Mine ventilation intelligent prediction method based on cloud computing platform

Abstract

The invention relates to the technical field of mine ventilation, and provides a mine ventilation intelligent prediction method based on a cloud computing platform, which is characterized in that in data acquisition and processing, an innovative sensor deployment and subnet cooperation mechanism is adopted, and data continuity and reliability are guaranteed; the data transmission network adopts the design of double-base-station hot standby, optical fiber double-ring network and the like, the stability is far better than that of a traditional network, and data loss is basically eradicated; in the aspect of prediction model, a unique CFD physical model and lightweight neural network hybrid architecture can accurately identify working conditions and switch prediction modes, the prediction precision is improved compared with that of a traditional single model, and the advantages are obvious under complex working conditions; the safety and fault processing mechanism is complete, data safety is ensured through data encryption, authority grading, remote disaster recovery and the like, rapid switching compensation can be achieved during faults, multiple regulation and control modes are supported, and different scene requirements are met.

Description

Technical Field

[0001] The present invention relates to the technical field of mine ventilation, and in particular to an intelligent prediction method for mine ventilation based on a cloud computing platform. Background Art

[0002] In the coal mining industry, the stable operation of the mine ventilation system is crucial, as it is directly related to the safety of underground workers and the smooth progress of production. However, existing mine ventilation prediction technologies have many difficult-to-overcome problems.

[0003] On the one hand, the prediction accuracy is difficult to meet actual needs. Traditional prediction models are mostly single models that cannot take into account the complex changes of the ventilation system under different working conditions. Under normal operating conditions, the response to slight fluctuations in ventilation parameters is slow; under abnormal operating conditions, such as abnormal gas concentration or sudden changes in wind speed, the prediction deviation is extremely large, resulting in the inability to issue early warnings and take effective measures in a timely and accurate manner, posing a serious safety hazard. On the other hand, the system reliability and stability are poor. The underground environment is harsh, and sensors are susceptible to electromagnetic interference, mechanical vibration, etc., resulting in data loss and frequent errors. In addition, the existing data transmission network has weak anti-interference capabilities, and network interruptions occur frequently. Once a failure occurs, data loss is serious, making it difficult to ensure the continued stable operation of the ventilation prediction system.

[0004] At the same time, energy efficiency and equipment management are low. Existing control strategies focus solely on ventilation safety, ignoring energy consumption and equipment lifespan. This results in high ventilation system energy consumption, increased equipment wear, and increased operating costs.

[0005] Therefore, an intelligent prediction method for mine ventilation based on a cloud computing platform is proposed to solve the above technical difficulties. Through innovative data collection, transmission, processing, prediction and control mechanisms, it can improve prediction accuracy, enhance system reliability, improve energy utilization efficiency and optimize equipment management, ensure the safe, efficient and stable operation of the mine ventilation system, and meet the urgent needs of the intelligent development of coal mines. Summary of the Invention

[0006] Technical problems solved

[0007] In view of the shortcomings of the existing technology, the present invention provides an intelligent prediction method for mine ventilation based on a cloud computing platform.

[0008] Technical Solution

[0009] To achieve the above-mentioned solution, the present invention provides the following technical solution: a method for intelligent prediction of mine ventilation based on a cloud computing platform, comprising the following steps:

[0010] S1 deploys mining intrinsically safe multi-parameter sensors every 50 meters at the mine excavation working face and 30 cm away from the roof in the center of the top of the mining working face.

[0011] S2 uses LoRa wireless modules to form a Mesh network, combined with the underground industrial ring network, to transmit the collected data to the ground cloud computing center.

[0012] S3 fuses measured data and simulation data to construct a hybrid data set, achieves time synchronization through the NTPv4 protocol, establishes a coordinate system conversion model based on the mine CAD drawings to achieve spatial alignment, and uses the weighted average method to fuse data.

[0013] S4 removes gross errors and repairs missing data from the original data. When removing gross errors, if there are more than 5 consecutive gross errors, or if 3 or more sensors in the same subnet trigger gross errors, a fault warning will be triggered and the data of the adjacent subnet will be interpolated and repaired.

[0014] S5 builds a hybrid prediction architecture, identifies abnormal operating conditions by ensuring that the gas concentration fluctuation rate is not less than 15% or the wind speed mutation rate is not less than 20%, and dynamically adjusts the model fusion weight: under normal operating conditions, when the weight α is not less than 0.8, neural network prediction is given priority; under abnormal conditions, when the weight α is not higher than 0.3, CFD fast simulation is initiated.

[0015] S6 implements data security protection and fault recovery based on a cloud computing platform. Data transmission uses AES-256-GCM encryption and dynamic key updates, and storage sets three levels of permissions based on the RBAC model. When a network interruption exceeds 2 hours, a local lightweight prediction model on the edge node is enabled.

[0016] S7 establishes a three-level early warning mechanism and multi-objective optimization control, triggering early warnings and implementing corresponding measures based on gas concentration.

[0017] Preferably, the multi-parameter sensor is KJ90X-F8 type, with explosion-proof grade ExibIMb and MA certification. The vibration sensor CA-YD-185 is pasted on the vertical surface of the bearing seat, and the damper status sensor GUD10 is installed at the damper hinge. The installation error does not exceed 2mm. Every 5 adjacent sensors form a monitoring subnet. When a single sensor fails, the subnet data is weighted fused, and the weight is calculated inversely according to the distance. Vibration sensors and damper status sensors are installed on the ventilator and damper respectively. A total of 10 parameters including wind speed, wind pressure, gas concentration, temperature, humidity, personnel position, equipment position, ventilator vibration, damper status, and dust concentration are collected.

[0018] Preferably, the LoRa network supports dual-base station hot standby and automatic routing reconstruction, and the industrial ring network adopts a 10G fiber-optic dual-ring network; the LoRaMesh network single base station covers a radius of 2km, the data upload delay does not exceed 500ms, and the penetration capability meets the transmission requirements of 5-layer tunnel walls; the industrial ring network supports IEEE802.1X authentication and link aggregation, and automatically switches to redundant links in the event of a single point failure.

[0019] Preferably, the deletion repair adopts a bidirectional LSTM neural network combined with a dynamic time window The historical data sequence is extracted, where T is the data collection time. The number of hidden layer neurons in the bidirectional LSTM neural network is 64. After 5-fold cross-validation optimization, the validation indicators are RMSE not exceeding 0.08 and RAE not exceeding 2.5%. The model training adopts a batch size of 32 and a learning rate of 0.001, combined with the early stopping method to avoid overfitting.

[0020] Preferably, the gas concentration fluctuation rate ΔC in the working condition identification is |C t -C t-10 | / C t-10 ×100%, the time window is 10 minutes, the wind speed mutation rate Δv=|v t -v t-5 | / v t-5 × 100%, the time window is 5 minutes, and the threshold is based on the coal mine safety regulations and historical accident data.

[0021] Preferably, the CFD fast simulation grid division accuracy is 0.2m, the computing resources for a single working condition are no less than 8 cores, and when multiple abnormal working conditions occur concurrently, scheduling is based on priority. Red warning scenarios are preferentially allocated computing resources of no less than 16 cores to ensure that the response time does not exceed 2 minutes.

[0022] Preferably, the fan speed adjustment range in the multi-objective optimization is limited to 40% to 90% of the rated speed, and is linked to the damper opening. For every 10% increase in speed, the corresponding damper opening is not less than 70%. It has been measured that the ventilation efficiency is increased by not less than 20% and the energy consumption is reduced by not less than 15%.

[0023] Preferably, the key in the data security protection is generated by a hardware security module and updated every 15 minutes through the Diffie-Hellman protocol. The core data is backed up to an off-site disaster recovery center in real time. The recovery time target does not exceed 20 minutes, and the recovery point target does not exceed 2 minutes.

[0024] Preferably, the cache capacity of the edge computing node is not less than 2 hours, and a local lightweight prediction model is enabled when the network is interrupted. The model is based on a simplified LSTM trained on historical 7 days of data, and the prediction delay does not exceed 10ms. After the network is restored, the data is automatically synchronized and the parameters are calibrated.

[0025] Preferably, the corresponding measures use the NSGA-II algorithm to solve the Pareto optimal solution of maximizing ventilation efficiency, minimizing energy consumption, and balancing equipment life, and output fan speed and damper opening control parameters.

[0026] Beneficial effects

[0027] Compared with the existing technology, the present invention provides an intelligent prediction method for mine ventilation based on a cloud computing platform, which has the following beneficial effects:

[0028] 1. This intelligent prediction method for mine ventilation based on a cloud computing platform uses innovative sensor deployment and subnet coordination mechanisms in data collection and processing to ensure continuous and reliable data. The data transmission network adopts dual-base station hot standby and optical fiber dual-ring network designs, which are far more stable than traditional networks and basically eliminate data loss. In terms of prediction models, the unique "CFD physical model + lightweight neural network" hybrid architecture can accurately identify working conditions and switch prediction modes. The prediction accuracy is improved compared to traditional single models, and it has obvious advantages under complex working conditions. The security and fault handling mechanisms are complete, and data encryption, permission classification, and off-site disaster recovery ensure data security. In the event of a fault, compensation can be quickly switched and multiple control modes are supported to meet the needs of different scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 Schematic diagram of the steps of the present invention. DETAILED DESCRIPTION

[0030] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0031] See also Figure 1 The present invention proposes an intelligent prediction method for mine ventilation based on a cloud computing platform, comprising the following steps:

[0032] S1 Data Collection

[0033] 1.1 Sensor Deployment Planning

[0034] In mining operations, the stable operation of the ventilation system is crucial to personnel safety and production efficiency, and accurate data collection is the foundation for intelligent ventilation prediction. KJ90X-F8 intrinsically safe multi-parameter sensors are deployed every 50 meters at the mine's excavation face, every 30 meters at the stope face, and at the center of the tunnel top (30 cm from the roof). These sensors strictly adhere to mining equipment safety standards and accurately measure 10 parameters: wind speed, wind pressure, gas concentration, temperature, humidity, personnel and equipment location, fan vibration, damper status, and dust concentration. These sensors provide data support for comprehensive ventilation system monitoring. For example, during the mining process at a large coal mine, the sensor detected abnormal fluctuations in gas concentration at the stope face in real time, providing a key basis for subsequent ventilation control.

[0035] For key equipment such as ventilators and dampers that play a decisive role in the operation of the ventilation system, CA-YD-185 vibration sensors and GUD10 damper status sensors are installed. The CA-YD-185 vibration sensor is attached to the vertical surface of the bearing seat using a high-temperature resistant coupling agent. With a sensitivity of up to 100mV / g, it can keenly detect subtle vibration changes during ventilator operation. By analyzing vibration data, potential ventilator faults such as bearing wear and blade imbalance can be detected in advance. The GUD10 damper status sensor is installed at the hinge between the damper door frame and the door leaf (error ≤ 2mm). With a switching signal response time of ≤ 20ms, it can quickly provide feedback on the opening and closing status of the damper, ensuring accurate and correct airflow control in the ventilation system. In actual applications, there have been cases where the damper failed to close completely due to mechanical failure. Thanks to this sensor, the abnormality was detected in time, preventing ventilation short-circuit accidents.

[0036] For multi-sensor collaboration scenarios, every five adjacent sensors form a monitoring subnet. When a single sensor fails, the data of other sensors in the subnet are preferentially called for weighted fusion (the weight is calculated inversely proportional to the distance between the sensors) to ensure data continuity.

[0037] 1.2 Data Transmission Network Construction

[0038] The data transmission network is like the "nerve network" of the entire system. Its stability and transmission efficiency directly affect the real-time and accuracy of ventilation predictions. LoRa wireless modules are used to form a Mesh network as the underlying architecture for data transmission. A dual-base station hot standby mode is adopted (when the primary base station fails, the backup base station takes over communication within 500ms). The routing protocol supports automatic reconstruction (link quality is scanned every 2 minutes, and switching is triggered when the delay increases by ≥30%). The LoRa wireless module uses the SX1278 chip, which operates in the 470MHz frequency band. It has the characteristics of low power consumption and strong penetration. The coverage radius of a single base station can reach 2km, and the data upload delay is ≤500ms. The penetration capability meets the transmission requirements of 5-layer tunnel walls. This ensures stable data transmission even in the tortuous mine tunnels and complex environments. The underground industrial ring network uses Huawei S5720-32X-EI equipment, equipped with a 10G fiber dual ring network, supports IEEE802.1X authentication and link aggregation (IEEE802.3ad), and has a bandwidth of 1Gbps. In the event of a single point of failure, the bandwidth automatically switches to a redundant link, further ensuring the stable transmission of data from the underground to the ground cloud computing center, laying a solid foundation for subsequent data processing and analysis.

[0039] 1.3 Construction of hybrid training dataset

[0040] To improve the accuracy and generalization of the prediction model, a high-quality training dataset is required. Measured and simulated data are integrated through spatiotemporal registration. For time synchronization, the NTPv4 protocol is used to precisely synchronize measured and simulated data with an accuracy of ±1ms. The data is uniformly marked with a UTC timestamp in the format of YYYY-MM-DDHH:MM:SS.SSS. For data with timestamp deviations greater than 10ms, linear interpolation is used to correct the continuity of the time series to ensure data consistency across the temporal dimension.

[0041] Spatial mapping establishes a coordinate system conversion model based on the mine CAD drawings: by accurately calibrating the turning points of the tunnel (including intersections and slope change points), the rotation matrix and translation vector are determined to ensure that the conversion error is ≤0.3m, achieving accurate matching of the simulation data and the measured data in spatial position. Data fusion uses the weighted average method, and the formula is:

[0042] in, (the inverse of the measured data variance, reflecting the credibility of the data), w 仿真= 0.3 (fixed base weight to avoid overfitting of simulated data). This method fully leverages the authenticity of measured data and the theoretical advantages of simulated data, making it suitable for the fusion processing of continuous variables such as wind speed and wind pressure. Outliers are processed using a 5×5 spatiotemporal neighborhood median filter. When sensor data is missing, the median of three surrounding sensors within 50 meters is used to fill in the gaps, effectively improving data integrity and reliability.

[0043] Dynamic time window is obtained by formula (T is the data collection time, unit: minutes). For example, when the data collection time is 60 minutes, w=11, which means it covers 11 minutes of historical data. It has been verified through actual measurements that this window captures the temporal correlation of wind speed and gas concentration optimally (correlation coefficient ≥ 0.92), and can adaptively adjust the window size according to the data collection time to adapt to different data fluctuation scenarios.

[0044] Taking a typical mine as an example, after a year of actual data collection, 200,000 data items were obtained; by simulating 8 different working conditions, 150,000 simulation data items were generated, and finally a mixed data set containing 300,000 data items was formed. The training set and test set were scientifically divided in an 8:2 ratio, providing sufficient and high-quality data resources for subsequent model training.

[0045] S2 Data Processing

[0046] 2.1 Gross Error Removal

[0047] The raw data inevitably contains gross errors due to factors such as sensor failure and electromagnetic interference. If left unaddressed, these errors will severely impact subsequent analysis and prediction results. During the gross error removal phase, gross errors are identified by calculating the mean and standard deviation of single-sensor data. A rolling window size of 1000 is updated in real time, and data points that meet these criteria are identified as gross errors and removed. Missing values ​​resulting from gross error removal are filled using linear interpolation of the five preceding and following valid values.

[0048] If there are more than 5 consecutive gross errors or ≥3 sensors in the same subnet trigger gross errors at the same time, the sensor fault warning will be triggered and the subnet will be marked as a "suspicious area". The data of the adjacent subnet (within 50 meters) will be automatically retrieved for three-dimensional spatial interpolation repair to avoid data chain breaks caused by failure of a single sensor, and maintenance personnel will be notified to carry out inspections.

[0049] 2.2 Missing Data Interpolation Repair

[0050] Due to the complex and ever-changing underground environment, sensor data may be missing, necessitating effective data repair. A bidirectional LSTM neural network is used for interpolation repair of missing data. This neural network can simultaneously learn both forward and backward information from the data. For mine ventilation data with its time series characteristics, it can more accurately capture data features, achieving high-precision missing value repair.

[0051] The network structure consists of an input layer (dimension 1), a bidirectional LSTM layer (with 64 hidden neurons, optimal through 5-fold cross-validation using root mean square error (RMSE) and repair accuracy (RAE). When the number of neurons is 64, RMSE ≤ 0.08 and RAE ≤ 2.5%, respectively, are superior to configurations with 48, 80, and 96 neurons, reducing errors by 12%, 8%, and 5%, respectively, achieving a good balance between model complexity and generalization ability), and a fully connected layer (dimension 1). The dynamic time window is determined using the above formula.

[0052] The data processing steps are as follows: First, normalization is performed using the formula to map the data to the interval [0, 1] to eliminate the dimensionality differences between different data dimensions and improve the training efficiency and accuracy of the model; then, with the missing point as the center, w time points before and after are extracted as the input sequence; the model training adopts a batch size of 32, 50 iterations, a learning rate of 0.001, and combines the early stopping method (termination if the validation set loss does not decrease for 10 consecutive rounds) to avoid model overfitting and ensure that the model has good generalization ability; finally, by performing denormalization, the normalized values ​​output by the model are restored to the actual data to complete the missing value repair, and the repair error is ≤3%, effectively ensuring the integrity and accuracy of the data.

[0053] S3 prediction model implementation

[0054] 3.1 Hybrid Prediction Architecture Design

[0055] To address the complex variations in mine ventilation systems under varying operating conditions, a hybrid prediction architecture combining a "CFD physical model + lightweight neural network" was developed. This architecture combines the accuracy of physical models with the flexibility of neural networks, enabling efficient and accurate ventilation predictions under varying operating conditions.

[0056] Operating condition identification is a key link in the hybrid prediction architecture, which determines the operating status of the mine ventilation system based on the gas concentration fluctuation rate and wind speed mutation rate.

[0057] Gas concentration fluctuation rate ΔC=|C t -C t-10 | / C t-10 ×100% (time window 10 minutes), wind speed mutation rate Δv=|v t -vt-5 | / v t-5 When ΔC ≥ 15% or Δv ≥ 20% (thresholds based on coal mine safety regulations and historical accident data), an abnormal operating condition is identified. For example, during mining operations in a certain mine, ventilation resistance changes due to advancing working faces, and the wind speed mutation rate exceeds the threshold. The system promptly identifies the abnormal condition and switches to the prediction mode, providing strong support for ensuring ventilation safety.

[0058] 3.2 Model Fusion and Parameter Adjustment

[0059] The model fusion weights are dynamically adjusted using a formula, where E is the average prediction error rate over the preceding hour, and α is in the range [0.3, 0.8]. Under normal operating conditions, α ≥ 0.8, and lightweight neural network predictions are prioritized. This network, deployed on an NVIDIA A100 GPU node (40GB of video memory), leverages its powerful computing power to achieve single-sample prediction latency of ≤ 30ms, enabling rapid response to normal changes in the ventilation system and meeting real-time requirements.

[0060] Under abnormal working conditions, with α ≤ 0.3, the ANSYS Fluent solver is used for rapid CFD simulation, with a meshing accuracy of 0.2m and a single-condition simulation time of ≤ 2 minutes (for parallel computing nodes ≥ 8 cores). When multiple abnormal working conditions occur concurrently, a priority queue is used to allocate computing resources: working faces with gas concentrations ≥ 1.5% (red alert) are prioritized for CFD simulation (single-condition computing resources ≥ 16 cores), while other abnormal working conditions are queued, ensuring a response time of ≤ 2 minutes for high-risk scenarios. By performing detailed simulations of the physical processes of the ventilation system, more accurate prediction results are provided, providing a scientific basis for responding to complex ventilation conditions.

[0061] 3.3 Data Migration and Model Training Optimization

[0062] To address the problem of insufficient data for new mines, DANN technology was used to migrate historical data from established mines. Adversarial training reduces inter-domain differences, enabling new mines, even in the absence of sufficient measured data, to leverage data resources from established mines for model training. This reduced the initial training data volume by over 60%, and increased model convergence speed by 40%. During the model training phase, 100,000 sets of simulation data were generated using Ventsim. Combined with measured data, 20 features, such as wind pressure gradient and wind speed coefficient of variation, were extracted. A stacking ensemble learning approach was employed, with the underlying models comprising XGBoost, LightGBM, and CNN-LSTM. The top layer used logistic regression to fuse prediction results, leveraging the strengths of the different models and improving the model's predictive accuracy and generalization capabilities. The model also supports online incremental training, which is automatically triggered when the error exceeds 5% for three consecutive cycles. This allows for continuous optimization of model parameters based on real-time changes in the mine ventilation system, ensuring the model consistently maintains excellent predictive performance.

[0063] S4 safety and troubleshooting implementation

[0064] 4.1 Data Security Protection System

[0065] Data security is an important prerequisite for ensuring the stable operation of the intelligent mine ventilation prediction system. During data transmission, the AES-256-GCM encryption algorithm is used. The GCM mode provides authenticated encryption, which can effectively prevent data theft and tampering. The key is generated by a hardware security module (HSM) and updated every 15 minutes via the Diffie-Hellman protocol to ensure the security of data during transmission. In terms of data storage, three levels of permissions are set based on the RBAC model: administrators have read and write permissions and can fully manage and configure the system; engineers have read-only permissions and can analyze data and optimize models; operators can only view the visual interface to ensure the security and standardization of data access. Core data (including raw sensor data and prediction model parameters) is backed up in real time to an off-site disaster recovery center, 50km away from the main data center, with a recovery time objective (RTO) of ≤20 minutes and a recovery point objective (RPO) of ≤2 minutes. Even in the event of a failure in the main data center, data and system operation can be quickly restored to ensure the continuity of the ventilation prediction system.

[0066] 4.2 Fault Recovery Mechanism

[0067] To ensure the ventilation prediction system can quickly resume normal operation in the face of various faults, a comprehensive fault recovery mechanism has been established. When a sensor fails, it automatically switches to the average value of three adjacent sensors for compensation. For example, if a single wind speed sensor fails, the average value of the sensors 50 meters before and after it is used as a replacement. The compensation error is ≤5%, ensuring data continuity and availability.

[0068] In the event of a network outage, edge computing nodes are activated to cache data for ≥2 hours. Upon network restoration, data is automatically synchronized to the cloud, supporting resumable transmission. If the network outage exceeds 2 hours, the edge computing node activates a local lightweight prediction model (a simplified LSTM model trained on 7 days of historical data, with a prediction delay of ≤10ms). Upon network restoration, cached data is automatically synchronized and model parameters are calibrated to prevent data loss. In the event of an algorithm anomaly, the backup basic BP neural network is triggered to take over the prediction, and an alert is sent to the operation and maintenance system within 10 seconds for timely troubleshooting and repair, ensuring the stability and reliability of the ventilation prediction system.

[0069] S5 Prediction Decision Implementation

[0070] 5.1 Graded Early Warning Mechanism

[0071] The prediction and decision-making system has established a three-level early warning mechanism, which can issue early warnings and take corresponding measures in a timely manner according to the different states of gas concentration.

[0072] The yellow warning corresponds to a gas concentration of 0.75% to 1.0%. At this time, the local ventilation fan is started and an alert pops up in the dispatch center. The response time is ≤15 seconds, reminding staff to pay close attention to the ventilation conditions and take preliminary ventilation control measures.

[0073] The orange warning corresponds to a gas concentration of 1.0% to 1.5%. The main fan speed will automatically increase by 10% (the speed adjustment range is limited to 40% to 90% of the rated speed to avoid motor overload), and will be linked to the damper opening (every 10% increase in speed corresponds to a damper opening ≥70%). At the same time, a text message warning will be sent with a response time of ≤30 seconds to increase ventilation and reduce gas concentration.

[0074] The red alert corresponds to a gas concentration ≥1.5%. The power supply to the working face will be cut off immediately, the anti-wind plan will be activated and an evacuation notice will be broadcast. The response time is ≤60 seconds to ensure the safe evacuation of personnel and minimize accident losses.

[0075] 5.2 Multi-objective optimization and control

[0076] Multi-objective optimization and control aims to maximize ventilation efficiency, minimize energy consumption, and balance equipment life. The ventilation efficiency calculation formula is: where V is the average wind speed in the tunnel (calculated by the average of three adjacent sensors), S is the tunnel cross-sectional area, L is the tunnel length, λ is the friction coefficient (0.02 for concrete tunnels and 0.03 for bare rock tunnels, automatically matched according to the tunnel support type), ρ is the air density (calculated by: P is the absolute pressure, T is the thermodynamic temperature, and R is the gas constant), and P is the wind pressure. The fan energy consumption formula is: Q is the air volume (collected in real time by the fan inverter), H is the wind pressure, and is the fan efficiency (calculated in real time using the equipment nameplate parameters and the operating current / voltage; efficiency is equal to the output power divided by the input power). The NSGA-II algorithm is used to solve the Pareto optimal solution and output control parameters such as fan speed and damper opening. After three months of field testing in a 10 million ton mine, ventilation efficiency increased by 22.3% and energy consumption decreased by 16.7%. The generalization error under different working conditions was ≤4.5%. While ensuring ventilation safety, it also improved energy utilization efficiency and extended the service life of equipment.

[0077] 5.3 System Control Mode

[0078] The system supports three control modes to meet the ventilation control needs of different scenarios. In manual mode, managers can manually adjust equipment parameters and preview the results through the monitoring interface, which is suitable for manual intervention in special circumstances. Automatic mode executes the control plan according to preset rules and sends instructions to the PLC control system via the Modbus / TCP protocol to realize the automated operation of the ventilation system and improve work efficiency. Adaptive mode uses reinforcement learning algorithms to learn the ventilation system's responses in real time, updating control parameters every hour. It can automatically optimize the control strategy based on the real-time changes of the ventilation system, further improving the intelligence level of the ventilation system.

[0079] The system is compatible with OPC UA and Modbus TCP / IP dual protocols, and can be connected to existing security monitoring systems (such as KJ95X) through a data isolation gateway to ensure interface compatibility and avoid protocol conflicts and data silos.

[0080] S6. Verification and evaluation system

[0081] 6.1 Key technical indicator testing

[0082]

[0083]

[0084] 6.2 Extreme Scenario Reliability Test

[0085] Multi-sensor failure scenario: When simulating the simultaneous failure of three adjacent sensors, subnet data fusion is used for repair, with an error of ≤ 6% (better than the industry average of 10%), ensuring data continuity.

[0086] Long-term network interruption scenario: Within 4 hours of network interruption, the local model prediction error of the edge node is ≤8%, and the data synchronization accuracy is 100% after recovery.

[0087] Concurrent scenarios of complex working conditions: When abnormal working conditions are triggered on 5 working faces at the same time, the system prioritizes simulation calculations for high-risk scenarios (red alerts), with the response time controlled within 2 minutes. The waiting time for ordinary abnormal working conditions is ≤5 minutes.

[0088] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A method for intelligent prediction of mine ventilation based on a cloud computing platform, characterized by: The following steps are involved: S1 deploys mining intrinsically safe multi-parameter sensors every 50 meters in the mine excavation working face and 30 cm from the top of the center of the mining working face tunnel; S2 uses LoRa wireless modules to form a Mesh network, combined with the underground industrial ring network, to transmit the collected data to the ground cloud computing center; S3 fuses measured data and simulation data to construct a hybrid data set, achieves time synchronization through the NTPv4 protocol, establishes a coordinate system conversion model based on the mine CAD drawings to achieve spatial alignment, and uses the weighted average method to fuse data; S4 removes gross errors and repairs missing data from the original data. If there are more than five consecutive gross errors during gross error removal, or if three or more sensors in the same subnet trigger gross errors, a fault warning will be triggered and interpolation and repair will be performed using data from adjacent subnets. S5 builds a hybrid prediction architecture to identify abnormal operating conditions by ensuring that the gas concentration fluctuation rate is no less than 15% or the wind speed mutation rate is no less than 20%. It dynamically adjusts the model fusion weight: under normal operating conditions, when the weight α is no less than 0.8, neural network prediction is prioritized. Under abnormal conditions, when the weight α is no more than 0.3, CFD fast simulation is initiated. S6 implements data security protection and fault recovery based on a cloud computing platform. Data transmission uses AES-256-GCM encryption and dynamic key updates. Storage is based on a three-level permission system with an RBAC model. When the network interruption exceeds 2 hours, the local lightweight prediction model of the edge node is enabled; S7 establishes a three-level early warning mechanism and multi-objective optimization control, triggering early warnings and implementing corresponding measures based on gas concentration.

2. The intelligent prediction method for mine ventilation based on a cloud computing platform according to claim 1, characterized in that: The multi-parameter sensor is KJ90X-F8, with explosion-proof grade ExibIMb and MA certification. The vibration sensor CA-YD-185 is pasted on the vertical surface of the bearing seat, and the damper status sensor GUD10 is installed at the damper hinge. The installation error does not exceed 2mm. Every 5 adjacent sensors form a monitoring subnet. When a single sensor fails, the subnet data is weightedly fused, and the weight is calculated in inverse proportion to the distance. Vibration sensors and damper status sensors are installed on the ventilator and damper respectively. A total of 10 parameters including wind speed, wind pressure, gas concentration, temperature, humidity, personnel position, equipment position, ventilator vibration, damper status, and dust concentration are collected.

3. The intelligent prediction method for mine ventilation based on a cloud computing platform according to claim 1 is characterized by: The LoRa network supports dual-base station hot standby and automatic routing reconstruction, and the industrial ring network adopts a 10G fiber dual-ring network; the LoRaMesh network has a single base station coverage radius of 2km, the data upload delay does not exceed 500ms, and the penetration capability meets the transmission requirements of 5-layer tunnel walls; the industrial ring network supports IEEE802.1X authentication and link aggregation, and automatically switches to redundant links in the event of a single point failure.

4. The intelligent prediction method for mine ventilation based on a cloud computing platform according to claim 1, characterized in that: The missing repair adopts a bidirectional LSTM neural network combined with a dynamic time window The historical data sequence is extracted, where T is the data collection time. The number of hidden layer neurons in the bidirectional LSTM neural network is 64. After 5-fold cross-validation optimization, the validation indicators are RMSE not exceeding 0.08 and RAE not exceeding 2.5%. The model training adopts a batch size of 32 and a learning rate of 0.001, combined with the early stopping method to avoid overfitting.

5. The intelligent prediction method for mine ventilation based on a cloud computing platform according to claim 1, characterized in that: The gas concentration fluctuation rate ΔC in the working condition identification is t -C t-10 | / C t-10 ×100%, the time window is 10 minutes, the wind speed mutation rate Δv=|v t -v t-5 | / v t-5 × 100%, the time window is 5 minutes, and the threshold is based on the coal mine safety regulations and historical accident data.

6. The intelligent prediction method for mine ventilation based on a cloud computing platform according to claim 1, characterized in that: The CFD fast simulation grid division accuracy is 0.2m, the computing resources for a single working condition are no less than 8 cores, and when multiple abnormal working conditions occur concurrently, scheduling is based on priority. Red warning scenarios are prioritized to allocate no less than 16 cores of computing resources to ensure that the response time does not exceed 2 minutes.

7. The intelligent prediction method for mine ventilation based on a cloud computing platform according to claim 1, characterized in that: In the multi-objective optimization, the fan speed adjustment range is limited to 40% to 90% of the rated speed, and is linked to the damper opening. For every 10% increase in speed, the corresponding damper opening is not less than 70%. Actual measurements show that the ventilation efficiency is increased by not less than 20% and the energy consumption is reduced by not less than 15%.

8. The intelligent prediction method for mine ventilation based on a cloud computing platform according to claim 1, characterized in that: The key in the data security protection is generated by a hardware security module and updated every 15 minutes through the Diffie-Hellman protocol. The core data is backed up in real time to an off-site disaster recovery center. The recovery time target does not exceed 20 minutes, and the recovery point target does not exceed 2 minutes.

9. The intelligent prediction method for mine ventilation based on a cloud computing platform according to claim 1, characterized in that: The edge computing node has a cache capacity of no less than 2 hours. When the network is interrupted, a local lightweight prediction model is enabled. The model is based on a simplified LSTM trained on historical 7 days of data. The prediction delay does not exceed 10ms. After the network is restored, the data is automatically synchronized and the parameters are calibrated.

10. The intelligent prediction method for mine ventilation based on a cloud computing platform according to claim 1, characterized in that: The corresponding measures adopt the NSGA-II algorithm to solve the Pareto optimal solution of maximizing ventilation efficiency, minimizing energy consumption and balancing equipment life, and output the fan speed and damper opening control parameters.

Images