A coal mine underground early warning system and method

Abstract

This invention relates to the field of underground coal mine safety technology, and discloses an underground coal mine early warning system and method. The method involves collecting multi-dimensional monitoring data of the underground coal mine environment; dynamically dividing the multi-dimensional monitoring data into windows to generate a set of data segments; performing anomaly pattern recognition based on the data segment set to extract anomaly feature points exceeding dynamic thresholds in each data segment; inputting the anomaly feature points into a pre-constructed spatiotemporal correlation model to analyze their diffusion paths in the time and spatial dimensions; generating an anomaly propagation map containing the propagation direction and intensity attenuation gradient of the anomaly feature points based on the diffusion paths; performing hierarchical clustering on the anomaly propagation map to divide anomaly clusters with similar propagation characteristics; generating early warning signals containing priority labels using the spatiotemporal distribution characteristics of the anomaly clusters; synchronously verifying the early warning signals with underground equipment status data to filter out target early warning signals that require response; and triggering corresponding emergency control commands based on the target early warning signals.

Description

Technical Field

[0001] This invention relates to the field of underground safety technology in coal mines, specifically to an underground early warning system and method for coal mines. Background Technology

[0002] The underground working environment in coal mines is complex and harsh. Affected by geological conditions, equipment operating status and other factors, it is easy for gas concentrations to exceed standards, temperature and humidity to be abnormal, vibration intensity to be too high and equipment current to be abnormal. If these situations are not detected and dealt with in time, they may cause serious safety accidents, threatening the lives of workers and the production order of coal mines.

[0003] Traditional monitoring methods are commonly used in underground coal mines, relying on fixed sensors to collect environmental data. However, these methods have significant limitations. Firstly, the data collected by traditional methods is often limited in scope, focusing only on one or a few environmental parameters, making it difficult to comprehensively reflect the overall underground environment and leading to the omission of potential safety hazards. Secondly, the data processing is relatively simple, often involving static threshold comparisons. This means setting fixed thresholds and issuing warnings when the monitored data exceeds them, failing to adjust the thresholds according to dynamic changes in the underground environment, which easily leads to false alarms or missed alarms.

[0004] Traditional monitoring methods lack analysis of the temporal continuity and spatial correlation of data during data processing, making it difficult to identify the propagation patterns and diffusion paths of abnormal data. When anomalies occur, it is impossible to accurately determine the direction of propagation, the scope of impact, and changes in intensity, making it difficult for staff to formulate targeted response measures and delaying the best handling opportunity. Furthermore, traditional methods do not correlate and verify abnormal early warning signals with underground equipment status data, potentially leading to mismatches between early warning signals and actual equipment operating status. This results in some early warning signals that do not require immediate response consuming significant emergency resources, while truly necessary early warning signals fail to receive timely processing, further reducing the effectiveness and reliability of underground safety early warning systems in coal mines and failing to meet the high requirements for safety early warning in the complex environment of underground coal mines. Summary of the Invention

[0005] The purpose of this invention is to provide an underground early warning system and method for coal mines to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides a method for early warning in coal mines, the method comprising:

[0007] Collect multi-dimensional monitoring data of the underground environment in coal mines, including gas concentration, temperature and humidity, vibration intensity and equipment current parameters;

[0008] Dynamically divide multidimensional monitoring data into windows to generate a set of data segments with temporal continuity;

[0009] Anomaly pattern recognition is performed based on a set of data segments, and abnormal feature points exceeding dynamic thresholds are extracted from each data segment.

[0010] Anomaly feature points are input into a pre-built spatiotemporal correlation model, and the diffusion path of anomaly feature points in the time and space dimensions is analyzed through the spatiotemporal correlation model.

[0011] An anomaly propagation map is generated based on the diffusion path, and the anomaly propagation map includes the propagation direction and intensity attenuation gradient of the anomaly feature points;

[0012] Hierarchical clustering of the anomaly propagation map was performed to divide anomaly clusters with similar propagation characteristics;

[0013] An early warning signal is generated by utilizing the spatiotemporal distribution characteristics of anomaly clusters, and the early warning signal includes a priority label for the anomaly clusters;

[0014] The warning signals are synchronously verified with the downhole equipment status data to filter out the target warning signals that need to be responded to.

[0015] The corresponding emergency control command is triggered based on the target warning signal.

[0016] Preferably, the abnormal pattern recognition based on the data fragment set includes:

[0017] Perform multi-scale sliding window analysis on a set of data segments to calculate the mean shift and standard deviation rate of change of the data within the window;

[0018] When the mean offset exceeds the historical baseline mean and the rate of change of standard deviation is greater than the dynamic fluctuation threshold, the current window is marked as an abnormal window.

[0019] Extract the instantaneous gradient of all data points within the anomaly window, and take the data point with the largest absolute gradient value as the anomaly feature point.

[0020] Preferably, the step of analyzing the diffusion path of anomalous feature points in the time and space dimensions through a spatiotemporal correlation model includes:

[0021] Establish a spatiotemporal neighborhood centered on the anomalous feature points, wherein the spatiotemporal neighborhood includes a time span and a spatial radius;

[0022] Calculate the energy transfer matrix of the anomalous feature points in the spatiotemporal neighborhood, where the energy transfer matrix reflects the coupling strength between the anomalous feature points;

[0023] The main propagation direction of the diffusion path is determined by the attenuation slope of the energy transfer matrix.

[0024] Preferably, generating the abnormal propagation map based on the diffusion path includes:

[0025] Map the main propagation direction of the diffusion path to the downhole three-dimensional spatial coordinate system to generate directed edges with directional arrows;

[0026] Set the length of the directed edge to the reciprocal of the attenuation slope of the corresponding energy transfer matrix;

[0027] The starting and ending points of the directed edges are marked with the spatial coordinates and timestamps of the anomalous feature points, respectively.

[0028] Preferably, the hierarchical clustering of the abnormal propagation map includes:

[0029] Calculate the directional similarity and length similarity between directed edges, and merge directed edges whose similarity exceeds the clustering threshold;

[0030] Anomaly clusters are divided based on the merged directed edge connections, and each anomaly cluster contains at least two directed edges.

[0031] Preferably, the step of generating an early warning signal using the spatiotemporal distribution characteristics of anomaly clusters includes:

[0032] The total length and average direction of directed edges within statistically abnormal clusters are consistent.

[0033] Compare the total length with a preset length threshold, and compare the average directional consistency with a preset directional threshold;

[0034] When the total length exceeds the length threshold and the average directional consistency is lower than the directional threshold, a high-priority label is generated.

[0035] Preferably, the step of synchronizing and verifying the early warning signal with the downhole equipment status data includes:

[0036] Obtain the operating parameters of the downhole equipment in the current time window, including current fluctuation rate and load change.

[0037] If the operating parameters of the equipment in the area corresponding to the target warning signal exceed the safety threshold, the target warning signal will be retained.

[0038] Otherwise, the target warning signal will be downgraded to an observation signal.

[0039] Preferably, the step of triggering the corresponding emergency control command based on the target warning signal includes:

[0040] Match the preset response strategy according to the priority label of the target warning signal;

[0041] If the priority label is high, then the area power-off command and ventilation acceleration command will be triggered.

[0042] If the priority tag is low, a local device production restriction command will be triggered.

[0043] Preferably, the method further includes:

[0044] The dynamic threshold and dynamic fluctuation threshold are updated in real time, and the update is based on the occurrence frequency of historical abnormal feature points and the average decay value of the energy transfer matrix.

[0045] Preferably, the present invention also includes a coal mine underground early warning system, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor, when executing the computer program, implements the steps of the above-described coal mine underground early warning method.

[0046] Compared with the prior art, the beneficial effects of the present invention are:

[0047] By dynamically dividing multidimensional monitoring data from underground coal mines into windows, a set of data segments with temporal continuity is generated, providing a structured data foundation for subsequent analysis. Based on this data segment set, anomaly pattern recognition is performed, and a multivariate statistical process control algorithm is used to extract anomaly feature points exceeding dynamic thresholds, accurately capturing abnormal states in the underground environment.

[0048] Anomaly feature points are input into a pre-constructed spatiotemporal correlation model to analyze their diffusion paths in both time and space dimensions. The spatiotemporal correlation model employs a graph neural network structure, combined with tunnel topology and ventilation parameters, to accurately deduce the anomaly propagation patterns. Based on the diffusion paths, an anomaly propagation map containing propagation direction and intensity attenuation gradients is generated, visually demonstrating the dynamic propagation characteristics of the anomaly.

[0049] Hierarchical clustering of the anomaly propagation map is performed, and density clustering algorithm is used to divide anomaly clusters with similar propagation characteristics. Anomaly clusters represent different types of risk events, supporting classified handling. Early warning signals are generated using the spatiotemporal distribution characteristics of anomaly clusters, and priority labels are dynamically assessed based on the impact range and spread speed.

[0050] The system synchronizes and verifies early warning signals with downhole equipment status data in real time to confirm the feasibility of the early warning signals and eliminate false alarms caused by equipment malfunctions. An intelligent verification mechanism filters out target early warning signals that require a response, ensuring the accuracy and operability of the warnings. Based on the target early warning signal, corresponding emergency control commands are triggered to achieve rapid and effective risk management.

[0051] This method achieves accurate early warning and rapid response to safety risks in coal mines through the synergistic effects of dynamic window analysis, anomaly feature identification, spatiotemporal propagation modeling, intelligent clustering for early warning, and multi-source data verification. Dynamic window segmentation ensures data continuity, anomaly identification improves detection sensitivity, spatiotemporal modeling reveals propagation patterns, intelligent clustering enables risk classification, and multi-source verification enhances the reliability of early warnings. The system significantly improves the accuracy and timeliness of underground coal mine safety monitoring. Attached Figure Description

[0052] Figure 1 This is a schematic diagram illustrating the working principle of the underground early warning method for coal mines as described in this invention.

[0053] Figure 2 This is a flowchart for anomaly pattern recognition based on a set of data fragments;

[0054] Figure 3 This is a flowchart for analyzing the spatiotemporal diffusion path of anomalous feature points. Detailed Implementation

[0055] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0056] Please see Figure 1This invention provides an underground early warning system and method for coal mines. The method includes: collecting multi-dimensional monitoring data of the underground coal mine environment; dynamically dividing the data into windows to generate a set of data segments with temporal continuity; performing anomaly pattern recognition based on the data segment set to extract anomaly feature points; inputting the anomaly feature points into a pre-constructed spatiotemporal correlation model to analyze their diffusion paths in the time and spatial dimensions; generating anomaly propagation maps based on the diffusion paths; performing hierarchical clustering on the anomaly propagation maps to divide anomaly clusters; generating early warning signals using the spatiotemporal distribution characteristics of the anomaly clusters; synchronously verifying the early warning signals with underground equipment status data to filter target early warning signals; and triggering corresponding emergency control commands based on the target early warning signals. The multi-dimensional monitoring data includes gas concentration, temperature and humidity, vibration intensity, and equipment current parameters. These data are collected in real time through a sensor network deployed in various underground areas. Sensor nodes record environmental parameters at a fixed sampling frequency and transmit them to a central processing unit. The dynamic window partitioning employs an adaptive time window mechanism, dynamically adjusting the window size based on data fluctuation characteristics. For example, the window shrinks to capture details when the data change rate increases and expands to smooth noise when the change rate decreases, thereby generating a set of data segments with temporal continuity. The anomaly pattern recognition process analyzes the data segment set in real time, identifying anomalies exceeding a dynamic threshold. This dynamic threshold is calculated based on historical data statistical characteristics and is adaptively updated with environmental changes. The spatiotemporal correlation model is constructed based on graph theory, mapping anomaly feature points to nodes in the graph and representing the spatiotemporal relationships between points through edges. The model considers the downhole spatial topology and time series correlation when analyzing the propagation path of anomaly feature points. The anomaly propagation map visualizes the propagation direction and intensity attenuation of anomaly points. A hierarchical clustering algorithm groups anomaly points based on the similarity of directed edges in the graph, forming anomaly clusters with common propagation characteristics. The early warning signal generation module assigns priority labels based on the spatiotemporal distribution characteristics of the anomaly clusters; higher priority labels indicate anomalies with wide propagation ranges or low directional consistency.

[0057] Example 1: See Figure 2The anomaly pattern recognition process is based on a set of data segments, generated by dynamic window partitioning, exhibiting continuous time-series characteristics. Anomaly pattern recognition first employs multi-scale sliding window analysis. This multi-scale sliding window is implemented by configuring windows with different time spans; for example, short-term windows detect instantaneous fluctuations, while long-term windows identify trend changes. The window size is adaptively adjusted based on the data sampling frequency and the typical duration of the anomaly event. The system advances the window in a sliding manner, moving a fixed step each time to ensure the continuity of data coverage. The mean offset of the data within the window is calculated using a moving average algorithm. The mean offset is defined as the difference between the current window's data mean and the historical baseline mean. The historical baseline mean is obtained by statistically analyzing the moving average of data over a past period and dynamically updated over time to reflect environmental changes. The rate of change of standard deviation is calculated by comparing the standard deviation of the data within the window with the historical standard deviation, which is derived from long-term data statistics.

[0058] When the mean offset exceeds the historical baseline mean and the rate of change of standard deviation is greater than the dynamic fluctuation threshold, the system marks the current window as an anomaly window. The dynamic fluctuation threshold is set based on the historical fluctuation characteristics of the downhole environment and incorporates an adaptive mechanism to adjust the threshold size according to the recent frequency of anomalies to avoid false alarms. After marking the anomaly window, the system extracts the instantaneous gradient of all data points within the window. The instantaneous gradient is obtained by calculating the first difference or numerical derivative of the data points, reflecting the instantaneous rate of data change. The data point with the largest absolute gradient value is selected as an anomaly feature point. This selection is based on the gradient value representing the severity of change; high gradient values ​​usually correspond to sudden anomaly events. The extraction process also includes filtering low-significance points, retaining only points with gradient values ​​exceeding a preset threshold to ensure the representativeness of the anomaly feature points.

[0059] In the specific implementation of multi-scale sliding window analysis, the window scale is set considering the data type and application scenario. For example, for high-frequency sampled vibration data, a second-level window is used to capture transient anomalies, while for low-frequency gas concentration data, a minute-level window is used to analyze cumulative effects. The window sliding step size is balanced between computational efficiency and real-time requirements, and is usually set as a proportion of the window size. The calculation of the mean offset introduces a weighted average method, assigning higher weights to recent data points to enhance sensitivity to the latest changes. The historical baseline mean is updated using exponential smoothing, with the smoothing factor adjusted according to data stability. The dynamic fluctuation threshold of the standard deviation rate of change is optimized by monitoring the coefficient of variation of historical data; when the coefficient of variation is large, the threshold is appropriately increased to reduce noise interference. The anomaly window marking logic integrates a real-time verification mechanism; when multiple consecutive windows are marked, the system merges adjacent windows to avoid fragmented processing. The calculation of the instantaneous gradient uses the central difference method to reduce boundary errors, and the gradient value is also normalized for easy comparison across data sources. The selection of anomaly feature points is based not only on the absolute value of the gradient but also on the spatial distribution and temporal persistence of the points; for example, points that appear in multiple scale windows are preferentially selected.

[0060] The anomaly pattern recognition module's overall workflow is based on streaming data processing, supporting high-throughput real-time analysis. After a set of data segments is input, the system processes multiple window scales in parallel to improve computational efficiency. Incremental algorithms are used to calculate mean offset and standard deviation rate of change, avoiding recalculation of the entire dataset and adapting to the needs of continuous downhole monitoring. The maintenance of historical baseline mean and dynamic fluctuation thresholds involves storing recent data in a circular buffer and periodically discarding old values ​​to ensure the timeliness of statistics. After extracting anomaly feature points, timestamps and sensor location information are appended to provide context for subsequent spatiotemporal analysis. This implementation enhances the robustness of anomaly detection through multi-scale analysis and a dynamic threshold mechanism, effectively distinguishing between genuine anomalies and environmental noise.

[0061] During window analysis, the system monitors the overall quality of the data fragment set, such as detecting missing or outliers, and uses interpolation or elimination methods to ensure the integrity of the analysis input. The synchronous advancement of multi-scale sliding windows ensures consistency in the time dimension and avoids temporal discrepancies. The calculation of mean offset also considers the data distribution pattern; for non-normally distributed data, median offset is used as a supplementary indicator to improve adaptability. The evaluation of the standard deviation change rate is combined with rolling window statistics, and the adjustment of dynamic fluctuation thresholds is based on machine learning models. The model training data comes from historical anomaly patterns, achieving intelligent thresholding. The labeling of anomaly windows not only relies on a single condition but also introduces composite indicators, such as a consistency score combining mean offset and variance change, to reduce misjudgments. The calculation of instantaneous gradients optimizes numerical stability; smoothing preprocessing is introduced for abruptly changing data, and the comparison of absolute gradient values ​​uses a local ranking method, selecting the top percentage of points within the window as candidate feature points. The extraction process of anomaly feature points integrates verification steps, such as checking the temporal persistence of feature points through backtracking analysis and eliminating instantaneous noise points. Feature points are also associated with sensor type and data source, supporting multi-dimensional data fusion analysis. The entire identification process is designed with computational lightweight considerations, employing a distributed processing framework to adapt to the limited resources of the downhole environment. The module's output interfaces are standardized for easy integration with subsequent spatiotemporal models. Through these detailed improvements, the anomaly pattern recognition module achieves efficient capture of downhole anomalies, providing reliable input for the early warning system.

[0062] The implementation of multi-scale sliding window analysis also includes window overlap control. The overlap degree is configured according to application requirements. High overlap enhances detection sensitivity but increases computational load. The system dynamically adjusts the overlap degree to balance performance. The historical baseline mean update mechanism for mean offset introduces a decay factor, gradually weakening the influence of old data and maintaining baseline currentity. The calculation of the standard deviation rate of change uses an unbiased estimation method to avoid small sample bias. The dynamic fluctuation threshold setting references industry safety standards and allows for manual calibration, integrating expert experience. The processing after anomaly window marking includes window merging and splitting logic. For long-duration anomalies, adjacent windows are automatically merged to form anomaly regions, improving analysis efficiency. The calculation of instantaneous gradient changes supports multiple difference step sizes to adapt to data with different sampling rates. Gradient value post-processing includes filtering and smoothing to reduce the impact of random fluctuations. The algorithm for selecting anomaly feature points introduces a weighting mechanism, such as weighting gradient values ​​according to sensor importance, prioritizing data in key areas. After feature point extraction, metadata is generated, including gradient magnitude, direction, and confidence level, for use by subsequent modules.

[0063] Example 2: See Figure 3After the anomalous feature points are output from the pattern recognition module, they are fed into the spatiotemporal correlation model for analysis. The core task of this model is to analyze the intrinsic connections and diffusion patterns of these feature points in the temporal and spatial dimensions. The system first establishes a spatiotemporal neighborhood for each anomalous feature point. The definition of the spatiotemporal neighborhood includes two key parameters: time span and spatial radius. The time span specifies the range of time intervals to be examined, usually extending forward and backward based on the timestamp of the anomalous point. The spatial radius defines a spherical or cubic region centered on the spatial coordinates of the anomalous point. The specific size of the neighborhood is not fixed, but is preset according to the layout characteristics of the underground roadways in the coal mine, the density of the sensor network, and the properties of the monitored physical quantities. For example, for anomalies in carbon monoxide concentration, due to the rapid diffusion rate of the gas, its spatial radius may be set relatively large in order to capture potential correlations over greater distances. Within the defined spatiotemporal neighborhood, the model begins to calculate the energy transfer matrix between anomalous feature points. This matrix is ​​a square matrix, with its rows and columns corresponding to each anomalous feature point in the neighborhood. Each element in the matrix numerically represents the coupling strength between two points. The calculation of coupling strength comprehensively considers spatiotemporal proximity and numerical similarity. Spatial proximity is evaluated by calculating the three-dimensional Euclidean distance between the two points, while temporal proximity is measured by the absolute difference between the timestamps of the two points. Numerical correlation may be reflected by calculating the correlation coefficient or covariance of the monitoring data values ​​of the two points. Finally, these factors are combined into a scalar value characterizing the coupling strength through a predefined kernel function.

[0064] After calculating the energy transfer matrix, the model needs to analyze the propagation modes implied by the matrix. The key is calculating the attenuation slope of the matrix. The attenuation slope describes the rate at which the coupling strength weakens with increasing spatiotemporal distance. Its calculation is typically achieved by analyzing the changing trends of matrix element values ​​with the spatiotemporal distance between corresponding point pairs. For example, the coupling strength of all point pairs can be fitted as a curve to the spatiotemporal distance; the slope of the fitted curve is the attenuation slope. This slope value has a clear physical meaning. A large negative slope indicates that the anomalous energy attenuates rapidly with increasing propagation distance, suggesting that the anomalous event may be localized and have a limited impact range. Conversely, a small negative slope, or even a slope close to zero, indicates that the anomalous energy can propagate to a greater spatiotemporal range, potentially indicating a persistent or diffusing anomalous source. The model determines the main propagation direction of the diffusion path based on the characteristics of the attenuation slope. The identification of the main propagation direction can be achieved using vector field analysis, treating each anomalous feature point as a potential source point and determining it by analyzing the dominant direction of energy transfer to surrounding points. For example, the composite direction of energy transfer vectors from that point to other points in the neighborhood can be calculated.

[0065] When constructing the spatiotemporal neighborhood, the system references underground geographic information system data, automatically avoiding physical barriers such as rock walls or sealed walls. This ensures the neighborhood's shape conforms to the actual tunnel spatial structure, rather than a simple geometric shape. The selection of the time span also considers the historical duration patterns of anomaly types, using shorter time windows for transient anomalies and longer observation periods for slow anomalies. The calculation of the energy transfer matrix includes a normalization step to eliminate the influence of different physical dimensions and numerical ranges, ensuring the comparability of coupling strength. The choice of kernel function is crucial; commonly used kernel functions include Gaussian or exponential kernel functions, which ensure that the coupling strength decays smoothly with increasing spatiotemporal distance. When analyzing the attenuation slope, the system may use weighted least squares for curve fitting, assigning higher weights to closer point pairs to improve the accuracy of estimating local propagation characteristics. Determining the main propagation direction may involve calculating the principal eigenvector of the energy transfer matrix, whose direction indicates the main trend of energy diffusion in space. The spatiotemporal correlation model relies on an efficient numerical computing library to handle matrix operations involving a large number of anomalous feature points. Model parameters, such as neighborhood size and kernel bandwidth, can be optimized through learning from historical data or fine-tuned by domain experts based on specific downhole geological conditions and production activities. By quantifying the spatiotemporal interactions between anomalous feature points, the model connects isolated anomalous points into meaningful propagation paths, providing crucial insights into the dynamic development of anomalous events. Its output directly serves the generation of subsequent anomalous propagation maps.

[0066] In a specific working area of ​​the west wing transport roadway in an underground coal mine, assuming around 10:05 AM, the system, through an anomaly pattern recognition module, captured a set of anomalous feature points from continuous monitoring data. These points mainly originated from gas concentration sensors and vibration sensors deployed in the area, whose readings showed significant fluctuations deviating from the normal baseline within a short period. These anomalous feature points were marked with precise spatial coordinates (e.g., point A is located at coordinates (1050, 200, -750), point B at (1080, 210, -755), and point C at (1020, 195, -740), all in meters) and timestamps accurate to the second (e.g., point A at 10:05:03, point B at 10:05:12, and point C at 10:05:25). The spatiotemporal correlation model began processing this set of anomalous feature points, first establishing a spatiotemporal neighborhood for each point. Taking point A as an example, the system defines its neighborhood range based on preset parameters (e.g., for gas anomaly events, a time span of 5 minutes and a spatial radius of 50 meters). This means that the system will filter out all other anomalous feature points within a spherical range that are between 10:00:03 and 10:10:03 in time and 50 meters away from point A in space. Points B and C, being within this spatial range, are included in the neighborhood of point A for analysis.

[0067] Within the spatiotemporal neighborhood of point A, the model calculates the energy transfer matrix centered on point A. This matrix aims to quantify the coupling strength between point A and other points in the neighborhood (primarily points B and C). The calculation of coupling strength comprehensively considers spatiotemporal distance and numerical correlation. Spatially, the system calculates the three-dimensional Euclidean distance between point A and point B to be approximately 32 meters, and the distance between point A and point C to be approximately 36 meters. Temporally, the time difference between point A and point B is 9 seconds, and the time difference between point A and point C is 22 seconds. Numerically, there is a strong positive correlation between the peak gas concentration at point A and the peak concentration at point B. The model uses a predefined decay function (e.g., a Gaussian kernel function) to map spatiotemporal distance to coupling strength; the closer the distance, the shorter the time interval, and the higher the numerical correlation, the greater the calculated coupling strength value. It is assumed that the calculated coupling strength between point A and point B is high, while the coupling strength with point A and point C is relatively low. The model analyzes the decay trend reflected by this energy transfer matrix. It fits a decay curve by examining the attenuation of coupling strength from point A to its neighboring points as time-space distance increases. The slope of this curve, the decay slope, characterizes the rate at which anomalous energy weakens as it diffuses outward from point A. The analysis suggests that the coupling strength decays relatively slowly (with a smaller absolute value of the decay slope) from point A to point B, while decaying very rapidly (with a larger absolute value of the decay slope) in other directions, such as to point C. Based on the decay slope characteristics of the energy transfer matrix, the model determines the main propagation direction of the diffusion path. In this example, the slow decay indicates that the anomalous energy is more likely to propagate towards point B. Therefore, the system determines that the main propagation direction from point A roughly points to the coordinates of point B, which roughly coincides with the axial direction of the transport tunnel.

[0068] The model performs the same analysis process for each anomalous feature point (such as point B and point C). For example, it establishes a spatiotemporal neighborhood for point B, calculates its energy transfer matrix, and determines its main propagation direction. Through this point-to-point analysis, the system can outline the potential diffusion path of the anomaly in the downhole spatial and temporal dimensions, linking isolated anomaly points into a logical sequence, providing crucial evidence for determining the source and spread trend of the anomaly. For example, the analysis results may show that the main propagation direction of the anomalous feature points exhibits a pattern of radiation outward from a certain local area, which may suggest that this area is the origin of the anomaly. All these analysis results based on spatiotemporal neighborhood and energy transfer matrix will be passed to the next stage to generate anomaly propagation maps.

[0069] Example 3: After obtaining the propagation path of the anomalous feature points, the system begins to generate an anomaly propagation map. The construction of this map begins by mapping the main propagation direction of each propagation path to a three-dimensional spatial coordinate system underground. The mapping process relies on a pre-established underground spatial model, which includes the three-dimensional coordinate information of the main structures such as roadways, working faces, and chambers. Each anomalous feature point is assigned precise spatial coordinates (x, y, z) and a timestamp t based on the sensor or positioning tag associated with it during acquisition. The main propagation direction is then converted into a three-dimensional direction vector originating from that point. Based on this, the system generates a directed edge to represent the propagation path. The starting point of the directed edge is the spatial coordinates of the anomalous feature point itself, and the ending point is determined by extending a calculated length along the direction vector from the starting point. The length L of the directed edge is determined by the reciprocal of the attenuation slope η of its corresponding energy transfer matrix, specifically defined by the following formula:

[0070]

[0071] In this relationship, It is a scaling factor set according to the actual spatial scale of the mine, used to convert the calculated relative length into metric units in the actual coordinate system. It is the decay slope of the energy transfer matrix, characterizing the rate of energy decay. When the decay slope... When the absolute value is large, it indicates that the energy decays rapidly and the effective propagation distance is short, thus the length of the generated directed edge is... Shorter; conversely, when A smaller value means the impact of the anomaly may propagate further, resulting in a longer directed edge. Each directed edge is assigned rich attribute information, including metadata such as start coordinates, end coordinates, direction vector, timestamp, decay slope value, and anomaly source type.

[0072] After the initial framework of the anomaly propagation map is established, the system performs hierarchical clustering analysis to group paths with similar propagation characteristics in space. The clustering process first calculates the similarity between any two directed edges, measuring both directional and length similarity. Directional similarity is obtained by calculating the cosine of the angle between the direction vectors of the two directed edges; the closer the cosine is to 1, the closer the directions of the two edges are to being parallel. Length similarity is evaluated by comparing the ratio of the lengths of the two directed edges, using a normalization process. Ultimately, the overall similarity between two edges is a weighted combination of directional and length similarity. The system presets a clustering threshold; when the overall similarity between two edges is greater than or equal to this threshold, they are considered similar and merged into the same candidate cluster.

[0073] The clustering algorithm employs a bottom-up hierarchical aggregation strategy, initially treating each directed edge as an independent cluster. The similarity between all pairs of edges is calculated, and the pair with the highest similarity exceeding a threshold is merged to form a new, larger cluster. The representative features of the new cluster (such as average direction and average length) are recalculated based on all its contained edges and used for subsequent similarity calculations with other clusters. This process is iterative, continuously merging the most similar clusters until the similarity between all clusters is below a preset threshold. Ultimately, each formed anomalous cluster must contain at least two directed edges that share commonalities in direction and approximate length in the graph space. The clustering results simplify the originally discrete and complex path network into several anomalous clusters with clear spatiotemporal characteristics, providing a structured foundation for subsequent assessment of the severity and impact of anomalous events. The entire graph generation and clustering process is completed within the computation engine, which also supports graphical rendering of the resulting graph.

[0074] Example 4: In the early warning signal generation stage, the system processes anomalous clusters formed after hierarchical clustering. Each anomalous cluster contains several directed edges with similar propagation characteristics. The system first needs to quantify the spatiotemporal distribution characteristics of each anomalous cluster, which is achieved by calculating two core indicators: the total length of directed edges and the average directional consistency. The total length is calculated by summing the lengths of all directed edges within the cluster. This value intuitively reflects the cumulative spatial impact range of the anomalous event. The calculation of the average directional consistency is more complex. It requires assessing the concentration of the directed edge directions within the cluster. During the calculation, the system first converts the direction vector of each directed edge into a unit vector, and then calculates the average vector of all these unit vectors. The value of the average directional consistency is the magnitude of this average vector, and its value range is between 0 and 1. The closer the value is to 1, the more consistent the propagation direction of the edges within the cluster, which may point to a common anomalous source or a clear propagation path. The closer the value is to 0, the more dispersed the propagation direction is, which may mean that there are multiple anomalous sources or that the propagation process has been strongly interfered with by a complex environment. Assuming that cluster analysis identifies three anomalous clusters that need to be evaluated in the current period, their preliminary spatiotemporal characteristics are shown in Table 1.

[0075] Table 1: Spatiotemporal characteristics of anomaly clusters and generation of early warning signals

[0076] Abnormal cluster number Total length of directed edges (meters) Average directional consistency Preset length threshold comparison results Preset direction threshold comparison results Automatically generated initial priority label Cluster-A 125.6 0.32 Exceeding the threshold (85 meters) Below the threshold (0.6) High priority Cluster-B 78.9 0.75 Not exceeding the threshold Exceeding the threshold low priority Cluster-C 215.4 0.68 Exceeding the threshold Exceeding the threshold Medium priority

[0077] According to preset rules, when the total length of directed edges exceeds a length threshold (e.g., 85 meters) and the average directional consistency is lower than a direction threshold (e.g., 0.6), the system generates a high-priority label. As shown in Table 1, cluster-A meets both conditions, indicating that it has an impact over a large area downhole, and the propagation direction is chaotic, which may indicate a dangerous situation of rapid spread and multi-point outbreak, so it is marked as high priority. The total length of cluster-B does not exceed the threshold, and although the directional consistency is high, it indicates that the scope of the event's impact is currently relatively limited, so it is marked as low priority. The total length of cluster-C far exceeds the threshold, and the directional consistency is also high, indicating that there is an anomaly source with a wide impact range but a relatively clear propagation direction, and its risk level is between the two, so it is marked as medium priority. After generating the initial warning signal, the process does not end. These signals must be synchronously verified with the downhole equipment status data within the same time window to confirm whether the anomaly actually constitutes a substantial interference or threat to safe production. The system will obtain the main equipment operating parameters in the physical area corresponding to the warning signal from the equipment monitoring system. The core parameters include current fluctuation rate and load change. Current fluctuation rate measures the stability of the motor current of a device and is usually calculated as the ratio of the standard deviation to the average value. Load variation reflects the deviation of the device's output power from normal operating conditions.

[0078] Continuing with the three anomalous clusters in Table 1 as examples, the synchronous verification process is as follows: For cluster-A, marked as high priority, the system retrieves the operating data of equipment such as coal mining machines and scraper conveyors within its area (e.g., the "West Wing Coal Mining Face"). If it is found that the current fluctuation rate of the coal mining machine in this area reaches 15% (the safety threshold is set to 10%), and the load change exceeds 20% of the rated value, these abnormal equipment parameters highly coincide with the warning signal in space and time, strongly indicating that the abnormal event has had a real impact on the equipment operation. Therefore, the system decides to retain the high priority status of this target warning signal. For cluster-B, located in the "Central Transport Roadway," the verification found that the current fluctuation rate of the belt conveyor in this area is only 4%, the load change is within the normal range, and the equipment is operating smoothly. Given that the anomalous cluster itself has a small impact range and the equipment status is normal, the system downgrades this target warning signal to an observation signal, only recording it and continuously monitoring it, without initiating an emergency response. For cluster-C, which covers a large area of ​​the "East Wing Return Air Lane," verification data shows a slight increase in the current fluctuation rate of local ventilation fans in this area, at 8%, close to but not exceeding the safety threshold, and the load change slightly exceeds the normal range. After weighted judgment, the system believes that the risk does exist, but its urgency is not as high as that of cluster-A. Therefore, it may maintain its priority but shorten the monitoring cycle for this cluster. Through this mechanism of cross-validating anomaly propagation characteristics with real-time equipment status, the system can effectively filter out "false alarms" that show anomalies at the data level but do not have an actual impact on the production system. This ensures that the final triggered target warning signal has high accuracy and necessity for response, improving the reliability and credibility of the warning system.

[0079] Example 5: After confirming the validity of the target early warning signal through synchronous verification, the system enters the emergency control command triggering stage. The core of this stage is to execute a preset response strategy based on the priority label of the early warning signal. The response strategy library contains a set of specific operation commands corresponding to different priority labels. The formulation of these commands references coal mine safety regulations, historical emergency response experience, and the actual underground production layout. Assume that the target early warning signal currently being processed by the system involves two areas: one is the "cluster-A" signal marked as high priority in the west wing coal mining face, and the other is the "cluster-C" signal marked as medium priority in the east wing return airway.

[0080] For high-priority "Cluster-A" signals, the system first matches the highest-level response strategy from the strategy library. This strategy contains a series of ordered operational instructions, the primary one being the triggering of a "regional power outage instruction." This instruction is sent through the centralized control system to the power control center of the "West Wing Coal Mining Face," which will cut off power to large electrical equipment such as the coal mining machine, scraper conveyor, and transfer machine according to preset logic. Almost simultaneously, the system triggers a "ventilation acceleration instruction," which is sent to the mine's main ventilation fan control system and the local ventilation fan control system of the working face, requiring the ventilation volume to be increased to the emergency level to quickly dilute any potentially accumulated harmful gases or dust in the area. The instruction issuance is not a one-time action; the system continuously monitors the feedback status of the power outage area and ventilation equipment to confirm whether the instruction has been executed correctly. For example, the system checks whether the power monitoring module of the working face reports zero voltage and current, and verifies whether the ventilation fan frequency or wind pressure sensor data has reached the preset emergency level. For medium-priority "Cluster-C" signals, the system matches a response strategy that focuses on risk control rather than immediate production shutdown. The strategy might trigger a "local equipment production restriction order," which aims to reduce risk while maintaining a certain production schedule. The system will send instructions to the equipment controllers in the "East Wing Return Airway" and related areas, requiring the belt conveyor speed to be reduced by 20%, or the traction speed of the coal mining machine to be limited to a safe range. This production restriction mode aims to reduce the amount of heat and dust generated by the equipment, buying time to handle abnormal situations. The system will also monitor parameters such as equipment load and current to confirm that the production restriction order has been executed.

[0081] During command triggering and execution, the system updates its internal key threshold parameters in real time, namely the dynamic threshold and the dynamic fluctuation threshold. The update process is not a simple replacement, but a smoothing calculation based on recent historical data. The dynamic threshold update examines the frequency of all identified anomalous feature points over a past period (e.g., 24 hours). If the frequency shows an upward trend, the system may moderately increase the dynamic threshold to avoid generating excessive interfering alarms during periods of frequent anomalous events; conversely, it will remain unchanged or slightly decrease. The dynamic fluctuation threshold update is based on the average attenuation slope calculated from the recent energy transfer matrix. If the average attenuation slope indicates that the propagation distance of the anomaly is increasing, the system may adjust the dynamic fluctuation threshold accordingly to capture anomalous patterns with wide-area propagation characteristics earlier. The entire command triggering and threshold update form a closed-loop feedback system. For example, after an emergency response to the "Cluster-A" event, the system will include the complete processing log of the event, including the initial characteristics of the anomaly, the propagation map, the triggered commands, and their effects, as a data sample in the case library. These case data will be used to periodically optimize response strategies and adjust threshold update algorithms, enabling the system to learn from actual handling experience and continuously improve the accuracy of early warnings and responses. This dynamic adaptive mechanism ensures that the early warning system can continuously evolve with changes in downhole environmental conditions and production status, avoiding false alarms and missed alarms caused by rigid thresholds, and ensuring that emergency control measures are as close as possible to the actual risk level.

[0082] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0083] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for early warning in coal mines, characterized in that, Includes the following steps: Collect multi-dimensional monitoring data of the underground environment in coal mines, including gas concentration, temperature and humidity, vibration intensity and equipment current parameters; Dynamically divide multidimensional monitoring data into windows to generate a set of data segments with temporal continuity; Anomaly pattern recognition is performed based on a set of data segments, and abnormal feature points exceeding dynamic thresholds are extracted from each data segment. Anomaly feature points are input into a pre-built spatiotemporal correlation model, and the diffusion path of anomaly feature points in the time and space dimensions is analyzed through the spatiotemporal correlation model. An anomaly propagation map is generated based on the diffusion path, and the anomaly propagation map includes the propagation direction and intensity attenuation gradient of the anomaly feature points; The step of generating an abnormal propagation map based on the diffusion path includes: Map the main propagation direction of the diffusion path to the downhole three-dimensional spatial coordinate system to generate directed edges with directional arrows; Set the length of the directed edge to the reciprocal of the attenuation slope of the corresponding energy transfer matrix; The starting and ending points of the directed edges are marked with the spatial coordinates and timestamps of the anomalous feature points, respectively. Hierarchical clustering of the anomaly propagation map was performed to divide anomaly clusters with similar propagation characteristics; The hierarchical clustering of the abnormal propagation map includes: Calculate the directional similarity and length similarity between directed edges, and merge directed edges whose similarity exceeds the clustering threshold; Anomaly clusters are divided based on the merged directed edge connections, and each anomaly cluster contains at least two directed edges; An early warning signal is generated by utilizing the spatiotemporal distribution characteristics of anomaly clusters, and the early warning signal includes a priority label for the anomaly clusters; The warning signals are synchronously verified with the downhole equipment status data to filter out the target warning signals that need to be responded to. The corresponding emergency control command is triggered based on the target warning signal.

2. The coal mine underground early warning method according to claim 1, characterized in that, The abnormal pattern recognition based on the data fragment set includes: Perform multi-scale sliding window analysis on the data fragment set to calculate the mean shift and standard deviation rate of change of the data within the window; When the mean offset exceeds the historical baseline mean and the rate of change of standard deviation is greater than the dynamic fluctuation threshold, the current window is marked as an abnormal window. Extract the instantaneous gradient of all data points within the anomaly window, and take the data point with the largest absolute gradient value as the anomaly feature point.

3. The coal mine underground early warning method according to claim 2, characterized in that, The analysis of the diffusion path of abnormal feature points in the time and space dimensions using a spatiotemporal correlation model includes: Establish a spatiotemporal neighborhood centered on the anomalous feature points, wherein the spatiotemporal neighborhood includes a time span and a spatial radius; Calculate the energy transfer matrix of the anomalous feature points in the spatiotemporal neighborhood, where the energy transfer matrix reflects the coupling strength between the anomalous feature points; The main propagation direction of the diffusion path is determined by the attenuation slope of the energy transfer matrix.

4. The coal mine underground early warning method according to claim 3, characterized in that, The method of generating early warning signals by utilizing the spatiotemporal distribution characteristics of abnormal clusters includes: The total length and average direction of directed edges within statistically abnormal clusters are consistent. Compare the total length with a preset length threshold, and compare the average directional consistency with a preset directional threshold; When the total length exceeds the length threshold and the average directional consistency is lower than the directional threshold, a high-priority label is generated.

5. The underground early warning method for coal mines according to claim 4, characterized in that, The process of synchronizing and verifying the early warning signal with the downhole equipment status data includes: Obtain the operating parameters of the downhole equipment in the current time window, including current fluctuation rate and load change. If the operating parameters of the equipment in the area corresponding to the target warning signal exceed the safety threshold, the target warning signal will be retained. Otherwise, the target warning signal will be downgraded to an observation signal.

6. The coal mine underground early warning method according to claim 5, characterized in that, The step of triggering the corresponding emergency control command based on the target warning signal includes: Match the preset response strategy according to the priority label of the target warning signal; If the priority label is high, then the area power-off command and ventilation acceleration command will be triggered. If the priority tag is low, a local device production restriction command will be triggered.

7. The coal mine underground early warning method according to claim 6, characterized in that, Also includes: The dynamic threshold and dynamic fluctuation threshold are updated in real time, and the update is based on the occurrence frequency of historical abnormal feature points and the average decay value of the energy transfer matrix.

8. A coal mine underground early warning system, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the underground early warning method for coal mines as described in any one of claims 1 to 7.

Images