A method and system for recognizing and intelligently warning gas concentration mode of a coal mining face

Abstract

The application discloses a kind of coal face gas concentration pattern recognition and intelligent early warning method and system, belong to coal mine safety monitoring technical field.The method is carried out microsecond clock synchronization to distributed sensor network by accurate time synchronization protocol, and gas concentration signal with time stamp is collected;Based on the physical topology of ventilation system, the spatial correlation weight is calculated, the adjacency matrix is constructed by fusing time delay cross-correlation coefficient;Real-time acquisition of coal winning machine position coding is spatial feature vector, and the concentration sequence is combined with the adjacency matrix and attached spatial feature vector, to construct enhanced spatio-temporal feature matrix;Input spatio-temporal graph convolution network model extracts spatio-temporal evolution characteristics, and combines spatio-temporal physical consistency condition to confirm gas accumulation evolution mode;Based on the control instruction generated according to mode, automatic closed-loop control is carried out to actuating mechanism.The application realizes accurate identification and hierarchical control to gas risk.

Description

Technical Field

[0001] This application belongs to the field of coal mine safety monitoring technology, specifically relating to a method and system for identifying gas concentration patterns and providing intelligent early warning in coal mining faces. Background Technology

[0002] With the deepening of intelligent mine construction, coal mine safety monitoring systems play a core role in preventing major gas accidents and ensuring mine production safety. Real-time monitoring and accurate early warning of gas concentration, as the foundation of ventilation safety management, directly affect the continuity and safety of coal mining operations. Especially in modern high-yield and high-efficiency coal mining faces, building a dynamic monitoring system that can cover the entire site and respond in real time has become an inevitable requirement for the intelligent transformation of coal mine safety production.

[0003] Among these, feature extraction and pattern recognition of gas concentration in coal mining faces are key pathways to achieving intelligent early warning. In the complex underground production environment, gas emission is subject to real-time adjustments from multiple sources, including mining progress, geological structure, and ventilation parameters, exhibiting complex coupled correlations in both temporal scale and spatial distribution. To accurately grasp the safety status, the monitoring system needs to synchronize data from dispersed sensor nodes and, by mining the inherent logic between the data from each monitoring point, achieve real-time identification of gas field distribution patterns and risk levels.

[0004] However, traditional monitoring and early warning systems mostly rely on static threshold discrimination methods. This isolated analysis of single-point concentration values ​​makes it difficult to distinguish between concentration fluctuations caused by normal production activities and abnormal mutations induced by gas outbursts. Simultaneously, sensors are susceptible to electromagnetic interference or dust obstruction, generating data noise. Due to the lack of effective constraints on the spatiotemporal physical laws between sensors, the system is highly prone to false alarms at single points. Furthermore, existing technologies lack the capacity to process multi-source heterogeneous monitoring data, making it difficult to establish dynamic matrix models reflecting gas migration characteristics and capture the nonlinear evolution trends during gas accumulation, resulting in significant lags in risk identification. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for identifying and intelligently warning about gas concentration patterns in coal mining faces, which can effectively solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: Firstly, a method for identifying gas concentration patterns and providing intelligent early warning in coal mining faces includes: Through a distributed sensor network, clock synchronization is performed on each sensor node based on a precise time synchronization protocol, gas concentration signals are collected synchronously at fixed periods, and a timestamp is added to each sampling point. Obtain the physical topology map of the ventilation system, calculate the spatial correlation weight based on the ventilation path length between sensor nodes, and perform time-domain correlation analysis on the concentration sequences of sensor node pairs with upstream and downstream causal relationships to calculate the time delay cross-correlation coefficient. The two are then fused to construct an adjacency matrix representing the directed transmission relationship. The location information of the coal mining machine is acquired in real time and encoded into a spatial feature vector. The concentration sequence of all sensor nodes within the preset time window is combined with the adjacency matrix, and the spatial feature vector is appended to the features of each sensor node to construct an enhanced spatiotemporal feature matrix. The enhanced spatiotemporal feature matrix is ​​input into the spatiotemporal graph convolutional network model. Spatial graph convolution and temporal causal convolution are performed sequentially through multiple spatiotemporal convolutional blocks to extract the spatiotemporal evolution features of the gas concentration field. The preliminary pattern classification results are output, which include at least one of the following: gas accumulation evolution mode, sudden warning mode, and sensor failure mode. When the preliminary model classification result is a gas accumulation evolution model, check whether the spatiotemporal physical consistency condition is met, including that the concentration of the sensor node shows an upward trend and the maximum time delay cross-correlation coefficient of the concentration sequence of its downstream sensor node and the concentration sequence of the sensor node increases synchronously. If the condition is met, the model is confirmed. Based on the confirmed pattern, the system retrieves handling rules from the preset risk response strategy library and generates control instructions to automatically control the early warning execution agency in a closed loop.

[0007] Preferably, clock synchronization is performed on each sensor node based on a precise time synchronization protocol, gas concentration signals are collected synchronously at fixed periods, and a timestamp is added to each sampling point, including: All sensor nodes are synchronized with the master clock of the central processing server via the mining industrial Ethernet and perform signal acquisition synchronously under the unified clock tick. Each sensor node converts the acquired raw analog voltage signal into a digital signal sequence, and the clock module of the sensor node adds a high-precision timestamp to each sampled data point in the digital signal sequence to record the absolute moment when the sampling occurred.

[0008] Preferably, temporal correlation analysis is performed on the concentration sequences of sensor node pairs with upstream and downstream causal relationships to calculate the time-delay cross-correlation coefficient, including: Before calculating the time delay cross-correlation coefficient, wavelet decomposition is performed on the concentration sequence of the target sensor node pair to remove high-frequency detail components corresponding to the coal mining machine cutting frequency and extract low-frequency profile signals that reflect the natural migration trend of gas. For the low-frequency profile signal, within a preset time lag range, the time series of one of the sensor nodes is moved step by step with the sampling period as the step size, and the correlation coefficient under each lag step size is calculated. The maximum value of the correlation coefficient sequence is searched. If the maximum value exceeds the preset correlation threshold, it is determined that there is a significant fluid transport causal relationship between the pair of sensor nodes.

[0009] Preferably, spatial graph convolution and temporal causal convolution are performed sequentially through multiple spatiotemporal convolution blocks, including: The spatiotemporal graph convolutional network model contains multiple spatiotemporal convolutional blocks stacked in series. Within each spatiotemporal convolutional block, spatial graph convolution is performed first, followed by temporal convolution. The spatial graph convolution operation is based on the Chebyshev polynomial approximation method of the normalized Laplacian matrix, and the features of each sensor node and its multi-order neighboring sensor nodes are weighted and aggregated. The temporal convolution operation employs a causal dilated convolution structure, and as the number of spatiotemporal convolutional block layers increases, the dilation rate of the causal dilated convolution is multiplied layer by layer to exponentially expand the temporal receptive field and capture the nonlinear evolution trend of gas concentration from the second to the minute level.

[0010] Preferably, based on the confirmed mode, the system retrieves handling rules from a pre-defined risk response strategy library and generates control instructions, specifically including: When the confirmed mode is the emergency warning mode, the highest priority control command is generated and pushed. The control command is used to trigger the parallel execution of power outage isolation operation, ventilation system air increase adjustment operation and emergency broadcast operation. When the confirmed mode is the gas accumulation and evolution mode, the corresponding level of control command is generated and pushed. The control command is used to trigger the coal mining machine electrical control system to reduce the traction speed to below a preset ratio, and to send a prompt operation containing alarm information to the explosion-proof mobile terminal device.

[0011] Preferably, the spatial correlation weight is calculated based on the ventilation path length between sensor nodes, specifically including: The spatial correlation weight w between any two sensor nodes i and j is calculated using a Gaussian kernel function. ij The calculation formula is: , where d(i,j) is the actual length along the ventilation path between the two sensor nodes, and σ is the spatial attenuation constant; The spatial attenuation constant σ is dynamically optimized through an online adaptive correction mechanism. Specifically, the online adaptive correction mechanism involves: extracting concentration monitoring data from the coal face shutdown and maintenance period in the historical database, using the Gaussian plume model inversion algorithm to fit the measured attenuation characteristic value under the current air volume conditions, and updating the preset σ in a rolling manner.

[0012] Preferably, the real-time acquisition of the coal mining machine's position information and its encoding into a spatial feature vector, and the appending of the spatial feature vector to the features of each sensor node to construct an enhanced spatiotemporal feature matrix, specifically includes: The precise three-dimensional position information of the coal mining machine in the working face spatial coordinate system can be obtained in real time by a positioning transmitter installed on the coal mining machine and a receiver deployed along the working face, or by a rotary encoder installed on the traveling part of the coal mining machine. When constructing the enhanced spatiotemporal feature matrix, the three-dimensional precise location information is encoded into a multi-dimensional spatial feature vector, and then appended to the concentration sequence features of each sensor node by feature concatenation to form an enhanced sensor node feature vector, so that the spatiotemporal graph convolutional network model can synchronously learn the spatiotemporal coupling relationship between the coal mining machine position and the gas concentration distribution.

[0013] Preferably, the spatiotemporal physical consistency condition further includes: When the preliminary pattern classification results show that only a single sensor node produces drastic numerical fluctuations, and none of the downstream neighboring sensor nodes of that sensor node, as determined by the adjacency matrix, detect a correlation response consistent with the causal relationship of fluid transport, the signal is determined to be a sensor failure mode.

[0014] Preferred options also include: During system operation, historical pattern recognition records are continuously accumulated to build a structured historical pattern database containing production intensity parameters, ventilation parameters, and gas concentration evolution trajectories. When the cumulative number of newly confirmed pattern samples added to the historical pattern database reaches the preset update trigger condition, an incremental training task is started. Based on the parameters of the currently running spatiotemporal graph convolutional network model as the initial values, iterative updates are performed only on the newly added sample set, and the updated model parameters are replaced in the online inference model file.

[0015] Secondly, a coal mining face gas concentration pattern recognition and intelligent early warning system, used to execute the above method, includes a central processing server, and: A distributed sensor network is deployed at key locations in the coal mining face to synchronize the clocks of each sensor node based on a precise time synchronization protocol. It collects gas concentration signals synchronously at a fixed sampling period and adds a timestamp to each sampling data point. The data transmission subsystem is deployed between the distributed sensor network and the central processing server, and is used to upload the gas concentration signals and corresponding timestamps collected by each sensor node to the central processing server with a deterministic delay based on the mining industrial Ethernet. The spatiotemporal feature matrix construction module, deployed in the central processing server, is used to acquire the physical topology map of the ventilation system of the coal mining face, calculate the spatial correlation weight based on the ventilation path length between sensor nodes and calculate the time delay cross-correlation coefficient through temporal correlation analysis, and construct an adjacency matrix by fusing the two. It also acquires the coal mining machine position information in real time and encodes it into a spatial feature vector, combines the concentration sequence within a preset time window with the adjacency matrix, and appends the spatial feature vector to the features of each sensor node to construct an enhanced spatiotemporal feature matrix. The pattern recognition module, deployed in the central processing server, is used to load the trained spatiotemporal graph convolutional network model, perform spatiotemporal convolution operations on the enhanced spatiotemporal feature matrix through the model to extract the spatiotemporal evolution features of the gas concentration field and output preliminary pattern classification results. The preliminary pattern classification results include at least one of the following: gas accumulation evolution mode, sudden warning mode, and sensor failure mode. When the preliminary pattern classification result is a gas accumulation evolution mode, it is verified whether it meets the preset spatiotemporal physical consistency condition to confirm the final mode. The intelligent decision-making and early warning module, deployed in the central processing server, is used to retrieve matching disposal rules from the preset risk response strategy library based on the confirmed pattern classification results, and generate control instructions to automatically control the early warning execution mechanism in a closed loop.

[0016] In summary, this application includes at least one of the following beneficial technical effects: 1. This invention achieves microsecond-level clock synchronization for a distributed sensor network based on a precise time synchronization protocol. It also integrates the physical topology of the ventilation system with the temporal delay causal relationships between sensor nodes to construct an adjacency matrix representing directed transmission relationships, thereby generating an enhanced spatiotemporal feature matrix. This method not only ensures absolute alignment of multi-point monitoring data on the time axis but also explicitly encodes the gas transport patterns in the tunnel into the data model. This effectively solves the problems of spurious correlations and pattern mismatches caused by asynchronous data and neglect of spatial physical constraints in traditional methods, providing a highly accurate data foundation for subsequent precise pattern recognition.

[0017] 2. This invention inputs an enhanced spatiotemporal feature matrix into a spatiotemporal graph convolutional network model that includes spatial graph convolution and temporal causal dilated convolution. The model automatically extracts multi-scale spatiotemporal evolution features of the gas concentration field, and then performs dual verification on the identified gas accumulation evolution pattern by combining it with preset spatiotemporal physical consistency conditions. This scheme overcomes the limitations of existing technologies that rely solely on data-driven approaches or simple thresholds. It utilizes deep learning to capture complex nonlinear trends that change from second-level abrupt changes to minute-level gradual variations, and it eliminates interference from sensor malfunctions or production artifacts through fluid dynamics causality, thereby improving the accuracy and robustness of gas accumulation risk identification and effectively reducing false alarms and false negatives.

[0018] 3. This invention achieves refined identification of gas emission sources by acquiring and injecting the three-dimensional position information of the coal mining machine into an enhanced spatiotemporal feature matrix in real time. Based on this, a hierarchical closed-loop control strategy matching the sudden warning and accumulation evolution modes is constructed. Unlike the single alarm output of traditional solutions, this invention can automatically execute differentiated control according to the risk type and source. For example, during accumulation evolution, it can actively limit the traction speed of the coal mining machine through protocol-level commands to suppress gas release from the source, or trigger power outage, ventilation increase, and broadcast in parallel during sudden warnings. This forms an automated closed-loop management system from intelligent perception and accurate identification to source and path coordinated control, improving the intelligence level of coal mine safety management and emergency response efficiency. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the method for identifying and intelligently warning about gas concentration patterns in coal mining faces, as described in this application.

[0020] Figure 2 This is a schematic diagram illustrating the core principle of gas outburst pattern recognition based on spatiotemporal graph convolutional networks in this application.

[0021] Figure 3 This is a schematic diagram illustrating the construction logic of the spatiotemporal feature matrix that integrates physical topological weights and temporal cross-correlation in this application.

[0022] Figure 4 This is a flowchart of the multi-level early warning and linkage decision-making logic based on the risk response strategy library in this application.

[0023] Figure 5 This is a schematic diagram of the distributed monitoring network and multi-level data flow of the early warning system of this application. Detailed Implementation

[0024] To further illustrate the technical means and effects adopted by the present invention to achieve the intended purpose, the following detailed description of specific embodiments based on the present invention is provided in conjunction with the accompanying drawings and preferred embodiments.

[0025] Reference Figures 1 to 5As shown in this embodiment of the invention, a coal mining face gas concentration pattern recognition and intelligent early warning system addresses the nonlinear and sudden characteristics of gas outburst patterns in complex mine environments. The system employs a distributed deployment model in physical space, specifically including a multi-point sensing module, a data transmission subsystem, a central processing server, and an early warning execution mechanism. The multi-point sensing module deploys high-precision gas concentration sensors at specific physical intervals at key locations in the coal mining face, such as the intake airway, the face head, the middle of the face, the face tail, the upper corner, and the return airway, constructing a fully covered distributed monitoring network. The data transmission subsystem is based on mining industrial Ethernet and adopts a redundant ring network architecture to ensure that the raw electrical signals collected by each sensor node, after digital conversion, can be uploaded to the central processing server with a deterministic delay of milliseconds.

[0026] The central processing server integrates a spatiotemporal feature matrix construction module, a pattern recognition module, and an intelligent decision-making and early warning module. The spatiotemporal feature matrix construction module transforms multi-source, heterogeneous sensor data into feature matrices with physical topological information. The pattern recognition module loads a pre-trained spatiotemporal graph convolutional network model to perform deep convolution operations and extract evolutionary trends from the feature matrices. The intelligent decision-making and early warning module, based on the identified pattern results, drives the early warning execution mechanism through a pre-set risk response strategy library, achieving automated closed-loop control of the coal mining machine's operating status, ventilation equipment parameters, and underground power supply.

[0027] Specifically, the present invention provides a method for identifying gas concentration patterns and providing intelligent early warning in coal mining faces, comprising the following specific steps: Step S1: Synchronously acquire and preprocess the gas concentration time series data.

[0028] The core problem to be solved in this step is to acquire a gas concentration observation sequence with strict time alignment, clean values, and uniform units under the strong electromagnetic interference and multi-source heterogeneous sensing environment in the mine. The specific steps are as follows: In step S101, multiple gas concentration sensors deployed in the intake airway, head, middle, tail, upper corner and return airway of the coal mining face form a distributed monitoring network. All sensor nodes are synchronized with the master clock of the IEEE 1588 precision time synchronization protocol in the central processing server through the mining industrial Ethernet. The clock deviation between the sensor nodes in the whole network is controlled within the microsecond level, providing a unified time reference for the strict alignment of data from multiple sensor nodes on the time axis.

[0029] In step S102, all sensor nodes synchronously acquire gas concentration signals under the drive of a unified clock beat, according to a fixed sampling period of 500 milliseconds.

[0030] During each sampling period, the electrochemical catalytic combustion sensing element inside the gas concentration sensor responds to the concentration of gas molecules in the air at the current location and outputs a weak analog current signal proportional to the volume fraction. After being conditioned by a preamplifier, the current signal forms a raw analog voltage signal in the range of 0 to 5 volts.

[0031] It should be noted that the 500-millisecond sampling period is set to balance the need for dynamic capture of gas concentration changes with the data transmission bandwidth limitations under the redundant ring network architecture of mining industrial Ethernet.

[0032] In step S103, the analog-to-digital converter unit built into each sensor node converts the original analog voltage signal obtained in step S102 into a digital signal sequence with 24-bit resolution; at the same time, the clock module of the sensor node adds a high-precision timestamp to each sampled data point to accurately record the absolute moment when the sampling occurs.

[0033] The digital signal sequence with timestamps is encapsulated into data frames according to the mining industrial Ethernet application layer protocol and uploaded to the real-time database of the central processing server through a redundant ring network.

[0034] In step S104, after receiving the raw digital observation stream, the central processing server immediately performs median filtering on each sensor data stream to remove random impulse noise caused by electromagnetic interference and mechanical vibration.

[0035] The specific processing mechanism is as follows: For each sensor channel, the system maintains a first-in-first-out sliding window containing 5 consecutive sampled values ​​in memory. The window width parameter 5 is determined by frequency domain analysis experiments based on the sensor response frequency and the typical electromagnetic noise spectrum distribution in the mine. Whenever a new sampled value arrives, the window slides forward one position. The system sorts the 5 values ​​in the window and extracts the median of the 3rd position after sorting as the effective gas concentration observation value at the current moment. Extreme jump values ​​in the window are naturally discarded after sorting. Thus, while maintaining the integrity of the slow change trend of gas concentration, it effectively suppresses instantaneous spike interference caused by the start of the coal mining machine cutting motor and the operation of the frequency converter switch.

[0036] Step S105: For the effective gas concentration observations after median filtering, the system further performs a minimum-maximum normalization transformation to uniformly map the numerical distribution deviations at different deployment points, caused by differences in sensor ranges and baseline background concentrations, to a closed interval of 0 to 1. The normalization transformation follows the formula: ;

[0037] Where x represents the current effective gas concentration observation value output in step S104, in volume fraction percentage; The preset minimum parameter is determined based on the background concentration in the intake roadway under normal mine ventilation conditions. This parameter is set through on-site measurement during system deployment. The full-scale alarm threshold set by the sensor at the factory is read from the sensor device description file; The normalized concentration characteristic value ranges from 0 to 1.

[0038] After normalization transformation, the numerical features output by the low-concentration reference sensor on the air inlet side and the high-concentration sensor on the air return side are scaled to the same order of magnitude, eliminating the static numerical deviation introduced by the difference in physical properties of the sensor installation location. This allows the subsequently loaded spatiotemporal graph convolutional network model to be free from the influence of inconsistent input feature scales during the training and inference phases, accelerating model convergence and improving the numerical stability of pattern recognition.

[0039] Through the above step S1, this method completes the entire process from the acquisition of physical signals from distributed sensors to the generation of standardized digital feature sequences. It not only ensures the accurate synchronization of data from multiple sensor nodes in terms of timestamps, but also enhances the authenticity and consistency of the data through median filtering and normalization operations. This ensures that the enhanced spatiotemporal feature matrix constructed in the subsequent step S2 can accurately reflect the dynamic evolution of the gas field in the coal mining face under real physical constraints.

[0040] Step S2: Construct the enhanced spatiotemporal feature matrix.

[0041] In this step, the normalized gas concentration sequence output from step S1 is integrated with the inherent ventilation physical topology of the coal mining face to construct a unified mathematical expression that simultaneously carries temporal evolution information and spatial correlation information, namely, an enhanced spatiotemporal feature matrix. This matrix is ​​then used by the spatiotemporal graph convolutional network model in step S3 for deep feature extraction. The specific steps are as follows: In step S201, the system defines the time observation window length parameter as the historical data range of the past 5 minutes. Based on the fixed sampling period of 500 milliseconds established in step S1, each sensor node accumulates 600 continuous sampling points within the 5-minute time window.

[0042] The system extracts the concentration observation sequence of all N sensor nodes within the time window from the real-time database of the central processing server, and organizes it in memory to form an initial observation matrix of dimension N×600. The i-th row of the initial observation matrix corresponds to the concentration change trajectory of the i-th sensor node in the past 5 minutes, and the j-th column corresponds to the instantaneous spatial distribution state of all sensor nodes at the j-th time step.

[0043] In step S202, the system loads the pre-stored physical topology map file of the ventilation system of the coal mining face. The file uses a sensor node-edge data structure to describe the roadway connectivity between the installation locations of each sensor.

[0044] For any two sensor nodes i and j, the system calculates the ventilation path length d(i,j) between them along the centerline of the roadway, in meters. The path length calculated here follows the cumulative length of the directed path in the actual flow direction of the underground airflow. For example, from the sensor node in the intake airway to the sensor node in the return airway, the distances of each segment are accumulated after passing through the intermediate sensor nodes such as the head, middle, tail, and upper corner of the working face in sequence along the airflow direction.

[0045] Step S203: To quantify the spatial correlation strength between sensor nodes caused by gas diffusion and wind carryover, the system uses a Gaussian kernel function to calculate the spatial distance weight w. ij The calculation formula is as follows: ;

[0046] In the formula, w ij d(i,j) represents the spatial correlation weight between sensor node i and sensor node j, with a value ranging from 0 to 1. The closer the value is to 1, the stronger the potential for spatial correlation of gas concentration between the two sensor nodes. d(i,j) is the actual length along the ventilation path between the two sensor nodes calculated in step S202, in meters. σ is a preset spatial attenuation constant parameter, in meters, the same length unit as d(i,j). exp represents an exponential function with the natural constant e as the base.

[0047] In a preferred embodiment of the present invention, the initial value of σ is predetermined based on the average cross-sectional wind speed and gas diffusion coefficient of the mine roadway using an empirical formula; while during the long-term operation of the system, the spatial attenuation constant σ is further dynamically optimized through an online adaptive correction mechanism.

[0048] The specific correction mechanism is as follows: the system extracts the upper corner and return airway concentration monitoring data of the coal face during the shutdown and maintenance period (i.e., the period when there is no mining disturbance and the gas outburst is relatively stable) from the historical database, uses the Gaussian plume model inversion algorithm to fit the measured attenuation characteristic value under the current air volume condition, and updates the preset σ in a rolling manner.

[0049] In this way, the spatial attenuation constant σ can automatically adapt to changes in resistance caused by adjustments to the working face ventilation system or roadway deformation, thereby ensuring the spatial weight w calculated by the Gaussian kernel function. ij It always conforms to the actual physical diffusion laws downhole, avoiding model mismatch problems caused by using fixed empirical values.

[0050] The physical law revealed by the above calculation formula is that as the ventilation path length between the two sensor nodes increases, the spatial correlation weight decreases exponentially and rapidly, which is consistent with the actual transport law of gas concentration field under the action of airflow, which weakens with increasing distance.

[0051] In step S204, in addition to the static spatial distance weights, the system further calculates the dynamic cross-correlation coefficients between the sensor node pairs in the time domain. Specifically, for the pair of sensor nodes—the upper corner sensor and the return airway sensor—which have a clear upstream-downstream causal relationship, the system gradually moves the time series of one of the sensor nodes within a time lag range of 5 to 60 seconds, using the sampling period as the step size, and calculates the Pearson correlation coefficient for each lag step.

[0052] Specifically, to avoid spurious peak interference caused by the periodic cutting operations of the coal mining machine in the correlation calculation, the system prioritizes wavelet decomposition of the concentration sequences from the upper corner sensor and the return airway sensor before performing cross-correlation calculations. The number of wavelet decomposition levels is determined based on the sampling frequency and the typical cutting frequency range of the coal mining machine. In one specific implementation, the Daubechies wavelet basis is used to perform a three-level decomposition of the original signal, removing high-frequency detail components corresponding to the coal mining machine cutting frequency (typically 0.05Hz to 0.2Hz), and extracting low-frequency profile signals reflecting the natural migration trend of gas for subsequent cross-correlation calculations.

[0053] The system searches for the maximum value of the correlation coefficient sequence and the corresponding optimal delay time step. The physical meaning of the optimal delay time step is to quantify the average transit time of the main energy peak required for the abnormal gas concentration signal to propagate from the upstream sensor node to the downstream sensor node under the current ventilation conditions. If the maximum correlation value exceeds the preset correlation threshold of 0.85, it indicates that there is a significant fluid transport causal relationship between the two sensor nodes. The system marks the connection weight of the sensor node pair in the adjacency matrix as a preset high value of 0.9 or 1.0, and stores the optimal delay time step as a dynamic feature parameter for use as an auxiliary input feature in the spatiotemporal graph convolutional network model in subsequent step S3 or for predicting the arrival time of the abnormal gas peak in step S4.

[0054] In step S205, the system performs a weighted fusion of the spatial distance weights calculated in step S203 and the temporal cross-correlation coefficients calculated in step S204. The fusion strategy adopts either multiplication or weighted summation.

[0055] The combined weight values ​​after fusion are filled into an N×N adjacency matrix A, and the matrix elements A ij This represents the directed connection strength from sensor node i to sensor node j. If there is no direct ventilation path connection between the two sensor nodes and no significant time delay correlation is detected, the corresponding matrix element is set to zero.

[0056] In step S206, the system combines and encapsulates the N×600 initial observation matrix generated in step S201 with the N×N adjacency matrix generated in step S205 to form a complete enhanced spatiotemporal feature matrix data structure.

[0057] The data structure is formally represented as a graph signal: the topology of the graph is defined by the adjacency matrix A, and the feature vector of each vertex in the graph consists of the normalized concentration sequence of the sensor node over 600 time steps, reflecting the evolution structure of the gas concentration field in the coal mining face on the full spatial scale and the recent historical time scale.

[0058] Through step S2 above, this method reconstructs the originally isolated multi-channel sensor data streams that only reflect single-point numerical changes into a graph-structured spatiotemporal feature matrix with embedded ventilation physical topology constraints. This not only preserves the detailed information of gas concentration evolution over time, but also explicitly encodes the spatial transmission relationship based on airflow path and the causal following relationship based on concentration sequence time delay between each monitoring point through the adjacency matrix.

[0059] Step S3: Perform gas outburst pattern recognition.

[0060] In this step, the enhanced spatiotemporal feature matrix constructed in step S2 is used as input data and fed into a pre-trained spatiotemporal graph convolutional network model. Through multi-layer nonlinear transformation, the spatiotemporal evolution features of the gas concentration field are automatically extracted, and the final output is a classification result of the gas emission mode at the current working face, distinguishing between different working conditions such as normal production fluctuations, gas accumulation and evolution, sensor failure, and sudden anomalies. The specific steps are as follows: Step S301: Load the pre-trained spatiotemporal graph convolutional network model file from the model library of the central processing server. The spatiotemporal graph convolutional network model file contains all weight parameters and bias parameters of four cascaded and stacked spatiotemporal convolutional blocks. Each spatiotemporal convolutional block performs transformation processing on the input features in the order of spatial domain convolution followed by temporal domain convolution. The output of the previous convolutional block is directly used as the input of the next convolutional block, forming a deep feature extraction pipeline that abstracts layer by layer.

[0061] In step S302, in each spatiotemporal convolution block, spatial graph convolution operation is first performed, which is based on the Chebyshev polynomial approximation method of the Laplace matrix.

[0062] Specifically, the spatial graph convolutional layer receives the graph signal feature matrix from the previous layer, and combines it with the ventilation network topology contained in the adjacency matrix A generated in step S205 to perform weighted aggregation of the features of each sensor node and its multi-order neighboring sensor nodes.

[0063] In this embodiment, the order K of the Chebyshev polynomial is set to 3. The receptive field of the spatial convolution kernel covers the 1st, 2nd, and up to 3rd order neighborhoods of each sensor node. This allows for the capture of the mid-range coupling correlation of gas concentration in the working face space within a single convolution layer. The spatial graph convolution operation follows the formula below: ;

[0064] In the formula: x is the feature vector or feature matrix input to the current spatial graph convolutional layer, with dimensions N×C. in N is the total number of sensor nodes, C in The number of input feature channels; The normalized Laplacian matrix is ​​obtained by normalizing the degree matrix from the adjacency matrix A. The specific calculation method is as follows: , where I is the identity matrix and D is the degree matrix of the adjacency matrix A; T k Let be the k-th order Chebyshev polynomial, its recursive definition is , And for k≥2, ; θ k The trainable weight coefficients learned by the model during the training phase for the k-th order polynomial components determine the degree of contribution of each order of neighborhood features to the final convolution result. σ is a nonlinear activation function; in this embodiment, a modified linear unit function is used, mathematically expressed as follows: ; y is the feature matrix output after spatial graph convolution, with dimensions N×C. out C out This determines the number of output feature channels.

[0065] Through the above calculations, the spatiotemporal graph convolutional network model can effectively extract concentration change patterns with clear spatial orientation, such as the movement of gas from the front to the rear of the turbine and the diffusion from the upper corner to the return airway.

[0066] Step S303: After the spatial dimension feature aggregation is completed in the spatial graph convolutional layer, the system continues to perform temporal convolution operations within the same spatiotemporal convolutional block.

[0067] The temporal convolutional layer employs a causal dilated convolutional structure. Causal convolution ensures that the output at the current moment depends only on the input at the current and past moments, avoiding the leakage of future information and satisfying the causal constraints in real-time early warning scenarios. Dilated convolution introduces a fixed sampling interval, i.e., a dilation rate, between the weights of the convolutional kernel, causing the receptive field of the convolution to expand exponentially along the time axis. Specific parameter settings are as follows: the dilation rate of the temporal convolutional layer in the first spatiotemporal convolutional block is set to 1, the second to 2, the third to 4, and the fourth to 8.

[0068] By using a block-by-block doubling expansion rate design, after stacking four spatiotemporal convolutional blocks, the theoretical temporal receptive field of the top layer of the spatiotemporal graph convolutional network model can cover hundreds of time steps, thereby simultaneously capturing the nonlinear evolution trend of gas concentration, which is characterized by instantaneous changes at the second level and slow accumulation at the minute level.

[0069] Step S304: Each spatiotemporal convolutional block is configured with a residual connection structure at its output.

[0070] The residual connection structure performs an element-wise addition operation on the original input features of the current spatiotemporal convolutional block and the output features after concatenated spatial and temporal convolutions. Let the input be X, and the composite mapping function of spatial and temporal convolutions be F, then the output of the residual connection is... This effectively alleviates the gradient vanishing problem during deep network training, allowing the backpropagation error signal to be directly transmitted to shallow parameters across multiple convolutional layers, ensuring stable updates of model weights and training convergence.

[0071] In step S305, after deep feature extraction is completed in four spatiotemporal convolutional blocks, the system flattens the high-dimensional spatiotemporal feature map output by the last spatiotemporal convolutional block into a one-dimensional feature vector and sends it into a fully connected layer for feature dimensionality reduction; subsequently, the output vector of the fully connected layer is passed to the Softmax classification layer.

[0072] The Softmax classification layer contains four output neurons, corresponding to the normal mode, accumulation and evolution mode, sensor failure mode, and sudden warning mode, respectively. The Softmax function maps the raw scores output by the fully connected layer to four probability values ​​between 0 and 1, and the sum of the four probability values ​​is always equal to 1. The calculation formula is as follows: ;

[0073] In the formula, z i Let be the output score of the fully connected layer corresponding to the i-th type of pattern, and exp be an exponential function with the natural constant e as the base. Given an input enhanced spatiotemporal feature matrix X, this is the probability estimate of the system determining that the current state belongs to the i-th type of pattern.

[0074] Step S306: Perform discriminative logic processing on the probability distribution output by the Softmax classification layer to determine the final gas emission pattern recognition result. When the output probability value of the corresponding accumulation and evolution pattern exceeds a preset confidence threshold of 90%, and simultaneously meets the following spatiotemporal physical consistency conditions, the system officially confirms the pattern as an accumulation and evolution pattern: Condition 1 is that the normalized concentration value of at least one sensor node shows a linear upward trend over the past 60 seconds, and the slope of its linear fit, after being converted to the actual concentration, is greater than a preset slope threshold of 0.05% per minute.

[0075] Condition 2 is a downstream sensor node of the sensor node, such as a return airway sensor node. The maximum cross-correlation coefficient of the concentration sequence of the sensor node and the concentration sequence of the sensor node increase synchronously within the same time period, indicating that there is a following relationship between the two that conforms to the fluid transport law.

[0076] In step S307, when the Softmax classification layer output shows that only a single sensor node in the air intake lane has a drastic numerical fluctuation, and the downstream sensor nodes that have a direct ventilation connection with this sensor node, including the working face head sensor node, the middle sensor node, and the upper corner sensor node, do not detect any correlation response that conforms to the laws of fluid dynamics, the system determines the signal as a sensor failure mode through the built-in topology constraint discrimination mechanism.

[0077] The topology constraint discrimination mechanism uses the sensor node connection relationship defined in the adjacency matrix A constructed in step S205 to check whether the abnormal fluctuation signal propagates downstream along the directed ventilation path. That is, if the abnormal signal only appears in an isolated sensor node and its first-order neighboring sensor nodes do not respond, it is identified as a sensor fault or local electromagnetic interference.

[0078] Through step S3 above, this method inputs the enhanced spatiotemporal feature matrix constructed in step S2 into a spatiotemporal graph convolutional network model containing four spatiotemporal convolutional blocks, residual connections, and a Softmax classification layer. This model uses Chebyshev graph convolution to capture the diffusion and migration features of gas in a multi-order neighborhood in the spatial dimension, and uses causal dilation convolution to capture the concentration evolution trend from the second level to the minute level in the temporal dimension. Finally, it outputs a gas emission pattern recognition result that meets the physical consistency constraint.

[0079] Step S4: Implement intelligent decision-making and early warning.

[0080] In this step, based on the gas outburst pattern recognition results output in step S3, matching handling rules are retrieved from the preset risk response strategy library. The early warning execution mechanism then implements automated closed-loop control of the coal mining machine's operating status, ventilation equipment parameters, and underground power supply, transforming the identified pattern semantic information into specific, executable on-site safety management actions. The specific steps are as follows: Step S401: Maintain a preset risk response strategy library in the non-volatile storage area of ​​the central processing server. The strategy library is organized in the form of structured data tables. Each record contains a pattern type field, a trigger condition field, an instruction priority field, a target early warning execution mechanism field, and a specific control parameter field.

[0081] The strategy library was constructed based on relevant mandatory clauses of the Coal Mine Safety Regulations, the experience and knowledge of mine ventilation safety experts, and the inversion analysis conclusions of historical accident cases at the working face.

[0082] After step S3 completes the pattern recognition and outputs the determined pattern category label, the system uses the pattern category label as the search keyword to perform an exact match query in the strategy library and extract the corresponding complete response strategy record.

[0083] Step S402: Perform branch judgment on the mode category label output in step S3, and enter the emergency warning mode handling process or the accumulation and evolution mode handling process respectively.

[0084] If the mode category label is "Sudden Warning Mode", the system further verifies the state persistence condition of this mode. The specific verification logic is as follows: within a preset duration threshold of no less than 6 consecutive sampling periods, which corresponds to a 500-millisecond sampling period and a total of 3 seconds, at least one sensor node outputs a normalized concentration value that is continuously higher than the preset sudden concentration threshold and shows a monotonically increasing trend. The preset sudden concentration threshold is set to 0.375 of the normalized value corresponding to the sensor's full-scale alarm threshold when the system is deployed, which is equivalent to a reference value of 1.5% of the actual volume fraction.

[0085] If the persistent condition is met, the system will push the Level 1 warning instruction marked with the highest execution priority in the response strategy record to the message queue head of the warning execution mechanism. The Level 1 warning instruction will be executed immediately with priority over all other control instructions.

[0086] In step S403, after the Level 1 warning command is parsed by the warning execution mechanism, the following three parallel control actions are triggered: The first control signal cuts off the power supply circuit of all non-intrinsically safe electrical equipment in the coal face and return airway through the intrinsically safe relay output module, thereby achieving forced power-off isolation and cutting off possible electrical ignition sources.

[0087] The second control signal establishes data exchange with the programmable logic controller of the mine ventilation system through the industrial Ethernet protocol. After receiving the adjustment command, the ventilation controller calculates the required adjustment amount according to the current ventilation network resistance distribution, and drives the return air side damper actuator to change the damper opening angle or sends a frequency increase command to the local ventilation fan inverter to increase the ventilation air volume flowing through the dangerous area to improve the gas dilution ratio.

[0088] The third control signal triggers the preset emergency broadcast program of the underground emergency broadcast system, and the broadcast terminal automatically plays the corresponding emergency evacuation voice prompts for sudden gas anomalies in a loop.

[0089] In step S404, if the branch judgment result in step S402 is the accumulation evolution mode, the system does not perform a forced power-off operation, but instead retrieves the response strategy for the corresponding Level 2 warning command. The Level 2 warning command triggers the following two parallel control actions: The first control signal is sent by the central processing server to the coal mining machine's electrical control system via an industrial Ethernet interface, instructing the modification of the speed setpoint. The instruction message follows the coal mining machine's communication protocol, modifying the target setpoint of the coal mining machine's traction speed to below 50% of the current rated speed. This 50% is a specific value of the preset ratio. Since the total amount of gas emitted from the coal face is approximately linearly positively correlated with the volume of coal broken by the coal mining machine per unit time, reducing the cutting traction speed directly reduces the newly exposed coal face area and the amount of broken coal, thereby suppressing the gas release rate.

[0090] The second control signal is sent by the central processing server to the push interface of the mine information release system. The structured message, which includes the alarm level of the accumulation evolution mode, the time of occurrence, the location of the monitoring area involved, and the suggested handling measures, is sent in real time through the underground wireless communication network to the explosion-proof mobile terminal device held by the on-duty safety monitoring officer. This prompts the officer to conduct intensive manual inspections of key accumulation areas such as the upper corner.

[0091] In step S405, during the execution of a Level 1 or Level 2 warning command, the system continuously receives the real-time gas concentration data stream from step S1 and periodically triggers incremental inference calculations from step S2 to step S3 to dynamically track and evaluate the warning effect.

[0092] If the pattern category label returns to the normal mode in several consecutive identification cycles, and the normalized concentration values ​​of all sensor nodes fall below the preset safety threshold (which is set to 0.1 of the normalized value corresponding to the full-scale alarm threshold and 0.4% of the actual volume fraction during system deployment), the system will send a reset command through the early warning execution mechanism to gradually remove the speed limit of the coal mining machine or allow manual restoration of power supply, thus completing a complete closed-loop management cycle of early warning, intervention, and recovery.

[0093] Through step S4, this method transforms the semantic information of the gas outburst pattern output in step S3 into multi-dimensional and differentiated control commands that act on mining equipment, ventilation facilities, and personnel alarm systems. The sudden warning mode corresponds to a rigid response strategy of forced power outage and emergency ventilation, while the accumulation and evolution mode corresponds to a flexible adjustment strategy of coal mining machine deceleration and manual inspection prompts. Together, they constitute a complete hierarchical response closed-loop control system. This enables the overall technical solution to be implemented as an automated and precise management and control of the safety situation at the coal mine site after completing the entire process from sensor data acquisition and spatiotemporal feature extraction to intelligent pattern recognition. This effectively solves the technical problems of high false alarm rate, delayed response, and inability to differentiate between different risk levels for differentiated handling in traditional static threshold discrimination methods.

[0094] During its long-term operation, this system continuously accumulates historical pattern recognition records and builds a knowledge base based on these records to study the evolution of gas in the working face. At the same time, it uses an incremental learning algorithm to periodically fine-tune the parameters of the spatiotemporal graph convolutional network model in step S3, so that the model's recognition ability evolves synchronously with changes in geological conditions and production environment.

[0095] A structured historical schema database is maintained in a dedicated storage partition of the central processing server, which contains more than 10,000 complete records of gas outburst events.

[0096] Each record serves as a data sample, containing three categories of feature fields. The first category comprises production intensity parameters, including the daily advance speed of the coal face (m / day), the average cutting current of the coal mining machine (A), and the estimated coal drop volume for the shift (tons). The second category consists of ventilation parameters, including the airflow in the intake airway (m³ / min), the negative pressure in the return airway (Pa), and the operating frequency of the local ventilation fan in the upper corner (Hz). The third category is the gas concentration evolution trajectory, which is a complete data snapshot of the N×600 enhanced spatiotemporal feature matrix constructed in step S2 within a 5-minute window before and after the event, as well as the true pattern category label output in step S3 and manually verified. These multi-source heterogeneous data are stored together through a unified time-series index, forming a structured sample set that can be used by subsequent machine learning tasks.

[0097] The system employs an incremental learning algorithm to periodically optimize the spatiotemporal graph convolutional network model in step S3. The specific execution process is as follows: When the number of manually confirmed pattern samples added to the historical pattern database reaches a preset update trigger threshold of 500 new samples, the system initiates an incremental training task in the background computing thread of the central processing server. This task extracts all weight parameters and bias parameters from the currently running online inference model as initial values, calculates the gradient of the loss function with respect to each weight coefficient only for the newly added sample set, and performs a limited number of iterative updates using a small learning rate. The small learning rate is set to one-tenth of the initial training learning rate, typically configured as an initial learning rate of 0.001 corresponding to an incremental learning rate of 0.0001. The loss function used is the cross-entropy loss function, mathematically expressed as follows: , where y i This is the one-hot encoded vector of the real pattern label that has been manually verified. This is the probability estimate of the i-th class pattern output by the Softmax classification layer in step S305. After the update is complete, the new model parameters are atomically replaced in the online inference model file in the model library.

[0098] The aforementioned incremental learning mechanism enables the model to gradually adapt to the slow drift of gas outburst statistics caused by the forward movement of the working face, changes in coal seam occurrence conditions, or adjustments to the ventilation system. This avoids the model's recognition accuracy from decaying due to outdated training data, and enables the system to self-evolve throughout its entire life cycle.

[0099] The system also integrates a coal mining machine position tracking module, which is used to obtain the precise position information of the coal mining machine in the working face spatial coordinate system in real time, and inject this information as an additional spatial feature dimension into the enhanced spatiotemporal feature matrix constructed in step S2, so as to assist the pattern recognition module in step S3 to more finely distinguish the gas source type.

[0100] The coal mining machine position tracking module consists of an infrared positioning transmitter installed on the coal mining machine body and multiple infrared receivers arranged along the hydraulic supports of the working face. Real-time positioning is achieved through line-of-sight infrared light signal transmission. In another optional embodiment, a rotary encoder installed on the end of the coal mining machine's traveling gear shaft can be used to calculate the displacement coordinates of the coal mining machine relative to the end roadway of the working face by accumulating the number of traveling pulses. Regardless of the positioning method used, the coal mining machine position tracking module reports a set of three-dimensional spatial coordinate values ​​to the central processing server at a 1-second interval. These coordinates represent the coal mining machine's position along the length of the working face, its position relative to the cutting depth of the coal wall, and its height from the floor.

[0101] The location coordinates are encoded as a three-dimensional feature vector. In step S206, when generating the enhanced spatiotemporal feature matrix, the system appends this three-dimensional feature vector to the 600-dimensional normalized concentration sequence features of each vertex by feature concatenation, forming an enhanced sensor node feature vector, with the final dimension becoming 603. The spatiotemporal graph convolutional network model in step S3 has synchronously learned the spatiotemporal coupling relationship between the coal mining machine location and the gas concentration distribution during the training phase.

[0102] During the online inference phase, the system further refines the identification of gas sources by analyzing the spatial lag relationship between the current cutting position coordinates of the coal mining machine and the real-time monitored peak gas concentration. If the peak concentration lags behind the cutting position of the coal mining machine by 0 to 5 meters and the lag time is within 30 seconds, it is determined that the current gas outburst mainly originates from the dissipation of newly exposed coal face during the coal mining process. If the peak concentration is located on the goaf side and has no obvious following relationship with the current position of the coal mining machine, that is, the spatial distance between the sensor node corresponding to the peak concentration and the position coordinates of the coal mining machine exceeds 10 meters and the time lag relationship is not significant, it is determined that the gas outburst originates from the dissipation of residual coal in the goaf or the influx of gas from adjacent layers through fractures. Through this discrimination mechanism, the identification basis for the accumulation evolution mode and the sudden warning mode in step S3 not only includes the spatiotemporal variation trend of the concentration value, but also integrates the physical cause information of the gas source, thereby improving the overall accuracy of pattern recognition to over 95%.

[0103] The two extended technical features mentioned above—a knowledge base and incremental learning mechanism based on historical pattern data, and coal mining machine position tracking and gas source identification—enhance the baseline technical solution comprising steps S1 to S4 from two dimensions: continuous model optimization capability and multi-source physical information fusion capability. The former ensures the model's adaptability in long-term operation, avoiding performance degradation due to changes in geological and mining conditions; the latter, by introducing coal mining machine position parameters, makes the pattern recognition process more consistent with the actual physical mechanism of gas outbursts, improving the accuracy and interpretability of distinguishing different types of abnormal events. Together, they provide further technical support for achieving long-term, reliable, and intelligent management of gas risks in coal mining faces.

[0104] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention. Therefore, the embodiments should be regarded as exemplary and non-limiting in all respects.

[0105] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment includes only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for identifying gas concentration patterns and providing intelligent early warning in coal mining faces, characterized in that, include: Through a distributed sensor network, clock synchronization is performed on each sensor node based on a precise time synchronization protocol, gas concentration signals are collected synchronously at fixed periods, and a timestamp is added to each sampling point. Obtain the physical topology map of the ventilation system, calculate the spatial correlation weight based on the ventilation path length between sensor nodes, and perform time-domain correlation analysis on the concentration sequences of sensor node pairs with upstream and downstream causal relationships to calculate the time delay cross-correlation coefficient. The two are then fused to construct an adjacency matrix representing the directed transmission relationship. The location information of the coal mining machine is acquired in real time and encoded into a spatial feature vector. The concentration sequence of all sensor nodes within the preset time window is combined with the adjacency matrix, and the spatial feature vector is appended to the features of each sensor node to construct an enhanced spatiotemporal feature matrix. The enhanced spatiotemporal feature matrix is ​​input into the spatiotemporal graph convolutional network model. Spatial graph convolution and temporal causal convolution are performed sequentially through multiple spatiotemporal convolutional blocks to extract the spatiotemporal evolution features of the gas concentration field. The preliminary pattern classification results are output, which include at least one of the following: gas accumulation evolution mode, sudden warning mode, and sensor failure mode. When the preliminary model classification result is a gas accumulation evolution model, check whether the spatiotemporal physical consistency condition is met, including that the concentration of the sensor node shows an upward trend and the maximum time delay cross-correlation coefficient of the concentration sequence of its downstream sensor node and the concentration sequence of the sensor node increases synchronously. If the condition is met, the model is confirmed. Based on the confirmed pattern, the system retrieves handling rules from the preset risk response strategy library and generates control instructions to automatically control the early warning execution agency in a closed loop.

2. The method for gas concentration pattern recognition and intelligent early warning in coal mining faces according to claim 1, characterized in that, Clock synchronization is performed on each sensor node based on a precise time synchronization protocol. Gas concentration signals are collected synchronously at fixed intervals, and a timestamp is added to each sampling point, including: All sensor nodes are synchronized with the master clock of the central processing server via the mining industrial Ethernet, and perform signal acquisition synchronously under the unified clock tick. Each sensor node converts the acquired raw analog voltage signal into a digital signal sequence, and the clock module of the sensor node adds a high-precision timestamp to each sampled data point in the digital signal sequence to record the absolute moment when the sampling occurred.

3. The method for gas concentration pattern recognition and intelligent early warning in coal mining faces according to claim 1, characterized in that, Temporal correlation analysis was performed on the concentration sequences of sensor node pairs with upstream and downstream causal relationships to calculate the time-delay cross-correlation coefficient, including: Before calculating the time delay cross-correlation coefficient, wavelet decomposition is performed on the concentration sequence of the target sensor node pair to remove high-frequency detail components corresponding to the coal mining machine cutting frequency and extract low-frequency profile signals that reflect the natural migration trend of gas. For the low-frequency profile signal, within a preset time lag range, the time series of one of the sensor nodes is moved step by step with the sampling period as the step size, and the correlation coefficient under each lag step size is calculated. The maximum value of the correlation coefficient sequence is searched. If the maximum value exceeds the preset correlation threshold, it is determined that there is a significant fluid transport causal relationship between the pair of sensor nodes.

4. The method for gas concentration pattern recognition and intelligent early warning in coal mining faces according to claim 1, characterized in that, Spatial graph convolution and temporal causal convolution are performed sequentially through multiple spatiotemporal convolution blocks, including: The spatiotemporal graph convolutional network model contains multiple spatiotemporal convolutional blocks stacked in series. Within each spatiotemporal convolutional block, spatial graph convolution is performed first, followed by temporal convolution. The spatial graph convolution operation is based on the Chebyshev polynomial approximation method of the normalized Laplacian matrix, and the features of each sensor node and its multi-order neighboring sensor nodes are weighted and aggregated. The temporal convolution operation employs a causal dilated convolution structure, and as the number of spatiotemporal convolutional block layers increases, the dilation rate of the causal dilated convolution is multiplied layer by layer to exponentially expand the temporal receptive field and capture the nonlinear evolution trend of gas concentration from the second to the minute level.

5. The method for gas concentration pattern recognition and intelligent early warning in coal mining faces according to claim 1, characterized in that, Based on the confirmed pattern, the system retrieves handling rules from the pre-defined risk response strategy library and generates control instructions, specifically including: When the confirmed mode is the emergency warning mode, the highest priority control command is generated and pushed. The control command is used to trigger the parallel execution of power outage isolation operation, ventilation system air increase adjustment operation and emergency broadcast operation. When the confirmed mode is the gas accumulation and evolution mode, the corresponding level of control command is generated and pushed. The control command is used to trigger the coal mining machine electrical control system to reduce the traction speed to below a preset ratio, and to send a prompt operation containing alarm information to the explosion-proof mobile terminal device.

6. The method for gas concentration pattern recognition and intelligent early warning in coal mining faces according to claim 1, characterized in that, Spatial association weights are calculated based on the ventilation path length between sensor nodes, specifically including: The spatial correlation weight w between any two sensor nodes i and j is calculated using a Gaussian kernel function. ij The calculation formula is: , where d(i,j) is the actual length along the ventilation path between the two sensor nodes, and σ is the spatial attenuation constant; The spatial attenuation constant σ is dynamically optimized through an online adaptive correction mechanism. Specifically, the online adaptive correction mechanism involves: extracting concentration monitoring data from the coal face shutdown and maintenance period in the historical database, using the Gaussian plume model inversion algorithm to fit the measured attenuation characteristic value under the current air volume conditions, and updating the preset σ in a rolling manner.

7. The method for gas concentration pattern recognition and intelligent early warning in coal mining faces according to claim 1, characterized in that, The process of acquiring the coal mining machine's location information in real time and encoding it into a spatial feature vector, and then appending the spatial feature vector to the features of each sensor node to construct an enhanced spatiotemporal feature matrix, specifically includes: The precise three-dimensional position information of the coal mining machine in the working face spatial coordinate system can be obtained in real time by a positioning transmitter installed on the coal mining machine and a receiver deployed along the working face, or by a rotary encoder installed on the traveling part of the coal mining machine. When constructing the enhanced spatiotemporal feature matrix, the three-dimensional precise location information is encoded into a multi-dimensional spatial feature vector, and then appended to the concentration sequence features of each sensor node by feature concatenation to form an enhanced sensor node feature vector, so that the spatiotemporal graph convolutional network model can synchronously learn the spatiotemporal coupling relationship between the coal mining machine position and the gas concentration distribution.

8. The method for gas concentration pattern recognition and intelligent early warning in coal mining faces according to claim 1, characterized in that, The spatiotemporal physical consistency condition also includes: When the preliminary pattern classification results show that only a single sensor node produces drastic numerical fluctuations, and none of the downstream neighboring sensor nodes of that sensor node, as determined by the adjacency matrix, detect a correlation response consistent with the causal relationship of fluid transport, the signal is determined to be a sensor failure mode.

9. The method for gas concentration pattern recognition and intelligent early warning in coal mining faces according to claim 1, characterized in that, Also includes: During system operation, historical pattern recognition records are continuously accumulated to build a structured historical pattern database containing production intensity parameters, ventilation parameters, and gas concentration evolution trajectories. When the cumulative number of newly confirmed pattern samples added to the historical pattern database reaches the preset update trigger condition, an incremental training task is started. Based on the parameters of the currently running spatiotemporal graph convolutional network model as the initial values, iterative updates are performed only on the newly added sample set, and the updated model parameters are replaced in the online inference model file.

10. A coal mining face gas concentration pattern recognition and intelligent early warning system, used to execute the coal mining face gas concentration pattern recognition and intelligent early warning method according to any one of claims 1 to 9, characterized in that, Including a central processing server, and: A distributed sensor network is deployed at key locations in the coal mining face to synchronize the clocks of each sensor node based on a precise time synchronization protocol. It collects gas concentration signals synchronously at a fixed sampling period and adds a timestamp to each sampling data point. The data transmission subsystem is deployed between the distributed sensor network and the central processing server, and is used to upload the gas concentration signals and corresponding timestamps collected by each sensor node to the central processing server with a deterministic delay based on the mining industrial Ethernet. The spatiotemporal feature matrix construction module, deployed in the central processing server, is used to acquire the physical topology map of the ventilation system of the coal mining face, calculate the spatial correlation weight based on the ventilation path length between sensor nodes and calculate the time delay cross-correlation coefficient through temporal correlation analysis, and construct an adjacency matrix by fusing the two. It also acquires the coal mining machine position information in real time and encodes it into a spatial feature vector, combines the concentration sequence within a preset time window with the adjacency matrix, and appends the spatial feature vector to the features of each sensor node to construct an enhanced spatiotemporal feature matrix. The pattern recognition module, deployed in the central processing server, is used to load the trained spatiotemporal graph convolutional network model, perform spatiotemporal convolution operations on the enhanced spatiotemporal feature matrix through the model to extract the spatiotemporal evolution features of the gas concentration field and output preliminary pattern classification results. The preliminary pattern classification results include at least one of the following: gas accumulation evolution mode, sudden warning mode, and sensor failure mode. When the preliminary pattern classification result is a gas accumulation evolution mode, it is verified whether it meets the preset spatiotemporal physical consistency condition to confirm the final mode. The intelligent decision-making and early warning module, deployed in the central processing server, is used to retrieve matching disposal rules from the preset risk response strategy library based on the confirmed pattern classification results, and generate control instructions to automatically control the early warning execution mechanism in a closed loop.

Images