Coal mine roof dynamic monitoring and early warning method and system based on artificial intelligence

Abstract

The invention provides a coal mine roof dynamic monitoring and early warning method and system based on artificial intelligence, and relates to the field of coal mine safety, and the method comprises the steps: collecting a stress waveform and a displacement waveform of a coal mine roof, extracting stress loosening characteristics, and constructing a three-dimensional stress conduction strength field; identifying a stress sudden change transition point by using a recurrent neural network to form a stress conduction chain; analyzing a regional topology migration rule and extracting a stress field instability path; matching with historical data to determine a prediction window; and continuously verifying the instability path and generating a support reinforcement scheme. The coal mine roof instability risk can be predicted in advance, and the early warning accuracy and timeliness are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to coal mine safety technology, and in particular to a coal mine roof dynamic monitoring and early warning method and system based on artificial intelligence. BACKGROUND

[0002] In the process of coal mining, roof collapse accidents caused by roof instability are one of the main threats to coal mine safety production. Coal mine roof dynamic monitoring is a key means to prevent roof collapse accidents. Traditional roof monitoring technology is mainly based on single-point stress or displacement monitoring, which measures the stress change and deformation of the roof to determine the stability of the roof. With the development of sensor technology, multi-point distributed monitoring has gradually become the mainstream, which can obtain more comprehensive roof state information. In recent years, with the rapid development of artificial intelligence technology, applying intelligent algorithms such as deep learning to the field of coal mine safety monitoring has become a research hotspot. Through intelligent analysis of monitoring data, roof instability risks can be more accurately predicted.

[0003] However, existing technologies mainly focus on the change of a single parameter such as stress or displacement, lack comprehensive analysis of the phase relationship between stress and displacement, and are difficult to accurately capture the early characteristics of stress relaxation in the roof, resulting in insufficient sensitivity of the early warning. Traditional monitoring methods usually use static threshold judgment, which cannot effectively identify the stress conduction process and mutation characteristics under complex geological conditions, especially the stress migration law under the influence of complex geological structures such as faults, and cannot fully utilize historical experience data to construct accurate prediction models, resulting in insufficient practicality and pertinence of early warning information, and difficulty in providing effective technical support for support and reinforcement. SUMMARY

[0004] The present application provides a coal mine roof dynamic monitoring and early warning method and system based on artificial intelligence, which can solve the problems in the prior art.

[0005] In a first aspect of the present application, a coal mine roof dynamic monitoring and early warning method based on artificial intelligence is provided, comprising: Collecting coal mine roof stress waveforms and displacement waveforms at multiple monitoring points, extracting stress relaxation characteristics according to the phase decoupling relationship between the stress waveforms and the displacement waveforms, and constructing a three-dimensional stress conduction intensity field in combination with the spatial projection of the monitoring points and the faults; Performing spatio-temporal mapping of the stress relaxation characteristics in the three-dimensional stress conduction intensity field to obtain a dynamic characteristic sequence, identifying stress mutation transition points from the dynamic characteristic sequence using a recurrent neural network, and forming a stress conduction chain according to spatio-temporal correlation coefficients by density clustering of the stress mutation transition points; Analyzing the topological migration law of the area in the stress conduction chain whose correlation coefficient exceeds a preset correlation threshold, extracting the motion trajectory of the area in the three-dimensional space to obtain a stress field instability path; Geometric characteristics of the stress field instability path are analyzed and morphological matching is performed with instability paths in a historical working face database, a prediction window is determined according to the instability process of the working face with the optimal matching degree; The stress field instability path is continuously verified within the prediction window, when the stress field instability path is verified for multiple times in succession and the prediction window is less than the preset evacuation time, a support reinforcement scheme is generated according to the stress field instability path and a warning signal is sent.

[0006] According to the phase decoupling relationship between the stress waveform and the displacement waveform, stress relaxation characteristics are extracted, and a three-dimensional stress conduction intensity field is constructed by combining the monitoring points and the spatial projection of the fault, including: The stress waveform and the displacement waveform are subjected to Fourier transform to obtain waveform frequency spectra; Waveform phase characteristics are extracted from the waveform frequency spectra, the waveform phase characteristics are subjected to multi-scale wavelet decomposition to obtain a characteristic sequence, waveform periodic components and waveform trend components are separated from the characteristic sequence, and stress relaxation characteristics are obtained by phase decoupling of the waveform periodic components and the waveform trend components; The vertical distance and the projection angle of the monitoring points to the fault projection surface are calculated, a stress conduction coefficient is constructed, and a three-dimensional stress conduction intensity field is constructed by nonlinear mapping of the stress conduction coefficient and the stress relaxation characteristics.

[0007] The stress relaxation characteristics are subjected to spatiotemporal mapping in the three-dimensional stress conduction intensity field to obtain a dynamic characteristic sequence, a stress mutation transition point is identified from the dynamic characteristic sequence by using a recurrent neural network, and the stress mutation transition points are formed into a stress conduction chain according to a spatiotemporal correlation coefficient by density clustering, including: The stress relaxation characteristics are subjected to spatiotemporal mapping in the three-dimensional stress conduction intensity field, a wave component is extracted by time domain transformation, a distribution component is obtained by spatial domain transformation, and a spatiotemporal feature matrix is constructed by combining the wave component and the distribution component; A dynamic characteristic sequence is extracted from the spatiotemporal feature matrix, and the dynamic characteristic sequence is subjected to time window segmentation to obtain a characteristic segment; A recurrent neural network is constructed, a memory unit and a forgetting gate are set in the recurrent neural network, the characteristic segment is input into the recurrent neural network, a feature change law is extracted by the memory unit and the forgetting gate, a stress mutation transition point and a transition intensity are identified from an output end; A time correlation is calculated by calculating the time interval between the stress mutation transition points, a spatial correlation is calculated by calculating the spatial distance between the stress mutation transition points, and a spatiotemporal correlation coefficient is constructed by fusing the time correlation and the spatial correlation; The stress mutation transition points are subjected to density clustering based on the spatiotemporal correlation coefficient, the density clustering results are sorted according to the transition intensity, and a stress conduction chain is constructed according to the conduction direction by using the sorting results.

[0008] analyzing the topological migration law of the region in the stress conduction chain whose correlation coefficient exceeds a preset correlation threshold, and extracting a motion trajectory of the region in a three-dimensional space to obtain a stress field instability path, comprising: calculating time sequence similarity of adjacent conduction points in the stress conduction chain to obtain a time sequence correlation degree, calculating a spatial distance and an included angle between the conduction points to obtain a spatial correlation degree, performing multi-dimensional weighted fusion on the time sequence correlation degree and the spatial correlation degree to obtain a correlation coefficient, and constructing a conduction point correlation matrix based on the correlation coefficient; constructing a topological graph structure based on the conduction point correlation matrix, taking the conduction points whose correlation coefficients exceed a preset correlation threshold as topological nodes, constructing a conduction subgraph by extracting a connection relationship between the topological nodes, and extracting a topological migration law by structure evolution and connection change of the conduction subgraph; mapping the topological migration law to a three-dimensional space coordinate system, constructing a motion trajectory vector sequence according to a distribution position and a motion speed of the topological nodes in the space, and extracting a stress field instability path based on direction change and amplitude change of the motion trajectory vector sequence.

[0009] performing geometric feature analysis on the stress field instability path and performing shape matching with instability paths in a historical working face database, and determining a prediction window according to an optimal matching degree of a working face in a failure process, comprising: performing geometric feature analysis on the stress field instability path, calculating a curvature change sequence of a roof fracture line, a crack propagation direction sequence, and a roof subsidence length sequence, and constructing a roof instability shape curve; extracting periodic subsidence features and stress concentration points from the roof instability shape curve, constructing a roof instability shape feature vector and performing standardization processing; extracting roof instability path samples from the historical working face database, calculating shape feature vectors of the roof instability path samples, matching the shape feature vectors with the standardized roof instability shape feature vectors, constructing a matching degree score matrix, and calculating a comprehensive matching degree based on the matching degree score matrix; sorting and selecting the roof instability path samples according to the comprehensive matching degree, and selecting a working face with the optimal comprehensive matching degree as a reference working face; extracting a roof caving time point and a roof damage range point from the reference working face, determining a time span of the prediction window according to the roof caving time point, determining a spatial boundary of the prediction window according to the roof damage range point, and outputting a final prediction window.

[0010] continuously verifying the stress field instability path in the prediction window, generating a support reinforcement scheme and issuing a warning signal according to the stress field instability path when the stress field instability path is continuously verified for multiple times and the prediction window is less than a preset evacuation time, comprising: The stress field data is collected in a prediction window, time series and spatial distribution information are extracted from the stress field data, a fluctuation amplitude curve of the time series is calculated through wave analysis, a stress gradient value of the spatial distribution is calculated through regional block division, and the fluctuation amplitude curve and the stress gradient value are combined to generate a stress field instability path evaluation sequence; A verification parameter of the fluctuation amplitude curve and the stress gradient value in the stress field instability path evaluation sequence is calculated, the verification parameter is compared with a preset safety threshold for verification, a stress field instability path that passes the verification is recorded, and a corresponding influence area and development direction are calculated to generate a stress field instability path evolution feature; A support reinforcement range is determined according to the stress field instability path evolution feature, a stress distribution is calculated based on the stress field instability path evaluation sequence in the support reinforcement range, and a support reinforcement scheme is generated; A time difference between the prediction window and a preset evacuation time is calculated, when the time difference is less than a safety response time and a continuous verification pass number of the stress field instability path reaches a preset number, a warning signal is generated according to the stress field instability path evolution feature and the support reinforcement scheme.

[0011] The verification parameter of the fluctuation amplitude curve and the stress gradient value in the stress field instability path evaluation sequence is calculated, the verification parameter is compared with a preset safety threshold for verification, a stress field instability path that passes the verification is recorded, and a corresponding influence area and development direction are calculated to generate a stress field instability path evolution feature, including: The fluctuation amplitude curve and the stress gradient value are extracted from the stress field instability path evaluation sequence, a fluctuation amplitude curve is subjected to time series difference to obtain an amplitude change rate, and the stress gradient value is subjected to spatial gradient calculation to obtain a gradient change intensity; The amplitude change rate and the gradient change intensity are subjected to spatio-temporal coupling analysis, a cross-response coefficient of the fluctuation amplitude curve and the stress gradient value is extracted, and a verification parameter is calculated based on the cross-response coefficient; The verification parameter is tracked in a continuous time period through a sliding window, when the verification parameter continuously meets the preset safety threshold in the continuous time period, it is determined that the verification passes, and a stress field instability path that passes the verification is recorded; A spatial distribution node of the stress field instability path that passes the verification is extracted, an influence area is calculated by calculating a coverage boundary of the spatial distribution node, a development direction is obtained by direction fitting on the spatial distribution node, and the influence area and the development direction are combined with time change to generate the stress field instability path evolution feature.

[0012] In a second aspect of the embodiment of the present application, a coal mine roof dynamic monitoring and early warning system based on artificial intelligence is provided, including: The first unit is configured to collect coal mine roof stress waveforms and displacement waveforms of a plurality of monitoring points, extract stress relaxation characteristics according to a phase decoupling relationship of the stress waveforms and the displacement waveforms, and construct a three-dimensional stress conduction intensity field in combination with spatial projection of the monitoring points and a fault; The second unit is configured to perform time-space mapping of the stress relaxation characteristics in the three-dimensional stress conduction intensity field to obtain a dynamic characteristic sequence, identify stress mutation transition points from the dynamic characteristic sequence by using a recurrent neural network, and form a stress conduction chain according to time-space correlation coefficients by density clustering. The third unit is configured to analyze a topological migration law of a region with a correlation coefficient exceeding a preset correlation threshold in the stress conduction chain, extract a motion trajectory of the region in a three-dimensional space to obtain a stress field instability path. The fourth unit is configured to perform geometric feature analysis on the stress field instability path and perform shape matching with instability paths in a historical working face database, determine a prediction window according to a most optimal working face instability process, and perform shape matching with instability paths in a historical working face database. The fifth unit is configured to continuously verify the stress field instability path in the prediction window, generate a support reinforcement scheme according to the stress field instability path and issue a warning signal when the stress field instability path is verified for a plurality of times and the prediction window is less than a preset evacuation time.

[0013] In a third aspect, an electronic device is provided, including: a processor; a memory configured to store processor-executable instructions; The processor is configured to invoke the instructions stored in the memory to execute the method described above.

[0014] In a fourth aspect, a computer-readable storage medium is provided, which stores computer program instructions. When the computer program instructions are executed by a processor, the method described above is implemented.

[0015] In the embodiment, the stress relaxation characteristics are extracted through the phase decoupling relationship between the stress waveform and the displacement waveform, and a three-dimensional stress conduction intensity field is constructed by combining the monitoring points and the spatial projection of the fault, thereby realizing accurate characterization of the stress state of the coal mine roof and overcoming the limitation of the traditional monitoring method which only relies on a single parameter. The stress mutation transition point is identified by using the recurrent neural network, and the stress conduction chain is formed through density clustering, which can capture the small changes and correlations in the stress field, thereby improving the sensitivity and accuracy of the early warning. The stress field instability path is obtained by analyzing the topological migration law of the high correlation coefficient region in the stress conduction chain, thereby realizing the visual expression and quantitative analysis of the dynamic instability process of the roof. The current stress field instability path is matched with the historical working face database in shape, a prediction window is determined based on the instability process of the similar working condition, and the early warning system has the functions of adaptive learning and experience knowledge accumulation. The stress field instability path is continuously verified in the prediction window, and a multiple confirmation mechanism is set, thereby significantly reducing the false positive rate and automatically generating a targeted support reinforcement scheme according to the instability path, which provides sufficient preparation time for the safe evacuation of miners and the reinforcement of the roof. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 A flowchart of the coal mine roof dynamic monitoring and early warning method based on artificial intelligence according to the embodiment of the present application is shown in Figure 2 A flowchart of the stress field instability path verification and early warning generation in the prediction window according to the embodiment of the present application is shown in DETAILED DESCRIPTION

[0017] To make the objectives, technical solutions, and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described below in conjunction with the drawings of the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0018] The technical solutions of the present application will be described in detail below with specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes can not be described in some embodiments.

[0019] Figure 1 A flowchart of the coal mine roof dynamic monitoring and early warning method based on artificial intelligence according to the embodiment of the present application is shown in Figure 1 The method comprises: The stress relaxation characteristics are extracted according to the phase decoupling relationship of the stress waveform and the displacement waveform, and a three-dimensional stress conduction intensity field is constructed by combining the monitoring points and the spatial projection of the fault; The dynamic characteristic sequence is obtained by mapping the stress relaxation characteristics in the three-dimensional stress conduction intensity field in space-time, the stress mutation transition points are identified from the dynamic characteristic sequence by using a recurrent neural network, and the stress conduction chain is formed by the stress mutation transition points according to the space-time correlation coefficient through density clustering; The topological migration law of the region with a correlation coefficient exceeding a preset correlation threshold in the stress conduction chain is analyzed, the motion trajectory of the region in the three-dimensional space is obtained, and a stress field instability path is obtained; The stress field instability path is analyzed for geometric characteristics and is morphologically matched with the instability paths in a historical working face database, a prediction window is determined according to the optimal matching degree of the working face instability process; The stress field instability path is continuously verified within the prediction window, and when the stress field instability path is verified for multiple times and the prediction window is less than a preset evacuation time, a support reinforcement scheme is generated according to the stress field instability path and a warning signal is issued.

[0020] In an alternative embodiment, the stress relaxation characteristics are extracted according to the phase decoupling relationship of the stress waveform and the displacement waveform, and a three-dimensional stress conduction intensity field is constructed by combining the monitoring points and the spatial projection of the fault, which includes: The stress waveform and the displacement waveform are subjected to Fourier transform to obtain waveform frequency spectra; Waveform phase characteristics are extracted from the waveform frequency spectra, the waveform phase characteristics are subjected to multi-scale wavelet decomposition to obtain a characteristic sequence, waveform periodic components and waveform trend components are separated from the characteristic sequence, and the stress relaxation characteristics are obtained by phase decoupling of the waveform periodic components and the waveform trend components; The vertical distance and the projection angle of the monitoring points to the fault projection surface are calculated, a stress conduction coefficient is constructed, and the three-dimensional stress conduction intensity field is constructed by nonlinear mapping of the stress conduction coefficient and the stress relaxation characteristics.

[0021] After the coal mine roof monitoring device collects stress waveform and displacement waveform data, the two types of waveform data are processed to extract stress relaxation characteristics. Specifically, the collected stress waveform and displacement waveform are converted to frequency domain space by Fourier transform to obtain the corresponding waveform frequency spectra. During Fourier transform, fast Fourier algorithm is used to process time domain data, and the stress waveform and displacement waveform of each monitoring point are converted into frequency spectrum representation composed of amplitude and phase. For monitoring data with a sampling rate of 200 points per second, 1024-point fast Fourier transform can obtain the frequency spectrum distribution in the range of 0-100 Hz.

[0022] From the waveform spectrum, the waveform phase feature is the next key operation. The waveform phase feature reflects the timing relationship between the stress wave and the displacement wave, and has important significance for identifying the coal mine roof loosening state. The extracted waveform phase feature is further processed by multi-scale wavelet decomposition. The Daubechies wavelet basis function is used for 5-layer decomposition of the phase feature, and a series of wavelet coefficients at different scales are obtained to form a feature sequence. For coal mine roof stress monitoring, the decomposition level is selected as 5 considering the frequency characteristics of coal mine roof stress changes, which can effectively capture the stress change characteristics from high frequency to low frequency.

[0023] The feature sequence contains waveform information at different frequencies and time scales. From the feature sequence, the waveform periodic component and the waveform trend component are separated using the empirical mode decomposition method. The waveform periodic component reflects the periodic variation law of the roof stress, and the waveform trend component reflects the long-term trend. In specific implementation, the wavelet coefficients are recombined according to the frequency, and the wavelet coefficients with a frequency greater than 1 Hz are reconstructed as the periodic component, and the wavelet coefficients with a frequency less than 1 Hz are reconstructed as the trend component.

[0024] The separated waveform periodic component and waveform trend component are subjected to phase decoupling operation to obtain stress loosening features. In the phase decoupling process, the phase difference between the stress waveform and the displacement waveform on the periodic component and the trend component is calculated, and the stress loosening feature is defined according to the rate of change of the phase difference. When the coal mine roof appears loosening, the phase difference between the stress wave and the displacement wave will change significantly. In actual monitoring, when the phase difference change rate exceeds 0.2 rad / min, it indicates that the roof may have loosening signs.

[0025] To construct a three-dimensional stress conduction intensity field, the vertical distance and projection angle of the monitoring point to the fault projection surface need to be calculated to construct the stress conduction coefficient. By obtaining the three-dimensional geological model of the coal mine fault, the vertical distance of each monitoring point to the nearest fault surface and its projection position on the fault surface are calculated. For the vertical distance, the Euclidean distance calculation method is used; for the projection angle, the angle between the monitoring point position vector and the fault surface normal vector is calculated. In practical application, when the vertical distance of the monitoring point is 50 meters and the projection angle is 30 degrees, the stress conduction coefficient is about 0.65, indicating that the degree of influence of the monitoring point on the fault is moderate.

[0026] The stress conduction coefficient reflects the degree of influence of the monitoring point on the fault, is nonlinearly mapped with the stress relaxation feature, and a three-dimensional stress conduction intensity field is constructed. The nonlinear mapping is realized by using a radial basis function network, the stress conduction coefficient is taken as a weight factor, and after being multiplied by the stress relaxation feature, spatial interpolation is performed through the radial basis function to generate a continuous three-dimensional stress conduction intensity field. The intensity field is represented by a three-dimensional grid, and the grid resolution is 5 meters, covering the entire monitoring area. In a coal mine working face with a length of 150 meters and a width of 80 meters, the three-dimensional stress conduction intensity field constructed contains about 24,000 grid points, and each grid point has a value representing the stress conduction intensity.

[0027] In this embodiment, by performing phase decoupling analysis on the stress waveform and the displacement waveform, the relaxation feature of the coal mine roof can be accurately identified, and the sensitivity and accuracy of the roof instability early warning are improved. The three-dimensional stress conduction intensity field constructed in combination with the spatial relationship between the monitoring point and the fault realizes the visual expression of the stress conduction path, and provides a spatial reference framework for subsequent dynamic feature analysis. The safety level of coal mine production is greatly improved, and the risk of safety accidents caused by roof instability is reduced.

[0028] In an alternative embodiment, the stress relaxation feature is spatiotemporally mapped in the three-dimensional stress conduction intensity field to obtain a dynamic feature sequence, a recurrent neural network is used to identify stress mutation transition points from the dynamic feature sequence, and a stress conduction chain is formed from the stress mutation transition points according to the spatiotemporal correlation coefficient through density clustering. The stress relaxation feature is spatiotemporally mapped in the three-dimensional stress conduction intensity field, the wave component is extracted through time domain transformation, and the distribution component is obtained through space domain transformation. The wave component and the distribution component are combined to construct a spatiotemporal feature matrix; A dynamic feature sequence is extracted from the spatiotemporal feature matrix, and the dynamic feature sequence is divided into feature segments by a time window; A recurrent neural network is constructed, a memory unit and a forgetting gate are set in the recurrent neural network, the feature segments are input into the recurrent neural network, the feature change rule is extracted through the memory unit and the forgetting gate, and stress mutation transition points and transition intensities are identified from the output end; The time correlation is calculated by calculating the time interval between the stress mutation transition points, the space correlation is calculated by calculating the spatial distance between the stress mutation transition points, and the time correlation and the space correlation are fused to construct a spatiotemporal correlation coefficient; The stress mutation transition points are density clustered based on the spatiotemporal correlation coefficient, the density clustering results are sorted according to the transition intensities, and the stress conduction chain is constructed according to the conduction direction based on the sorting results.

[0029] The stress relaxation feature needs to be mapped in time domain and space domain in the three-dimensional stress conduction intensity field to obtain the complete space-time feature. During the time domain mapping process, the stress relaxation feature of each monitoring point is processed by sliding window, the window length is 60 minutes, and the window sliding step is 5 minutes, and the trend of the stress relaxation feature of each monitoring point changing with time is extracted. By discrete wavelet transform on the data in the window, the fluctuation component of the stress relaxation feature is extracted, reflecting the short-term change of the coal mine roof stress. In the space domain mapping process, the Kriging interpolation method is used to map the discrete stress relaxation features of the monitoring points to the grid of the three-dimensional stress conduction intensity field, to obtain the spatial distribution component of the stress relaxation feature, reflecting the distribution state of the coal mine roof stress in space. When a three-dimensional grid with a grid size of 5 meters is used, there are about 24000 grid points in a typical coal mine working face area, and each grid point has a stress relaxation feature value interpolated from the surrounding monitoring points.

[0030] In the process of combining the fluctuation component and the distribution component to construct the space-time feature matrix, the fluctuation component is arranged in time sequence to form the time dimension, and the distribution component is arranged according to the spatial position to form the space dimension, and the two are combined orthogonally to form the space-time feature matrix. Each element in the matrix represents the stress relaxation feature value at a specific time and a specific position, and the matrix size is related to the number of grid points in the monitoring area and the monitoring time. In a 24-hour monitoring period, the generated space-time feature matrix contains 288 time points (based on a sliding step of 5 minutes) and 24000 spatial points, and the total size of the matrix is 288x24000.

[0031] When extracting the dynamic feature sequence from the space-time feature matrix, the principal component analysis method is used for dimension reduction processing to extract the main change mode of the space-time feature matrix and reduce data redundancy. The principal components with a contribution rate of more than 85% are selected as the dynamic feature sequence, and usually the first 10-15 principal components are retained. The dynamic feature sequence reflects the main mode of the coal mine roof stress changing with time and space, and can effectively capture the dynamic evolution process of the stress field. Then the dynamic feature sequence is segmented according to a fixed time window to obtain feature segments, and the window length is set to 12 time points, i.e. 60 minutes, and the overlap rate of adjacent windows is 50%. The feature segments are used as the input units of the recurrent neural network for subsequent stress sudden change transition point identification.

[0032] The constructed recurrent neural network adopts a long short-term memory network structure, including an input layer, a hidden layer, and an output layer. The number of input layer nodes is consistent with the feature segment dimension, the hidden layer contains 64 long short-term memory units, each unit is provided with a memory unit and a forgetting gate. The memory unit is responsible for storing historical feature information, and the forgetting gate controls the retention degree of historical information, enabling the network to learn long-term sequential dependencies. The number of output layer nodes is 2, representing the stress mutation probability and the transition intensity, respectively. The network training uses known stress mutation events in historical monitoring data, and uses the stochastic gradient descent algorithm to optimize the network parameters, with a learning rate of 0.01 and a training iteration number of 1000 times. After the model converges, new feature segments are input into the trained recurrent neural network, and the output end gives the stress mutation probability and transition intensity at each time point. When the stress mutation probability exceeds the threshold value of 0.75, the point is identified as a stress mutation transition point, and the corresponding transition intensity, time and spatial position information are recorded.

[0033] When calculating the time correlation between stress mutation transition points, the time interval of each pair of transition points is calculated, and the time interval is converted into time correlation through an exponential decay function. The shorter the time interval, the higher the correlation. When the time interval is 30 minutes, the time correlation is about 0.7; when the time interval is 120 minutes, the time correlation decreases to 0.3. In the calculation of spatial correlation, the Euclidean distance is used to measure the spatial distance between transition points, and the spatial distance is converted into spatial correlation through a Gaussian kernel function. The closer the spatial distance, the higher the correlation. When the spatial distance is 10 meters, the spatial correlation is about 0.9; when the spatial distance is 50 meters, the spatial correlation decreases to 0.4. The fusion of time correlation and spatial correlation adopts a weighted average method, with a weight ratio of 6:4, to obtain the integrated spatiotemporal correlation coefficient. The spatiotemporal correlation coefficient reflects the correlation degree between stress mutation transition points, with a coefficient value ranging from 0 to 1.

[0034] When clustering the stress mutation transition points based on the spatiotemporal correlation coefficient, a density-based spatial clustering algorithm is used. The neighborhood radius is set to 0.6, the minimum number of points is 4, and the spatiotemporal correlation coefficient is used as the distance measure for clustering. The neighborhood radius represents the spatiotemporal correlation threshold for classifying two transition points into the same class, and the minimum number of points represents the minimum number of transition points required to form a class cluster. The clustering results are sorted according to the number of transition points contained and the average transition intensity, with higher transition intensity ranking higher. The sorted clustering results are connected in the order of stress conduction time, and the stress conduction chain is constructed. The conduction direction is from the early occurring transition point to the late occurring transition point, forming a complete stress conduction chain structure.

[0035] In this embodiment, the precise identification and conduction chain construction of coal mine roof stress mutation are realized by spatio-temporal mapping and deep learning method, which has significant technical effect. The recurrent neural network is used to process dynamic characteristic sequence, which can effectively capture the time sequence variation law of stress field and identify the tiny stress mutation transition point that cannot be found by traditional methods. The density clustering method based on spatio-temporal correlation coefficient realizes the automatic discovery of stress conduction path and reveals the internal law of coal mine roof stress conduction. The stress conduction chain constructed intuitively displays the evolution path of stress field, which provides a key basis for predicting the instability of coal mine roof.

[0036] In an optional implementation, the topological migration law of the region in the stress conduction chain whose correlation coefficient exceeds the preset correlation threshold is analyzed, and a stress field instability path is obtained by extracting the motion trajectory of the region in a three-dimensional space, comprising: The time sequence similarity of adjacent conduction points in the stress conduction chain is calculated to obtain the time sequence correlation degree, the spatial distance and the included angle between the conduction points are calculated to obtain the spatial correlation degree, the time sequence correlation degree and the spatial correlation degree are multi-dimensionally weighted and fused to obtain the correlation coefficient, and the conduction point correlation matrix is constructed based on the correlation coefficient; The topological graph structure is constructed based on the conduction point correlation matrix, the conduction points whose correlation coefficients exceed the preset correlation threshold are taken as topological nodes, the connection relationship between the topological nodes is extracted to construct a conduction subgraph, and the topological migration law is extracted through the structure evolution and connection change of the conduction subgraph; The topological migration law is mapped to a three-dimensional space coordinate system, a motion trajectory vector sequence is constructed according to the distribution position and motion speed of the topological nodes in space, and a stress field instability path is extracted based on the direction change and amplitude change of the motion trajectory vector sequence.

[0037] Firstly, the time sequence data and spatial position information of each conduction point in the stress conduction chain are obtained. The stress conduction chain is usually composed of a series of monitoring points distributed in space, and each monitoring point records the stress data changing with time. These monitoring points can be stress sensors buried in rock-soil mass, or stress conduction points obtained by displacement monitoring inversion.

[0038] For any two conduction points in the stress conduction chain, the time window of the stress value change sequence is extracted, and the similarity of the two sequences is calculated by using the dynamic time warping algorithm to obtain the time correlation. The dynamic time warping algorithm solves the problem of time axis misalignment caused by different sampling frequencies by finding the best matching path between the two time sequences. In coal mine roof monitoring, the length of the stress value sequence of each conduction point is 60 sampling points, and the time correlation calculation result of the adjacent two points falls between 0 and 1. The larger the value, the stronger the correlation. In practical application, when the stress change trend of two conduction points is basically the same, the time correlation is usually more than 0.85; when the stress change trend is significantly different, the time correlation may be as low as 0.3.

[0039] The spatial correlation calculation includes two dimensions of spatial distance and angle. The spatial distance calculates the straight-line distance between the conduction points by using the Euclidean distance. The closer the distance, the higher the spatial correlation. The angle calculation considers the position relationship of the conduction points relative to the working face advancing direction, and evaluates the spatial relationship between the conduction points by calculating the angle between the connecting line of the two points and the working face advancing direction. In coal mine roof monitoring, when the distance between the conduction points is 10 meters and the angle is 15 degrees, the spatial correlation is about 0.92; when the distance increases to 50 meters and the angle increases to 45 degrees, the spatial correlation decreases to 0.38. When the time correlation and the spatial correlation are multi-dimensionally weighted and fused, the weighted average method is used, the weight of the time correlation is 0.6, and the weight of the spatial correlation is 0.4. The correlation coefficient is calculated by comprehensive calculation. The correlation coefficient reflects the comprehensive correlation between the conduction points, which is used to construct the conduction point correlation matrix. The conduction point correlation matrix is a symmetric matrix, the matrix size is the same as the number of conduction points, and the matrix element value is the correlation coefficient between the corresponding conduction points.

[0040] When constructing the topology graph structure based on the conductive point correlation matrix, the preset correlation threshold is set to 0.75, the conductive point pairs with correlation exceeding the threshold are taken as the topology nodes, and the connection relationship between these nodes is extracted to construct the conductive subgraph. The conductive subgraph is represented by an undirected weighted graph structure, the nodes represent the conductive points, the edges represent the connection relationship between the conductive points, and the weight of the edge is the correlation coefficient. In coal mine roof monitoring, a typical conductive subgraph contains about 25-40 nodes, reflecting the high correlation area in the stress conduction network. The evolution of the structure of the conductive subgraph is analyzed by comparing the changes of the subgraph in different time windows, including the changes of the number of nodes, the connection mode of the edges, the density of the graph, etc. The connection changes are described by the increase and decrease of the edges and the weight changes. In the process of stress conduction of the coal mine roof, the conductive subgraph usually presents an evolution process from loose to tight and then to decomposition, reflecting the dynamic changes of the stress field from stable to concentrated and then to release. The topology migration law extraction uses the graph embedding algorithm to map the conductive subgraph to a low-dimensional feature space and extract the main mode of the change of the graph structure. The trajectory in the feature space reflects the evolution path of the conductive subgraph over time, and the turning point of the trajectory corresponds to the turning point of the stress field state.

[0041] When mapping the topology migration law to the three-dimensional space coordinate system, it is necessary to convert the abstract trajectory in the graph feature space into the actual space position sequence. According to the distribution position of the topology nodes in space, the spatial center position of the conductive subgraph in each time window is calculated, which is obtained by weighted average of the spatial coordinates of all nodes in the subgraph, and the weight is the degree centrality of the node. The degree centrality represents the connection importance of the node in the graph, and the nodes with more connections have higher degree centrality. By connecting the spatial center positions of different time windows, the motion trajectory vector sequence is formed. The motion trajectory vector sequence describes the moving path of the stress conduction center in the three-dimensional space, and each vector contains the starting position, the ending position, the direction and the amplitude information. In coal mine roof monitoring, a typical motion trajectory vector sequence contains about 20-30 vectors, spanning the entire monitoring area.

[0042] The stress field instability path is extracted based on the direction change and amplitude change of the sequence of motion trajectory vectors. The direction change is calculated by the angle between adjacent vectors, and the amplitude change is calculated by the rate of change of vector length. When the direction change is greater than 30 degrees or the amplitude change rate exceeds 50%, it is marked as a trajectory inflection point, which reflects the significant change of the stress field conduction direction or intensity. Connecting all trajectory inflection points and the paths between them forms the stress field instability path. The stress field instability path is the key motion trajectory of the stress conduction center in space, reflecting the evolution process of the stress field from stability to instability. In the application of coal mine roof monitoring, the stress field instability path usually appears as a curve that moves from the edge to the center of the working face and then extends along a certain direction. The path length is usually between 80-120 meters, containing 5-8 key inflection points. The geometric characteristics of the instability path, such as curvature, length, direction change frequency, etc., are important indicators for evaluating the risk of roof instability.

[0043] In this embodiment, by analyzing the topological migration law in the stress conduction chain and extracting three-dimensional space motion trajectories, the accurate identification of the stress field instability path is achieved, which has significant technical effects. This method breaks through the limitations of traditional single-point monitoring, integrates discrete conduction points into a coherent topological structure, and reveals the overall evolution law of the stress field. Through multi-dimensional correlation coefficient calculation and topological graph analysis, the stress conduction pattern that is difficult to discover by conventional methods is captured. The extracted stress field instability path directly displays the spatial trajectory of the roof stress conduction, providing accurate basis for predicting the position of roof instability.

[0044] In an alternative embodiment, the stress field instability path is analyzed for geometric characteristics and matched with the instability path in the historical working face database, and the prediction window is determined according to the instability process of the working face with the optimal matching degree, including: The stress field instability path is analyzed for geometric characteristics, the curvature change sequence of the roof fracture line, the crack propagation direction sequence, and the roof subsidence length sequence are calculated, and the roof instability morphology curve is constructed; Periodic subsidence features and stress concentration points are extracted from the roof instability morphology curve, a roof instability morphology feature vector is constructed and standardized, and the roof instability morphology feature vector is matched with the standardized roof instability morphology feature vector; The roof instability path samples are extracted from the historical working face database, the morphology feature vectors of the roof instability path samples are calculated, the morphology feature vectors are matched with the standardized roof instability morphology feature vectors, a matching degree score matrix is constructed, and a comprehensive matching degree is calculated based on the matching degree score matrix; The roof instability path samples are sorted according to the comprehensive matching degree, and the working face with the optimal comprehensive matching degree is selected as the reference working face; The roof falling time point and the roof failure range point are extracted from the reference working face, the time span of the prediction window is determined according to the roof falling time point, the spatial boundary of the prediction window is determined according to the roof failure range point, and the final prediction window is output.

[0045] When the geometric characteristics of the stress field instability path are analyzed, the curvature variation sequence of the roof fracture line, the crack propagation direction sequence and the roof subsidence length sequence need to be calculated. The curvature variation sequence calculation method is to discretely sample the stress field instability path in three-dimensional space, take a sampling point every 5 meters, calculate the curvature of the circle formed by the adjacent three points, and obtain the curvature sequence distributed along the path. In coal mine roof monitoring, the curvature variation sequence reflects the bending degree of the roof fracture line, and the greater the curvature value, the more acute the turning of the path and the higher the roof stress concentration degree. The crack propagation direction sequence is obtained by calculating the angle between the connecting line of the adjacent two points on the path and the working face trend, recording the main direction change of the crack propagation. In the working face monitoring of a certain coal mine, the angle between the crack propagation direction and the working face trend is 15° to 75°, and the average is 45°, indicating that the crack mainly expands along the seam trend. The roof subsidence length sequence is calculated according to the vertical displacement of each point on the path, reflecting the spatial distribution of the roof subsidence. In a typical working face, the roof subsidence length has obvious differences at different positions of the path, ranging from a few millimeters to a few centimeters, forming a wavy distribution. The combination of the three sequences constitutes the roof instability form curve, which comprehensively describes the geometric characteristics of the roof instability.

[0046] When the periodic subsidence feature and stress concentration point are extracted from the roof instability form curve, wavelet transform is used for time-frequency analysis of the roof subsidence length sequence to extract the periodic subsidence feature. For a 120-meter long instability path, usually 2-4 subsidence periods can be identified, with a period length of 20-40 meters. The stress concentration point is determined by analyzing the extreme points of the curvature variation sequence, and the position corresponding to the local maximum curvature is the stress concentration point. In a typical working face, there are usually 3-5 stress concentration points, distributed at the turning points of the instability path. The periodic subsidence feature and the stress concentration point jointly constitute the roof instability form feature vector, with a vector dimension of usually 10-15 feature values, containing period length, period amplitude, stress concentration point position and intensity, etc. Standardization processing is to normalize each element in the feature vector to the range of 0-1, eliminating the dimension influence, and facilitating subsequent matching calculation.

[0047] When extracting the roof instability path sample from the historical working face database, the working face sample with complete monitoring record is selected, and the database contains the roof instability path data of 100 historical working faces. For each historical sample, the morphological feature vector is calculated according to the same method, and the historical feature library is constructed. The standardized roof instability morphological feature vector of the current working face is matched with each vector in the historical feature library, and the cosine similarity method is used to calculate the similarity between the feature vectors. The similarity value ranges from 0 to 1, and the larger the value, the higher the matching degree. For a coal mine working face, the matching degree with the most similar working face in the historical library is 0.92, indicating that the two have high similarity. Based on the matching calculation results of all historical samples, a matching score matrix is constructed. The row of the matrix represents the features of the current working face, and the column represents the historical samples, and the matrix element value is the matching degree of the corresponding feature. The comprehensive matching degree is calculated by weighted summation of the score matrix, and the weight distribution is determined according to the importance of the feature. The periodic subsidence feature weight is usually 0.5, and the stress concentration point weight is 0.5.

[0048] When the comprehensive matching degree is used to screen and sort the roof instability path sample, the historical working faces are arranged from high to low according to the comprehensive matching degree. In the sorting result, the working face with a comprehensive matching degree exceeding 0.8 is considered as a high matching sample, and usually 5-10 historical working faces meet the condition. From the high matching sample, the working face with the highest comprehensive matching degree is selected as the reference working face. The comprehensive matching degree of the reference working face is usually above 0.85, ensuring the reliability of the prediction result. If the matching degrees of multiple working faces are similar, factors such as geological condition similarity and mining technology similarity can be considered for secondary screening to select the most suitable reference working face.

[0049] When extracting the roof caving time point and the roof failure range point from the reference working face, the historical monitoring data of the reference working face is analyzed to determine the time interval from the formation of the stress field instability path to the roof caving. In a case of a working face in a certain coal mine, the time interval from the formation of the stress field instability path to the roof caving of the reference working face is 4.5 hours. According to the time point of the formation of the stress field instability path of the current working face, the time interval is added to determine the time span of the prediction window. The time span of the prediction window is usually set to 0.8-1.2 times of the historical reference time interval to consider the uncertainty caused by the difference in geological conditions. The roof failure range point is determined by analyzing the spatial distribution of the roof caving of the reference working face, including the starting position, expansion direction and final range of the caving. The roof failure range of the reference working face is mapped into the coordinate system of the current working face, considering the differences in working face size, strike, etc. to determine the spatial boundary of the prediction window. The spatial boundary usually includes a polygonal region covering all areas where roof caving is expected to occur, with a certain safety margin. The final output prediction window includes two dimensions of time span and spatial boundary, providing a basis for subsequent verification and early warning.

[0050] In the embodiment, the accurate determination of the roof instability prediction window is realized by analyzing the geometric characteristics of the stress field instability path and matching the shape with the historical working face data, which has a significant technical effect. The method fully utilizes historical experience data, extracts multi-dimensional geometric characteristics such as curvature change, crack propagation direction and roof subsidence length, and fully describes the morphological characteristics of roof instability. The prediction mechanism based on shape matching effectively solves the problem of relying on expert experience and lacking objective basis in traditional prediction methods, making the prediction result more reliable. By selecting the historical working face with the optimal matching degree as the reference, the accuracy and practicality of the prediction window are ensured.

[0051] As shown in Figure 2 , a flow chart of the prediction window stress field instability path verification and early warning generation process for the embodiment of the present application is shown in Figure 2 .

[0052] In an alternative embodiment, the stress field instability path is continuously verified within the prediction window, and when the stress field instability path is continuously verified for multiple times and the prediction window is less than the preset evacuation time, a support reinforcement scheme is generated according to the stress field instability path and an early warning signal is issued, which includes: Stress field data is collected within the prediction window, time series and spatial distribution information are extracted from the stress field data, the fluctuation amplitude curve of the time series is calculated by fluctuation analysis, the stress gradient value of the spatial distribution is calculated by region blocking, and the fluctuation amplitude curve and the stress gradient value are combined to generate a stress field instability path evaluation sequence; The verification parameter of the fluctuation amplitude curve and the stress gradient value in the stress field instability path evaluation sequence is calculated, the verification parameter is compared with a preset safety threshold, a stress field instability path that passes the verification is recorded, and a corresponding influence area and development direction are calculated to generate stress field instability path evolution characteristics; A support reinforcement range is determined according to the stress field instability path evolution characteristics, and a stress distribution is calculated based on the stress field instability path evaluation sequence in the support reinforcement range to generate a support reinforcement scheme; A time difference between a prediction window and a preset evacuation time is calculated, when the time difference is less than a safety response time and a continuous verification pass number of the stress field instability path reaches a preset number, a warning signal is generated according to the stress field instability path evolution characteristics and the support reinforcement scheme.

[0053] Collecting stress field data in the prediction window is a basic step of continuous verification. The stress sensor network arranged on the coal mine roof collects stress data at a frequency of 5 times per second, and the data collection covers the entire prediction window time period. The stress field data includes stress values, displacement values and acoustic emission signals of each monitoring point, which collectively reflect the stress state changes of the coal mine roof. The time series extracted from the stress field data is the stress value of each monitoring point changing with time, and the continuous stress time series is obtained. The spatial distribution information is obtained by spatial interpolation of the stress values of different monitoring points at the same time, and a stress distribution map covering the entire working face is generated.

[0054] When calculating the fluctuation amplitude curve of the time series through wave analysis, the sliding window method is adopted, the window length is 10 minutes, and the sliding step is 1 minute. The standard deviation of the data in each window is calculated to obtain the amplitude value of stress fluctuation, and the fluctuation amplitude curve is formed during the window sliding process. The fluctuation amplitude curve reflects the trend of the stress field fluctuation degree changing with time, and the fluctuation intensification usually means the decrease of the roof stability. For the collected stress data, the fluctuation amplitude gradually increases from the initial 0.15 MPa to 0.45 MPa, indicating that the roof stress fluctuation is obviously enhanced. When calculating the stress gradient value of the spatial distribution through regional block, the working face is divided into 5m x 5m grid units, and the stress gradient value of each grid unit is calculated, that is, the stress change rate between adjacent grid points. The stress gradient value reflects the degree of non-uniformity of the stress spatial distribution, and the greater the gradient value, the more obvious the stress concentration, and the higher the risk of roof instability. In a specific working face, the stress gradient value reaches 0.08 MPa / m near the potential fracture line of the roof, which is 3-4 times that of the normal area. The fluctuation amplitude curve and the stress gradient value are combined to generate the stress field instability path evaluation sequence, which contains the stress field characteristics of time and space dimensions, and comprehensively reflects the stress state of the roof.

[0055] The verification parameters of the fluctuation amplitude curve and the stress gradient value in the stress field instability path evaluation sequence include the fluctuation amplitude growth rate, the fluctuation duration, the gradient value change rate, and the gradient high-value area. The fluctuation amplitude growth rate is the relative change rate of the fluctuation amplitude of adjacent time windows, reflecting the speed of stress fluctuation intensification. The fluctuation duration is the time length during which the fluctuation amplitude is continuously higher than the background value, reflecting the stability of stress fluctuation. The gradient value change rate is the change speed of the stress gradient value over time, reflecting the development trend of the stress concentration area. The gradient high-value area is the area where the stress gradient value exceeds a certain threshold, reflecting the range of stress concentration. These verification parameters are compared with preset safety thresholds, which are determined according to historical data statistics. The fluctuation amplitude growth rate threshold is set to 30% / hour, the fluctuation duration threshold is 45 minutes, the gradient value change rate threshold is 50% / hour, and the gradient high-value area threshold is 15% of the working face area. When the above parameters all exceed their respective thresholds, the stress field instability path is recorded as verified. For the verified instability path, its influence area and development direction are calculated. The influence area is the size of the stress anomaly area, obtained by integrating the gradient high-value area. The development direction is the expansion direction of the stress gradient high-value area, obtained by connecting the gradient high-value centers at consecutive time points. The stress field instability path evolution characteristics include the expansion rate of the influence area and the change trend of the development direction, which are the basis for generating the support reinforcement scheme.

[0056] When determining the support reinforcement range according to the stress field instability path evolution characteristics, the current instability path influence area is taken as the basis, and the development direction is considered to predict the area that may expand in the future. The reinforcement range is usually set as the current influence area plus a safety buffer zone extending along the development direction, and the buffer zone width is 50% of the influence area width. In a specific working face, when the influence area is 200 square meters and the development direction deviates 15 degrees along the working face advancing direction, the support reinforcement range covers an area of about 300 square meters. Based on the stress field instability path evaluation sequence in the support reinforcement range, the stress distribution is calculated, and the stress concentration degree and principal stress direction of each area are analyzed. The stress distribution is obtained by spatial interpolation of the stress gradient values in the evaluation sequence, forming a continuous stress distribution map. According to the stress distribution, the support reinforcement scheme is generated, including support location, support density, and support strength. The support location is preferentially considered in the area with the highest stress gradient, the support density is proportional to the stress gradient, and the support strength is proportional to the fluctuation amplitude. For areas with a stress gradient value exceeding 0.07 MPa / m, the support density is increased to 1.5 times the standard density; for areas with a fluctuation amplitude exceeding 0.4 MPa, the single support strength is increased to 300 kN. The generated support reinforcement scheme is presented in the form of a working face coordinate map, indicating the support parameters of different areas, facilitating on-site implementation.

[0057] The time difference between the prediction window and the preset evacuation time is the key basis for issuing a warning signal. The preset evacuation time is the shortest time required for safe evacuation calculated based on the number of personnel in the working face, the length of the evacuation route, and the evacuation speed, and is usually set to 30 minutes. The time difference is the difference between the end time of the prediction window and the current time, reflecting the remaining time available for response. The safe response time is the shortest time required for support reinforcement and personnel evacuation, and is set to 45 minutes. When the time difference is less than the safe response time, i.e., the remaining time of the prediction window is insufficient to complete the safe response, and the number of consecutive verifications of the stress field instability path reaches the preset number, the system will issue a warning signal. The preset number is usually set to 3 to avoid false positives due to single abnormal data. The warning signal includes three levels: yellow warning indicating the need for enhanced monitoring, orange warning indicating the need for support reinforcement, and red warning indicating the immediate evacuation of personnel. The warning signal is issued simultaneously through the control center display screen, on-site sound and light alarm, and staff mobile devices, ensuring that information is timely conveyed to all relevant personnel.

[0058] In this embodiment, by continuously verifying the stress field instability path within the prediction window, accurate early warning of coal mine roof instability is achieved. Combining dynamic monitoring with continuous verification, the stress field instability state is comprehensively evaluated through fluctuation analysis and regional block calculation, improving the reliability of the warning. The support reinforcement scheme generated based on the evolution characteristics of the stress field instability path is highly targeted and resource-efficient, avoiding the resource waste caused by blind reinforcement in traditional methods. The warning mechanism considers the relationship between the prediction window and the evacuation time, maximizing production time while ensuring safety.

[0059] In an alternative embodiment, the verification parameters of the fluctuation amplitude curve and the stress gradient value in the stress field instability path evaluation sequence are calculated, the verification parameters are compared with the preset safety threshold, the stress field instability paths that pass the verification are recorded, and the corresponding influence area and development direction are calculated to generate the evolution characteristics of the stress field instability path, including: The fluctuation amplitude curve and the stress gradient value are extracted from the stress field instability path evaluation sequence, the amplitude change rate is obtained by time difference of the fluctuation amplitude curve, and the gradient change intensity is obtained by spatial gradient calculation of the stress gradient value; The amplitude change rate and the gradient change intensity are analyzed in time and space, the cross-response coefficient of the fluctuation amplitude curve and the stress gradient value is extracted, and the verification parameters are calculated based on the cross-response coefficient; The verification parameters are continuously tracked in consecutive time periods by sliding window, and when the verification parameters continuously meet the preset safety threshold in consecutive time periods, it is determined that the verification is passed, and the stress field instability paths that pass the verification are recorded; The spatial distribution nodes of the stress field instability path passing the verification are extracted, the coverage boundary of the spatial distribution nodes is calculated to obtain an influence area, the development direction of the spatial distribution nodes is fitted to obtain a development direction, and the influence area and the development direction are combined with time variation to generate stress field instability path evolution characteristics.

[0060] First, the fluctuation amplitude curve and the stress gradient value are extracted from the stress field instability path evaluation sequence. The fluctuation amplitude curve reflects the fluctuation intensity change of the stress field with time, and its data is derived from the stress sensor network arranged on the coal mine roof. The stress sensors are arranged in a 5m x 5m grid, with a sampling frequency of 5 times per second, and the fluctuation amplitude curve is generated by sliding window processing. The window length is set to 10 minutes, and the step length is 1 minute. The standard deviation of the data in each window is calculated as the fluctuation amplitude. The stress gradient value reflects the degree of uneven distribution of stress in space, and is obtained by spatial difference calculation of stress values at different positions in the working face area. In the actual monitoring of the coal mine working face, the fluctuation amplitude curve shows a gradually increasing trend with time, from the initial 0.12MPa to 0.48MPa before the warning; the stress gradient value increases significantly near the potential roof fracture area, with a maximum of 0.085MPa / m. The time series difference of the fluctuation amplitude curve is obtained by calculating the difference of the fluctuation amplitude at adjacent time points, which reflects the change rate of the fluctuation intensity. The time series difference uses the forward difference method to calculate the increment of the fluctuation amplitude per minute, obtaining the amplitude change rate time series. The spatial gradient calculation of the stress gradient value is to analyze the second-order change of the stress gradient in space, to calculate the spatial change rate of adjacent gradient values, and to obtain the gradient change intensity distribution. The spatial gradient calculation uses the central difference method to approximate the second-order differential of the stress gradient field in the working face area, and the result reflects the spatial variation intensity of the stress gradient.

[0061] The time-space coupling analysis of amplitude change rate and gradient change intensity is a key step to identify the precursors of roof instability. The time-space coupling analysis uses cross-correlation method to calculate the correlation between amplitude change rate at different time points and gradient change intensity at corresponding spatial locations. The cross-correlation calculation is based on sliding window, with window length of 30 minutes and step length of 5 minutes. For each window, the cross-correlation coefficient is calculated between the amplitude change rate sequence and the gradient change intensity sequence at corresponding locations, which reflects the coupling degree between the two. The cross-correlation coefficient ranges from -1 to 1, with positive value indicating positive correlation between the two, and higher value indicating higher coupling degree. When the roof is about to lose stability, the cross-correlation coefficient usually exceeds 0.75, indicating that stress fluctuation and spatial gradient change are highly synchronized. The verification parameters based on cross-correlation coefficient calculation include three indicators: cross-correlation duration, cross-correlation peak value, and cross-correlation growth rate. Cross-correlation duration is the time length during which the cross-correlation coefficient exceeds the threshold value of 0.6; cross-correlation peak value is the maximum value of cross-correlation coefficient within the observation period; cross-correlation growth rate is the time variation rate of cross-correlation coefficient. These three indicators together constitute the verification parameters, which are used to judge the reliability of stress field instability path.

[0062] The continuous period tracking of verification parameters through sliding window is a specific implementation method of continuous verification. The sliding window length is set to 60 minutes, with step length of 10 minutes. The verification parameters within each window are evaluated. The preset safety thresholds include three aspects: cross-correlation duration threshold is set to 30 minutes, cross-correlation peak value threshold is set to 0.75, and cross-correlation growth rate threshold is set to 0.01 / minute. When the verification parameters within a window meet the above three threshold conditions simultaneously, it is recorded as a successful verification. The continuous period tracking requires that the verification is successful in three consecutive windows, and then the stress field instability path is verified to be successful. This multi-window continuous verification mechanism effectively reduces the false positive probability caused by short-term fluctuations. In the monitoring of a certain coal mine working face, this mechanism successfully filters the short-term stress fluctuations caused by blasting operations, avoiding false positives. The stress field instability paths that pass the verification are recorded completely, including the time period, spatial distribution, and parameter characteristics of the verification. The recorded content is stored in the monitoring database, providing data support for subsequent analysis and model optimization.

[0063] The spatial distribution nodes of the stress field instability path passing the verification are the basis for calculating the influence area and development direction. The spatial distribution nodes refer to the grid points in the stress field whose cross-response coefficients exceed the threshold value 0.7, and these points constitute the spatial skeleton of the stress field instability path. The extraction of the spatial distribution nodes uses the threshold segmentation method to screen the discrete point set meeting the condition from the stress field. The convex hull algorithm is used to calculate the coverage boundary of the spatial distribution nodes, and the discrete point set is converted into a continuous polygon area, and the area of the polygon is the influence area of the stress field instability path. When the number of spatial distribution nodes is 32, the calculated influence area is about 240 square meters, covering about 15% of the area of the working face. The direction fitting of the spatial distribution nodes is a method for determining the development direction of the instability path, and the principal component analysis technology is used to extract the main distribution direction of the point set. The principal component analysis extracts the eigenvector corresponding to the maximum eigenvalue of the covariance matrix of the spatial distribution nodes as the development direction. The development direction is expressed by the angle with the trend of the working face, for example, the development direction of the instability path of a working face is 38 degrees with the trend of the working face. The influence area and the development direction are combined with the time variation to generate the evolution characteristics of the stress field instability path, including the expansion rate of the influence area, the stability of the development direction and the change trend of the spatial distribution node density. The influence area expansion rate reflects the speed of the instability path expansion, the development direction stability reflects the consistency of the instability path direction, and the spatial distribution node density change trend reflects the aggravation of the instability degree.

[0064] In the embodiment, the precise verification of the coal mine roof instability risk is realized by calculating the verification parameters of the fluctuation amplitude curve and the stress gradient value in the stress field instability path evaluation sequence, and the weak signal before the roof instability is captured through the cross-response coefficient. The design of the multi-index verification parameter and the multi-window continuous verification mechanism significantly improve the reliability of the early warning and effectively reduce the false positive rate. The extraction of the spatial distribution nodes and the application of the convex hull algorithm enable the influence range of the stress field instability path to be accurately defined. The development direction extraction method based on the principal component analysis accurately predicts the expansion trend of the instability path, and provides accurate guidance for support reinforcement.

[0065] In a second aspect, the embodiment of the present application provides a coal mine roof dynamic monitoring and early warning system based on artificial intelligence, which comprises: A first unit is configured to collect coal mine roof stress waveforms and displacement waveforms of a plurality of monitoring points, extract stress relaxation features according to the phase decoupling relationship between the stress waveforms and the displacement waveforms, and construct a three-dimensional stress conduction intensity field in combination with the spatial projection of the monitoring points and the fault. The second unit is configured to map the stress relaxation feature in a three-dimensional stress conduction intensity field in space-time to obtain a dynamic feature sequence, identify a stress mutation transition point from the dynamic feature sequence by using a recurrent neural network, and form a stress conduction chain according to a space-time correlation coefficient by density clustering. The third unit is configured to analyze a topological migration rule of a region in the stress conduction chain whose correlation coefficient exceeds a preset correlation threshold, and extract a motion trajectory of the region in a three-dimensional space to obtain a stress field instability path. The fourth unit is configured to perform geometric feature analysis on the stress field instability path and perform shape matching with instability paths in a historical working face database, determine a prediction window according to a working face with an optimal matching degree. The fifth unit is configured to continuously verify the stress field instability path in the prediction window, generate a support reinforcement scheme according to the stress field instability path and issue a warning signal when the stress field instability path passes the verification for multiple times and the prediction window is less than a preset evacuation time.

[0066] In a third aspect, an electronic device is provided, including: a processor; a memory for storing processor-executable instructions; The processor is configured to invoke the instructions stored in the memory to execute the method described above.

[0067] In a fourth aspect, a computer-readable storage medium is provided, which stores computer program instructions, and the computer program instructions are executed by a processor to implement the method described above.

[0068] The present application can be a method, device, system and / or computer program product. The computer program product can include a computer readable storage medium having computer readable program instructions stored therein, which are used to perform various aspects of the present application.

[0069] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement to part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A coal mine roof dynamic monitoring and early warning method based on artificial intelligence, characterized in that, The method comprises the following steps: Collecting coal mine roof stress waveforms and displacement waveforms at multiple monitoring points, extracting stress relaxation characteristics according to the phase decoupling relationship between the stress waveforms and the displacement waveforms, and constructing a three-dimensional stress conduction intensity field in combination with the spatial projection of the monitoring points and faults; Mapping the stress relaxation characteristics in the three-dimensional stress conduction intensity field to obtain a dynamic characteristic sequence, identifying stress mutation transition points from the dynamic characteristic sequence by using a recurrent neural network, and forming stress conduction chains according to the spatiotemporal correlation coefficients of the stress mutation transition points by density clustering; Analyzing the topological migration law of the region whose correlation coefficient exceeds a preset correlation threshold in the stress conduction chain, extracting the motion trajectory of the region in the three-dimensional space to obtain a stress field instability path; Performing geometric feature analysis on the stress field instability path and morphological matching with the instability paths in a historical working face database, determining a prediction window according to the instability process of the working face with the optimal matching degree, and continuously verifying the stress field instability path in the prediction window; When the stress field instability path is continuously verified for multiple times and the prediction window is less than a preset evacuation time, generating a support reinforcement scheme according to the stress field instability path and issuing a warning signal.

2. The method of claim 1, wherein, Extracting stress relaxation characteristics according to the phase decoupling relationship between the stress waveforms and the displacement waveforms, and constructing a three-dimensional stress conduction intensity field in combination with the spatial projection of the monitoring points and faults comprises: Obtaining waveform frequency spectra by Fourier transform of the stress waveforms and the displacement waveforms; Extracting waveform phase features from the waveform frequency spectra, obtaining a feature sequence by multi-scale wavelet decomposition of the waveform phase features, separating waveform periodic components and waveform trend components from the feature sequence, and obtaining stress relaxation characteristics by phase decoupling of the waveform periodic components and the waveform trend components; Calculating the vertical distance and the projection angle of the monitoring points to the fault projection surface, constructing a stress conduction coefficient, and constructing a three-dimensional stress conduction intensity field by nonlinear mapping of the stress conduction coefficient and the stress relaxation characteristics.

3. The method of claim 1, wherein, Mapping the stress relaxation characteristics in the three-dimensional stress conduction intensity field to obtain a dynamic characteristic sequence, identifying stress mutation transition points from the dynamic characteristic sequence by using a recurrent neural network, and forming stress conduction chains according to the spatiotemporal correlation coefficients of the stress mutation transition points by density clustering comprises: Mapping the stress relaxation characteristics in the three-dimensional stress conduction intensity field in the time domain and the space domain, extracting wave components by time domain transformation, obtaining distribution components by space domain transformation, combining the wave components and the distribution components to construct a spatiotemporal feature matrix; Extracting a dynamic characteristic sequence from the spatiotemporal feature matrix, and obtaining feature segments by time window segmentation of the dynamic characteristic sequence; Constructing a recurrent neural network, setting a memory unit and a forgetting gate in the recurrent neural network, inputting the feature segments into the recurrent neural network, extracting feature change rules through the memory unit and the forgetting gate, identifying stress mutation transition points and transition intensities from the output end; Calculating the time correlation between the stress mutation transition points to obtain a time correlation, calculating the spatial correlation between the stress mutation transition points to obtain a spatial correlation, and fusing the time correlation and the spatial correlation to construct a spatiotemporal correlation coefficient. Density clustering is performed on the stress mutation transition points based on the spatiotemporal correlation coefficient, the density clustering results are sorted according to the transition intensity, and the sorted results are constructed into a stress conduction chain according to the conduction direction.

4. The method of claim 1, wherein, The topological migration law of a region in the stress conduction chain where the correlation coefficient exceeds a preset correlation threshold is analyzed, and a stress field instability path is obtained by extracting the motion trajectory of the region in a three-dimensional space, including: A time sequence similarity of adjacent conduction points in the stress conduction chain is calculated to obtain a time sequence correlation, a spatial distance and an included angle between the conduction points are calculated to obtain a spatial correlation, the time sequence correlation and the spatial correlation are multi-dimensionally weighted and fused to obtain a correlation coefficient, and a conduction point correlation matrix is constructed based on the correlation coefficient; A topological graph structure is constructed based on the conduction point correlation matrix, the conduction points where the correlation coefficient exceeds the preset correlation threshold are taken as topological nodes, a conduction subgraph is constructed by extracting the connection relationship between the topological nodes, and the topological migration law is extracted by structure evolution and connection change of the conduction subgraph; The topological migration law is mapped to a three-dimensional space coordinate system, a motion trajectory vector sequence is constructed according to the distribution position and motion speed of the topological nodes in the space, and a stress field instability path is extracted based on the direction change and amplitude change of the motion trajectory vector sequence.

5. The method of claim 1, wherein, Geometric feature analysis is performed on the stress field instability path, and a form matching is performed with a failure path in a historical working face database, a prediction window is determined according to a failure process of a working face with an optimal matching degree, including: Geometric feature analysis is performed on the stress field instability path, a curvature change sequence of a roof fracture line, a crack propagation direction sequence, and a roof subsidence length sequence are calculated, and a roof failure form curve is constructed; Periodic subsidence features and stress concentration points are extracted from the roof failure form curve, a roof failure form feature vector is constructed and standardized, and a roof failure path sample is extracted from a historical working face database; The form feature vector of the roof failure path sample is matched with the standardized roof failure form feature vector, a matching degree score matrix is constructed, and a comprehensive matching degree is calculated based on the matching degree score matrix; The roof failure path sample is screened and sorted according to the comprehensive matching degree, and a working face with an optimal comprehensive matching degree is selected as a reference working face; A roof caving time point and a roof damage range point are extracted from the reference working face, a time span of the prediction window is determined according to the roof caving time point, a spatial boundary of the prediction window is determined according to the roof damage range point, and a final prediction window is output.

6. The method of claim 1, wherein, The stress field instability path is continuously verified within the prediction window, a support reinforcement scheme is generated according to the stress field instability path when the stress field instability path is verified continuously for multiple times and the prediction window is less than a preset evacuation time, and a warning signal is issued, including: Stress field data is collected within the prediction window, time sequence and spatial distribution information are extracted from the stress field data, a fluctuation amplitude curve of the time sequence is calculated by fluctuation analysis, stress gradient values of the spatial distribution are calculated by region blocking, and a stress field instability path evaluation sequence is generated by combining the fluctuation amplitude curve and the stress gradient values; Geometric feature analysis is performed on the stress field instability path, and a form matching is performed with a failure path in a historical working face database, a prediction window is determined according to a failure process of a working face with an optimal matching degree, including: Geometric feature analysis is performed on the stress field instability path, a curvature change sequence of a roof fracture line, a crack propagation direction sequence, and a roof subsidence length sequence are calculated, and a roof failure form curve is constructed; Periodic subsidence features and stress concentration points are extracted from the roof failure form curve, a roof failure form feature vector is constructed and standardized, and a roof failure path sample is extracted from a historical working face database; The form feature vector of the roof failure path sample is matched with the standardized roof failure form feature vector, a matching degree score matrix is constructed, and a comprehensive matching degree is calculated based on the matching degree score matrix; The roof failure path sample is screened and sorted according to the comprehensive matching degree, and a working face with an optimal comprehensive matching degree is selected as a reference working face; A roof caving time point and a roof damage range point are extracted from the reference working face, a time span of the prediction window is determined according to the roof caving time point, a spatial boundary of the prediction window is determined according to the roof damage range point, and a final prediction window is output. The stress field instability path is continuously verified within the prediction window, a support reinforcement scheme is generated according to the stress field instability path when the stress field instability path is verified continuously for multiple times and the prediction window is less than a preset evacuation time, and a warning signal is issued, including: Stress field data is collected within the prediction window, time sequence and spatial distribution information are extracted from the stress field data, a fluctuation amplitude curve of the time sequence is calculated by fluctuation analysis, stress gradient values of the spatial distribution are calculated by region blocking, and a stress field instability path evaluation sequence is generated by combining the fluctuation amplitude curve and the stress gradient values; The verification parameter of the fluctuation amplitude curve and the stress gradient value in the stress field instability path evaluation sequence is calculated, the verification parameter is compared and verified with a preset safety threshold, a stress field instability path that passes the verification is recorded, and a corresponding influence area and development direction are calculated to generate stress field instability path evolution characteristics; A support reinforcement range is determined according to the stress field instability path evolution characteristics, and a stress distribution is calculated based on the stress field instability path evaluation sequence in the support reinforcement range to generate a support reinforcement scheme; A time difference between a prediction window and a preset evacuation time is calculated, and when the time difference is less than a safety response time and a continuous verification pass number of the stress field instability path reaches a preset number, a warning signal is generated according to the stress field instability path evolution characteristics and the support reinforcement scheme.

7. The method of claim 6, wherein, The verification parameter of the fluctuation amplitude curve and the stress gradient value in the stress field instability path evaluation sequence is calculated, the verification parameter is compared and verified with a preset safety threshold, a stress field instability path that passes the verification is recorded, and a corresponding influence area and development direction are calculated to generate stress field instability path evolution characteristics include: The fluctuation amplitude curve and the stress gradient value are extracted from the stress field instability path evaluation sequence, the amplitude change rate is obtained by time difference of the fluctuation amplitude curve, and the gradient change strength is obtained by spatial gradient calculation of the stress gradient value; The amplitude change rate and the gradient change strength are analyzed by time-space coupling, the cross-response coefficient of the fluctuation amplitude curve and the stress gradient value is extracted, and the verification parameter is calculated based on the cross-response coefficient; The verification parameter is tracked in a continuous time period by a sliding window, and when the verification parameter continuously meets the preset safety threshold in the continuous time period, it is determined that the verification passes, and the stress field instability path that passes the verification is recorded; The spatial distribution node of the stress field instability path that passes the verification is extracted, the influence area is obtained by calculating the coverage boundary of the spatial distribution node, the development direction is obtained by direction fitting of the spatial distribution node, and the stress field instability path evolution characteristics are generated by combining the influence area and the development direction with time changes.

8. The coal mine roof dynamic monitoring and early warning system based on artificial intelligence, for realizing the method of any one of the preceding claims 1-7, characterized in that, It includes: The first unit is used for collecting coal mine roof stress waveforms and displacement waveforms of a plurality of monitoring points, extracting stress relaxation characteristics according to the phase decoupling relationship of the stress waveforms and the displacement waveforms, and constructing a three-dimensional stress conduction intensity field by combining the monitoring points and the spatial projection of the fault; The second unit is used for time-space mapping of the stress relaxation characteristics in the three-dimensional stress conduction intensity field to obtain a dynamic characteristic sequence, identifying stress mutation transition points from the dynamic characteristic sequence by using a recurrent neural network, and forming stress conduction chains by density clustering according to the time-space correlation coefficients of the stress mutation transition points; The third unit is used for analyzing the topological migration law of a region in the stress conduction chain whose correlation coefficient exceeds a preset correlation threshold, and extracting the motion trajectory of the region in the three-dimensional space to obtain a stress field instability path; The fourth unit is used for geometric feature analysis of the stress field instability path and shape matching with instability paths in a historical working face database, and determining a prediction window according to the instability process of a working face with the optimal matching degree. A fifth unit for continuously verifying the stress field instability path within a prediction window, when the stress field instability path is continuously verified for multiple times and the prediction window is less than a preset evacuation time, generating a support reinforcement scheme according to the stress field instability path and issuing a warning signal.

9. An electronic device, comprising: Comprise: a processor; a memory for storing processor-executable instructions; wherein the processor is configured to invoke the instructions stored in the memory to perform the method of any one of claims 1 to 7.

10. A computer-readable storage medium having stored thereon computer program instructions, wherein, The computer program instructions, when executed by the processor, implement the method of any one of claims 1 to 7. The computer program instructions, when executed by the processor, implement the method of any one of claims 1 to 7.

Images