step S4 includes:

[0060] Step 4.1: Construct a hierarchical threshold system;

[0061] For each type of disaster risk probability With comprehensive risk index Four risk levels are set, corresponding to three judgment thresholds:

[0062] As the three risk thresholds for the c-th type of disaster, satisfying ;

[0063] Based on the comprehensive risk index c=2 represents the number of disaster types, which are divided into four comprehensive risk levels;

[0064] As three thresholds for the comprehensive risk index, satisfying ;

[0065] To ensure the scientific rigor and practicality of risk level classification, a multi-objective optimization method is employed to optimize the single-hazard threshold. With comprehensive threshold Perform joint optimization;

[0066] With the goal of maximizing early warning accuracy and minimizing false alarm and false negative rates, a weighted comprehensive loss function is constructed as follows:

[0067] ;

[0068] Accuracy refers to the accuracy of the early warning system. , , Let be the weighting coefficient, satisfying + + =1, these coefficients can be dynamically adjusted according to mine safety management requirements; FAR is the false alarm rate, MAR is the missed alarm rate;

[0069] The threshold space is optimized using grid search, Bayesian optimization, or particle swarm optimization algorithms.

[0070] ;

[0071] The following constraints must be met during the optimization process:

[0072] ;

[0073] By using historical coal mine accident records and monitoring data for the corresponding time periods, a labeled sample set is constructed. The risk status within a certain time window before the accident occurs is labeled as a high-risk positive sample, and the rest of the time period is labeled as a negative sample, which is used for threshold optimization training.

[0074] Step 4.2: Dynamic threshold adjustment;

[0075] To adapt to the different risk characteristics of different mines, mining areas, and mining stages, a dynamic threshold adjustment mechanism is introduced:

[0076] Step 4.2.1: Optimize and obtain a set of global baseline thresholds based on historical data: ;

[0077] Step 4.2.2: Based on the risk distribution characteristics of recent monitoring data, dynamically shift the baseline threshold:

[0078] ;

[0079] in, This is the offset. The optimal dynamic threshold for disaster type c at the k-th warning level;

[0080] offset Determined by the difference between the moving average of recent risk probabilities and the baseline distribution:

[0081] ;

[0082] in, This represents the average risk probability of the c-th type of disaster within the recent time window. Historical benchmark average wind

[0083] Risk probability; To adjust the step size coefficient;

[0084] When an actual accident occurs, the threshold reset mechanism is automatically triggered, and the threshold is re-optimized using the latest accident data to achieve a closed loop of experience feedback.

[0085] Step 4.3: Generation and tiered delivery of early warning information;

[0086] The system simultaneously outputs single-hazard graded early warning and combined two-hazard graded early warning:

[0087] Single disaster early warning mainly displays the current risk level, risk probability, and main contributing monitoring indicators of gas and roof disasters;

[0088] The comprehensive early warning system can display the overall risk level of the mine, the comprehensive risk index, the dominant disaster type, and the sources of coupled impacts;

[0089] The system employs attention weight visualization and feature contribution analysis to provide interpretable evidence for early warning results, displaying the attention weights among disaster nodes in the graph attention network. It presents the intensity of disaster coupling; and the heat map of convolutional neural network kernel activation to locate high-risk periods and key monitoring indicators; and realizes multi-terminal collaborative push, with risk level classification pushed to different management levels.

[0090] Compared with existing technologies, this patented method has significant advantages: by fusing multi-source monitoring data, it effectively compensates for the shortcomings of insufficient information utilization in existing technologies and improves the completeness of disaster information expression; by using convolutional neural networks to extract complex temporal features, it solves the problem of incomplete disaster feature capture in existing early warning methods and greatly improves disaster identification capabilities; by constructing a disaster coupling relationship model through graph attention networks, it realizes the correlation perception between roof and gas disasters and overcomes the shortcomings of existing technologies that do not consider the mutual influence between disasters; by constructing a fusion risk probability model, it can quantitatively describe disaster risks and solve the problem that traditional methods cannot accurately quantify risks; by realizing collaborative disaster early warning and hierarchical identification, it effectively reduces the false alarm rate and missed alarm rate of early warning, and ultimately improves the overall accuracy of coal mine disaster risk prediction and the reliability of early warning, meeting the safety control needs of deep coal mining. Attached Figure Description

[0091] Figure 1 A flowchart of a dual-network fusion intelligent identification and early warning method for coal mine roof gas disasters provided in an embodiment of this application;

[0092] Figure 2 A flowchart illustrating data acquisition and preprocessing provided for embodiments of this application;

[0093] Figure 3 A schematic diagram of the construction architecture of the multi-hazard fusion prediction model for roof and gas disasters provided in the embodiments of this application;

[0094] Figure 4 A schematic diagram of the architecture of the roof and gas disaster fusion early warning model provided for embodiments of this application; Detailed Implementation

[0095] In the embodiments of this application, the terms "exemplary" or "for example" are used to indicate examples, illustrations, or descriptions. Any embodiment or design described as "exemplary" or "for example" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of terms such as "exemplary" or "for example" is intended to present the relevant concepts in a specific manner.

[0096] In the embodiments of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined with "first" and "second" may explicitly or implicitly include one or more of that feature.

[0097] In the description of the embodiments of this application, the term "at least one" means one or more, and "more than one" means two or more. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of a single item or a plurality of items. For example, at least one of a, b, or c can mean: a, b, c, a and b, a and c, b and c, or a and b and c, where a, b, and c can be a single item or a plurality of items.

[0098] In the description of the embodiments of this application, the term "and / or" refers to and covers any and all possible combinations of one or more of the associated listed items. The term "and / or" describes an association relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this application generally indicates that the preceding and following related objects have an "or" relationship.

[0099] In the description of the embodiments of this application, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, "linking" can be a detachable connection or a non-detachable connection; it can be a direct connection or an indirect connection through an intermediate medium. "Fixed connection" refers to a connection where the relative positional relationship remains unchanged after the connection. Furthermore, the directional terms mentioned in the embodiments of this application, such as "inner" and "outer," are only for reference to the directions in the accompanying drawings. Therefore, the directional terms used are for better and clearer explanation and understanding of the embodiments of this application, and are not intended to indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the embodiments of this application.

[0100] In the description of embodiments of this application, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0101] Please see Figure 1 , Figure 2 , Figure 1 Dual-network fusion intelligence for coal mine roof gas disasters provided in embodiments of this application

[0102] Flowchart of the identification and early warning method; Figure 2 A flowchart illustrating data acquisition and preprocessing provided for embodiments of this application;

[0103] This application provides a dual-network fusion intelligent identification and early warning method for coal mine roof gas disasters, which includes the following steps:

[0104] Step S1: Data Acquisition and Preprocessing: Collect multi-source monitoring data related to gas and roof disasters in the coal mine working face, establish an initial dataset, identify outliers in the initial dataset, remove outliers, imput missing values, and normalize the data to form preprocessed data.

[0105] Specifically, the initial dataset includes gas concentration data, carbon monoxide concentration data, fully mechanized mining resistance data, surrounding rock delamination data, and time series data of anchor cable stress.

[0106] The collected monitoring data is divided into training set, validation set, and test set.

[0107] Outlier identification and removal employs the 3σ principle based on statistical distribution to identify anomalous jumps and outliers in each monitoring indicator. Specifically, for the time series data of any monitoring indicator, its mean μ and standard deviation σ are calculated. Data points outside the interval (μ-3σ, μ+3σ) are identified as outliers, uniformly marked, and removed. The resulting time series gaps are then used for subsequent missing data imputation.

[0108] Missing data imputation employs a time-series neighborhood-based interpolation method; min-max normalization is used for normalization, mapping all monitoring indicator data to the [0, 1] interval. Normalization is a mature technique, which will be briefly described here.

[0109] For any monitoring indicator sequence x, its normalized calculation formula is: ;

[0110] in, and These are the minimum and maximum values ​​of the indicator in the historical training set, respectively. The normalized data has no dimensional bias and can be directly used as the standard input for the fusion prediction model.

[0111] Please see Figure 3 , Figure 3 A schematic diagram of the architecture for constructing a multi-hazard fusion prediction model for roof and gas disasters provided in the embodiments of this application; Step S2, construction of multi-hazard fusion prediction model: Integrating graph attention network (GAT) into convolutional neural network model, and extracting time variation features and local pattern features of multi-source monitoring data based on convolutional neural network (CNN), and then constructing a correlation network of gas disaster, roof disaster and corresponding monitoring indicators through graph attention network, thereby learning the dependency relationship between nodes and the coupling effect of disasters, and finally outputting the dual-hazard coupling feature representation.

[0112] Step S2 includes step 2.1, temporal and local feature extraction based on convolutional neural networks, and step 2.2, disaster coupling modeling based on graph attention networks. Using the standardized data preprocessed in step S1 as input data, a one-dimensional convolutional neural network is used to extract local features from the time series data, and a graph attention network is used to model the coupling relationship between multiple disasters, thus achieving a deep fusion of monitoring features and disaster correlation.

[0113] Specifically, step 2.1, the temporal and local feature extraction based on the convolutional neural network, is as follows:

[0114] The preprocessed multi-source monitoring data related to roof and gas disasters were constructed into a time window sequence, with each time window containing multiple monitoring indicators; a one-dimensional convolutional neural network was used to extract local features from the sequence data within each time window.

[0115] Let the input data be: Where T is the length of the time window and N is the number of monitoring indicators. Represents the set of real numbers; after nonlinear mapping using the ReLU activation function, it outputs high-dimensional deep features: ;

[0116] in, Here are the convolution kernel parameters; The bias is indicated by *; * indicates a convolution operation; d represents the output feature dimension. Temporal-spatial features extracted by a convolutional neural network.

[0117] Step 2.2 Disaster Coupling Modeling Based on Graph Attention Network is as follows: The main disaster types in coal mines and their corresponding multi-source monitoring data are abstracted into nodes in a graph structure, and a disaster association graph is constructed: G=(V,E);

[0118] Among them, the node set m represents the total number of nodes; specifically, it includes gas disaster nodes, roof disaster nodes, and key monitoring indicator nodes such as gas concentration, fully mechanized mining resistance, surrounding rock delamination, and anchor cable stress.

[0119] The edge set E represents the relationships between nodes, predefined based on data correlation, focusing on characterizing the coupling influence path between gas and the roof, and outputting features using a convolutional neural network. Based on this, an initial feature vector is generated for each graph node through a linear mapping layer, forming a node feature matrix. ;

[0120] The disaster association graph G and the node feature matrix F are input into the graph attention network layer. The influence weights of neighboring nodes on the central node are adaptively learned through the self-attention mechanism to quantify the disaster coupling strength.

[0121] Attention coefficient between node i and its neighbor node j The calculation is as follows:

[0122] ; ;

[0123] in, For node features; For learnable weight matrix, For attention parameter vectors, This represents vector concatenation. Let i be the set of neighboring nodes. These are the normalized attention weights;

[0124] Then, the node features are updated as follows: ;

[0125] A multi-head attention mechanism is used to concatenate and aggregate the outputs of K independent attention heads:

[0126] ;

[0127] Where K is the number of attention heads, For splicing operations;

[0128] Finally, the node features output by the graph attention network are... Flattening or global pooling yields the globally coupled feature vector: ;

[0129] The feature vector Z integrates the temporal variation characteristics of gas and roof multi-source monitoring data and the coupling relationship between the two hazards, and serves as the input for subsequent risk assessment and graded early warning modules.

[0130] Please see Figure 4 , Figure 4 This is a schematic diagram of the architecture of the roof and gas disaster fusion early warning model provided in the embodiments of this application. Step S3: Construction of the fusion early warning model: Based on the coupling characteristics, a risk probability calculation model is constructed to calculate the probability of gas disaster occurrence, the probability of roof disaster occurrence, and the comprehensive risk index. A dynamic model update mechanism is introduced to realize dynamic risk assessment.

[0131] Step S3 includes the following steps: Step 3.1 Single disaster risk probability calculation, Step 3.2 Comprehensive risk index calculation, and Step 3.3 Dynamic model update; Based on the global coupled feature vector, the single disaster risk probability calculation, gas and roof disaster comprehensive risk index calculation are completed, and an online learning mechanism is introduced to realize dynamic model update, ensuring the real-time performance and adaptability of the early warning model.

[0132] Step 3.1 The calculation of single disaster risk probability is as follows:

[0133] Based on the dual-hazard coupling feature vector Z output in step S2.2, a hazard risk probability calculation model is constructed to comprehensively quantify the single-hazard risks of gas and roof collapse, as well as the coupling effects between the two hazards. A dynamic update mechanism is introduced to enable the early warning results to respond in real time to changes in the mine's safety status. Assuming there are C=2 hazard types, for the c-th hazard, the probability of occurrence of the c-th hazard is... The calculation formula is: ;

[0134] in, , For learnable weights, , For bias, It is a Sigmoid activation function, with an output range of [0,1]. This represents the probability of the c-th type of disaster occurring. The higher the value, the higher the risk of the disaster.

[0135] Step 3.2 The calculation of the comprehensive risk index is as follows:

[0136] There are mutual induction and coupling enhancement effects between gas and roof disasters, requiring the construction of a coupling influence matrix. Quantify the intensity of the mutual influence between two disasters;

[0137] The coupling effect can be seen in the attention weights output by the graph attention layer. By aggregating the results, independent coupled learning modules can be designed: ;

[0138] in, The node features are those corresponding to disaster i and disaster j; MLP stands for Multilayer Perceptron. This represents the coupling influence coefficient of disaster j on disaster i;

[0139] Taking into account the basic risks of gas and roof respectively, as well as the coupling effect between the two hazards, a comprehensive risk index is constructed. The formula for calculating the comprehensive risk index R is: ;

[0140] Alternatively, a weighted summation method can be used: ;

[0141] in, For adjustable weighting coefficients, satisfying Overall Risk Index The higher the value, the higher the overall risk of gas and roof. Let i be the probability of disaster i occurring. Let be the probability of disaster j occurring.

[0142] Step 3.3, the dynamic update of the model, is as follows:

[0143] To adapt to changes in mine operating conditions and drift in monitoring data distribution, an online learning and incremental update mechanism is introduced to achieve dynamic correction of the risk probability model; during the training phase, a weighted cross-entropy loss function is used, with a focus on high-risk samples. ;

[0144] in, For the true label of the c-th type of disaster; These are class weights used to address the problem of imbalanced samples.

[0145] During the model deployment phase, new batches of real-time monitoring data are used. To update samples, sliding window retraining or online gradient update mechanisms are used to perform real-time parameter fine-tuning: ;

[0146] in, These are the current model parameters. For learning rate, For new batches of monitoring data within the current time window, the model is updated periodically or triggered to adapt to new data distributions and disaster evolution patterns.

[0147] Comprehensive risk index based on the current moment Single disaster risk probability It outputs real-time early warning information.

[0148] when At that time, a single warning for the c-th type of disaster is triggered;

[0149] when When this occurs, a comprehensive gas-roof warning is triggered. The warning information includes explanatory details such as the risk level, the dominant hazard type (gas-dominant / roof-dominant / dual-hazard coupled-dominant), and the sources of coupled impacts.

[0150] in As a dynamically adjustable risk threshold, it needs to be calibrated based on historical accident data and expert experience.

[0151] Step S4, Disaster Classification and Early Warning Output: Based on the single disaster occurrence probability and comprehensive risk index, a risk level classification standard is constructed to realize disaster risk classification and early warning and output the early warning results.

[0152] Specifically, step S4 includes step 4.1 constructing a hierarchical threshold system, step 4.2 dynamic threshold adjustment, and step 4.3 generating and hierarchically pushing early warning information.

[0153] Step 4.1 Construct a hierarchical threshold system: for each type of disaster risk probability With comprehensive risk index Four risk levels are set, corresponding to three judgment thresholds, as shown in Table 1 below:

[0154]

[0155] Level 4 risk level corresponds to Level 3 discrimination threshold. For the three risk thresholds of the c-th type of disaster, satisfying ;

[0156] Based on the comprehensive risk index c=2 represents the number of disaster types, which are divided into four comprehensive risk levels, as shown in Table 2 below:

[0157]

[0158] in, As three thresholds for the comprehensive risk index, satisfying ;

[0159] To ensure the scientific rigor and practicality of risk level classification, a multi-objective optimization method is employed to optimize the single-hazard threshold. With comprehensive threshold Perform joint optimization;

[0160] With the goal of maximizing early warning accuracy and minimizing false alarm and false negative rates, a weighted comprehensive loss function is constructed as follows:

[0161] ;

[0162] Accuracy refers to the accuracy of the early warning system. Let be the weighting coefficient, satisfying These coefficients can be dynamically adjusted according to mine safety management requirements; FAR is the false alarm rate, and MAR is the false alarm rate.

[0163] The threshold space is optimized using grid search, Bayesian optimization, or particle swarm optimization algorithms.

[0164] ;

[0165] The following constraints must be met during the optimization process: ;

[0166] By using historical coal mine accident records and monitoring data for the corresponding time periods, a labeled sample set is constructed. The risk status within a certain time window before the accident occurs is labeled as a high-risk positive sample, and the rest of the time period is labeled as a negative sample, which is used for threshold optimization training.

[0167] Step 4.2: Dynamic threshold adjustment;

[0168] To adapt to the different risk characteristics of different mines, mining areas, and mining stages, a dynamic threshold adjustment mechanism is introduced:

[0169] Step 4.2.1: Optimize and obtain a set of global baseline thresholds based on historical data:

[0170] Step 4.2.2: Based on the risk distribution characteristics of recent monitoring data, dynamically shift the baseline threshold:

[0171] ;

[0172] in, This is the offset. The optimal dynamic threshold for disaster type c at the k-th warning level;

[0173] offset Determined by the difference between the moving average of recent risk probabilities and the baseline distribution: ;

[0174] in, This represents the average risk probability of the c-th type of disaster within the recent time window. Historical benchmark average wind

[0175] Risk probability; To adjust the step size coefficient;

[0176] When an actual accident occurs, the threshold reset mechanism is automatically triggered, and the threshold is re-optimized using the latest accident data to achieve a closed loop of experience feedback.

[0177] Step 4.3: Generation and tiered delivery of early warning information;

[0178] The system simultaneously outputs single-hazard graded early warning and combined two-hazard graded early warning:

[0179] Single disaster early warning mainly displays the current risk level, risk probability, and main contributing monitoring indicators of gas and roof disasters;

[0180] The comprehensive early warning system can display the overall risk level of the mine, the comprehensive risk index, the dominant disaster type, and the sources of coupled impacts;

[0181] The system employs attention weight visualization and feature contribution analysis to provide interpretable evidence for early warning results, displaying the attention weights among disaster nodes in the graph attention network. It presents the disaster coupling intensity; and uses a heatmap of convolutional neural network kernel activation to locate high-risk periods and key monitoring indicators; and enables multi-terminal collaborative push, with risk level-based push notifications to different management levels. See Table 3 below:

[0182]

[0183] At the same time, it is also necessary to establish an evaluation index system for the early warning effect, and to continuously track and optimize the performance of the hierarchical early warning model, as shown in Table 4 below:

[0184]

[0185] Based on the evaluation results, closed-loop optimization is performed periodically or triggered (after an incident) on the feature extraction module, coupled modeling module, risk probability model, and threshold classification criteria.

[0186] The early warning method provided in the embodiments of this application improves the completeness of disaster information expression by fusing multi-source monitoring data; improves disaster identification capability by extracting complex temporal features using convolutional neural networks; establishes a disaster coupling relationship model based on graph attention network to realize the perception of roof and gas disaster association; constructs a fusion risk probability model that can quantitatively describe the degree of disaster risk; realizes collaborative disaster early warning and hierarchical identification, reduces false alarm rate and false alarm rate, and thus improves the overall accuracy of coal mine disaster risk prediction and early warning reliability.

[0187] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

Claims

1. A dual-network fusion intelligent identification and early warning method for coal mine roof gas disasters, characterized in that, include: Step S1: Data Acquisition and Preprocessing: Collect multi-source monitoring data related to gas and roof disasters in the coal mine working face, establish an initial dataset, identify outliers in the initial dataset, remove outliers, imput missing values, and normalize the data to form preprocessed data. Step S2, Construction of Multi-Disaster Fusion Prediction Model: Integrate graph attention network into convolutional neural network model, and extract time variation features and local pattern features of multi-source monitoring data based on convolutional neural network. Then, construct the association network of gas disaster, roof disaster and corresponding monitoring indicators through graph attention network, and learn the dependency relationship between nodes and the coupling effect of disasters. Finally, output the dual-disaster coupling feature representation. Step S3: Construction of the integrated early warning model: Based on the coupling features, a risk probability calculation model is constructed to calculate the probability of gas disaster occurrence, the probability of roof disaster occurrence, and the comprehensive risk index. A dynamic model update mechanism is introduced to realize dynamic risk assessment. Step S4, Disaster Classification and Early Warning Output: Based on the single disaster occurrence probability and comprehensive risk index, a risk level classification standard is constructed to realize disaster risk classification and early warning and output the early warning results.

2. The dual-network fusion intelligent identification and early warning method for coal mine roof gas disasters according to claim 1, characterized in that, The initial dataset mentioned in step S1 includes gas concentration data, carbon monoxide concentration data, fully mechanized mining resistance data, surrounding rock delamination data, and time series data of anchor cable stress. The collected monitoring data is divided into training set, validation set, and test set; Outlier identification and removal adopts the 3σ principle based on statistical distribution to identify abnormal jump points and outliers in each monitoring indicator; Missing data imputation adopts the interpolation imputation method based on time series neighborhood; the min-max normalization method is used for normalization processing to uniformly map the data of each monitoring indicator to the interval [0, 1].

3. The dual-network fusion intelligent identification and early warning method for coal mine roof gas disasters according to claim 1, characterized in that, Step S2 includes step 2.1 temporal and local feature extraction based on convolutional neural network and step 2.2 disaster coupling modeling based on graph attention network; using the standardized data preprocessed in step S1 as input data, time series local feature extraction is realized based on one-dimensional convolutional neural network, and multi-disaster coupling relationship modeling is realized based on graph attention network, thus completing the deep fusion of monitoring features and disaster correlation.

4. The dual-network fusion intelligent identification and early warning method for coal mine roof gas disasters according to claim 3, characterized in that, Step 2.1, temporal and local feature extraction based on convolutional neural networks, is as follows: the preprocessed multi-source monitoring data related to roof and gas disasters are constructed into a time window sequence, with each time window containing multiple monitoring indicators; a one-dimensional convolutional neural network is used to extract local features from the sequence data within each time window; Let the input data be: Where T is the time window length and N is the number of monitoring indicators; after convolutional layers, a high-dimensional feature representation is obtained: ; in, These are the convolution kernel parameters; The bias is indicated by *; * indicates a convolution operation; d represents the output feature dimension. Temporal-spatial features extracted by a convolutional neural network.

5. The dual-network fusion intelligent identification and early warning method for coal mine roof gas disasters according to claim 4, characterized in that, The specific steps of step 2.2, disaster coupling modeling based on graph attention network, are as follows: the main disaster types in coal mines and their corresponding multi-source monitoring data are abstracted into nodes in a graph structure, and a disaster association graph is constructed: G=(V,E); Among them, the node set m represents the total number of nodes; specifically, it includes gas disaster nodes, roof disaster nodes, and key monitoring indicator nodes such as gas concentration, fully mechanized mining resistance, surrounding rock delamination, and anchor cable stress. The edge set E represents the relationships between nodes, predefined based on data correlation, focusing on characterizing the coupling influence path between gas and the roof, and outputting features using a convolutional neural network. Based on this, an initial feature vector is generated for each graph node through a linear mapping layer, forming a node feature matrix. ; The disaster association graph G and the node feature matrix F are input into the graph attention network layer. The influence weights of neighboring nodes on the central node are adaptively learned through the self-attention mechanism to quantify the disaster coupling strength. Attention coefficient between node i and its neighbor node j The calculation is as follows: ; ; in, For node features; For learnable weight matrix, For attention parameter vectors, This represents vector concatenation. Let be the set of neighboring nodes of node i. These are the normalized attention weights; Then, the node features are updated as follows: ; A multi-head attention mechanism is used to concatenate and aggregate the outputs of K independent attention heads: ; Where K is the number of attention heads, For splicing operations; Finally, the node features output by the graph attention network are... Flattening or global pooling yields the globally coupled feature vector: ; The feature vector Z integrates the temporal variation characteristics of gas and roof multi-source monitoring data and the coupling relationship between the two hazards, and serves as the input for subsequent risk assessment and graded early warning modules.

6. The dual-network fusion intelligent identification and early warning method for coal mine roof gas disasters according to claim 1, characterized in that, Step S3 includes the following steps: Step 3.1 Single disaster risk probability calculation, Step 3.2 Comprehensive risk index calculation, and Step 3.3 Dynamic model update. Based on the global coupled feature vector, the single disaster risk probability calculation, gas and roof disaster comprehensive risk index calculation are completed, and an online learning mechanism is introduced to realize dynamic model update, ensuring the real-time performance and adaptability of the early warning model.

7. The dual-network fusion intelligent identification and early warning method for coal mine roof gas disasters according to claim 6, characterized in that, Step 3.1 The calculation of single disaster risk probability is as follows: Based on the dual-hazard coupling feature vector Z output in step S2.2, a hazard risk probability calculation model is constructed to achieve comprehensive quantification of the single hazard risk of gas and roof and the coupling influence between the two hazards. A dynamic update mechanism is introduced to enable the early warning results to respond in real time to changes in the mine's safety status. Suppose there are C=2 types of disasters. For the c-th type of disaster, what is the probability of the c-th type of disaster occurring? The calculation formula is: ; in, , For learnable weights, , For bias, It is a Sigmoid activation function, with an output range of [0,1]. This represents the probability of the c-th type of disaster occurring. The higher the value, the higher the risk of the disaster.

8. The dual-network fusion intelligent identification and early warning method for coal mine roof gas disasters according to claim 7, characterized in that, Step 3.2 The calculation of the comprehensive risk index is as follows: There are mutual induction and coupling enhancement effects between gas and roof disasters, requiring the construction of a coupling influence matrix. Quantify the intensity of the mutual influence between two disasters; The coupling effect can be seen in the attention weights output by the graph attention layer. By aggregating the results, independent coupled learning modules can be designed: ; in, The node features are those corresponding to disaster i and disaster j; MLP stands for Multilayer Perceptron. This represents the coupling influence coefficient of disaster j on disaster i; Taking into account the basic risks of gas and roof respectively, as well as the coupling effect between the two hazards, a comprehensive risk index is constructed. The formula for calculating the comprehensive risk index R is: ; Alternatively, a weighted summation method can be used: ; in, For adjustable weighting coefficients, satisfying Overall Risk Index The higher the value, the higher the overall risk of gas and roof. Let i be the probability of disaster i occurring. Let be the probability of disaster j occurring.

9. The dual-network fusion intelligent identification and early warning method for coal mine roof gas disasters according to claim 8, characterized in that, Step 3.3, the dynamic update of the model, is as follows: To adapt to changes in mine operating conditions and drift in monitoring data distribution, an online learning and incremental update mechanism is introduced to achieve dynamic correction of the risk probability model; during the training phase, a weighted cross-entropy loss function is used, with a focus on high-risk samples. ; in, For the true label of the c-th type of disaster; These are class weights used to address the problem of imbalanced samples. During the model deployment phase, new batches of real-time monitoring data are used. To update samples, sliding window retraining or online gradient update mechanisms are used to perform real-time parameter fine-tuning: ; in, These are the current model parameters. For learning rate, For new batches of monitoring data within the current time window, the model is updated periodically or triggered to adapt to new data distributions and disaster evolution patterns.

10. The dual-network fusion intelligent identification and early warning method for coal mine roof gas disasters according to claim 1, characterized in that, Step S4 includes: Step 4.1: Construct a hierarchical threshold system; For each type of disaster risk probability With comprehensive risk index Four risk levels are set, corresponding to three judgment thresholds: As the three risk thresholds for the c-th type of disaster, satisfying ; Based on the comprehensive risk index c=2 represents the number of disaster types, which are divided into four comprehensive risk levels; As three thresholds for the comprehensive risk index, satisfying ; To ensure the scientific rigor and practicality of risk level classification, a multi-objective optimization method is employed to optimize the single-hazard threshold. With comprehensive threshold Perform joint optimization; With the goal of maximizing early warning accuracy and minimizing false alarm and false negative rates, a weighted comprehensive loss function is constructed as follows: ; Accuracy refers to the accuracy of the early warning system. , , Let be the weighting coefficient, satisfying + + =1, these coefficients can be dynamically adjusted according to mine safety management requirements; FAR is the false alarm rate, MAR is the missed alarm rate; The threshold space is optimized using grid search, Bayesian optimization, or particle swarm optimization algorithms. ; The following constraints must be met during the optimization process: ; By using historical coal mine accident records and monitoring data for the corresponding time periods, a labeled sample set is constructed. The risk status within a certain time window before the accident occurs is labeled as a high-risk positive sample, and the rest of the time period is labeled as a negative sample, which is used for threshold optimization training. Step 4.2: Dynamic threshold adjustment; To adapt to the different risk characteristics of different mines, mining areas, and mining stages, a dynamic threshold adjustment mechanism is introduced: Step 4.2.1: Optimize and obtain a set of global baseline thresholds based on historical data: ; Step 4.2.2: Based on the risk distribution characteristics of recent monitoring data, dynamically shift the baseline threshold: ; in, This is the offset. The optimal dynamic threshold for disaster type c at the k-th warning level; offset Determined by the difference between the moving average of recent risk probabilities and the baseline distribution: ;in, This represents the average risk probability of the c-th type of disaster within the recent time window. The historical average risk probability; To adjust the step size coefficient; When an actual accident occurs, the threshold reset mechanism is automatically triggered, and the threshold is re-optimized using the latest accident data to achieve a closed loop of experience feedback. Step 4.3: Generation and tiered delivery of early warning information; The system simultaneously outputs single-hazard graded early warning and combined two-hazard graded early warning: Single disaster early warning mainly displays the current risk level, risk probability, and main contributing monitoring indicators of gas and roof disasters; The comprehensive early warning system can display the overall risk level of the mine, the comprehensive risk index, the dominant disaster type, and the sources of coupled impacts; The system employs attention weight visualization and feature contribution analysis to provide interpretable evidence for early warning results, displaying the attention weights among disaster nodes in the graph attention network. It presents the intensity of disaster coupling; and the heat map of convolutional neural network kernel activation to locate high-risk periods and key monitoring indicators; and realizes multi-terminal collaborative push, with risk level classification pushed to different management levels.

Images