Regional gas disaster early warning system based on deep learning and early warning method thereof

Abstract

The invention relates to a regional gas disaster early warning system based on deep learning and an early warning method thereof, and the system comprises a data collection module, a data preprocessing module, a deep learning model training module, a real-time early warning analysis module, an early warning information issuing module, an early warning result visualization module, and a data iteration optimization module. The data acquisition module comprises a distributed multi-source sensor network and a data acquisition substation; the data acquisition module is deployed to an underground air inlet way, an air return way and a working face; the data preprocessing module, the deep learning model training module and the data iterative optimization module are deployed to a ground GPU server; the real-time early warning analysis module and the early warning release module are deployed to a ground early warning server; and the early warning result visualization module is deployed to a personal terminal. According to the method, through multi-source heterogeneous data fusion and deep learning combination model construction, early recognition and accurate early warning of the gas disaster precursor are realized, and sufficient emergency disposal time is won for coal mine safety production.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of coal mine safety monitoring, and in particular to a regional gas disaster early warning system and method based on deep learning, which is suitable for real-time monitoring and early warning of disasters such as gas explosion, gas outburst and gas overrun in coal mines. BACKGROUND

[0002] Gas disasters have always been a major threat to the safety of coal production. Traditional gas early warning systems mainly rely on single parameter threshold judgment or simple machine learning models, which have the following key technical problems: (1) Weak feature extraction capability: the feature combination of gas concentration, pressure and other parameters needs to be designed manually, it is difficult to capture the complex nonlinear correlation and time evolution law between parameters, the spatiotemporal feature compatibility is poor, and the multi-modal data fusion is not strong; (2) Poor model generalization: the model trained for a specific mine has a significant decrease in early warning accuracy when the geological conditions and mining technology change; (3) Obvious early warning lag: only alarm when the parameters approach the danger threshold, lacking early identification ability of disaster precursors; (4) Low system integration: data acquisition, processing, early warning and feedback are relatively independent, and a closed-loop intelligent early warning system has not been formed.

[0003] With the breakthrough of deep learning technology in the field of time series data processing and feature self-adaptive extraction, its application potential in safety monitoring field gradually appears. However, in the existing technology, due to the need to process real-time data of sensors, the data has strong time series, much noise and may be affected by many factors, so a single model often cannot balance the prediction accuracy and robustness. The application of deep learning model in gas early warning has problems such as insufficient data fusion, biased model training samples, and difficulty in balancing real-time and accuracy, and a complete and reliable regional early warning system and method solution has not been formed. SUMMARY

[0004] The purpose of the present application is to overcome the above-mentioned deficiencies, provide a regional gas disaster early warning system and method based on deep learning, overcome the defects of weak spatiotemporal feature extraction capability, poor generalization and early warning lag of existing gas early warning systems, and provide a regional gas disaster early warning system and method based on deep learning, which is constructed by combining multi-source heterogeneous data fusion and deep learning model, realizes early identification and accurate early warning of gas disaster precursors, and wins sufficient emergency disposal time for coal mine safety production.

[0005] The purpose of the present application is achieved as follows: The application discloses a regional gas disaster early warning system based on deep learning, which comprises a data acquisition module, a data preprocessing module, a deep learning model training module, a real-time early warning analysis module, an early warning information publishing module, an early warning result visualization module and a data iteration optimization module, wherein the data acquisition module comprises a distributed multi-source sensor network and a data acquisition substation; the data acquisition module is arranged in an air inlet lane, an air return lane and a working face underground; the data preprocessing module, the deep learning model training module and the data iteration optimization module are arranged in a ground GPU server; the real-time early warning analysis module and the early warning publishing module are arranged in a ground early warning server; the early warning result visualization module is arranged in a personal terminal; the distributed multi-source sensor network is connected to the data acquisition substation; the data acquisition module is connected to the data preprocessing module and the data iteration optimization module respectively; the data preprocessing module and the data iteration optimization module are connected to the deep learning model training module respectively; the deep learning model training module is connected to the real-time early warning analysis module; the real-time early warning analysis module is connected to the early warning publishing module and the data iteration optimization module respectively; and the early warning publishing module is connected to the early warning result visualization module.

[0006] Further, the distributed multi-source sensor network comprises a gas concentration sensor, a wind speed sensor, a temperature sensor, a flow sensor, a stress sensor and a gas pressure sensor, wherein the gas concentration sensor, the wind speed sensor, the temperature sensor, the flow sensor, the stress sensor and the gas pressure sensor are connected to the data acquisition substation respectively.

[0007] Further, the data acquisition module acquires parameters including gas concentration, gas pressure, gas flow, wind speed, temperature, coal seam stress signals and adjacent information of positions of the sensors; sensor data is transmitted to the data acquisition substation through RS485 shielded twisted pair lines, and the transmission delay is less than or equal to 100 ms.

[0008] Further, the data preprocessing module adopts an isolated forest algorithm to eliminate outliers, adopts an SMOTE-ENN algorithm to process sample imbalance, and adopts a Z-score standardization algorithm to map data to the interval [-1, 1].

[0009] Further, the deep learning model training module constructs a Graph Transformer-LSTM-Attention fusion early warning model, wherein the Graph Transformer-LSTM-Attention fusion early warning model comprises: 2-layer Graph Transformer: 2 graph perception multi-head attention mechanisms, a total of 14 linear changes; 2-layer bidirectional LSTM: 128 hidden layer nodes, a dropout rate of 0.3, and capturing of time evolution rules; Attention layer: scaled dot-product attention mechanism, to strengthen the weight of key features of disaster precursor; Fully connected layer: hidden layer with 64 neurons and 4 classification output layer, output 4 risk levels of no risk, low risk, medium risk and high risk.

[0010] Further, the risk level output by the real-time early warning analysis module is divided into no risk, low risk, medium risk and high risk, and the risk dominant factor is analyzed through the weight of the attention layer. The no risk level: probability <0.2, low risk level: 0.2≤probability<0.4, medium risk level: 0.4≤probability<0.7, high risk level: probability≥0.7.

[0011] Further, the graded early warning of the early warning publishing module includes: no risk for blue early warning state, low risk for yellow early warning message, medium risk for orange sound and light alarm and ventilation adjustment instruction, and high risk for red early warning, emergency evacuation and power-off operation.

[0012] Further, the early warning result visualization module displays the raw data, prediction results, regional early warning state and model iteration in real time; the data iteration optimization module collects new data every week, fine-tunes the model using incremental learning, and automatically replaces the original model when the accuracy of the new model improves by≥3%.

[0013] An early warning method of a regional gas disaster early warning system based on deep learning, based on the above-mentioned regional gas disaster early warning system based on deep learning, including the following contents: Step one, data collection step: collect coal mine underground gas concentration, gas pressure, gas flow, wind speed, temperature, coal seam stress parameters and sensor location information through distributed multi-source sensor network; Step two, data preprocessing step: clean, align and enhance the collected multi-source data, including using the isolation forest algorithm to remove outliers, implementing millisecond-level alignment of different sampling frequency data based on timestamp, intercepting time series segments through sliding window, processing sample imbalance problem through SMOTE-ENN algorithm, and finally mapping the data to the interval [-1, 1] through Z-score standardization; Step three, model training step: build a Graph Transformer-LSTM-Attention fusion early warning model, train the model through historical monitoring data, and output 4 risk levels of no risk, low risk, medium risk and high risk; Step four, real-time early warning analysis step: receive preprocessed real-time data, input the trained fusion early warning model, output risk level every second, and calculate the contribution of each parameter to locate the risk dominant factor; Step 5: Warning Issuance Procedures: Trigger a tiered warning mechanism based on the risk level, including displaying normal status, sending warning text messages, activating audible and visual alarms, pushing out regional production restriction and ventilation adjustment instructions, and triggering emergency evacuation instructions, etc. Step Six: Data Iteration and Optimization: Regularly collect new monitoring data, fine-tune the model using incremental learning, and automatically replace the original model when the accuracy of the new model reaches a set threshold. Step 7: Visualization of Early Warning Results: Real-time display of real-time curves for each parameter, predicted values, regional early warning status, model iteration status, and system linkage information.

[0014] Furthermore, in regional multi-sensor scenarios, step three also includes constructing a multi-source time-series data fusion prediction model, the construction method of which includes: Data acquisition: The historical concentration sequence of each sensor is input in the time dimension; the distance and location relationship between sensors are extracted in the spatial dimension, and the spatial correlation features are represented by an adjacency matrix; Feature extraction: Bidirectional LSTM is used to process time-series data from a single sensor; Graph Transformer is used to model spatial dependencies between sensors; Data prediction: The Graph Transformer multi-head attention mechanism is used to fuse temporal and spatial features and output the concentration prediction of each sensor to perform regional sensor data prediction.

[0015] Compared with the prior art, the beneficial effects of the present invention are: This invention provides a regional gas disaster early warning system and method based on deep learning. The system includes seven modules: data acquisition, preprocessing, model training, real-time early warning analysis, early warning issuance, data iterative optimization, and result visualization. The method includes seven steps: data acquisition, preprocessing, model training, real-time early warning analysis, early warning issuance, data iterative optimization, and result visualization. It achieves deep fusion of multi-source data and simultaneous analysis of spatiotemporal features through a Graph Transformer-LSTM-Attention fusion model, enabling the identification of disaster risks 3-5 minutes before gas parameters reach a dangerous threshold, with an early warning accuracy rate exceeding 92%, providing effective technical support for safe coal mine production. It possesses the following advantages: (1) Strong feature extraction capability: The Graph Transformer-LSTM-Attention fusion model realizes the collaborative extraction of global spatial correlation features and temporal evolution law. The attention mechanism strengthens the key features of disaster precursors. Compared with traditional machine learning models, the feature extraction accuracy is improved by more than 40%, and it can capture early precursor information such as small fluctuations in gas concentration and slow changes in coal seam stress.

[0016] (2) Good early warning capability: Through in-depth analysis of multi-source time series data, disaster risks can be identified 3-5 minutes before gas parameters reach the traditional threshold. The early warning lead time is greatly improved compared with the existing system, providing sufficient time for emergency response.

[0017] (3) Excellent generalization ability: The incremental learning mechanism enables the model to be dynamically adapted to the mine conditions. Under scenarios such as changes in geological conditions and increased mining depth, the early warning accuracy rate remains above 92%, solving the limitation of the traditional model of "one adjustment per mine".

[0018] (4) Accurate risk positioning: By analyzing the dominant risk factors through attention weight analysis, it helps on-site personnel to quickly locate potential hazards (such as abnormal gas pressure or insufficient ventilation), greatly improving the efficiency of handling.

[0019] (5) High system integration: It forms a complete closed-loop system from data acquisition, processing, early warning to feedback, with each module working together to achieve intelligent and automated gas disaster early warning. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the early warning system of the present invention.

[0021] Figure 2 This is a network diagram of the deep learning training model of the present invention.

[0022] Figure 3 This is a sensor deployment diagram of Embodiment 1 of the present invention.

[0023] Figure 4 This is a flowchart illustrating the early warning method of the present invention.

[0024] in: 1. Gas concentration sensor; 2. Wind speed sensor; 3. Temperature sensor; 4. Flow sensor; 5. Stress sensor; 6. Gas pressure sensor; 7. Data acquisition module; 8. Data preprocessing module; 9. Deep learning model training module; 10. Real-time early warning analysis module; 11. Early warning information release module; 12. Early warning result visualization module; 13. Data iteration optimization module. Detailed Implementation

[0025] To better understand the technical solution of the present invention, a detailed description will be provided below in conjunction with relevant illustrations. It should be understood that the specific embodiments described below are not intended to limit the specific implementation of the technical solution of the present invention, but are merely possible implementations of the technical solution of the present invention. It should be noted that the descriptions of the positional relationships of the components herein, such as component A being located above component B, are based on the relative positions of the components in the illustrations and are not intended to limit the actual positional relationships of the components. Example 1

[0026] SeeFigures 1-4 , Figure 1 A schematic diagram of the structure of the deep learning-based regional gas disaster early warning system of Embodiment 1 is shown. As shown in the figure, the deep learning-based regional gas disaster early warning system of Embodiment 1 includes a data acquisition module 7, a data preprocessing module 8, a deep learning model training module 9, a real-time early warning analysis module 10, an early warning information release module 11, an early warning result visualization module 12, and a data iteration optimization module 13. The data acquisition module 7 includes a distributed multi-source sensor network and data acquisition substations. The distributed multi-source sensor network includes a gas concentration sensor 1, a wind speed sensor 2, a temperature sensor 3, a flow sensor 4, a stress sensor 5, and a gas pressure sensor 6.

[0027] The data acquisition module 7 is deployed in the underground intake airway, return airway, and working face; the data preprocessing module 8, deep learning model training module 9, and data iterative optimization module 13 are deployed on the ground GPU server; the real-time early warning analysis module 10 and early warning release module 11 are deployed on the ground early warning server; and the early warning result visualization module 12 is deployed on the personal terminal; the modules are connected in sequence to form a closed-loop early warning system.

[0028] The gas concentration sensor 1, wind speed sensor 2, temperature sensor 3, flow sensor 4, stress sensor 5, and gas pressure sensor 6 are respectively connected to the data acquisition substation. The data acquisition module 7 is respectively connected to the data preprocessing module 8 and the data iteration optimization module 13. The data preprocessing module 8 and the data iteration optimization module 13 are respectively connected to the deep learning model training module 9. The deep learning model training module 9 is connected to the real-time early warning analysis module 10. The real-time early warning analysis module 10 is respectively connected to the early warning release module 11 and the data iteration optimization module 13. The early warning release module 11 is connected to the early warning result visualization module 12.

[0029] The data acquisition module 7 is used to collect gas and environmental parameters in a certain area of ​​the coal mine. The collected parameters include gas concentration, gas pressure, gas flow rate, wind speed, temperature, coal seam stress signal, and proximity information of each sensor location. Sensor data is transmitted to the data acquisition substation via RS485 shielded twisted-pair cable with a transmission delay ≤100ms. Gas concentration is collected by gas concentration sensor 1, which uses a laser absorption sensor with a sampling frequency of 2Hz. Gas pressure is collected by gas pressure sensor 6, which uses a piezoelectric sensor. The gas flow rate is collected by flow sensor 4, which is a vortex flow meter with an accuracy of ±1.5%; the wind speed is collected by wind speed sensor 2, which is a thermal anemometer with a wind speed range of 0-25 m / s; the temperature is collected by temperature sensor 3, which is a thermocouple sensor with a temperature sensing range of -20-150℃; and the coal seam stress is collected by stress sensor 5, which is a fiber optic grating sensor with a range of 0-50 MPa.

[0030] The data preprocessing module 8 cleans, aligns, and enhances the collected multi-source data; it uses the isolated forest algorithm to remove outliers in parameters such as gas concentration and pressure, achieving an anomaly detection accuracy of ≥98%; for samples with small data volumes, it uses One-Class SVM (Single Class Support Vector Machine) for outlier detection; it achieves millisecond-level alignment of data from different sampling frequencies based on timestamps; it uses a sliding window algorithm (window size 10s, step size 2s) to extract time-series segments, and uses the SMOTE-ENN (Synthetic Minority Oversampling and Edit Nearest Neighbor) algorithm to handle sample imbalance, adjusting the sample class ratio to 1:5; finally, it uses Z-score standardization to map the data to the [-1, 1] interval, outputting a standardized time-series dataset; it generates an adjacency matrix based on the location coordinates and adjacency relationships of each sensor, and connects the time-series dataset with the adjacency matrix to form the input feature vector.

[0031] The deep learning model training module 9 constructs a Graph Transformer-LSTM-Attention (Graph Transformer-Long Short-Term Memory Network-Attention Mechanism) fusion early warning model for feature extraction and risk level classification. The Graph Transformer-LSTM-Attention model includes: The system consists of two Graph Transformer layers, including two graph-aware multi-head attention mechanisms (one per layer), totaling 14 linear transformations. Each attention mechanism module contains four linear transformations, two feedforward neural network (FNN) linear transformations, and two input and output projection linear transformations, extracting global spatial correlation features of each parameter. LSTM layer (2-layer bidirectional LSTM, 128 hidden nodes, dropout rate 0.3) to capture temporal evolution patterns; Attention layer (scaled dot product attention mechanism) strengthens the weight of key features of disaster precursors; The fully connected layer (containing a hidden layer with 64 neurons and a 4-classification output layer) outputs four levels of risk: no risk, low risk, medium risk, and high risk. The model uses the Adam optimizer (learning rate 0.001, decay rate 1e-5), cross-entropy loss function, and is trained using historical monitoring data (including three types of samples: normal working conditions, disaster precursors, and minor anomalies, with a total of ≥100,000 samples). The ratio of training set to test set is 7:3. The model achieves an accuracy of ≥95% on the test set and an F1-score of ≥94%.

[0032] The real-time early warning analysis module 10 receives preprocessed real-time data and inputs it into the trained GraphTransformer-LSTM-Attention model, outputting the risk level once per second. It has built-in risk level determination rules: no risk (output probability < 0.2), low risk (0.2 ≤ probability < 0.4), medium risk (0.4 ≤ probability < 0.7), and high risk (probability ≥ 0.7). At the same time, it calculates the contribution of each parameter and locates the dominant risk factor through attention layer weight analysis.

[0033] The early warning information release module 11 triggers a graded early warning mechanism based on the risk level. When there is no risk, the area displays a blue normal status. When the risk is low, a yellow early warning SMS and system pop-up are sent to the coal mine monitoring center, prompting for strengthened inspection of key areas. When the risk is medium, the underground audible and visual alarm (orange light plus intermittent buzzer) is activated, and regional production restriction and ventilation adjustment instructions are pushed to the working face controller. When the risk is high, a red early warning is issued, triggering the emergency evacuation instruction of the underground broadcast system, automatically cutting off the power supply and transportation equipment in the risk area, and simultaneously transmitting the early warning information to safety supervision platforms at all levels.

[0034] The early warning result visualization module 12 is mainly used to display the real-time curves of various parameters, the magnitude of the predicted value, the regional early warning status, and the model iteration status in real time. At the same time, it displays the linkage information between the system and equipment such as underground coal mining machines, tunneling machines, and ventilation fans. When a risk warning occurs, it can display the instructions sent by the early warning release module, the location information of relevant technical personnel, and the progress of the implementation of relevant rectification measures in real time.

[0035] The data iteration optimization module 13 constructs a model iteration update mechanism, automatically collecting new monitoring data (including verification data after early warning) every week. After preprocessing, the data is added to the training set, and the Graph Transformer-LSTM-Attention model is fine-tuned using incremental learning. Each fine-tuning iteration is performed 5-10 times, and the model performance is verified through A / B testing. When the accuracy of the new model is ≥3% higher than that of the original model, the real-time early warning model is automatically replaced to ensure that the system adapts to changes in mine working conditions.

[0036] This embodiment 1 relates to an early warning method for a regional gas disaster early warning system based on deep learning, which includes the following steps: Step 1: Data Acquisition Steps: Collect data on underground coal mine gas concentration, gas pressure, gas flow rate, wind speed, temperature, coal seam stress parameters, and the location information of each sensor through a distributed multi-source sensor network; Step 2, Data Preprocessing: The collected multi-source data is cleaned, aligned and enhanced, including using the Isolation Forest algorithm to remove outliers, achieving millisecond-level alignment of data with different sampling frequencies based on timestamps, extracting time-series segments through a sliding window, using the SMOTE-ENN algorithm to handle the sample imbalance problem, and finally mapping the data to the [-1,1] interval through Z-score standardization. Step 3: Model Training Steps: Construct a Graph Transformer-LSTM-Attention fusion early warning model, train the model using historical monitoring data, and output four risk levels: no risk, low risk, medium risk, and high risk. Step 4, Real-time Early Warning Analysis: Receive preprocessed real-time data, input it into the trained fusion early warning model, output the risk level once per second, calculate the contribution of each parameter, and locate the dominant risk factors. Step 5: Warning Issuance Procedures: Trigger a tiered warning mechanism based on the risk level, including displaying normal status, sending warning text messages, activating audible and visual alarms, pushing out regional production restriction and ventilation adjustment instructions, and triggering emergency evacuation instructions, etc. Step Six: Data Iteration and Optimization: Regularly collect new monitoring data, fine-tune the model using incremental learning, and automatically replace the original model when the accuracy of the new model reaches a set threshold. Step 7: Visualization of Early Warning Results: Real-time display of real-time curves for each parameter, predicted values, regional early warning status, model iteration status, and system linkage information.

[0037] In regional multi-sensor scenarios, step three also includes constructing a multi-source time-series data fusion prediction model, specifically including: (1) Data acquisition: Time dimension: Input the historical concentration sequence for each sensor; Spatial dimension: Extract the distance and positional relationships between sensors, and represent spatial association features using an adjacency matrix; (2) Feature extraction: Timing characteristics: Bidirectional LSTM is used to process the timing data of a single sensor; Spatial characteristics: Spatial dependencies between sensors are modeled using Graph Transformer; (3) Data prediction: The Graph Transformer multi-head attention mechanism is used to fuse temporal and spatial features, and simultaneously capture the "historical changes of a certain sensor" and the "real-time influence of surrounding sensors", outputting the concentration prediction of each sensor to perform regional sensor data prediction.

[0038] Specific implementation examples: The following uses a coal mine with an annual output of 12 million tons as an application scenario to explain in detail the implementation process of this system: 1. Sensors are deployed in key areas such as the No. 5 coal seam longwall face, intake airway, and return airway of the mine. The deployment of gas concentration sensors, wind speed sensors, and temperature sensors is detailed in the appendix. Figure 3 As shown, there are a total of 17 sensors: one gas pressure sensor at each end and the middle of the working face (3 in total); two gas flow sensors deployed 72 meters from the working face in the intake and return airways; and one coal seam stress sensor deployed every 60 meters along the coal seam (6 in total). All sensors are connected to the data acquisition substation via RS485 shielded twisted-pair cables and connected to the mine's industrial ring network. The real-time sensor data and the adjacency matrix data obtained from the spatial adjacency information of the sensors are fused and transmitted to the ground monitoring center server. The server is equipped with data preprocessing and deep learning model training modules, and the remaining modules are deployed to personal terminals as needed by the mine.

[0039] 2. Data Preprocessing Module: The preprocessing module in the server processes the collected data. It uses the Isolation Forest algorithm to remove abnormal values ​​of gas concentration jumps caused by sensor loosening (approximately 200 abnormal data points are processed per day). It aligns the gas concentration (2Hz) and coal seam stress (0.5Hz) data using timestamps. It uses the SMOTE-ENN algorithm to expand the historical disaster precursor samples from 500 to 2500, and combines them with normal operating condition samples (80,000) to form a training set. After Z-score standardization, it outputs the model input feature vector data.

[0040] 3. Deep learning model module training: The GraphTransformer-LSTM-Attention model was trained on a GPU server (NVIDIA 4070) for 8 hours. The training set accuracy reached 96.2%, the test set accuracy reached 95.1%, and the F1-score reached 94.5%. The model was saved in a lightweight ONNX format (approximately 80MB in size) and deployed to a real-time alert server.

[0041] 4. The real-time early warning analysis module receives a pre-processed data stream every second and inputs the trained and deployed model to calculate the risk level. At 14:32 on a certain day, the model outputs a risk level of medium risk (probability 0.58), and the attention weight shows that the contribution of gas outburst is 65%, which is judged to be caused by insufficient local ventilation. The system immediately issues an orange warning, pushes the ventilation fan speed-up command to the downhole controller, and notifies the inspection personnel to conduct on-site verification.

[0042] 5. The data iteration and optimization module automatically collects the week's monitoring data (approximately 140,000 records) every Sunday at 2:00 AM. After preprocessing, the data is added to the training set, and the model is incrementally fine-tuned (8 iterations). After one fine-tuning, the accuracy of the model on the test set is improved to 96.8%, and the original early warning model is automatically replaced, continuously optimizing system performance.

[0043] 6. Visualize the early warning results, displaying real-time curves of data from each sensor, predicted values, regional early warning results, model iteration counts, etc., while also displaying the linkage information between the system and underground coal mining machines, tunneling machines, ventilation fans, and other equipment. When a risk warning occurs, the instructions sent by the early warning release module and the location information of relevant technical personnel, as well as the progress of relevant rectification measures, can be displayed in real time.

[0044] Through the application of this embodiment, the false alarm rate of gas disaster warning in the mine was reduced from 8% in the traditional system to 2.3%, and the average warning lead time reached 4.2 minutes, effectively avoiding three potential gas outburst risks.

[0045] The above are merely specific application examples of the present invention and do not constitute any limitation on the scope of protection of the present invention. All technical solutions formed by equivalent transformations or substitutions fall within the scope of protection of the present invention.

Claims

1. A deep learning-based regional gas disaster early warning system, characterized in that: The application relates to a deep learning-based real-time mine gas disaster early warning system, which comprises a data acquisition module (7), a data preprocessing module (8), a deep learning model training module (9), a real-time early warning analysis module (10), an early warning information publishing module (11), an early warning result visualization module (12) and a data iteration optimization module (13), the data acquisition module (7) comprises a distributed multi-source sensor network and a data acquisition substation; the data acquisition module (7) is arranged in an air inlet lane, an air return lane and a working face underground; the data preprocessing module (8), the deep learning model training module (9) and the data iteration optimization module (13) are arranged in a ground GPU server; the real-time early warning analysis module (10) and the early warning publishing module (11) are arranged in a ground early warning server; the early warning result visualization module (12) is arranged in a personal terminal; the distributed multi-source sensor network is connected to the data acquisition substation, the data acquisition module (7) is connected to the data preprocessing module (8) and the data iteration optimization module (13) respectively, the data preprocessing module (8) and the data iteration optimization module (13) are connected to the deep learning model training module (9) respectively, the deep learning model training module (9) is connected to the real-time early warning analysis module (10), the real-time early warning analysis module (10) is connected to the early warning publishing module (11) and the data iteration optimization module (13) respectively, and the early warning publishing module (11) is connected to the early warning result visualization module (12).

2. The regional gas disaster early warning system based on deep learning according to claim 1, characterized in that: The distributed multi-source sensor network comprises a gas concentration sensor (1), a wind speed sensor (2), a temperature sensor (3), a flow sensor (4), a stress sensor (5) and a gas pressure sensor (6), and the gas concentration sensor (1), the wind speed sensor (2), the temperature sensor (3), the flow sensor (4), the stress sensor (5) and the gas pressure sensor (6) are connected to the data acquisition substation respectively. 3.The regional gas disaster early warning system based on deep learning according to claim 1, characterized in that: The parameters collected by the data acquisition module (7) include gas concentration, gas pressure, gas flow, wind speed, temperature, coal seam stress signals and adjacent information of the positions of the sensors; sensor data is transmitted to the data acquisition substation through RS485 shielded twisted pair lines, and the transmission delay is less than or equal to 100 ms.

4. The regional gas disaster early warning system based on deep learning according to claim 1, characterized in that: The data preprocessing module (8) adopts an isolated forest algorithm to eliminate abnormal values, adopts an SMOTE-ENN algorithm to process sample imbalance, and adopts Z-score standardization to map data to the [-1, 1] interval.

5. The regional gas disaster early warning system based on deep learning according to claim 1, characterized in that: The deep learning model training module (9) constructs a Graph Transformer-LSTM-Attention fusion early warning model, and the Graph Transformer-LSTM-Attention fusion early warning model comprises: 2-layer Graph Transformer: two graph perception multi-head attention mechanisms, a total of 14 linear changes; 2-layer bidirectional LSTM: the number of hidden layer nodes is 128, the dropout rate is 0.3, and the time sequence evolution law is captured; attention layer: scaling dot product attention mechanism, which strengthens the weight of disaster precursor key features; Full connection layer: hidden layer containing 64 neurons and 4 classification output layer, output 4 risk levels of no risk, low risk, medium risk and high risk. 6.The regional gas disaster early warning system based on deep learning according to claim 1, characterized in that: The risk level output by the real-time early warning analysis module (10) is divided into no risk, low risk, medium risk and high risk, and the risk dominant factor is analyzed through attention layer weight, no risk level: probability <0.2, low risk level: 0.2≤probability <0.4, medium risk level: 0.4≤probability <0.7, high risk level: probability ≥0.

7. 7.The regional gas disaster early warning system based on deep learning according to claim 1, characterized in that: The hierarchical early warning of the early warning release module (11) includes: no risk is blue early warning state, low risk sends yellow early warning message, medium risk starts orange sound and light alarm and ventilation adjustment instruction, high risk triggers red early warning, emergency evacuation and power-off operation. 8.The regional gas disaster early warning system based on deep learning according to claim 1, characterized in that: The early warning result visualization module (12) displays the original data, prediction result, regional early warning state and model iteration in real time; the data iteration optimization module (13) collects new data every week, fine tunes the model using incremental learning, and automatically replaces the original model when the accuracy of the new model improves by ≥3%.

9. A warning method of a deep learning-based regional gas disaster warning system, based on the deep learning-based regional gas disaster warning system according to any one of claims 1-8, characterized in that, It includes the following contents: Step one, data collection step: collect coal mine underground gas concentration, gas pressure, gas flow, wind speed, temperature, coal seam stress parameters and sensor position information through distributed multi-source sensor network; Step two, data preprocessing step: clean, align and enhance the collected multi-source data, including using isolated forest algorithm to remove outliers, implementing millisecond-level alignment of different sampling frequency data based on timestamp, intercepting time series segments through sliding window, processing sample imbalance problem through SMOTE-ENN algorithm, and finally mapping data to [-1, 1] interval through Z-score standardization; Step three, model training step: build a Graph Transformer-LSTM-Attention fusion early warning model, train the model through historical monitoring data, and output 4 risk levels of no risk, low risk, medium risk and high risk; Step four, real-time early warning analysis step: receive real-time data after preprocessing, input the trained fusion early warning model, output risk level every second, and calculate the contribution of each parameter to locate the risk dominant factor; Step five, early warning release step: trigger hierarchical early warning mechanism according to risk level, including displaying normal state, sending early warning message, starting sound and light alarm, pushing regional production limiting and ventilation adjustment instruction, triggering emergency evacuation instruction, etc.; Step six, data iteration optimization step: periodically collect new monitoring data, fine tune the model using incremental learning, and automatically replace the original model when the accuracy of the new model improves to the set threshold; Step seven, early warning result visualization step: real-time display of each parameter real-time curve, prediction value size, regional early warning condition, model iteration state and system linkage information.

10. The warning method of the regional gas disaster early warning system based on deep learning according to claim 9, characterized in that: In the regional multi-sensor scene, step three also includes building a multi-source time series data fusion prediction model, and its construction method includes: Data acquisition: the time dimension input the historical concentration sequence of each sensor; the space dimension extracts the distance and position relationship between sensors, and represents the space correlation characteristics with an adjacency matrix; Feature extraction: use bidirectional LSTM to process the time series data of a single sensor; use Graph Transformer to model the spatial dependence between sensors; Data prediction: use the multi-head attention mechanism of Graph Transformer to fuse time and space features, output the concentration prediction of each sensor, and perform regional sensor data prediction.

Images