Intelligent sensing early warning and linkage emergency system for multiple disasters in coal mine

Abstract

The invention provides a coal mine multi-disaster intelligent sensing early warning and linkage emergency system, and relates to the technical field of coal mine safety, multi-source data collected by a hollow probe sensor group is uploaded to an edge gateway through an industrial ring network for primary processing, and the multi-source data after primary processing by the edge gateway is uploaded to a data layer through the industrial ring network; the processed multi-source data is classified by a time sequence database and a space database of the data layer, the classified multi-source data is analyzed by the analysis layer, and an analysis result of the analysis layer is uploaded to the early warning layer. According to the invention, three-in-one monitoring of deep part-surface-environment is realized through the hollow probe sensor group, and real-time and reliable data is guaranteed by combining edge calculation and dual-channel redundancy transmission; and the space-time database and the disaster chain rule base support analysis layer are fused with long and short-term memory network time sequence modeling and graph convolutional network space correlation and dynamic coupling correction to realize accurate quantification of the composite disaster probability.

Description

Technical Field

[0001] The present invention relates to the field of coal mine safety technology, and in particular to a coal mine multi-hazard intelligent perception early warning and linkage emergency response system. Background Art

[0002] During the coal mining process, due to the complex underground composition, various types of underground disasters occur in coal mines, such as gas accumulation causing suffocation of workers, water inrush in the tunnel causing changes in the coal seam structure or damage to equipment, dust causing harm to workers' health or the risk of dust explosion, fire causing casualties or equipment loss, etc.

[0003] The environments in which coal mine disasters such as gas accumulation, water disasters, dust, and fire occur have the following common characteristics: complexity and concealment of geological structures, dynamic imbalance of environmental parameters, coupling effects of multiple disasters, etc. Among them, the coupling effect of multiple disasters is more common in coal mine disasters. For example: the shock wave of a gas explosion stirs up deposited dust and triggers a secondary explosion; the high temperature of a fire accelerates the thermal decomposition of the coal body and releases more gas; sudden water damages closed facilities and causes abnormal gas outflow; high-pressure gas may accumulate in water-sealed areas, etc.; and when the coupling effect of multiple disasters exists, geological structure is also an important reference factor for the coupling of multiple disasters.

[0004] However, in existing technologies, traditional sensors are mostly deployed on the surface of tunnels or fixed points. They can only monitor environmental parameters and cannot penetrate deep into the coal body to obtain advanced indicators, resulting in delayed warnings and susceptibility to local interference. In addition, existing coal mine disaster perception and warning systems mostly analyze single disasters independently, ignore the interactive effects of disaster chains, and do not consider the differences in geological units. The false alarm rate is high and there are defects. Summary of the Invention

[0005] The purpose of the present invention is to address the shortcomings of the existing technology and propose an intelligent perception warning and linkage emergency system for multiple disasters in coal mines. Through a hollow probe rod sensor group, the "deep-surface-environment" three-in-one monitoring is realized, and the spatiotemporal database and the disaster chain rule library support analysis layer are combined to integrate long-term and short-term memory network time series modeling, graph convolution network spatial association and dynamic coupling correction to achieve accurate quantification of the probability of compound disasters.

[0006] To achieve the above objectives, the present invention adopts the following technical solutions: a coal mine multi-hazard intelligent perception warning and linkage emergency response system, comprising:

[0007] A sensing layer, comprising a hollow probe sensor group, which acquires the temperature within the tunnel wall, the gas concentration within the tunnel wall, the water pressure at the tunnel wall, the gas concentration within the tunnel, the temperature within the tunnel, and the dust concentration within the tunnel;

[0008] The transport layer includes an edge gateway for filtering the data collected by the hollow probe sensor group and calculating characteristic parameters in real time, and a redundant link for automatically switching and ensuring the upload of key data when the optical fiber is disconnected;

[0009] A data layer, comprising a time series database for storing raw data of the hollow probe sensor group, a spatial database for storing and recording probe positions, geological unit classifications, and coordinates of historical disaster points, and a disaster case database for storing disaster data;

[0010] An analysis layer, comprising a disaster coupling model for combining multi-source data with disaster chain rules to perform disaster prediction;

[0011] An early warning layer, comprising a hierarchical early warning platform for hierarchical response, an audible and visual alarm device, and an underground broadcast linkage module;

[0012] The disaster coupling model includes a time feature extraction module, a space feature extraction module, a disaster chain rule base and a risk probability fusion module;

[0013] The multi-source data collected by the hollow probe rod sensor group is uploaded to the edge gateway through the industrial ring network for preliminary processing. The multi-source data preliminarily processed by the edge gateway is then uploaded to the data layer through the industrial ring network. The processed multi-source data is classified by the time series database and spatial database of the data layer. The classified multi-source data is then analyzed by the analysis layer. The analysis results of the analysis layer are uploaded to the early warning layer. The hierarchical early warning platform of the early warning layer implements corresponding early warning content according to the analysis results, and broadcasts the early warning content through the sound and light alarm device and the underground broadcast linkage module.

[0014] As a preferred embodiment, the hollow probe rod sensor group includes an intermediate tube, one end of the intermediate tube is connected to end head one, a temperature sensor one and a gas sensor one are installed in end head one, the probes of the temperature sensor one and the gas sensor one are facing the outside of end head one, an end head two is provided on the end of the intermediate tube away from end head one, a water pressure sensor is installed on the end head two, the probe of the water pressure sensor is facing the inside of end head two, an end head three is provided on the end of the end two away from the intermediate tube, a gas sensor two, a temperature sensor two and a dust sensor are installed on the outside of the end head three through a mounting bracket, and the temperature sensor one, the gas sensor one, the water pressure sensor, the dust sensor, the temperature sensor two and the gas sensor two are connected to the industrial ring network via a transmission cable.

[0015] First, a hole is drilled at the selected tunnel wall position, and the drilling depth is greater than the sum of the lengths of the intermediate tube and end one. The intermediate tube pushes end one into the tunnel wall, and the entire intermediate tube is inside the tunnel wall, end two is on the tunnel wall surface, and end three is inside the tunnel. After the hollow probe rod sensor group is inserted into the drilled hole, expansion glue is filled between the hollow probe rod sensor group and the hole wall for sealing.

[0016] As a preferred embodiment, the time feature extraction module is constructed based on the long short-term memory network time series model, specifically:

[0017] S1.1.1. Split the time series sliding window: First, define the window length and step size, and then split the continuous data into overlapping time series segments;

[0018] S1.1.2. Construction of Long Short-Term Memory Network Time Series Model: Set the goal to predict the trend of multi-source parameter changes in the next ten minutes, set the loss function to mean absolute error, set the number of hidden units to 128, and the dropout rate to 0.2 for a single-layer long short-term memory network. Input the time series window data of a single hollow probe sensor group, fully connected layer, and generate the time feature vector

[0019] S1.1.3. Set key time features: First, calculate the rate of change of parameters within the window, then use FFT to extract the main frequency components of each multi-source parameter, identify abnormal fluctuation patterns, and then detect mutation points and mark them as potential risk events.

[0020] As a preferred embodiment, the spatial feature extraction module specifically includes:

[0021] S1.2.1. Construct spatial topology: Generate a triangular mesh based on the physical position coordinates (XYZ) of the probes and define the proximity relationship of the probes. If two probes are adjacent in the triangulated mesh, the adjacency matrix element A ij =1, otherwise 0;

[0022] S1.2.2. Build a graph convolutional network: Input the current feature vector of each hollow probe sensor group and output the spatial feature vector through the propagation formula Characterizing the comprehensive risk of the area where the hollow probe sensor group n is located, the propagation formula is:

[0023]

[0024] in:

[0025] H (l) Provide evidence for the node features of layer l;

[0026] is the adjacency matrix with self-loops added;

[0027] is the degree matrix after adding the self-loop;

[0028] is the inverse square root of the degree matrix;

[0029] W (l) is the trainable weight matrix of layer l;

[0030] σ(·) nonlinear activation function;

[0031] S1.2.3. Detection of regional anomalies: If one of the multi-source parameters of three adjacent hollow probe sensor groups exceeds the threshold at the same time, it is marked as a regional risk.

[0032] As a preferred embodiment, the specific steps of fusing the time features and the spatial features of the time feature extraction module and the spatial feature extraction module are as follows:

[0033] S1.3.1. Concatenate the temporal feature vector and the spatial feature vector into a joint feature vector. By vertical concatenation, a 96-dimensional joint feature vector is formed, preserving the temporal and spatial dimension information, which can be expressed as:

[0034]

[0035] S1.3.2. First, perform attention mechanism weighting:

[0036]

[0037] Then perform weighted fusion:

[0038]

[0039] In the attention mechanism weighted formula and weighted fusion formula:

[0040] α n is the attention weight, which indicates the importance of the spatiotemporal features of the nth probe to the final fusion result. The larger the weight, the more critical the risk signal of the hollow probe sensor group.

[0041] W a It is a trainable weight matrix that maps high-dimensional features to attention scores and captures the importance of different features through learning;

[0042] is a joint eigenvector, which is composed of the time eigenvector and the space eigenvector, and represents the comprehensive risk status of the hollow probe sensor group n;

[0043] Softmax is a normalization function that converts the attention score into a probability distribution to ensure that the sum of all weights is 1;

[0044] N is the total number of probe rods, which indicates the number of probe rods deployed in the tunnel area currently analyzed;

[0045] h fused It is the fusion feature vector, which represents the weighted sum of all probe features, reflects the overall risk status of the region, and is used for the final warning decision;

[0046] S1.3.3, Spatiotemporal coupling analysis: Combine the time series input with the spatial topology, iteratively update the node status, and then output the spatiotemporal risk score s for each hollow probe sensor group. n ∈[0,1], which represents the probability of a compound disaster occurring at position n of the hollow probe sensor group.

[0047] As a preferred implementation method, the construction process of the disaster chain rule library is as follows: first, geological data, monitoring data and engineering data are collected, and the entities of the geological data, monitoring data and engineering data are defined; then, preliminary association rules are formulated through manual screening based on the changes in geological data and monitoring data in disaster events; then, association rules are used to mine the association between disaster events and monitoring data and geological data; with disaster events as nodes and preliminary association rules as the basis, association rules of multi-source data in each disaster event are mined; then, unsupervised learning of multi-source data of disaster precursors is performed through the DBSCAN algorithm to extract common time series features; finally, disaster event data is input, the disaster chain rules are verified, and corrections are made through manual re-inspection.

[0048] As a preferred embodiment, the steps of data input and preprocessing of the risk probability fusion module are as follows:

[0049] S3.1.1. Single-hazard risk prediction: Use the long short-term memory network time series model to predict gas risk, use random forest to predict water inrush risk, use XGBoost to predict dust risk, and use Bayesian network to predict fire risk;

[0050] S3.1.2. Extract the temporal characteristics of the results of single-hazard risk prediction: Extract the temporal trend of the risk probability of a single-hazard risk;

[0051] S3.1.3. Extract spatial features of single-hazard risk prediction results: Extract the topological correlation of the hollow probe sensor group and the regional geological type;

[0052] S3.1.4. Setting the disaster coupling coefficient: Based on the disaster chain rule base, the disaster chain rule establishes the correlation intensity matrix: M∈R 4×4 , where: M is the disaster chain coupling coefficient matrix, which quantifies the interaction between disasters and is the core parameter of multi-hazard risk modeling; R represents the real number space, which is used to define the mathematical properties of the matrix and is not an independent parameter.

[0053] As a preferred embodiment, the risk probability fusion module adopts a two-level fusion architecture, combining basic probability fusion and coupling effect correction. The construction steps are as follows:

[0054] S3.2.1. Using weighted average and nonlinear transformation to perform basic probability fusion:

[0055]

[0056] in:

[0057] P base It represents the comprehensive probability output by the activation function after weighted fusion of the single disaster probabilities, ranging from [0,1];

[0058] σ represents compressing the weighted sum to the probability interval, and the threshold b = 0.6 suppresses low-probability noise;

[0059] w k represents the importance of disaster type k, which is determined through historical data training;

[0060] P k Indicates the independent prediction probability range of each disaster [0,1];

[0061] k represents the disaster type index, where k = 1, 2, 3, and 4 correspond to gas, water inrush, dust, and fire disaster types respectively.

[0062] S3.2.2. Coupling effect correction: Input base probability P base , spatiotemporal characteristics F ST And the correlation strength matrix M, the modified formula is:

[0063] P final =P base +α*ReLU(W c *(M*P)⊙F ST )

[0064] in:

[0065] P final The final risk probability after correction is the sum of the basic probability and the coupling correction term. The range may exceed [0, 1], but it is ultimately mapped to the warning level through threshold grading.

[0066] α is the coupling effect strength coefficient, which is a scalar that represents the magnitude of the influence of the control correction term on the final probability; ReLU is the rectified linear unit activation function, which ensures that the correction term is non-negative and only enhances the risk;

[0067] W c is the trainable coupling weight matrix;

[0068] M is the disaster chain coupling coefficient matrix;

[0069] P is the probability vector of a single disaster;

[0070] ⊙ is element-by-element multiplication;

[0071] F ST is the spatiotemporal eigenvector;

[0072] S3.2.3. Set up the weight learning mechanism: Calculate the loss function:

[0073]

[0074] in:

[0075] L is a scalar, representing the total loss value. The goal of model optimization is to minimize L;

[0076] N is a positive integer, indicating the total number of training samples and the number of disaster events participating in the current batch training;

[0077] y i Is a binary label, indicating the true label of the i-th sample: y i =1 means a compound disaster actually occurred, y i =0 means no disaster occurred;

[0078] is a scalar, which represents the final predicted probability of the model for the i-th sample, indicating the predicted probability of a compound disaster;

[0079] Log is the natural logarithm function, which is used to calculate the log likelihood of the predicted probability and amplify the penalty when the prediction is wrong;

[0080] λ is a non-negative real number, representing the regularization coefficient, which controls the strength of the regularization term. The larger λ is, the stronger the model complexity penalty is.

[0081] W c is a matrix, which represents the trainable coupling weight matrix and is used to adjust the impact intensity of the disaster chain coupling effect;

[0082] ||W c ||2 is a scalar, representing the L2 regularization term, and the calculation matrix W c Take the square root of the sum of the squares of all elements.

[0083] As a preferred embodiment, the final risk probability obtained by the risk probability fusion module is sent to the graded warning platform, and a graded warning plan is formulated according to the value of the final risk probability: the final risk probability P final When it is in the range of [0.8, 1.0], it is set as a red warning; the final risk probability P final When it is in the range of [0.6, 0.8], it is set as orange warning; the final risk probability Pfinal When it is in the range of [0.4, 0.6], it is set as a yellow warning.

[0084] Compared with the prior art, the advantages and positive effects of the present invention are:

[0085] The present invention realizes the "deep-surface-environment" three-in-one monitoring through the hollow probe rod sensor group, and combines edge computing with dual-channel redundant transmission to ensure real-time and reliable data; the spatiotemporal database and the disaster chain rule library support the analysis layer to integrate long-term and short-term memory network temporal modeling, graph convolution network spatial association and dynamic coupling correction to achieve accurate quantification of the probability of complex disasters; finally, through the hierarchical early warning platform, the sound and light alarm, equipment control and disaster avoidance guidance are linked to achieve the effects of advanced early warning, rapid emergency response and low false alarm rate, which significantly improves the safety efficiency of mines and provides full-chain technical guarantees for disaster prevention and control under complex geological conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0086] Figure 1 The module architecture diagram of the intelligent perception warning and linkage emergency response system for multiple disasters in coal mines proposed by the present invention is as follows;

[0087] Figure 2 The present invention provides a structural schematic diagram of the hollow probe rod sensor group of the coal mine multi-hazard intelligent perception warning and linkage emergency system. DETAILED DESCRIPTION

[0088] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0089] Example 1

[0090] like Figure 1 and Figure 2 As shown, the present invention provides a technical solution: a coal mine multi-hazard intelligent perception warning and linkage emergency response system, comprising:

[0091] A sensing layer, comprising a hollow probe sensor group, which acquires the temperature within the tunnel wall, the gas concentration within the tunnel wall, the water pressure at the tunnel wall, the gas concentration within the tunnel, the temperature within the tunnel, and the dust concentration within the tunnel;

[0092] Wherein, the hollow probe rod sensor group includes an intermediate tube, one end of the intermediate tube is connected to end head 1, temperature sensor 1 and gas sensor 1 are installed in end head 1, the probes of temperature sensor 1 and gas sensor 1 are facing the outside of end head 1, end head 2 is provided on the end of the intermediate tube away from end head 1, a water pressure sensor is installed on end head 2, the probe of water pressure sensor is facing the inside of end head 2, end head 2 is provided on the end away from the intermediate tube, gas sensor 2, temperature sensor 2 and dust sensor are installed on the outside of end head 3 through a mounting bracket, and the temperature sensor 1, gas sensor 1, water pressure sensor, dust sensor, temperature sensor 2 and gas sensor 2 connect data to the industrial ring network through a transmission cable;

[0093] Furthermore, a hole is first drilled at a selected tunnel wall position, the drilling depth being greater than the sum of the lengths of the intermediate tube and the end head 1, the intermediate tube pushes the end head 1 into the tunnel wall, and the entire intermediate tube is in the tunnel wall, the end head 2 is on the tunnel wall surface, and the end head 3 is in the tunnel. After the hollow probe rod sensor group is inserted into the drill hole, expansion glue is filled between the hollow probe rod sensor group and the hole wall for sealing.

[0094] The transport layer includes an edge gateway for filtering the data collected by the hollow probe sensor group and calculating characteristic parameters in real time, and a redundant link for automatically switching and ensuring the upload of key data when the optical fiber is disconnected;

[0095] Among them, the edge gateway can use an intelligent gateway equipped with NVIDIA Jetson AGX Xavier to implement real-time data filtering and feature parameter calculation, and use redundant links to ensure priority transmission of critical data when the transmission network is interrupted;

[0096] Furthermore, by reducing cloud load through edge intelligent pre-processing and combining it with a highly reliable communication mechanism, the real-time and security of multi-source data in complex geological environments is ensured;

[0097] A data layer, comprising a time series database for storing raw data of the hollow probe sensor group, a spatial database for storing recorded probe positions, geological unit classifications, and historical disaster point coordinates, and a disaster case database for storing disaster data;

[0098] The disaster data in the disaster case database includes sensor time-series clips, video recordings, emergency response records, and geological reports during disaster events. It uses semantic coding to achieve multimodal retrieval and supports case version management and difference comparison.

[0099] An analysis layer, comprising a disaster coupling model for combining multi-source data with disaster chain rules for disaster prediction, the disaster coupling model comprising a temporal feature extraction module, a spatial feature extraction module, a disaster chain rule base, and a risk probability fusion module;

[0100] An early warning layer, comprising a hierarchical early warning platform for hierarchical response, an audible and visual alarm device, and an underground broadcast linkage module;

[0101] The final risk probability calculated by the risk probability fusion module is input into the hierarchical warning platform, and warning information is sent to the sound and light alarm device and the underground broadcast linkage module according to the value of the final risk probability.

[0102] Furthermore, the multi-source data collected by the hollow probe rod sensor group is uploaded to the edge gateway through the industrial ring network for preliminary processing. The multi-source data preliminarily processed by the edge gateway is then uploaded to the data layer through the industrial ring network. The processed multi-source data is classified by the time series database and spatial database of the data layer. The classified multi-source data is then analyzed by the analysis layer. The analysis results of the analysis layer are uploaded to the early warning layer. The hierarchical early warning platform of the early warning layer implements corresponding early warning content according to the analysis results, and broadcasts the early warning content through the sound and light alarm device and the underground broadcast linkage module.

[0103] In this embodiment, a three-dimensional sensing network is constructed through a hollow probe rod sensor group. One end part is used to monitor coal gas and temperature, the second end part detects water pressure, and the third end part collects dust and environmental parameters. The data is transmitted to the edge gateway through the industrial ring network for filtering and feature extraction. Combined with a dual-channel redundant link, the data is ensured to be uploaded in real time and reliably. The data layer relies on the time series database, spatial database and disaster case library to realize multi-source data storage and semantic retrieval. The analysis layer integrates risk probability through long-short-term memory network spatiotemporal modeling and disaster chain rule library, triggering the hierarchical early warning platform to link sound and light alarms and underground broadcast directional disaster avoidance guidance, effectively improving the mine disaster prevention and control capabilities.

[0104] Example 2

[0105] like Figure 1 and Figure 2 As shown in the figure, firstly, the temporal and spatial features of the hollow probe sensor data are extracted through the time feature extraction module and the spatial feature extraction module. Then, the risk probability fusion module is constructed in combination with the disaster chain rule base. Specifically:

[0106] S1. The time feature extraction module and the space feature extraction module are used to extract the time feature and the space feature of the hollow probe sensor data, and then the extracted time feature and space feature are spliced and fused. Specifically:

[0107] S1.1. The temporal feature extraction module is constructed based on the long short-term memory network temporal model. Specifically:

[0108] S1.1.1. Split the time series sliding window: First, define the window length and step size, and then split the continuous data into overlapping time series segments. After the sliding window segmentation, each window can be independently input into the long short-term memory network model to capture local time series patterns while maintaining global continuity;

[0109] S1.1.2. Construction of long short-term memory network time series model: Set the goal to predict the trend of multi-source parameter changes in the next ten minutes, set the loss function to mean absolute error, set the number of hidden units to 128, and set a single-layer long short-term memory network with a dropout rate of 0.2 to capture long-term dependencies. You can set it to randomly discard 20% of neurons to prevent overfitting. Input the time series window data of a single hollow probe sensor group. The fully connected layer maps the output of the long short-term memory network to low-dimensional features, retaining key time series information, and generating a time feature vector.

[0110] S1.1.3. Set key time characteristics: First, calculate the rate of change of the parameters within the window. Use the gradient characteristics to reflect the short-term trend of the parameters, such as a 0.5% increase in gas concentration per minute. Use the acceleration characteristics to detect sudden changes in the trend, such as accelerated gas outburst. Then, perform a fast Fourier transform on the window data using FFT to extract the frequency component with the highest energy. Extract the main frequency components of each multi-source parameter. Identify abnormal patterns and abnormal fluctuation patterns through changes in frequency domain energy distribution. Then, set a threshold for the parameter change rate to detect sudden changes and mark them as potential risk events.

[0111] S1.2, the spatial feature extraction module specifically includes:

[0112] S1.2.1. Construct spatial topology: Based on the physical position coordinates (XYZ) of the probe, generate a triangular mesh and define the proximity relationship of the probes. If two probes are adjacent in the triangular mesh, the adjacency matrix element A ij =1, otherwise 0, and the geological unit to which the probe rod belongs can be marked based on geological exploration data;

[0113] S1.2.2. Build a graph convolutional network: Input the current feature vector of each hollow probe sensor group and output the spatial feature vector through the propagation formula Characterizing the comprehensive risk of the area where the hollow probe sensor group n is located, the propagation formula is:

[0114]

[0115] in:

[0116] H (l)Provide evidence for the node features of layer l;

[0117] is the adjacency matrix with self-loops added;

[0118] is the degree matrix after adding the self-loop;

[0119] is the inverse square root of the degree matrix;

[0120] W (l) is the trainable weight matrix of layer l;

[0121] σ(·) nonlinear activation function;

[0122] The architecture of the graph convolutional network can be designed as follows:

[0123] Input layer: the current feature vector of the probe, dimension 6;

[0124] Graph convolutional network layer: number of hidden layers: 2 layers; number of hidden units: 64→32; Dropout rate: 0.3;

[0125] Output layer: spatial feature vector, representing the comprehensive risk of the area where the probe is located;

[0126] S1.2.3. Detection of regional anomalies: If one of the multi-source parameters of three adjacent hollow probe sensor groups exceeds the threshold at the same time, it is marked as a regional risk;

[0127] S1.3. The specific steps of fusing the temporal features and spatial features of the temporal feature extraction module and the spatial feature extraction module are as follows:

[0128] S1.3.1. Concatenate the temporal feature vector and the spatial feature vector into a joint feature vector. By vertical concatenation, a 96-dimensional joint feature vector is formed, preserving the temporal and spatial dimension information, which can be expressed as:

[0129]

[0130] 1.3.2. First, perform attention mechanism weighting:

[0131]

[0132] Then perform weighted fusion:

[0133]

[0134] In the attention mechanism weighted formula and weighted fusion formula:

[0135] α nis the attention weight, which indicates the importance of the spatiotemporal features of the nth probe to the final fusion result. The larger the weight, the more critical the risk signal of the hollow probe sensor group.

[0136] W a It is a trainable weight matrix that maps high-dimensional features to attention scores and captures the importance of different features through learning;

[0137] is a joint eigenvector, which is composed of the time eigenvector and the space eigenvector, and represents the comprehensive risk status of the hollow probe sensor group n;

[0138] Softmax is a normalization function that converts the attention score into a probability distribution to ensure that the sum of all weights is 1;

[0139] N is the total number of probe rods, which indicates the number of probe rods deployed in the tunnel area currently analyzed;

[0140] h fused It is the fusion feature vector, which represents the weighted sum of all probe features, reflects the overall risk status of the region, and is used for the final warning decision;

[0141] S1.3.3, Spatiotemporal coupling analysis: Combine the time series input with the spatial topology, iteratively update the node status, and then output the spatiotemporal risk score s for each hollow probe sensor group. n ∈[0,1], represents the probability of a compound disaster occurring at position n of the hollow probe sensor group;

[0142] S2. The construction process of the disaster chain rule library is as follows: first, geological data, monitoring data and engineering data are collected, and entities such as geological units, monitoring points, disaster types and equipment are defined for the geological data, monitoring data and engineering data; then, preliminary association rules are formulated through manual screening based on the changes in geological data and monitoring data in disaster events; then, association rules are used to mine associations between disaster events and monitoring data and geological data; with disaster events as nodes and preliminary association rules as the basis, association rules of multi-source data in each disaster event are mined; then, unsupervised learning of multi-source data of disaster precursors is performed through the DBSCAN algorithm to extract common time series features; finally, disaster event data is input, disaster chain rules are verified, and corrections are made through manual re-inspection;

[0143] S3. First, the data of the risk probability fusion module is input and preprocessed. Then, a two-level fusion architecture is adopted, combining basic probability fusion and coupling effect correction. Specifically:

[0144] S3.1. The steps of data input and preprocessing of the risk probability fusion module are as follows:

[0145] S3.1.1. Single-hazard risk prediction: Use the long short-term memory network time series model to predict gas risk, use random forest to predict water inrush risk, use XGBoost to predict dust risk, and use Bayesian network to predict fire risk;

[0146] S3.1.2. Extract the temporal characteristics of the results of single-hazard risk prediction: Extract the temporal trend of the risk probability of a single-hazard risk;

[0147] S3.1.3. Extract spatial features of single-hazard risk prediction results: Extract the topological correlation of the hollow probe sensor group and the regional geological type;

[0148] S3.1.4. Setting the disaster coupling coefficient: Based on the disaster chain rule base, the disaster chain rule establishes the correlation intensity matrix: M∈R 4×4 , where: M is the disaster chain coupling coefficient matrix, which quantifies the interaction between disasters and is the core parameter of multi-hazard risk modeling; R represents the real number space, which is used to define the mathematical properties of the matrix and is not an independent parameter;

[0149] S3.2. The risk probability fusion module adopts a two-level fusion architecture, combining basic probability fusion and coupling effect correction. The construction steps are as follows:

[0150] S3.2.1. Using weighted average and nonlinear transformation to perform basic probability fusion:

[0151]

[0152] in:

[0153] P base It represents the comprehensive probability output by the activation function after weighted fusion of the single disaster probabilities, ranging from [0,1];

[0154] σ represents compressing the weighted sum to the probability interval, and the threshold b = 0.6 suppresses low-probability noise;

[0155] w k The importance of disaster type k is determined by historical data training, and w is trained by minimizing the cross entropy loss function through the gradient descent algorithm. k ;

[0156] P k Indicates the independent prediction probability range of each disaster [0,1];

[0157] k represents the disaster type index, where k = 1, 2, 3, and 4 correspond to gas, water inrush, dust, and fire disaster types respectively.

[0158] S3.2.2. Coupling effect correction: Input base probability P base , spatiotemporal characteristics F STAnd the correlation strength matrix M, the modified formula is:

[0159] P final =P base +α*ReLU(W c *(M*P)⊙F ST )

[0160] in:

[0161] P final The final risk probability after correction is the sum of the basic probability and the coupling correction term. The range may exceed [0, 1], but it is ultimately mapped to the warning level through threshold grading.

[0162] α is the coupling effect strength coefficient, which is a scalar that represents the magnitude of the influence of the control correction term on the final probability; ReLU is the rectified linear unit activation function, which ensures that the correction term is non-negative and only enhances the risk;

[0163] W c is the trainable coupling weight matrix;

[0164] M is the disaster chain coupling coefficient matrix;

[0165] P is the probability vector of a single disaster;

[0166] ⊙ is element-by-element multiplication;

[0167] F ST is the spatiotemporal eigenvector;

[0168] S3.2.3. Set up the weight learning mechanism: Calculate the loss function:

[0169]

[0170] in:

[0171] L is a scalar, representing the total loss value. The goal of model optimization is to minimize L;

[0172] N is a positive integer, indicating the total number of training samples and the number of disaster events participating in the current batch training;

[0173] y i Is a binary label, indicating the true label of the i-th sample: y i =1 means a compound disaster actually occurred, y i =0 means no disaster occurred;

[0174] is a scalar, which represents the final predicted probability of the model for the i-th sample, indicating the predicted probability of a compound disaster;

[0175] Log is the natural logarithm function, which is used to calculate the log likelihood of the predicted probability and amplify the penalty when the prediction is wrong;

[0176] λ is a non-negative real number, representing the regularization coefficient, which controls the strength of the regularization term. The larger λ is, the stronger the model complexity penalty is.

[0177] W c is a matrix, which represents the trainable coupling weight matrix and is used to adjust the impact intensity of the disaster chain coupling effect;

[0178] ||W c ||2 is a scalar, representing the L2 regularization term, and the calculation matrix W c The square root of the sum of the squares of all elements;

[0179] Furthermore, the final risk probability obtained by the risk probability fusion module is sent to the graded warning platform, and a graded warning plan is formulated according to the value of the final risk probability: the final risk probability P final When it is in the range of [0.8, 1.0], it is set as a red warning; the final risk probability P final When it is in the range of [0.6, 0.8], it is set as orange warning; the final risk probability P final When it is in the range of [0.4, 0.6], it is set as yellow warning

[0180] In this embodiment, accurate risk warning is achieved through the fusion of spatiotemporal features and dynamic disaster chain modeling:

[0181] Long short-term memory networks are used to analyze time series, and graph convolutional networks are combined to construct spatial topological relationships of probes to detect regional anomalies. Spatiotemporal features are weighted and fused through the attention mechanism to generate a joint risk vector. A dynamic rule base is constructed based on geological data and historical cases. Common precursor patterns are extracted through association rule mining and unsupervised learning, and thresholds of different geological units are adapted. The weighted average probability of a single disaster is used to form a basic probability. A two-level architecture of coupling the disaster chain matrix and the spatiotemporal feature correction results is used to output a composite risk value. Combined with the triggering of red / orange / yellow three-level warnings according to risk probability, and the linkage of sound and light alarms and equipment control, it can achieve advanced warning, high composite disaster identification accuracy, low latency throughout the entire process, and low false alarm rate.

[0182] The above description is merely a preferred embodiment of the present invention and does not constitute any other form of limitation to the present invention. Any person skilled in the art may utilize the technical contents disclosed above to change or modify them into equivalent embodiments with equivalent changes for application in other fields. However, any simple modification, equivalent change, and modification of the above embodiments made in accordance with the technical essence of the present invention without departing from the technical solution of the present invention shall still fall within the scope of protection of the technical solution of the present invention.

Claims

1. The coal mine multi-hazard intelligent perception warning and linkage emergency response system is characterized by: include: A sensing layer, comprising a hollow probe sensor group, which acquires the temperature within the tunnel wall, the gas concentration within the tunnel wall, the water pressure at the tunnel wall, the gas concentration within the tunnel, the temperature within the tunnel, and the dust concentration within the tunnel; The transport layer includes an edge gateway for filtering the data collected by the hollow probe sensor group and calculating characteristic parameters in real time, and a redundant link for automatically switching and ensuring the upload of key data when the optical fiber is disconnected; A data layer, comprising a time series database for storing raw data of the hollow probe sensor group, a spatial database for storing and recording probe positions, geological unit classifications, and coordinates of historical disaster points, and a disaster case database for storing disaster data; An analysis layer, comprising a disaster coupling model for combining multi-source data with disaster chain rules to perform disaster prediction; An early warning layer, comprising a hierarchical early warning platform for hierarchical response, an audible and visual alarm device, and an underground broadcast linkage module; The disaster coupling model includes a time feature extraction module, a space feature extraction module, a disaster chain rule base and a risk probability fusion module; The multi-source data collected by the hollow probe rod sensor group is uploaded to the edge gateway through the industrial ring network for preliminary processing. The multi-source data preliminarily processed by the edge gateway is then uploaded to the data layer through the industrial ring network. The processed multi-source data is classified by the time series database and spatial database of the data layer. The classified multi-source data is then analyzed by the analysis layer. The analysis results of the analysis layer are uploaded to the early warning layer. The hierarchical early warning platform of the early warning layer implements corresponding early warning content according to the analysis results, and broadcasts the early warning content through the sound and light alarm device and the underground broadcast linkage module.

2. The coal mine multi-hazard intelligent perception warning and linkage emergency response system according to claim 1 is characterized by: The hollow probe rod sensor group includes an intermediate tube, one end of the intermediate tube is connected to end head 1, a temperature sensor 1 and a gas sensor 1 are installed in end head 1, the probes of the temperature sensor 1 and the gas sensor 1 are facing the outside of end head 1, an end head 2 is provided on the end of the intermediate tube away from end head 1, a water pressure sensor is installed on the end head 2, the probe of the water pressure sensor is facing the inside of end head 2, an end head 3 is provided on the end of the end 2 away from the intermediate tube, a gas sensor 2, a temperature sensor 2 and a dust sensor are installed on the outside of the end head 3 through a mounting bracket, and the temperature sensor 1, gas sensor 1, water pressure sensor, dust sensor, temperature sensor 2 and gas sensor 2 connect data to the industrial ring network through a transmission cable; First, a hole is drilled at the selected tunnel wall position, and the drilling depth is greater than the sum of the lengths of the intermediate tube and end one. The intermediate tube pushes end one into the tunnel wall, and the entire intermediate tube is inside the tunnel wall, end two is on the tunnel wall surface, and end three is inside the tunnel. After the hollow probe rod sensor group is inserted into the drilled hole, expansion glue is filled between the hollow probe rod sensor group and the hole wall for sealing.

3. The coal mine multi-hazard intelligent perception warning and linkage emergency response system according to claim 1 is characterized by: The temporal feature extraction module is constructed based on the long short-term memory network temporal model. Specifically: S1.1.

1. Split the time series sliding window: First, define the window length and step size, and then split the continuous data into overlapping time series segments; S1.1.

2. Construction of Long Short-Term Memory Network Time Series Model: Set the goal to predict the trend of multi-source parameter changes in the next ten minutes, set the loss function to mean absolute error, set the number of hidden units to 128, and the dropout rate to 0.2 for a single-layer long short-term memory network. Input the time series window data of a single hollow probe sensor group, fully connected layer, and generate the time feature vector S1.1.

3. Set key time features: First, calculate the rate of change of parameters within the window, then use FFT to extract the main frequency components of each multi-source parameter, identify abnormal fluctuation patterns, and then detect mutation points and mark them as potential risk events.

4. The coal mine multi-hazard intelligent perception warning and linkage emergency response system according to claim 1 is characterized by: The spatial feature extraction module specifically includes: S1.2.

1. Construct spatial topology: Based on the physical position coordinates (XYZ) of the probe, generate a triangular mesh and define the proximity relationship of the probes. If two probes are adjacent in the triangular mesh, the adjacency matrix element A ij =1, otherwise 0; S1.2.

2. Build a graph convolutional network: Input the current feature vector of each hollow probe sensor group and output the spatial feature vector through the propagation formula Characterizing the comprehensive risk of the area where the hollow probe sensor group n is located, the propagation formula is: in: H (l) Provide evidence for the node features of layer l; is the adjacency matrix with self-loops added; is the degree matrix after adding the self-loop; is the inverse square root of the degree matrix; W (l) is the trainable weight matrix of layer l; σ(·) nonlinear activation function; S1.2.

3. Detection of regional anomalies: If one of the multi-source parameters of three adjacent hollow probe sensor groups exceeds the threshold at the same time, it is marked as a regional risk.

5. The coal mine multi-hazard intelligent perception warning and linkage emergency response system according to claim 4 is characterized by: The specific steps of fusing the time features and spatial features of the time feature extraction module and the spatial feature extraction module are as follows: S1.3.

1. Concatenate the temporal feature vector and the spatial feature vector into a joint feature vector. By vertical concatenation, a 96-dimensional joint feature vector is formed, preserving the temporal and spatial dimension information, which can be expressed as: S1.3.

2. First, perform the attention mechanism weighting: Then perform weighted fusion: In the attention mechanism weighted formula and weighted fusion formula: α n is the attention weight, which indicates the importance of the spatiotemporal features of the nth probe to the final fusion result. The larger the weight, the more critical the risk signal of the hollow probe sensor group. W a It is a trainable weight matrix that maps high-dimensional features to attention scores and captures the importance of different features through learning; is a joint eigenvector, which is composed of the time eigenvector and the space eigenvector, and represents the comprehensive risk status of the hollow probe sensor group n; Softmax is a normalization function that converts the attention score into a probability distribution to ensure that the sum of all weights is 1; N is the total number of probe rods, which indicates the number of probe rods deployed in the tunnel area currently analyzed; h fused It is the fusion feature vector, which represents the weighted sum of all probe features, reflects the overall risk status of the region, and is used for the final warning decision; S1.3.3, Spatiotemporal coupling analysis: Combine the time series input with the spatial topology, iteratively update the node status, and then output the spatiotemporal risk score s for each hollow probe sensor group. n ∈[0,1], which represents the probability of a compound disaster occurring at position n of the hollow probe sensor group.

6. The coal mine multi-hazard intelligent perception warning and linkage emergency response system according to claim 1 is characterized by: The construction process of the disaster chain rule library is as follows: first, geological data, monitoring data and engineering data are collected, and the entities of geological data, monitoring data and engineering data are defined; then, preliminary association rules are formulated through manual screening based on the changes in geological data and monitoring data in disaster events; then, association rules are used to mine associations between disaster events and monitoring data and geological data; with disaster events as nodes and preliminary association rules as the basis, association rules of multi-source data in each disaster event are mined; then, unsupervised learning of multi-source data of disaster precursors is performed through the DBSCAN algorithm to extract common time series features; finally, disaster event data is input, the disaster chain rules are verified, and corrections are made through manual re-inspection.

7. The coal mine multi-hazard intelligent perception warning and linkage emergency response system according to claim 1 is characterized by: The steps of data input and preprocessing of the risk probability fusion module are as follows: S3.1.

1. Single-hazard risk prediction: Use the long short-term memory network time series model to predict gas risk, use random forest to predict water inrush risk, use XGBoost to predict dust risk, and use Bayesian network to predict fire risk; S3.1.

2. Extract the temporal characteristics of the results of single-hazard risk prediction: Extract the temporal trend of the risk probability of a single-hazard risk; S3.1.

3. Extract spatial features of single-hazard risk prediction results: Extract the topological correlation of the hollow probe sensor group and the regional geological type; S3.1.

4. Setting the disaster coupling coefficient: Based on the disaster chain rule base, the disaster chain rule establishes the correlation intensity matrix: M∈R 4 ×4 , where: M is the disaster chain coupling coefficient matrix, which quantifies the interaction between disasters and is the core parameter of multi-hazard risk modeling; R represents the real number space, which is used to define the mathematical properties of the matrix and is not an independent parameter.

8. The coal mine multi-hazard intelligent perception warning and linkage emergency response system according to claim 1 is characterized by: The risk probability fusion module adopts a two-level fusion architecture, combining basic probability fusion and coupling effect correction. The construction steps are as follows: S3.2.

1. Using weighted average and nonlinear transformation to perform basic probability fusion: in: P base It represents the comprehensive probability output by the activation function after weighted fusion of the single disaster probabilities, ranging from [0,1]; σ represents compressing the weighted sum to the probability interval, and the threshold b = 0.6 suppresses low-probability noise; w k represents the importance of disaster type k, which is determined through historical data training; P k Indicates the independent prediction probability range of each disaster [0,1]; k represents the disaster type index, where k = 1, 2, 3, and 4 correspond to gas, water inrush, dust, and fire disaster types respectively. S3.2.

2. Coupling effect correction: Input base probability P base , spatiotemporal characteristics F ST And the correlation strength matrix M, the modified formula is: P final =P base +α*ReLU(W c *(M*P)⊙F ST ) in: P final The final risk probability after correction is the sum of the basic probability and the coupling correction term. The range may exceed [0, 1], but it is ultimately mapped to the warning level through threshold grading. α is the coupling effect strength coefficient, which is a scalar that represents the magnitude of the influence of the control correction term on the final probability; ReLU is the rectified linear unit activation function, which ensures that the correction term is non-negative and only enhances risk; W c is the trainable coupling weight matrix; M is the disaster chain coupling coefficient matrix; P is the probability vector of a single disaster; ⊙ is element-by-element multiplication; F ST is the spatiotemporal eigenvector; S3.2.

3. Set up the weight learning mechanism: Calculate the loss function: in: L is a scalar, representing the total loss value. The goal of model optimization is to minimize L; N is a positive integer, indicating the total number of training samples and the number of disaster events participating in the current batch training; y i Is a binary label, indicating the true label of the i-th sample: y i =1 means a compound disaster actually occurred, y i =0 means no disaster occurred; is a scalar, which represents the final predicted probability of the model for the i-th sample, indicating the predicted probability of a compound disaster; Log is the natural logarithm function, which is used to calculate the log likelihood of the predicted probability and amplify the penalty when the prediction is wrong; λ is a non-negative real number, representing the regularization coefficient, which controls the strength of the regularization term. The larger λ is, the stronger the model complexity penalty is. W c is a matrix, which represents the trainable coupling weight matrix and is used to adjust the impact intensity of the disaster chain coupling effect; ||W c ||2 is a scalar, representing the L2 regularization term, and the calculation matrix W c Take the square root of the sum of the squares of all elements.

9. The coal mine multi-hazard intelligent perception warning and linkage emergency response system according to claim 8, characterized in that: The final risk probability obtained by the risk probability fusion module is sent to the graded warning platform, and a graded warning plan is formulated according to the value of the final risk probability: the final risk probability P final When it is in the range of [0.8, 1.0], it is set as a red warning; the final risk probability P final When it is in the range of [0.6, 0.8], it is set as orange warning; the final risk probability P final When it is in the range of [0.4, 0.6], it is set as a yellow warning.

Images