Search

CN-121563251-B - Four-dimensional GIS-based mine salient dangerous area dynamic identification and control decision method

CN121563251BCN 121563251 BCN121563251 BCN 121563251BCN-121563251-B

Abstract

The application discloses a method for dynamically identifying, controlling and deciding a mine salient dangerous area based on a four-dimensional GIS, and belongs to the technical field of mine safety control. The method comprises the steps of obtaining mine multi-source data, enhancing image class data in the multi-source data by using HSMix frames, mapping the enhanced image class data to a four-dimensional tile structure, constructing a space-time index, extracting space risk features by using a three-dimensional U-Net network, extracting multi-scale space-time features by combining an EMD-GCN model, fusing the space risk features and the multi-scale space-time features to obtain a four-dimensional risk identification result, constructing a multi-task full-connection network to output cost, risk and feasibility scores, and recommending a prevention and treatment scheme by using a multi-objective optimization method. The method disclosed by the application can realize dynamic identification and scientific decision of the mine outburst danger, and improves the accuracy and feasibility of prevention and control.

Inventors

  • LI QINGSONG
  • ZHU QUANJIE
  • ZOU QUANLE
  • ZHANG PENG
  • WANG CHUNHUA
  • ZHANG SHUJIN
  • SHEN ZHENHUA

Assignees

  • 贵州省矿山安全科学研究院有限公司
  • 贵州省煤矿设计研究院有限公司

Dates

Publication Date
20260512
Application Date
20260126

Claims (9)

  1. 1. The method for dynamically identifying, controlling and deciding the mine salient dangerous area based on the four-dimensional GIS is characterized by comprising the following steps: S1, acquiring mine multi-source data, wherein the mine multi-source data comprises space static basic data, space-time dynamic monitoring data and numerical simulation auxiliary data, wherein image data are subjected to enhancement processing by adopting a HSMix framework to obtain enhanced image data, and the method specifically comprises the following substeps: S1.1, respectively processing two source images to be enhanced by adopting an SLIC super-pixel algorithm to generate a first super-pixel grid and a second super-pixel grid which are respectively corresponding to each other; s1.2, randomly selecting a super pixel area from the second super pixel grid based on Bernoulli distribution, and generating a binary hard mixed mask; s1.3, fusing the two source images and the labels corresponding to the two source images according to the following formula to obtain hard mixed image data and hard mixed label data: Wherein, the For the image data after hard-blending, For the first one of the source images, For the second one of the source images, For the element-wise multiplication, In order to have a hard mix mask, To be the full size consistent with the hard mix mask The mask is used to indicate that the mask, In order to obtain the tag data after hard mixing, For the label corresponding to the first source image, A label corresponding to the second source image; s1.4, respectively processing the two source images by adopting an OpenCV static saliency algorithm to generate a corresponding first normalized saliency map and a corresponding second normalized saliency map; s1.5, calculating a relative saliency map reflecting the saliency ratio of the same pixel of two source images based on the first normalized saliency map and the second normalized saliency map; S1.6, verifying and confirming semantic consistency of a splicing area based on the hard mixed mask and the hard mixed image data, and fusing the first super-pixel grid and the second super-pixel grid to generate a mixed super-pixel grid; S1.7, calculating the average value of corresponding values of all pixels in the relative saliency map for each super pixel in the mixed super pixel grid to obtain the average mixing proportion of the super pixels, and further generating a soft mixing mask; S1.8, according to the same fusion logic as that of the step S1.3, combining the two source images and the labels corresponding to the two source images by combining the soft mixing mask to obtain soft mixed image data and soft mixed label data, wherein the soft mixed image data is the enhanced image data; s2, mapping the enhanced image class data to a four-dimensional tile structure, wherein the four-dimensional tile structure comprises a space dimension and a time dimension; S3, constructing a space-time index based on the four-dimensional tile structure, wherein the space-time index distributes unique identifiers for each four-dimensional tile through time stamps and space coordinates; S4, processing the space dimension data mapped in the four-dimensional tiles by adopting a three-dimensional U-Net network to obtain three-dimensional space risk characteristics; S5, processing the sensor time sequence data in the time-space dynamic monitoring data by adopting an EMD-GCN model to obtain multi-scale time-space characteristics; s6, fusing the three-dimensional space risk features and the multi-scale space-time features to obtain a four-dimensional risk identification result of the mine salient risk region; s7, constructing a multi-task full-connection network, inputting control strategy parameters into the multi-task full-connection network, and processing and outputting cost scores, risk scores and feasibility scores by the multi-task full-connection network; And S8, recommending a mine outburst risk control scheme through a multi-objective optimization method based on the cost score, the risk score and the feasibility score.
  2. 2. The method according to claim 1, characterized in that the calculation of the relative saliency map in step S1.5 comprises the following sub-steps: S1.5.1 determining a pixel coordinate range of the first normalized saliency map and the second normalized saliency map, the pixel coordinates being in Representation of wherein For the image line index(s), Indexing for the image column; S1.5.2, for each pixel in the pixel coordinate range, assigning the pixel coordinates in the first normalized saliency map to be The pixel value of (2) is recorded as The pixel coordinates in the second normalized saliency map are given as The pixel value of (2) is recorded as ; S1.5.3 calculating the corresponding value of the pixel in the relative saliency map according to the following formula : Wherein, the Representing the second source image at coordinates The significance ratio of the position relative to the first source image is in the range of ; S1.5.4 traversing all pixel coordinates of the first normalized saliency map and the second normalized saliency map, repeating steps S1.5.2 to S1.5.3, generating the complete relative saliency map.
  3. 3. The method according to claim 2, characterized in that said step S1.7 of calculating the average blending ratio of the superpixels and generating a soft blending mask comprises the following sub-steps: S1.7.1 identifying each superpixel in the hybrid superpixel grid, and marking each superpixel as Wherein Index for super-pixel; S1.7.2, counting each super pixel The total number of pixels included, noted as The said Is super pixel All coordinates in The sum of the corresponding pixel numbers; S1.7.3 for each superpixel Traversing all pixel coordinates contained therein Extracting the corresponding pixels in the relative saliency map Values and for these Summing the values; S1.7.4, calculate each superpixel according to the following formula Average mixing ratio of (3) : Wherein, the The range of the values is as follows Representing super-pixels Fusing the weight of the saliency features in the second source image, wherein the saliency features in the second source image are the image features extracted from the second source image through the OpenCV static saliency algorithm in the step S1.4 and corresponding to the area marked by the second normalized saliency map; S1.7.5, determining the size of the soft mixing mask to be consistent with the sizes of the two source images; s1.7.6, for each pixel coordinate of the source image Judging the super pixel to which the pixel belongs Setting the corresponding value of the pixel in the soft mixing mask as the super pixel to which the pixel belongs Average mixing ratio of (3) And traversing all pixels to generate the complete soft mixing mask.
  4. 4. The method according to claim 3, wherein the mapping of the enhanced image class data in step S2 to a four-dimensional tile structure, the four-dimensional tile structure comprising a spatial dimension and a temporal dimension, comprises the following sub-steps: S2.1, classifying the mine multi-source data, and dividing the mine multi-source data into enhanced image data, original numerical value data and original time sequence data, wherein the enhanced image data corresponds to geological map and logging profile image data in the mine multi-source data, the original numerical value data corresponds to gas pressure field and stress field distribution data in the mine multi-source data, and the original time sequence data corresponds to sensor time sequence data in the mine multi-source data; S2.2, respectively extracting corresponding space coordinate information of the enhanced image type data and the original numerical value type data, wherein the space coordinate information is consistent with a mine coordinate system associated with the space static basic data in the mine multi-source data; s2.3, extracting corresponding time stamp information of the original time sequence data, wherein the time stamp information is consistent with a unified time reference anchored by the time-space dynamic monitoring data in the mine multi-source data; S2.4, determining dimension parameters of the four-dimensional tile structure, wherein the dimension parameters comprise space dimensions Time dimension Wherein the spatial dimension Corresponding to the three-dimensional coordinates of the mine coordinate system, the time dimension Corresponds to the time stamp information, wherein For acquisition or analog time stamping of the four-dimensional tile correspondence data, Three-dimensional space coordinates of the four-dimensional tile corresponding data in a mine coordinate system are respectively obtained; s2.5, associating the enhanced image class data and the original numerical class data to the space dimension of the four-dimensional tile structure according to the space coordinate information Associating the original time sequence class data to the time dimension of the four-dimensional tile structure according to time stamp information thereof And simultaneously establishing association relations of different types of the enhanced image class data in the four-dimensional tile structure, and generating the four-dimensional tile containing complete space-time dimension information.
  5. 5. The method according to claim 4, wherein the constructing a space-time index based on the four-dimensional tile structure in step S3 comprises the following sub-steps: S3.1, determining a unique identification composition rule of the space-time index, wherein the unique identification is marked as Is composed of the time dimension information and the space dimension information of the four-dimensional tile structure, namely ; S3.2, traversing all the four-dimensional tiles, and distributing corresponding unique identifiers for each four-dimensional tile according to the unique identifier composition rule ; S3.3, establishing a space-time index library, and identifying the unique identifier of each four-dimensional tile Performing associated storage with the enhanced image class data contained in the four-dimensional tile to form a corresponding relation table of 'unique identification-tile data'; s3.4, realizing time sequence slice backtracking based on the space-time index library, namely when the target time input by the user is acquired When the unique identification is selected from the corresponding relation table In (a) Extracting said enhanced image class data contained by these four-dimensional tiles, loading and visually rendering the target time And (5) setting state data of each area of the mine, and finishing time sequence slicing backtracking.
  6. 6. The method according to claim 5, wherein said step S3 further comprises the sub-step of implementing a time sliding based on said spatio-temporal index base, in particular comprising the following sub-steps: s3.5, acquiring a time sliding parameter set by a user, wherein the time sliding parameter comprises initial time And time step The initial time For the starting point of time of the time slide, said time step Time interval for loading two adjacent times of data; S3.6, according to the initial time And the time step Generating a time series corresponding to the time sliding, wherein the time series is expressed as' 、 、 、...、 "Rule generation in which Is a non-negative integer number of the number, The value of (2) is determined by the time range that the user needs to view; s3.7 traversing each time point in the time series Screening out unique identifiers based on a corresponding relation table of unique identifier-tile data of the space-time index library In (a) Is used to determine the position of the tile in the four dimensions, extracting each time point The enhanced image class data contained in the corresponding four-dimensional tile; s3.8, loading and switching each time point in turn according to the sequence of the time sequence And the corresponding enhanced image data dynamically and visually displays the distribution change trend of the mine salient risk factors in space, so as to finish the realization of time sliding, wherein the mine salient risk factors comprise gas concentration distribution and stress value distribution.
  7. 7. The method according to claim 6, wherein in the step S4, the spatial data in the four-dimensional tile is processed by using a three-dimensional U-Net network to obtain a three-dimensional spatial risk feature, and specifically includes the following sub-steps: S4.1 extracting the spatial dimension from the four-dimensional tile Corresponding spatial data, preprocessing the spatial data to obtain the three-dimensional U-Net network Input data in format, dimensions Respectively correspond to the trend of the mine space trend and depth dimension; S4.2, controlling an encoder part of the three-dimensional U-Net network to execute 3 times of downsampling operation, and gradually extracting high-level abstract spatial features in the input data; s4.3, controlling a decoder part of the three-dimensional U-Net network to execute up-sampling operation for 3 times, gradually recovering the spatial resolution and supplementing local detail characteristics; s4.4, setting an output layer at the output end of the decoder of the three-dimensional U-Net network, classifying the feature map after the third up-sampling through the output layer, and outputting a three-dimensional risk segmentation map, wherein the three-dimensional risk segmentation map is the three-dimensional space risk feature and is used for being fused with the multi-scale space-time feature subsequently.
  8. 8. The method according to claim 1, wherein the step S5 uses an EMD-GCN model to process the sensor time sequence data in the spatio-temporal dynamic monitoring data to obtain multi-scale spatio-temporal features, and specifically includes the following sub-steps: S5.1, extracting sensor time sequence data from the space-time dynamic monitoring data, and carrying out denoising pretreatment on the sensor time sequence data to obtain pretreated sensor time sequence data; s5.2, using the sensor-time step combination corresponding to the preprocessed sensor time sequence data as a node of a time-space diagram Wherein each node Is marked as , For the sensor index to be used, For the time step index (tsf), Represent the first The first sensor is at The connection between different sensor nodes at the same time step is used as a space edge, and the connection between different time step nodes of the same sensor is used as a time edge to jointly form the edge of a time-space diagram Forming a complete space-time diagram ; S5.3 for each node Initial characteristics of (a) The EMD algorithm is adopted to decompose the EMD algorithm into 4 intrinsic mode functions IMF and 1 residual term The decomposition formula is: Wherein, the 、 Corresponding to high-frequency dynamic characteristics, wherein the high-frequency dynamic characteristics comprise gas concentration mutation characteristics and stress value instantaneous fluctuation characteristics; 、 corresponding to low-frequency background characteristics, wherein the low-frequency background characteristics comprise gas concentration baseline characteristics and stress value overall distribution trend characteristics, and the residual items Representing the long-term change trend of the initial characteristics, and temporarily not participating in subsequent characteristic processing; S5.4, for each of the natural mode functions Performing linear transformations alone Will be of different scales The feature is mapped to feature dimension matched with the convolution of the subsequent graph in a unified way, and then the feature nonlinear expression capacity is enhanced through a ReLU activation function, wherein the processing formula is as follows: Wherein, the Is the first The linear transformation weight matrix of the individual IMFs, Is the first The linear transformation bias terms of the individual IMFs, Is the first The features of the IMFs after mapping and activation; s5.5, adopting characteristic splicing function Will 4 Splicing along the dimension of the feature channel to obtain the multi-scale fusion feature of the node Acquiring the space-time diagram Final adjacency matrix of (2) Constructing ChebyNet graph convolution operator based on Chebyshev polynomial, and fusing the multi-scale fusion features With the final adjacency matrix Inputting ChebyNet, extracting node local association features and global space-time dependence features, and outputting an updated node feature matrix after being activated by a ReLU; And S5.6, performing feature integration on the updated node feature matrix to obtain multi-scale space-time features which cover high-frequency dynamic features and low-frequency background features and contain space-time correlation information, and completing the processing process of the EMD-GCN model.
  9. 9. The method according to claim 8, wherein the step S6 of fusing the three-dimensional space risk feature and the multi-scale space-time feature to obtain a four-dimensional risk identification result of the mine salient dangerous area specifically comprises the following sub-steps: S6.1, adopting a characteristic channel splicing mode, and based on the space dimension of the three-dimensional space risk characteristic Time dimension with the multi-scale spatio-temporal features Establishing a space-time correlation dimension, splicing a space feature matrix formed by the three-dimensional space risk features and a time sequence feature matrix formed by the multi-scale space-time features along the space-time correlation dimension to form a space-containing dimension Dimension with time The four-dimensional feature matrix is normalized, and finally the four-dimensional risk identification result is output, wherein the four-dimensional risk identification result is used for representing mine salient risk levels at different time points and different spatial positions; in addition, in the step S7, a multi-task fully-connected network is constructed and adopted, the control policy parameters are input into the multi-task fully-connected network, and the cost score, the risk score and the feasibility score are output, which specifically comprises the following sub-steps: S7.1, constructing a multi-task full-connection network, wherein the input of the multi-task full-connection network is a control strategy parameter, and the control strategy parameter comprises extraction rate and drilling depth; S7.2, training the multi-task full-connection network by adopting mine history prevention and control data, wherein the history prevention and control data comprises prevention and control strategy records, corresponding actual cost, actual risk change and actual feasibility evaluation results, and a loss function in the training process The mean value of the mean square error of the predicted value and the actual value of the cost score, the risk score and the feasibility score is set as the formula: Wherein, the Output as network A class score prediction value is used for the class, Is the first in the historical data The actual value of the class score is calculated, The cost is indicated by the fact that, A risk is indicated as such, Representing feasibility; And, recommending a mine outburst risk control scheme by a multi-objective optimization method based on the cost score, the risk score and the feasibility score in the step S8, specifically comprising the following sub-steps: S8.1, constructing a reinforcement learning agent, wherein the action space of the agent is an alternative control strategy set, the state space is the current mine risk state, and the rewarding function is set as Training an agent to learn optimal strategy selection logic through a reinforcement learning algorithm; S8.2, executing Bayesian optimization, namely constructing a Gaussian process model to estimate the control strategy score distribution by taking the total strategy cost less than or equal to a preset cost threshold value, the risk score more than or equal to a preset risk threshold value and the feasibility score more than or equal to a preset feasibility threshold value as constraint conditions, and searching a pareto optimal strategy set meeting the constraint by adopting Expected Improvement criteria; And S8.3, screening a comprehensive optimal scheme from the pareto optimal strategy set, wherein the comprehensive optimal scheme meets constraint requirements, and the grading difference value of any two types does not exceed a preset difference value range, so that the comprehensive optimal scheme is used as a finally recommended mine outburst risk control scheme.

Description

Four-dimensional GIS-based mine salient dangerous area dynamic identification and control decision method Technical Field The application relates to the technical field of mine safety prevention and control, in particular to a four-dimensional GIS-based dynamic identification and prevention decision method for a mine salient dangerous area. Background Mine outburst risks (such as gas outburst and rock burst) prevention and control depend on accurate integration and dynamic analysis of multi-source data such as geological maps, logging profiles, sensor time sequences and the like, but the prior art has significant bottlenecks. In the data enhancement link, the methods such as CutMix and Mixup destroy key space information such as formation boundaries, coal seam contours and the like through square cutting or global pixel mixing, and do not conduct differentiation treatment on salient areas such as gas abnormal enrichment areas, logging abnormal sections and the like, so that the identification degree of high-risk areas is reduced, the existing masks are mostly random binary matrixes, so that accurate splicing of similar geological areas is realized through hard masks without profile protection, and fusion weights are dynamically adjusted through soft masks without importance. Meanwhile, the traditional three-dimensional segmentation technology cannot be generally fused with time dimension, and a graph neural network is difficult to directly adapt to non-stable time sequence data of a sensor, so that four-dimensional characteristics of space and time are difficult to extract, the problems of non-uniform multi-scale characteristic dimension, low spectral domain extraction efficiency and the like exist, in addition, a decision link generally lacks quantitative support of cost-risk-feasibility, and the requirements of dynamic identification and scientific prevention and treatment of a mine salient dangerous area cannot be met. Disclosure of Invention The application aims to provide a method for dynamically identifying and controlling a mine salient dangerous area based on a four-dimensional GIS so as to solve or at least partially solve the technical problems. The application provides a method for dynamically identifying, controlling and deciding a mine salient dangerous area based on a four-dimensional GIS, which comprises the following steps: s1, acquiring mine multi-source data, wherein the mine multi-source data comprises space static basic data, space-time dynamic monitoring data and numerical simulation auxiliary data, and enhancing the image data by adopting a HSMix framework to obtain enhanced image data; s2, mapping the enhanced image class data to a four-dimensional tile structure, wherein the four-dimensional tile structure comprises a space dimension and a time dimension; S3, constructing a space-time index based on the four-dimensional tile structure, wherein the space-time index distributes unique identifiers for each four-dimensional tile through time stamps and space coordinates; S4, processing the space dimension data mapped in the four-dimensional tiles by adopting a three-dimensional U-Net network to obtain three-dimensional space risk characteristics; S5, processing the sensor time sequence data in the time-space dynamic monitoring data by adopting an EMD-GCN model to obtain multi-scale time-space characteristics; s6, fusing the three-dimensional space risk features and the multi-scale space-time features to obtain a four-dimensional risk identification result of the mine salient risk region; s7, constructing a multi-task full-connection network, inputting control strategy parameters into the multi-task full-connection network, and processing and outputting cost scores, risk scores and feasibility scores by the multi-task full-connection network; And S8, recommending a mine outburst risk control scheme through a multi-objective optimization method based on the cost score, the risk score and the feasibility score. Optionally, in the step S1, the image class data is enhanced by adopting a HSMix framework, which specifically includes the following sub-steps: S1.1, respectively processing two source images to be enhanced by adopting an SLIC super-pixel algorithm to generate a first super-pixel grid and a second super-pixel grid which are respectively corresponding to each other; s1.2, randomly selecting a super pixel area from the second super pixel grid based on Bernoulli distribution, and generating a binary hard mixed mask; s1.3, fusing the two source images and the labels corresponding to the two source images according to the following formula to obtain hard mixed image data and hard mixed label data: Wherein, the For the image data after hard-blending,For the first one of the source images,For the second one of the source images,For the element-wise multiplication,In order to have a hard mix mask,To be the full size consistent with the hard mix maskThe mask is used to indicate that the mas