A method for coal mine ventilation optimization and air quality control based on adaptive learning

Abstract

The present invention discloses a method for coal mine ventilation optimization and air quality control based on adaptive learning, which relates to the technical field of coal mine ventilation optimization and includes the following steps: arranging multiple sensors in the mine to collect and monitor key environmental parameters in real time, and uploading the collected data to a central control system; comparing the collected real-time data with historical data, and dividing the collected data into normal operating condition data and abnormal operating condition data according to the environmental status. The present invention realizes precise ventilation control through real-time monitoring of multiple sensors and deep learning algorithms, quickly locates problem areas in emergencies, and dynamically adjusts airflow to ensure the safety of miners. The system uses an LSTM model to predict event trends in advance, take preventive control measures in a timely manner, and supports real-time adjustments to improve response capabilities. Through multi-stage ventilation modes, the system optimizes energy consumption and equipment operating efficiency, reduces unnecessary energy waste and maintenance costs, and achieves a balance between safety and economy.

Description

Technical Field

[0001] The present invention relates to the technical field of coal mine ventilation optimization, and in particular to a coal mine ventilation optimization and air quality control method based on adaptive learning. Background Art

[0002] Adaptive learning-based coal mine ventilation optimization and air quality control aims to achieve efficient ventilation and air quality management by dynamically adjusting the ventilation system within coal mines through intelligent technology. This approach relies on an adaptive learning algorithm that automatically identifies and predicts the optimal ventilation strategy under different environmental conditions by monitoring gas composition, temperature, humidity, and other data within the mine in real time. The system continuously learns from historical data and changes in the mine environment, gradually optimizing ventilation patterns to ensure worker safety, improve energy efficiency, and reduce the accumulation and emission of harmful gases. This approach reduces ventilation costs and improves mine production efficiency while ensuring mine safety.

[0003] The existing technology has the following deficiencies:

[0004] In coal mine ventilation optimization and air quality control based on adaptive learning, the adaptive learning algorithms of existing technologies may experience "overfitting" under extreme operating conditions. When the mine system operates under certain conditions for a long time (such as stable climate or constant production load), the adaptive learning system may mistakenly regard this state as a "standard operating condition" and optimize the ventilation strategy based on this. However, when an emergency occurs (such as a sudden increase in gas, fire or equipment failure), the system may not be able to quickly identify the anomaly and adapt to the new environmental requirements, resulting in delayed or failed ventilation response. This may aggravate the increase in harmful gas concentrations, the spread of fire, and even directly endanger the safety of miners. The severity of this problem lies in the fact that if the system's pre-judgment and emergency ventilation strategy are insufficient, critical rescue time will be missed, resulting in catastrophic consequences.

[0005] The above information disclosed in this Background section is only for enhancement of understanding of the background of the present disclosure and therefore it may contain information that does not form the prior art that is already known to a person of ordinary skill in the art. Summary of the Invention

[0006] The purpose of the present invention is to provide a method for coal mine ventilation optimization and air quality control based on adaptive learning, which realizes precise ventilation control and event emergency management by combining multiple sensors with deep learning and reinforcement learning algorithms. When an emergency such as a gas leak or fire occurs, the system quickly detects and locates the problem area, dynamically adjusts the airflow to concentrate on the affected area, avoids the shutdown of the entire mine, improves safety and prevents secondary disasters. At the same time, the system predicts the development trend of events through the LSTM model, takes preventive control in advance, reduces the occurrence of risks, and supports real-time adjustment strategies to improve adaptability. A multi-stage ventilation mode is adopted to reduce fan power consumption and airflow waste while ensuring safety, continuously optimize energy consumption levels, effectively reduce operating costs and improve the economy and sustainability of the system to solve the problems in the above-mentioned background technology.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a method for optimizing coal mine ventilation and controlling air quality based on adaptive learning, comprising the following steps:

[0008] Various sensors are deployed in the mine to collect and monitor key environmental parameters in real time, and the collected data is uploaded to the central control system;

[0009] Compare the collected real-time data with historical data, and divide it into normal operating condition data and abnormal operating condition data according to the environmental status. Use the K-means clustering classification algorithm to identify potential extreme conditions to ensure that subsequent algorithm optimization can cover the control requirements under different operating conditions;

[0010] Based on the mine's historical abnormal events, a library of ventilation plans and emergency response strategies is established, including multi-layer ventilation modes to deal with gas surges, fires, and equipment failures. The plan library is regularly updated based on new data.

[0011] Using reinforcement learning algorithms, a preliminary ventilation strategy is generated based on real-time data and recommendations from a pre-plan library. This strategy is continuously optimized through interactive feedback from historical data, and the effects of adjustments are simulated within the mine's airflow and air quality models to obtain the optimal preliminary strategy.

[0012] Multiple anomaly detection algorithms are used to analyze real-time mine data in parallel. Once a deviation from normal data patterns is detected, an alarm is triggered, the current preliminary ventilation strategy is interrupted, and emergency response mode is entered.

[0013] After an emergency occurs, a deep learning algorithm is activated to re-evaluate the current status and predict the development trend of the incident based on historical abnormal data. Based on the prediction results, the ventilation mode is adjusted dynamically in real time to ensure that the air quality in the mine is restored to a safe range in the shortest time possible while minimizing energy consumption.

[0014] Preferably, the specific steps of using the K-means clustering algorithm to classify the mine environment data are as follows:

[0015] Extract key features from real-time data collected by sensors, perform outlier detection and noise filtering on the sensor data, and eliminate data anomalies caused by sensor false alarms or failures. The preprocessed data must also be normalized to map data of different units and dimensions to a unified scale to ensure that the weights of various parameters are balanced in cluster analysis. Finally, the processed data is merged with historical data to form a complete data set.

[0016] Divide the data set into K clusters. After determining the value of K, select the initial cluster centers through random initialization or K-means++ algorithm to reduce the risk of the algorithm falling into a local optimal solution.

[0017] Assign each piece of data to the nearest cluster center based on the Euclidean distance, then recalculate the cluster center position based on all the data points assigned to the cluster, and repeat the process of data assignment and cluster center update until the position of the cluster center no longer changes significantly or the set number of iterations is reached;

[0018] After clustering is completed, different labels are assigned to each cluster, and the characteristic centers of each cluster are analyzed. Isolated data that cannot be classified into regular clusters are marked as abnormal conditions, triggering further detection and alarms.

[0019] Preferably, the specific steps for establishing a library of multiple ventilation plans and emergency response strategies are as follows:

[0020] Gather data on various abnormal events that have occurred in the mine in the past;

[0021] According to the severity and impact scope of different abnormal events, the events are divided into multiple levels;

[0022] In the process of building the emergency plan library, all ventilation modes and emergency response measures are coded in the form of rules to form a modular strategy library;

[0023] To ensure that the contingency plan library can adapt to the ever-changing mine environment, new data and the latest safety incident feedback are regularly incorporated into analysis and existing contingency plans are optimized.

[0024] Preferably, the specific steps of generating a preliminary ventilation strategy using a reinforcement learning algorithm are as follows:

[0025] In the reinforcement learning framework, the state space, action space, and reward mechanism are first defined. The state space includes the environmental parameters within the mine, the action space refers to the control instructions that can be executed by the ventilation system, and the reward mechanism is designed based on the ventilation effect and safety of the system.

[0026] Based on current real-time data, the reinforcement learning algorithm reads historical cases and emergency plans that are closest to the current environment from the plan library as a reference for generating preliminary strategies;

[0027] After generating the initial strategy, we conducted virtual tests in a mine airflow and air quality simulation model to verify the effectiveness of the strategy under various extreme scenarios by simulating different working conditions.

[0028] Once the preliminary strategy passes the simulation test, it is deployed to the actual mine ventilation system and the execution effect is monitored in real time. If the ventilation effect meets expectations and meets safety standards, the strategy is marked as qualified and recorded in the plan library for reference in similar scenarios in the future. If deviations occur during execution, the algorithm will adjust the strategy based on the new data to ensure that the ventilation system always remains in optimal condition.

[0029] Preferably, the specific steps of using multiple anomaly detection algorithms to analyze real-time mine data in parallel are as follows:

[0030] In the process of anomaly detection, the mine environment parameters X collected from the sensor are i}={x1,x2,…,x n} is input into the algorithm module, where x i Represents the real-time data of the i-th sensor. In order to ensure the stability and efficiency of the algorithm, the collected data is feature extracted and standardized. Let X∈R m×n Represents the data matrix, where m is the number of samples and n is the dimension of the monitored environmental parameter. The standardization formula is: j = 1, ..., n, where μ j and σ j are the mean and standard deviation of parameter j, and the standardized data z ij It is passed to the subsequent detection algorithm to eliminate the influence of different parameter scales;

[0031] The isolation forest algorithm is used to construct a random tree to evaluate the isolation degree of the sample points. For each sample z i , the isolation forest will generate T random trees, the depth of each tree represents the number of steps required to completely separate the sample points, let sample z i The depth of the t-th tree is H(z i ), then the average depth of all trees is: Combined with the expected tree depth c(n), the anomaly score S(z i ) is calculated as follows:

[0032]

[0033] , by setting the threshold α, such as S(z i)>α, it is marked as abnormal.

[0034] Preferably, a class of support vector machines is used to construct boundaries in high-dimensional space to detect abnormal data points. Given a data set Z = {z1, z2, ..., z m}, SVM separates normal data from abnormal data by finding a hyperplane w·z-ρ=0, and the optimization objective is: subject to:w·z i ≥ρ-ξ i ,ξ i ≥0, where v is the parameter that controls the abnormality rate, ξ i It is a slack variable used to process data points on the soft boundary. When new sample data falls outside the hyperplane, it is marked as an outlier.

[0035] Through the parallel analysis of isolation forest and SVM, two independent anomaly scores are calculated for each data, namely S IF (z i ) and S SVM (z i ), in order to achieve multi-algorithm fusion, a weighted anomaly scoring strategy is adopted. The specific expression is: S total (z i )=β·S IF (z i )+(1-β)·S SVM (z i ), where β is the algorithm weight parameter, reflecting the importance of isolation forest and SVM in different scenarios. total (z i )>α, the system immediately triggers an alarm, interrupts the current ventilation strategy, enters the emergency response mode, and executes the corresponding plan.

[0036] Preferably, the specific steps of using deep learning algorithms to reassess emergencies and dynamically adjust ventilation modes are as follows:

[0037] When an emergency is detected in the mine, the latest real-time data is collected immediately;

[0038] After the initial status assessment, predict the development trend of the emergency by using similar cases in historical data;

[0039] Generate and execute real-time ventilation adjustment instructions based on the prediction results;

[0040] After adjusting the ventilation mode, feedback data is continuously collected and analyzed to ensure that the air quality gradually returns to a safe range. If any abnormal changes are detected or the strategy execution does not meet expectations, the deep learning algorithm will immediately re-evaluate the situation and automatically optimize the adjustment instructions to achieve secondary or multiple iterative optimization. In addition, the data and response process of this emergency will be stored in the historical database for reference for similar events in the future and to improve the model's adaptive capabilities.

[0041] In the above technical solution, the technical effects and advantages provided by the present invention are:

[0042] The present invention uses a variety of sensors to monitor the mine environment in real time, and combines deep learning and reinforcement learning algorithms to achieve precise ventilation control. When an emergency (such as a gas leak or fire) occurs, the system can quickly detect and locate the problem area, and dynamically adjust the ventilation mode to concentrate the main airflow in the affected area. At the same time, through precise control of regional classification, the shutdown of the entire mine or the full evacuation of personnel is avoided, and local problems are quickly resolved. This rapid response capability significantly improves the safety of the mine, reduces the risk of the event worsening and spreading, and maximizes the protection of the lives of miners. In addition, the precise control strategy avoids unnecessary airflow disturbances, reduces the diffusion of gas in the mine, and helps prevent the occurrence of secondary disasters.

[0043] By adopting time-series deep learning models such as LSTM, the present invention enables the system to predict the development trend of events in advance and adopt preventive control strategies to reduce the occurrence of potential risks. For example, after detecting a decrease in wind speed and a slow increase in gas concentration, the system can increase the fan power in advance and open the backup damper for ventilation to prevent the gas concentration from rising further. The combination of prediction and prevention enables the system to take emergency measures before an event occurs, buying more time to deal with potential dangers. In addition, the system also supports real-time adjustment strategies. If the initial prediction deviates, the algorithm can quickly correct and redeploy the best solution. This capability makes the mine's response to emergencies more flexible and efficient, avoiding the limitations of traditional management with delayed reactions.

[0044] The present invention not only ensures the safe operation of the mine through a phased ventilation mode, but also significantly reduces energy consumption. At the beginning of the event, the system runs the ventilation equipment at high power to quickly reduce the concentration of harmful gases; when the air quality gradually recovers, the system will reduce the fan power and close some unnecessary dampers to reduce unnecessary airflow waste. By recording the energy consumption data and response effects of historical events, the energy consumption level of each stage is continuously optimized to achieve the best balance between safety and energy efficiency. Compared with the traditional full-power ventilation solution, this multi-stage mode significantly reduces electricity consumption and equipment wear, reduces the long-term operating costs of the mine, and improves the economy and sustainability of the ventilation system. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, a brief introduction to the drawings required for use in the embodiments will be given below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0046] Figure 1 This is a method flow chart of a method for coal mine ventilation optimization and air quality control based on adaptive learning of the present invention. DETAILED DESCRIPTION

[0047] Example embodiments will now be described more fully with reference to the accompanying drawings. However, example embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, these example embodiments are provided so that the description of this disclosure will be thorough and complete and will fully convey the concepts of the example embodiments to those skilled in the art.

[0048] The present invention provides Figure 1 The method for optimizing coal mine ventilation and controlling air quality based on adaptive learning is shown, comprising the following steps:

[0049] Various sensors are deployed in the mine to collect and monitor key environmental parameters in real time, and the collected data is uploaded to the central control system;

[0050] Various sensors are deployed in the mine to collect and monitor key environmental parameters in a comprehensive and real-time manner, and upload the data to the central control system to ensure the safety and efficiency of air quality and ventilation. These sensors mainly include the following categories:

[0051] Gas sensor: used to monitor methane concentration, prevent gas from exceeding the standard, and avoid the risk of explosion caused by gas accumulation.

[0052] Carbon Monoxide Sensor: Detects CO concentration in mines to prevent poisoning risks caused by insufficient ventilation, equipment failure, or fire.

[0053] Oxygen sensor: monitors oxygen content to ensure that the oxygen concentration in the mine is within a safe range and prevent the occurrence of oxygen deficiency.

[0054] Temperature and humidity sensors: Monitor temperature and humidity changes in mines to control air comfort and prevent excessive temperatures from overheating equipment or affecting worker health.

[0055] Dust sensor: monitors dust concentration in the air to prevent excessive dust accumulation from causing explosions or respiratory diseases.

[0056] Wind speed and direction sensors: Monitor the speed and direction of airflow within the mine to ensure the ventilation system operates along the intended path and avoid airflow turbulence.

[0057] Pressure sensor: Detects changes in air pressure in different areas and is used to predict gas leaks or abnormal damper operation.

[0058] Carbon dioxide sensor: monitors carbon dioxide concentration to prevent miners from suffocating or becoming fatigued due to excessive carbon dioxide.

[0059] Hydrogen sulfide sensor: Detects the presence of toxic hydrogen sulfide gas to prevent poisoning caused by gas leakage in mines.

[0060] Through the coordinated monitoring of these sensors, the system can fully grasp the environmental conditions inside the mine, provide reliable data support for intelligent ventilation and air quality control, and issue timely warnings and adjustments when abnormal working conditions occur.

[0061] The collected real-time data is compared with historical data and divided into normal operating condition data and abnormal operating condition data according to the environmental status. In this process, the K-means clustering classification algorithm is used to identify potential extreme conditions to ensure that subsequent algorithm optimization can cover the control requirements under different operating conditions;

[0062] The specific steps of using K-means clustering algorithm to classify mine environment data are as follows:

[0063] The system first extracts key features from the real-time data collected by the sensors, such as gas concentration, carbon monoxide concentration, temperature and humidity, wind speed, etc. To ensure data quality, the system needs to detect outliers and filter noise on the sensor data to eliminate data anomalies caused by sensor false alarms or failures. In addition, some data may be missing, in which case interpolation or average substitution is used to supplement it. The preprocessed data also needs to be normalized to map data of different units and dimensions to a unified scale (such as 0 to 1) to ensure that the weights of each parameter are balanced in the cluster analysis. Finally, the system merges the processed data with historical data to form a complete data set.

[0064] The core of the K-means algorithm is to partition a dataset into K clusters. Therefore, before applying the algorithm, it is necessary to determine an appropriate number of clusters, K. Mine managers can estimate the value of K based on the mine's historical operating records and different operating conditions (such as normal conditions, elevated gas levels, and fire precursors). To further refine the selection of K, the elbow method or silhouette coefficient method can be used to analyze the error under different numbers of clusters to ensure a more representative clustering result. After determining the value of K, the initial cluster centers are selected through random initialization or the K-means++ algorithm to reduce the risk of the algorithm falling into a local optimum.

[0065] The K-means algorithm consists of two main loops: first, assigning each data point to the nearest cluster center based on Euclidean distance; second, recalculating the cluster center location based on all data points assigned to that cluster. The system repeats this process of data assignment and cluster center update until the cluster center location no longer changes significantly or the set number of iterations is reached. As real-time mine data continues to flow into the system, it triggers the K-means algorithm's online update mechanism, incrementally updating the cluster centers based on the new data. This iterative update mechanism helps the system dynamically identify new data patterns and maintain consistency with historical data.

[0066] After clustering, the system assigns a different label to each cluster (such as "normal operating conditions," "abnormal gas concentration," "precursor of equipment failure," etc.) and analyzes the characteristic centers of each cluster. For example, when the gas concentration at the center of a cluster indicates that it is above the safety threshold, the system labels that cluster as "gas warning state." For isolated data that cannot be classified into conventional clusters, the system marks it as an abnormal operating condition, triggering further detection and alarms. This classification result not only provides a data basis for real-time control strategies, but also enriches historical data sets, allowing subsequent algorithm optimization to cover a variety of extreme situations, thereby enhancing the system's ability to respond to emergencies.

[0067] Based on the mine's historical abnormal events, a library of ventilation plans and emergency response strategies is established, including multi-layer ventilation modes to deal with gas surges, fires, and equipment failures. The library is regularly updated based on new data.

[0068] The specific steps for establishing a library of multiple ventilation plans and emergency response strategies are as follows:

[0069] The system first compiles data on various abnormal events that have occurred in the mine in the past, including detailed records of cases such as gas surges, fires, and equipment failures. This data comes from sensor logs, manual inspection reports, accident analysis reports, and historical archives in the mine safety management system. The analysis focuses on identifying the triggering conditions, scope of impact, duration, and degree of harm for each type of event. Through data analysis and retrospective statistics, the system summarizes the occurrence patterns of various abnormal events (such as specific seasons and equipment aging cycles), laying the foundation for the design of corresponding emergency plans. The comprehensiveness and accuracy of the data are crucial to ensure that the system can cover all potential risks when facing complex scenarios.

[0070] According to the severity and scope of impact of different abnormal events, events are divided into multiple levels, such as "mild abnormality", "moderate abnormality", "serious abnormality" and "critical state". Different levels of events correspond to different ventilation modes. For example, a slight increase in gas may only require the power of the ventilator in a local area to be increased, while a serious gas leak requires the dampers of the entire mine to be switched, the ventilators to be started to maximum power, and personnel evacuation orders to be issued. The system designs multi-layer ventilation modes based on the event level and mine layout to ensure that ventilation control from local to global can be achieved. The plan covers normal ventilation, gas emergency, fire handling and safe ventilation plans after equipment shutdown.

[0071] During the construction of the emergency plan library, the system encodes all ventilation modes and emergency response measures in the form of rules, forming a modular strategy library. These modular plans are managed by a rule engine, which can automatically match applicable plans based on real-time data and historical cases. For example, when the gas concentration in a certain area exceeds the set threshold and the wind speed slows, the system will select the plan related to the gas leak based on the rules and determine whether to enter the second or third level response mode. Each plan includes specific operating instructions, the startup sequence of the ventilation equipment, the opening and closing settings of the dampers, the evacuation routes of personnel, and other details to ensure that each emergency measure can be executed efficiently.

[0072] To ensure that the plan library can cope with the ever-changing mine environment, the system regularly incorporates new data and the latest safety incident feedback into its analysis and optimizes existing plans. Through adaptive learning algorithms, the system can identify which plans are most effective in actual events and make adjustments to any deficiencies. For example, if a delayed response to a gas leak leads to an escalation of danger, the system will optimize the alarm trigger time and ventilation equipment response sequence in the plan. Data review after each incident can prompt dynamic updates to the plan library for intelligent optimization. Mine managers can also manually adjust plans based on the latest regulations and management requirements to ensure that the strategy library remains efficient and compliant.

[0073] A reinforcement learning algorithm generates a preliminary ventilation strategy based on real-time data and recommendations from a database of pre-plans. This algorithm continuously optimizes through interactive feedback from historical data and simulates the effects of adjustments within the mine's airflow and air quality models to determine the optimal preliminary strategy.

[0074] The specific steps for generating a preliminary ventilation strategy using reinforcement learning algorithms are as follows:

[0075] In the reinforcement learning framework, the system first defines the state space (State Space), action space (Action Space), and reward mechanism (Reward Function). The state space contains the combined state of real-time data such as gas concentration, CO content, temperature and humidity, and wind speed for various environmental parameters within the mine. The action space refers to the control instructions that can be executed by the ventilation system, such as adjusting fan power, opening or closing dampers, and changing the direction of regional airflow. The reward mechanism is designed based on the ventilation effect and safety of the system. For example, positive rewards are given when gas concentration decreases and air quality improves, and negative rewards are given when ventilation energy consumption is too high or response is delayed. This mechanism ensures that the algorithm's goal is to achieve an efficient and low-energy ventilation strategy while maintaining safety.

[0076] Based on the current real-time data, the reinforcement learning algorithm will read the historical cases and emergency plans that are closest to the current environment from the plan library as a reference for generating preliminary strategies. At this time, the algorithm will select the best initial action combination under the current state, such as increasing the main fan power or opening the backup vents in advance when the gas concentration is high. These preliminary actions are determined by Q-learning or deep Q network (DQN) algorithms. The system will evaluate the short-term and long-term benefits of each action to formulate a preliminary ventilation strategy. The guidance of the plan library can effectively narrow the algorithm's exploration space, speed up the generation of strategies, and ensure a quick response in emergency situations.

[0077] After generating the preliminary strategy, the system conducts a virtual test in a mine's airflow and air quality simulation model. By simulating different working conditions (such as wind speed changes and sudden gas increases), the system verifies the effectiveness of the strategy in various extreme scenarios. If the simulation results show that the strategy has flaws, such as low ventilation efficiency or substandard air quality in some areas, the system will adjust the action combination and retest. During this process, the reinforcement learning algorithm is continuously optimized through interactive feedback with historical data, and the simulation results are fed back to the strategy model, so that the next round of strategy generation is closer to the optimal solution.

[0078] Once the initial strategy passes simulation testing, the system deploys it to the actual mine ventilation system and monitors its effectiveness in real time. If the ventilation results meet expectations and safety standards, the strategy is marked as qualified and recorded in the emergency plan library for future reference in similar scenarios. If deviations occur during execution, the algorithm adjusts the strategy based on the new data to ensure that the ventilation system remains in optimal condition. This parallel process of "exploration and utilization" enables the system to maintain dynamic learning and adaptability in a changing environment, achieving safe, efficient, and energy-saving ventilation control.

[0079] Multiple anomaly detection algorithms (such as the Isolation Forest algorithm and Support Vector Machine) are used to analyze real-time mine data in parallel. Once a data pattern deviates from the normal state, an alarm is triggered, the current preliminary ventilation strategy is interrupted, and emergency response mode is entered. During this process, the system marks the anomaly and stores the data samples for subsequent algorithm optimization.

[0080] The specific steps for using multiple anomaly detection algorithms to analyze real-time mine data in parallel are as follows:

[0081] In the process of anomaly detection, the mine environment parameters X collected from the sensor are i}={x1,x2,…,x n} is input into the algorithm module, where x i Represents the real-time data of the i-th sensor. In order to ensure the stability and efficiency of the algorithm, the system will extract features and standardize these data. Let X∈R m×n Represents a data matrix, where m is the number of samples (e.g., the amount of sensor data collected in the past 5 minutes) and n is the dimension of the monitored environmental parameter (e.g., gas concentration, wind speed, etc.). The normalization formula is: j = 1, ..., n, where μ j and σ j are the mean and standard deviation of parameter j respectively. The standardized data z ij It is passed to the subsequent detection algorithm to eliminate the impact of different parameter scales.

[0082] The isolation forest algorithm is used to construct a random tree to evaluate the isolation degree of the sample points. i , the isolation forest will generate T random trees, and the depth of each tree represents the number of steps required to completely separate the point. i The depth of the t-th tree is H(z i ), then the average depth of all trees is: Combined with the expected tree depth c(n) (adjusted for the number of tree samples n), the anomaly score S(z i ) is calculated as follows:

[0083]

[0084] , if S(z i ) is close to 1, the sample point is highly abnormal; if it is close to 0.5, it is within the normal range. By setting the threshold α, such as S(z i )>α, it is marked as abnormal.

[0085] One-Class SVM is used to construct boundaries in high-dimensional space to detect abnormal data points. Given a data set Z = {z1, z2, ..., zm}, SVM separates normal data from abnormal data by finding an optimal hyperplane w·z-ρ=0. Its goal is to maximize the margin while ensuring that most samples are in the normal region. The optimization goal is: subject to:w·z i ≥ρ-ξ i ,ξ i ≥0, where ν is the parameter that controls the abnormality rate, ξ i Is a slack variable used to handle data points on the soft boundary. When new sample data falls outside the hyperplane, it is marked as an outlier.

[0086] Through the parallel analysis of isolation forest and SVM, the system calculates two independent anomaly scores for each data, namely S IF (z i ) and S SVM (z i ). In order to achieve multi-algorithm fusion, the system adopts a weighted anomaly scoring strategy, the specific expression is: S total (z i )=β·S IF (z i )+(1-β)·S SVM (z i ), where β is the algorithm weight parameter, reflecting the importance of isolation forest and SVM in different scenarios. total (z i ) > α (the set global abnormality threshold), the system immediately triggers an alarm, interrupts the current ventilation strategy, enters emergency response mode, and executes the corresponding plan. At this time, the system sends instructions to all fans and dampers through the central control system and issues an evacuation notice in the personnel dispatch system to ensure the safety of miners in the shortest possible time.

[0087] After an emergency occurs, the system activates a deep learning algorithm to reassess the current status and, combined with historical anomaly data, predict the development trend of the incident. Based on the predicted results, the system dynamically adjusts the ventilation mode in real time to ensure that the air quality in the mine is restored to a safe range as quickly as possible while minimizing energy consumption. This adjustment includes detailed control of fan speed, damper opening and closing status, and regional airflow distribution.

[0088] The specific steps for using deep learning algorithms to reassess emergencies and dynamically adjust ventilation modes are as follows:

[0089] When an emergency event (such as a sudden increase in gas or a fire) is detected in the mine, the system immediately collects the latest real-time data, including key parameters such as gas concentration, temperature and humidity, wind speed, and oxygen content. This data, along with historical abnormal data, is input into the deep learning model for status assessment. The system uses time series models such as long short-term memory networks (LSTM) to analyze the changing trends of the current status to ensure that it can understand whether the development of the current event represents a continuation, deterioration, or alleviation of the warning state. In addition, the system also compares environmental differences in different areas to determine whether the problem is localized or may trigger a chain reaction throughout the entire mine.

[0090] After an initial status assessment, the system uses historical data from similar cases to predict the likely development of the emergency. Based on the rate of change of current environmental parameters, the deep learning model infers whether the incident will worsen (e.g., continued gas levels or fire spread) and how long it will take to return to normal. The system also considers the possibility of equipment failure or human intervention, and, based on historical emergency response results, simulates multiple possible scenarios. Each scenario is assigned a risk level, and the model selects the response strategy with the lowest risk and highest resource efficiency, providing a basis for subsequent adjustments.

[0091] Based on the prediction results, the system generates and executes real-time ventilation adjustment instructions. This includes adjusting the fan speed (such as increasing the fan power to a peak value or reducing it to an energy-saving state), the opening and closing status of the damper (such as closing certain ventilation paths to concentrate wind force), and optimizing regional airflow distribution (directing more wind force into the affected area). The system will gradually deploy these adjustments and continuously monitor the feedback effects of various parameters to ensure that the ventilation system can respond quickly to changes in events. At the same time, the control strategy will take into account the optimization of energy consumption and avoid running the fan at high power for a long time, thereby balancing safety and cost.

[0092] After adjusting the ventilation mode, the system continuously collects and analyzes feedback data to ensure that air quality gradually returns to a safe range. If any unusual changes are detected or if policy execution falls short of expectations, the deep learning algorithm immediately reassesses the situation and automatically optimizes the adjustment instructions, achieving secondary or multiple iterations of optimization. Furthermore, the system stores the data and response process for this emergency in a historical database for future reference and to enhance the model's adaptability. This entire process implements closed-loop management from event detection, prediction, adjustment, and feedback optimization, ensuring that the mine can quickly return to a stable state in the event of an emergency and minimizing unnecessary energy consumption.

[0093] Implementation method 1: Precise ventilation control based on regional classification;

[0094] In mines, the complexity of the environment and the diversity of risk areas dictate the need for precise ventilation control based on regional classification. Mines are divided into multiple independent areas, such as production areas, transportation areas, equipment maintenance areas, etc. Each area is equipped with multiple sets of sensors to monitor key environmental parameters (such as gas concentration, CO concentration, temperature, humidity and wind speed). When the system detects a sudden increase in gas concentration in a specific area, the deep learning model will quickly determine the type and risk level of the event based on the matching of the real-time data and historical data of the area. For example, if the increase in gas concentration is accompanied by a decrease in wind speed, this may indicate a failure of the regional ventilation equipment or a blockage in the ventilation path.

[0095] Once the type and risk level of the event are identified, the system immediately activates the response mechanism. The ventilation system will dynamically adjust the opening and closing status of the dampers and the wind speed of the ventilators in the area to concentrate the main airflow to the affected area. At the same time, to prevent unnecessary energy consumption, the system will reduce the ventilation supply to other non-critical areas. For example, when high gas concentrations appear in the production area, the system will close some dampers leading to the transportation area or maintenance area to ensure that the airflow is concentrated in key areas. In addition, if the incident is local and controllable (such as a problem in a branch tunnel), the system will give priority to using local ventilation equipment for processing without affecting the overall ventilation status of the mine.

[0096] The advantage of this implementation is the ability to quickly respond to localized incidents without shutting down the entire mine or evacuating all personnel. This hierarchical control ensures efficient allocation of system resources and avoids the energy waste caused by blindly increasing ventilation rates. Furthermore, precise zone control reduces wind disturbances and prevents gas from spreading between zones, effectively improving mine safety and response speed.

[0097] Implementation method 2: Combining dynamic prediction with preventive regulation;

[0098] Emergencies are often hidden and sudden, so predictive control is a crucial component of mine safety. This implementation utilizes time series models, such as long short-term memory (LSTM) networks, to analyze and predict real-time data and historical cases within the mine. The LSTM model can capture trends in parameter changes over time and identify potential early warning signs before an abnormal event occurs. For example, if wind speeds in a particular area decrease, temperatures rise, and gas concentrations slowly increase, the system predicts a potential fan failure and issues an early warning.

[0099] When the system predicts the possibility of an abnormal event, it automatically triggers preventive control strategies. In the early stages of a gas leak, the system might preemptively increase fan power and open backup dampers to increase ventilation in the area. If the model predicts an increased risk of fire, the system reroutes airflow to prevent gas accumulation or the spread of fire. Simultaneously, the personnel dispatch system issues advance evacuation notices to ensure workers avoid areas of potential incidents. These preventive measures effectively reduce the risk of an incident escalating and buy more time for emergency response.

[0100] This implementation also features dynamic updates as events unfold. Even if predictions are inaccurate, the deep learning model can continuously revise its judgment based on new real-time data and dynamically adjust its response strategy. For example, the system might discover that the initially predicted fire risk has decreased, halting emergency measures and switching to regular ventilation. This flexible, preventative control significantly improves the mine's resilience and management, preventing the possibility of an incident escalating.

[0101] Implementation method 3: Energy-efficient multi-stage ventilation mode;

[0102] Optimizing energy consumption while ensuring mine safety is a key aspect of ventilation system design. This implementation proposes a multi-stage ventilation model that balances safety and energy efficiency by gradually reducing fan power and adjusting damper status after an emergency. In the early stages of an emergency, to ensure air quality in the mine quickly returns to a safe range, the system commands all main and local fans to operate at maximum power to maximize air flow.

[0103] As air quality gradually improves, the system analyzes the recovery progress based on the deep learning model's predictions and enters the second phase of low-power operation. During this phase, the system gradually reduces fan power and closes unnecessary dampers to reduce energy consumption. For example, when gas concentrations drop below a critical value and remain stable, the system will switch to lower-power backup fans to maintain air circulation in the area, rather than continuing to use high-power fans. Simultaneously, the system closes dampers unrelated to the primary ventilation path to prevent wasted airflow.

[0104] This implementation also includes an intelligent energy optimization mechanism. The system records the effectiveness and energy consumption of each ventilation adjustment and continuously optimizes the timing and magnitude of fan power adjustments. For example, if the system finds that prolonged high-power operation is not significantly improving event handling, the high-power operation period will be shortened. Furthermore, after the event, the system uses feedback data to develop more energy-efficient ventilation strategies for similar events in the future.

[0105] This multi-stage ventilation mode ensures mine safety in emergency situations while minimizing long-term energy consumption. By combining safety and energy efficiency, this implementation not only improves the economic efficiency of mine management but also enhances the intelligence level of the ventilation system.

[0106] The present invention uses a variety of sensors to monitor the mine environment in real time, and combines deep learning and reinforcement learning algorithms to achieve precise ventilation control. When an emergency (such as a gas leak or fire) occurs, the system can quickly detect and locate the problem area, and dynamically adjust the ventilation mode to concentrate the main airflow in the affected area. At the same time, through precise control of regional classification, the shutdown of the entire mine or the full evacuation of personnel is avoided, and local problems are quickly resolved. This rapid response capability significantly improves the safety of the mine, reduces the risk of the event worsening and spreading, and maximizes the protection of the lives of miners. In addition, the precise control strategy avoids unnecessary airflow disturbances, reduces the diffusion of gas in the mine, and helps prevent the occurrence of secondary disasters.

[0107] By adopting time-series deep learning models such as LSTM, the present invention enables the system to predict the development trend of events in advance and adopt preventive control strategies to reduce the occurrence of potential risks. For example, after detecting a decrease in wind speed and a slow increase in gas concentration, the system can increase the fan power in advance and open the backup damper for ventilation to prevent the gas concentration from rising further. The combination of prediction and prevention enables the system to take emergency measures before an event occurs, buying more time to deal with potential dangers. In addition, the system also supports real-time adjustment strategies. If the initial prediction deviates, the algorithm can quickly correct and redeploy the best solution. This capability makes the mine's response to emergencies more flexible and efficient, avoiding the limitations of traditional management with delayed reactions.

[0108] The present invention adopts a staged ventilation mode. The above scheme not only ensures the safe operation of the mine, but also significantly reduces energy consumption. At the beginning of the event, the system runs the ventilation equipment at high power to quickly reduce the concentration of harmful gases; when the air quality gradually recovers, the system will reduce the fan power and close some unnecessary dampers to reduce unnecessary airflow waste. By recording the energy consumption data and response effects of historical events, this scheme continuously optimizes the energy consumption level of each stage to achieve the best balance between safety and energy efficiency. Compared with the traditional full-power ventilation scheme, this multi-stage mode significantly reduces electricity consumption and equipment wear, reduces the long-term operating costs of the mine, and improves the economy and sustainability of the ventilation system.

[0109] The above description is merely illustrative of certain exemplary embodiments of the present invention. It goes without saying that those skilled in the art will be able to modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the above drawings and description are illustrative in nature and should not be construed as limiting the scope of protection of the claims.

Claims

1. A method for coal mine ventilation optimization and air quality control based on adaptive learning, characterized in that: The steps include: Various sensors are deployed in the mine to collect and monitor key environmental parameters in real time, and the collected data is uploaded to the central control system; Compare the collected real-time data with historical data, and divide it into normal operating condition data and abnormal operating condition data according to the environmental status. Use the K-means clustering classification algorithm to identify potential extreme conditions to ensure that subsequent algorithm optimization can cover the control requirements under different operating conditions; Based on the mine's historical abnormal events, a library of ventilation plans and emergency response strategies is established, including multi-layer ventilation modes to deal with gas surges, fires, and equipment failures. The plan library is regularly updated based on new data. Using reinforcement learning algorithms, a preliminary ventilation strategy is generated based on real-time data and recommendations from a pre-plan library. This strategy is continuously optimized through interactive feedback from historical data, and the effects of adjustments are simulated within the mine's airflow and air quality models to obtain the optimal preliminary strategy. Multiple anomaly detection algorithms are used to analyze real-time mine data in parallel. Once a deviation from normal data patterns is detected, an alarm is triggered, the current preliminary ventilation strategy is interrupted, and emergency response mode is entered. After an emergency occurs, a deep learning algorithm is activated to reassess the current status and, combined with historical abnormal data, predict the development trend of the incident. Based on the predicted results, the ventilation mode is dynamically adjusted in real time to ensure that the air quality in the mine is restored to a safe range in the shortest possible time while minimizing energy consumption. The specific steps for generating a preliminary ventilation strategy using reinforcement learning algorithms are as follows: In the reinforcement learning framework, the state space, action space, and reward mechanism are first defined. The state space includes the environmental parameters within the mine, the action space refers to the control instructions that can be executed by the ventilation system, and the reward mechanism is designed based on the ventilation effect and safety of the system. Based on current real-time data, the reinforcement learning algorithm reads historical cases and emergency plans that are closest to the current environment from the plan library as a reference for generating preliminary strategies; After generating the initial strategy, we conducted virtual tests in a mine airflow and air quality simulation model to verify the effectiveness of the strategy under various extreme scenarios by simulating different working conditions. Once the initial strategy passes simulation testing, it is deployed to the actual mine ventilation system and its effectiveness is monitored in real time. If the ventilation results meet expectations and safety standards, the strategy is marked as qualified and recorded in the plan library for future reference. If deviations occur during execution, the algorithm adjusts the strategy based on the new data to ensure that the ventilation system always maintains optimal performance. The specific steps for using multiple anomaly monitoring algorithms to analyze real-time mine data in parallel are as follows: During the anomaly detection process, the mine environmental parameters collected from the sensors is input into the algorithm module, where Indicates the i In order to ensure the stability and efficiency of the algorithm, the collected data is subjected to feature extraction and standardization. represents the data matrix, where m is the sample size, n It is the dimension of the monitored environmental parameters, and the standardized formula is: ,in, and The parameters are j The mean and standard deviation of the standardized data It is passed to the subsequent detection algorithm to eliminate the influence of different parameter scales; The isolation forest algorithm is used to construct a random tree to evaluate the isolation degree of the sample points. , Isolation Forest will generate T random trees, the depth of each tree represents the number of steps required to completely separate the sample points. The depth in the t-th tree is , then the average depth of all trees is: , combined with the expected tree depth , anomaly score of isolation forest The calculation is as follows: , by setting the threshold ,like , it is marked as abnormal.

2. The method for coal mine ventilation optimization and air quality control based on adaptive learning according to claim 1, characterized in that: The specific steps of using K-means clustering algorithm to classify mine environment data are as follows: Extract key features from real-time data collected by sensors, perform outlier detection and noise filtering on sensor data, and eliminate data anomalies caused by sensor false alarms or failures; The preprocessed data also needs to be normalized to map data of different units and dimensions to a unified scale to ensure that the weights of various parameters are balanced in cluster analysis. Finally, the processed data is merged with historical data to form a complete data set. Divide the data set into K clusters. After determining the value of K, select the initial cluster centers through random initialization or K-means++ algorithm to reduce the risk of the algorithm falling into a local optimal solution. Assign each piece of data to the nearest cluster center based on the Euclidean distance, then recalculate the cluster center position based on all the data points assigned to the cluster, and repeat the process of data assignment and cluster center update until the position of the cluster center no longer changes significantly or the set number of iterations is reached; After clustering is completed, different labels are assigned to each cluster, and the characteristic centers of each cluster are analyzed. Isolated data that cannot be classified into regular clusters are marked as abnormal conditions, triggering further detection and alarms.

3. The method for coal mine ventilation optimization and air quality control based on adaptive learning according to claim 1, characterized in that: The specific steps for establishing a library of multiple ventilation plans and emergency response strategies are as follows: Gather data on various abnormal events that have occurred in the mine in the past; According to the severity and impact scope of different abnormal events, the events are divided into multiple levels; In the process of building the emergency plan library, all ventilation modes and emergency response measures are coded in the form of rules to form a modular strategy library; To ensure that the contingency plan library can adapt to the ever-changing mine environment, new data and the latest safety incident feedback are regularly incorporated into analysis and existing contingency plans are optimized.

4. The method for coal mine ventilation optimization and air quality control based on adaptive learning according to claim 1, characterized in that: A class of support vector machines is used to construct boundaries in high-dimensional space to detect abnormal data points. Given a data set , SVM finds a hyperplane Separate normal data from abnormal data, and optimize the goal to be: , ,in, is the parameter that controls the abnormality rate, It is a slack variable used to process data points on the soft boundary. When new sample data falls outside the hyperplane, it is marked as an outlier. Through parallel analysis of isolation forest and SVM, two independent anomaly scores are calculated for each data, namely and ,In order to achieve multi-algorithm fusion, a weighted anomaly scoring strategy is adopted. The specific expression is: ,in is the algorithm weight parameter, reflecting the importance of isolation forest and SVM in different scenarios. , the system immediately triggers an alarm, interrupts the current ventilation strategy, enters the emergency response mode, and executes the corresponding plan.

5. The method for coal mine ventilation optimization and air quality control based on adaptive learning according to claim 1, characterized in that: The specific steps for using deep learning algorithms to reassess emergencies and dynamically adjust ventilation modes are as follows: When an emergency is detected in the mine, the latest real-time data is collected immediately; After the initial status assessment, predict the development trend of the emergency by using similar cases in historical data; Generate and execute real-time ventilation adjustment instructions based on the prediction results; After adjusting the ventilation mode, feedback data is continuously collected and analyzed to ensure that the air quality gradually returns to a safe range. If any abnormal changes are detected or the strategy execution does not meet expectations, the deep learning algorithm will immediately re-evaluate the situation and automatically optimize the adjustment instructions to achieve secondary or multiple iterative optimization. In addition, the data and response process of this emergency will be stored in the historical database for reference in future events and to improve the model's adaptive capabilities.

Images