A deep learning-based intelligent regulation and control method for a mine ventilation system

Abstract

The present application belongs to the technical field of data acquisition and monitoring and control, and particularly relates to a mine ventilation system intelligent regulation and control method based on deep learning, which comprises the following steps: obtaining time series monitoring data and performing spatial sorting, obtaining harmful gas concentration data and wind speed data; obtaining a flow field response correction factor based on the extreme value and time accumulation characteristics of the wind speed data, and correcting the original conduction time; obtaining a topological dynamic complexity index based on the probability distribution of the corrected conduction time, and dynamically obtaining a real-time hidden layer capacity; inputting data into a hidden layer configured with the real-time hidden layer capacity to output a prediction value, which is used for triggering an early warning and generating a regulation and control instruction. The present application solves the technical problem that the existing model cannot adapt to the dynamic changes of the mine ventilation network topology, and improves the generalization ability and intelligent regulation and control precision of the model.

Description

Technical Field

[0001] This invention relates to the field of data acquisition and monitoring control technology. More specifically, this invention relates to an intelligent control method for mine ventilation systems based on deep learning. Background Technology

[0002] Mine ventilation systems are core infrastructure for ensuring safe underground production. Through the coordinated control of fans and air doors, they continuously deliver fresh air to the working face and roadways, diluting and expelling harmful substances such as methane. The control precision of the ventilation system directly affects the real-time control of harmful gas concentrations and energy efficiency. Therefore, integrating multi-source sensor monitoring information to achieve intelligent control of the ventilation network is of great significance for ensuring safety and improving production efficiency.

[0003] Existing intelligent control technologies for mine ventilation typically employ deep learning models, such as Long Short-Term Memory (LSTM) networks, for time-series prediction. These technologies utilize gating mechanisms in hidden layers to extract deep features from data collected by multiple sensors, learning their nonlinear patterns. Real-time monitoring data is then input into the model to predict air demand and automatically adjust ventilation actuators, attempting to achieve on-demand air supply and closed-loop control.

[0004] However, existing technologies have a core flaw: the models used all rely on a fixed hidden layer capacity to process time-series data. In actual mining processes, as the working face advances or roadways are connected, the physical topology of the ventilation network undergoes complex dynamic changes, leading to a weakening or even disappearance of the spatial dependencies between sensors. Faced with dynamic topology changes, models with fixed capacity cannot adaptively adjust the network structure and are prone to overfitting due to learning spurious correlations from historical topologies. This results in a significant decrease in the model's ability to extract features from core data, making it unable to provide accurate parameter predictions under complex working conditions and seriously affecting the reliability and safety of intelligent control systems. Summary of the Invention

[0005] To address the technical problem in existing technologies where models with fixed hidden layer capacity cannot adapt to dynamic changes in ventilation network topology, leading to decreased feature extraction capabilities and an inability to provide accurate predictions and reliable control, this invention provides a deep learning-based intelligent control method for mine ventilation systems. The method includes: acquiring time-series monitoring data covering the total time span of the coverage data and spatially sorting it; acquiring harmful gas concentration data and wind speed data; detecting the wind speed data, extracting local maxima and minima, and calculating the global mean, increasing interval time sum, and decreasing interval time sum of wind speed; and calculating the wind speed based on the local maxima, local minima, the global mean wind speed, the increasing interval time sum, and the decreasing interval time sum. The flow field response correction factor is obtained by summing the data; the original conduction time is obtained based on the cross-correlation values ​​of two sets of adjacent harmful gas concentration data at each time offset; the original conduction time is corrected by the flow field response correction factor to obtain the corrected conduction time; the topology dynamic complexity index is obtained based on the probability distribution value of the corrected conduction time, and the real-time hidden layer capacity is obtained based on the topology dynamic complexity index; the harmful gas concentration data and the wind speed data are input into the hidden layer configured with the real-time hidden layer capacity, and the required air volume prediction value and the harmful gas concentration prediction value are output; the alarm device is triggered based on the harmful gas concentration prediction value, and the fan speed control command and damper opening control command are generated based on the required air volume prediction value.

[0006] Preferably, the step of acquiring time-series monitoring data covering the total time span of the coverage data and spatially sorting it to obtain hazardous gas concentration data and wind speed data includes: deploying multiple types of sensors at equal intervals in key areas of the mine ventilation system, and using a collaborative layout to deploy wind speed sensors between adjacent hazardous gas concentration sensors; controlling each of the multiple types of sensors to continuously collect operating condition data to obtain time-series monitoring data; filling missing data using linear interpolation, removing outliers using a three-standard-deviation criterion, and standardizing all time-series monitoring data after removing outliers; and spatially sorting all standardized time-series monitoring data according to the distance of the sensor deployment location from the wellhead to obtain hazardous gas concentration data and wind speed data.

[0007] Preferably, obtaining the flow field response correction factor based on each local maxima, each local minima, the global mean wind speed, the cumulative sum of time over the increasing interval, and the cumulative sum of time over the decreasing interval includes: calculating an amplitude asymmetry index based on the cumulative difference between each local maxima, each local minima, and the global mean wind speed; calculating a trend fluctuation dominance index based on the difference between the cumulative sum of time over the increasing interval and the cumulative sum of time over the decreasing interval; and calculating the product of the amplitude asymmetry index and the trend fluctuation dominance index to obtain the flow field response correction factor.

[0008] Preferably, the calculation formulas for the amplitude asymmetry indicator and the trend fluctuation dominance indicator are as follows: ; In the formula, This is an indicator of amplitude asymmetry. It is the hyperbolic tangent function; This represents the total number of local maxima. Number the maximum points; For the first Local maxima; This represents the global average wind speed. This is the normalized wind speed constant; This represents the total number of local minima. Number the local minimum points; For the first Local minimum; Indicates taking the absolute value; It is a leading indicator of trend fluctuations; This is the cumulative sum over an increasing interval of time. This is the cumulative sum over a decreasing interval; This is the normalized time constant.

[0009] Preferably, the step of obtaining the original conduction time based on the cross-correlation value of two sets of adjacent harmful gas concentration data at each time offset, and correcting the original conduction time by a flow field response correction factor to obtain the corrected conduction time includes: calculating the cross-correlation value of two sets of adjacent harmful gas concentration data at each time offset; extracting the time offset corresponding to the cross-correlation value reaching its maximum value as the original conduction time; calculating the difference between I and the flow field response correction factor, and calculating the product of the difference and the original conduction time to obtain the corrected conduction time.

[0010] Preferably, the method for obtaining the probability distribution value of the corrected conduction time includes: dividing all global corrected conduction times into multiple equidistant intervals; and statistically analyzing the probability distribution value of the corrected conduction time falling into each interval.

[0011] Preferably, the topological dynamic complexity index is obtained based on the probability distribution value of the corrected propagation time, including: In the formula, This is a metric for topological dynamic complexity. It is the hyperbolic tangent function; This represents the total number of equally spaced intervals. For interval numbering; These are probability distribution values; This is the operation for the natural logarithm.

[0012] Preferably, obtaining the real-time hidden layer capacity based on the topology dynamic complexity index includes: calculating the sum of 1 and the topology dynamic complexity index, calculating the product of the sum and a preset initial hidden layer capacity, rounding the product to the nearest integer, and using the rounded result as the real-time hidden layer capacity.

[0013] Preferably, inputting the harmful gas concentration data and the wind speed data into a hidden layer configured with the real-time hidden layer capacity, and outputting the predicted air volume and the predicted harmful gas concentration, includes: constructing a sliding time window to segment the time-series data, generating multiple sets of historical operating condition input feature vectors, using the actual air volume at future times as the actual air volume label, and using the actual harmful gas concentration at future times as the actual harmful gas concentration label; initializing the network weights of the long short-term memory network, inputting the input feature vectors into the hidden layer configured with the real-time hidden layer capacity, and outputting the predicted air volume and the predicted harmful gas concentration.

[0014] Preferably, the alarm device is triggered based on the predicted value of the harmful gas concentration, and the fan speed control command and damper opening control command are generated based on the predicted air demand value, including: during the mine operation phase, inputting real-time monitoring data stream into the model that has completed model training, and outputting the predicted value of the harmful gas concentration at future times; triggering the alarm device when the predicted value of the harmful gas concentration exceeds a preset safety threshold; generating the fan speed control command and damper opening control command according to the predicted value of the air demand at future times, and sending them to the underground actuator.

[0015] The beneficial effects of this invention are as follows: (1) This invention acquires time-series monitoring data and performs spatial sorting, extracts extreme values ​​and time accumulation features of wind speed data to construct flow field response correction factors, and then combines cross-correlation values ​​to dynamically correct the original conduction time; thus, this invention can accurately quantify the driving or blocking effect of real airflow fluctuations on the propagation of harmful gases, eliminate the calculation error of conduction delay caused by ignoring the dynamic changes of the flow field in traditional methods, reflect the physical transmission process of gas between adjacent measuring points, and provide a reliable data basis for subsequently determining the topological changes of the ventilation network.

[0016] (2) The present invention evaluates the topology complexity index based on the probability distribution of the modified propagation time, and dynamically obtains the real-time hidden layer capacity of the deep learning model accordingly. Thus, the present invention fundamentally solves the technical problem that the model with fixed hidden layer capacity in the prior art cannot adapt to the dynamic changes in the topology of the mine ventilation network. By making the network structure complexity adaptively match the physical topology complexity of the current working condition, the false association generated by the model learning the historical failed topology is avoided, thereby eliminating the overfitting phenomenon and enhancing the deep feature extraction capability and generalization performance of the model under complex and variable working conditions.

[0017] (3) The present invention inputs the harmful gas concentration data and wind speed data into the hidden layer configured with real-time hidden layer capacity for prediction, and triggers early warning and generates control commands for fans and air doors based on the accurate prediction results; thus, the present invention relies on a highly adaptable prediction model that dynamically matches the physical topology to ensure the accuracy of prediction of key environmental parameters under complex working conditions in the mine, thereby realizing early warning of abnormalities in the ventilation system, on-demand air supply and closed-loop intelligent control, effectively improving the guarantee capability of mine safety production and the accuracy of ventilation control. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating an intelligent control method for a mine ventilation system based on deep learning, as described in this invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0021] This invention discloses an intelligent control method for mine ventilation systems based on deep learning, referring to... Figure 1 This includes steps S1 to S3: S1. Obtain time-series monitoring data covering the total time span of the coverage data and perform spatial sorting to obtain harmful gas concentration data and wind speed data.

[0022] It should be noted that, as the physical topology of the mine ventilation network changes continuously as the working face advances, the spatial dependence between sensors weakens or even disappears. Therefore, this invention acquires long-sequence data covering the entire coal mining cycle and performs spatial sorting to establish the correspondence between discrete data and physical topology, thereby providing a structured data foundation for the subsequent extraction of spatial transmission features.

[0023] Specifically, multiple types of sensors are deployed at equal intervals in key areas of the mine ventilation system. A collaborative layout is used, placing wind speed sensors between adjacent harmful gas concentration sensors. Each type of sensor is controlled to continuously collect operating data at a preset frequency to obtain time-series monitoring data covering the entire coal mining cycle. In this embodiment, the preset frequency is once every thirty seconds, and the complete coal mining cycle is 7 days. Missing data caused by equipment failure is filled using linear interpolation. A three-standard-deviation criterion is used to remove outliers caused by sensor electromagnetic interference and instantaneous fluctuations in underground operating conditions. All time-series monitoring data after removing outliers are standardized to eliminate numerical deviations caused by differences in monitoring dimensions and units between different types of sensors.

[0024] Furthermore, based on the distance of the sensor deployment location from the wellhead, all standardized time-series monitoring data are spatially sorted. The extracted continuous monitoring values ​​of harmful gases are arranged in chronological order to construct a one-dimensional time-series sequence, and this spatially sorted one-dimensional time-series sequence is recorded as the harmful gas concentration data. The extracted continuous wind speed monitoring values ​​are arranged in chronological order to construct a one-dimensional time series. This spatially sorted one-dimensional time series is then recorded as wind speed data. ,in The number of sensor locations. The sensors are numbered sequentially from closest to furthest from the wellhead. This provides a structured data foundation for subsequent analysis of spatial dependencies between sensors and quantification of dynamic changes in the ventilation network topology, serving as the sampling time.

[0025] S2. Extract extreme values ​​and time accumulation features based on wind speed data to obtain flow field response correction factors. Use the flow field response correction factors to correct the original conduction time. Evaluate the topology dynamic complexity index based on the probability distribution of the corrected conduction time. Dynamically obtain the real-time hidden layer capacity accordingly.

[0026] It should be noted that, since traditional data processing models use a fixed hidden layer capacity, they are prone to overfitting due to spurious correlations in historical time series data when the topology changes dynamically. Therefore, this invention constructs a flow field response correction factor to dynamically adjust the gas conduction time and uses the corrected spatiotemporal conduction features to quantify the topological complexity, so that the network structure can adaptively match the changes in the ventilation network and enhance the model's generalization performance.

[0027] Specifically, the wind speed data is calculated. For the first-order difference sequence between adjacent sampling times, extreme points are identified, local maxima and local minima are extracted, and the total number of local maxima is recorded. and the total number of local minima The global mean wind speed is obtained by summing the values ​​of all wind speed data within the total time span and dividing by the total number of sampling points. The time intervals during which the first-order difference sequence is continuously greater than zero are defined as increasing intervals. The durations of all increasing intervals are summed to obtain the sum of the increasing interval times. The time intervals during which the first-order difference sequence is continuously less than zero are defined as decreasing intervals. The durations of all decreasing intervals are summed to obtain the cumulative sum of the decreasing interval times. .

[0028] Furthermore, the amplitude asymmetry index is calculated. The formula for its calculation is: In the formula, This is an indicator of amplitude asymmetry. It is the hyperbolic tangent function; This represents the total number of local maxima. Number the maximum points; For the first Local maxima; This represents the global average wind speed. This is the normalized wind speed constant, with a value equal to 1 and the same unit as the wind speed data. This represents the total number of local minima. Number the local minimum points; For the first Local minimum; This indicates taking the absolute value.

[0029] The calculation formula uses the cumulative difference between the extreme points and the mean to map the amplitude fluctuation characteristics. When the total positive deviation of the local maximum being greater than the mean increases, the cumulative difference becomes larger, leading to a larger value input into the hyperbolic tangent function, thus increasing the amplitude asymmetry index. The increase means that the intensity of the positive speed increase fluctuation is much greater than the negative speed decrease fluctuation, which quantifies the strong driving effect of wind flow on gas propagation.

[0030] Furthermore, calculate the dominant indicator of trend fluctuation. The formula for its calculation is: In the formula, It is a leading indicator of trend fluctuations; It is the hyperbolic tangent function; This is the cumulative sum of time over an increasing interval, in seconds; This is the cumulative sum of time over a decreasing interval, in seconds; This is the normalized time constant, with a value of 3600 and a unit of seconds.

[0031] The calculation formula maps the wind speed evolution trend by calculating the distribution of time differences between increasing and decreasing states. When the cumulative time difference in the increasing interval is greater than the cumulative time difference in the decreasing interval, the time difference becomes larger, making the trend fluctuation dominant indicator. The fact that the wind speed is becoming positive and increasing in magnitude indicates that the change in wind speed is dominated by an increasing trend and has a long duration.

[0032] Furthermore, the flow field response correction factor is calculated. The formula for its calculation is: In the formula, This is a flow field response correction factor; This is an indicator of amplitude asymmetry. It is a leading indicator for trend fluctuations.

[0033] This calculation formula integrates amplitude characteristics and trend characteristics; when both the amplitude asymmetry indicator and the trend fluctuation dominance indicator are large positive numbers, their product increases, thus increasing the flow field response correction factor. An increase in wind speed indicates that the positive increase in wind speed is dominant and lasts for a long time, representing a strong driving state of the flow field.

[0034] It should be noted that, due to the physical time delay in the conduction of harmful gases in mine roadways, the one-dimensional time series collected by adjacent sensors will be misaligned in waveform, which will lead to quantification errors in the actual driving effect of airflow. Therefore, this invention introduces a time offset to dynamically shift the time series, and finds the shift scale with the highest waveform overlap through cross-correlation calculation, thereby accurately locking the real physical conduction delay.

[0035] Specifically, the time offset between two adjacent sets of harmful gas concentration data is calculated. Cross-correlation value under The formula for its calculation is: In the formula, Time offset The cross-correlation value under the following conditions; It is a cross-correlation function; , The first The first sensor collected the data. Harmful gas concentration data at each sampling time. For data arrive A one-dimensional time series consisting of harmful gas concentration data at all sampling times; , The first The data collected by the first sensor , Harmful gas concentration data at each sampling time. For data arrive A one-dimensional time series consisting of harmful gas concentration data at all sampling times.

[0036] The calculation formula uses a cross-correlation function to directly perform a global comparison of two sets of one-dimensional time series, quantifying the overall waveform similarity of the two sets of one-dimensional time series at a given time offset. When the waveform overlap of the two sets of one-dimensional time series at a specific time offset increases, the similarity increases, resulting in a higher cross-correlation value. An increase in the time offset means that the time offset precisely aligns with the overall gas transport delay between the two points; this invention iterates through the time offset. The range of values ​​to extract the cross-correlation values. The time offset corresponding to the maximum value As the original conduction time This reflects the basic time delay required for the propagation of concentration changes between two points.

[0037] Among them, time offset The value ranges from zero to the upper limit of the constraint, which is the time offset scaling factor. Total time span of data The product of time offset scaling factor Based on statistical data from historical gas diffusion experiments, the empirical range is: In this embodiment, the time offset scaling factor is A value of 0.6 can indirectly avoid calculation distortion caused by excessively short overlapping data while ensuring that the search range covers the effective lag interval; in other embodiments, implementers can set the time offset ratio coefficient according to the longest path size of the mine ventilation network.

[0038] Furthermore, based on the original conduction time Calculate the corrected conduction time The formula for its calculation is: In the formula, To correct for conduction time; This is the original conduction time; This is the flow field response correction factor.

[0039] The calculation formula dynamically scales the original physical delay using the flow field response characteristics; when the flow field is in a strongly driven state, it leads to a flow field response correction factor. When the value is greater than zero and increases, the scaling factor decreases, thus correcting the propagation time. The smaller size means that the physical airflow accelerates the actual propagation process of the gas, achieving calibration of the conduction time to the real working conditions.

[0040] It should be noted that since the dispersion of the correction transmission time directly reflects the physical stability of the overall network topology, the more dispersed the transmission time between adjacent nodes, the more disordered the global transmission relationship becomes. Therefore, this invention constructs a topology dynamic complexity index based on the information entropy principle to quantitatively characterize the structural complexity of the ventilation network under the current operating conditions.

[0041] Specifically, the global correction propagation time is divided into... The statistically corrected propagation time falls within the equidistant intervals. probability distribution values ​​of each interval , .

[0042] Furthermore, the dynamic complexity index of the topology is calculated. The formula for its calculation is: In the formula, This is a metric for topological dynamic complexity. It is the hyperbolic tangent function; This represents the total number of equally spaced intervals. For interval numbering; These are probability distribution values; This is an operation on the natural logarithm; it should be noted that when season .

[0043] The calculation formula measures the degree of disorder in the system by summing the negative logarithms. When the distribution of the corrected propagation time is more uniform and discrete across different intervals, the difference in probability distribution values ​​decreases, leading to a larger sum of information entropy and thus increasing the topological dynamic complexity index after mapping. The increase in size indicates that the ventilation network topology changes extremely frequently, enabling precise quantification of the global topology complexity.

[0044] Furthermore, calculate the real-time hidden layer capacity. The formula for its calculation is: In the formula, For the real-time hidden layer capacity; This is the initial hidden layer capacity; This is a metric for topological dynamic complexity. This is the rounding function.

[0045] This calculation formula couples topological features to network hyperparameters; when the topological dynamic complexity index When the value increases, the product terms increase overall, thus increasing the capacity of the real-time hidden layer. A significant increase means that, under complex and varied network structures, the model needs more neurons to expand the feature extraction space and avoid feature loss.

[0046] Wherein, the initial hidden layer capacity The values ​​were obtained based on a grid search method combined with historical data size assessment, and the range was [value range missing]. In this embodiment, the initial hidden layer capacity is... Setting it to 64 can balance computational efficiency and basic feature extraction capability of the model under the baseline steady-state conditions; in other embodiments, implementers can set the initial hidden layer capacity according to the dimension of the initial input features.

[0047] S3. Input the harmful gas concentration data and the wind speed data into the hidden layer configured with the real-time hidden layer capacity for prediction, and perform abnormal early warning and closed-loop control of the fan and damper based on the output harmful gas concentration prediction value and air volume prediction value.

[0048] It should be noted that, due to the nonlinear coupling characteristics of multiple fans and multiple air doors in the mine ventilation system, conventional static control strategies are difficult to achieve advance adjustment and on-demand air supply. Therefore, this invention is based on a predictive model that has completed structural adjustment, combined with a weighted loss function for backpropagation training, to output a high-precision future state prediction, thereby ensuring the reliable execution of closed-loop intelligent control.

[0049] Specifically, a sliding time window is constructed to segment the time-series data, generating multiple sets of historical operating condition input feature vectors. The actual air demand at future moments is used as the actual air demand label, and the actual harmful gas concentration at future moments is used as the actual harmful gas concentration label. The network weights of the Long Short-Term Memory network are initialized, and the input feature vectors are input into a network configured with real-time hidden layer capacity. In the hidden layer, the corresponding predicted air volume and predicted concentration of harmful gases are output.

[0050] Furthermore, during the mine operation phase, real-time monitoring data streams are input into the model that has completed model training, outputting predicted values ​​of harmful gas concentrations for future moments, and triggering an alarm device when the predicted value of harmful gas concentration exceeds a preset safety threshold; at the same time, the intelligent control decision module generates fan speed control commands and damper opening control commands based on the predicted air volume for future moments, and sends them to the underground actuators to achieve closed-loop control.

[0051] The empirical range of the preset security threshold is as follows: In this embodiment, the preset security threshold is set to Values As an early warning trigger line, it can reserve sufficient safety time margin for the mechanical response of the fan and damper and the physical transmission of airflow; in other embodiments, the implementers can set preset safety thresholds according to the mine gas level and the specific location of the return airway.

Claims

1. A method for intelligent control of mine ventilation systems based on deep learning, characterized in that, include: Acquire time-series monitoring data covering the total time span of the coverage data and perform spatial sorting to obtain harmful gas concentration data and wind speed data; The wind speed data is detected, local maxima and local minima are extracted, and the global mean of wind speed, the time sum of increasing intervals, and the time sum of decreasing intervals are calculated. Based on each local maximum, each local minimum, the global mean wind speed, the time summation of the increasing interval, and the time summation of the decreasing interval, the flow field response correction factor is obtained. The original conduction time is obtained based on the cross-correlation values ​​of two adjacent sets of harmful gas concentration data at each time offset; The original conduction time is corrected by a flow field response correction factor to obtain the corrected conduction time; The topological dynamic complexity index is obtained based on the probability distribution value of the corrected propagation time, and the real-time hidden layer capacity is obtained based on the topological dynamic complexity index. The harmful gas concentration data and the wind speed data are input into the hidden layer configured with the real-time hidden layer capacity, and the required air volume prediction value and the harmful gas concentration prediction value are output. The alarm device is triggered based on the predicted value of the harmful gas concentration, and the fan speed control command and damper opening control command are generated based on the predicted value of the required air volume.

2. The intelligent control method for a mine ventilation system based on deep learning according to claim 1, characterized in that, The process of acquiring time-series monitoring data covering the total time span of the coverage data and spatially sorting it to obtain harmful gas concentration data and wind speed data includes: Multiple types of sensors are deployed at equal intervals in key areas of the mine ventilation system, and wind speed sensors are placed between adjacent harmful gas concentration sensors using a collaborative layout method. Control each of the various types of sensors to continuously collect operating condition data and obtain time-series monitoring data; Linear interpolation was used to fill in the missing data, and outliers were removed using the three-standard-deviation criterion. All time-series monitoring data after removing outliers were standardized. Based on the distance of the sensor deployment location from the wellhead, all standardized time-series monitoring data are spatially sorted to obtain harmful gas concentration data and wind speed data.

3. The intelligent control method for a mine ventilation system based on deep learning according to claim 1, characterized in that, The flow field response correction factor is obtained based on each local maxima, each local minima, the global mean wind speed, the sum of time increments over increasing intervals, and the sum of time increments over decreasing intervals, including: The amplitude asymmetry index is calculated based on the cumulative difference between each local maximum, each local minimum and the global mean wind speed. The trend fluctuation dominance index is calculated based on the difference between the sum of time in the increasing interval and the sum of time in the decreasing interval. The flow field response correction factor is obtained by multiplying the amplitude asymmetry index and the trend fluctuation dominance index.

4. The intelligent control method for a mine ventilation system based on deep learning according to claim 3, characterized in that, The formulas for calculating the amplitude asymmetry indicator and the trend fluctuation dominance indicator are as follows: ; In the formula, This is an indicator of amplitude asymmetry. It is the hyperbolic tangent function; This represents the total number of local maxima. Number the maximum points; For the first Local maxima; This represents the global average wind speed. This is the normalized wind speed constant; This represents the total number of local minima. Number the local minimum points; For the first Local minimum; Indicates taking the absolute value; It is a leading indicator of trend fluctuations; This is the cumulative sum over an increasing interval of time. This is the cumulative sum over a decreasing interval; This is the normalized time constant.

5. The intelligent control method for a mine ventilation system based on deep learning according to claim 1, characterized in that, The original conduction time is obtained based on the cross-correlation value of two sets of adjacent harmful gas concentration data at each time offset; The original conduction time is corrected using a flow field response correction factor to obtain the corrected conduction time, including: Calculate the cross-correlation value of two sets of adjacent harmful gas concentration data at each time offset; extract the time offset corresponding to the maximum cross-correlation value as the original conduction time; Calculate the difference between 1 and the flow field response correction factor, and calculate the product of the difference and the original conduction time to obtain the corrected conduction time.

6. The intelligent control method for a mine ventilation system based on deep learning according to claim 1, characterized in that, The method for obtaining the probability distribution value of the corrected propagation time includes: Divide all global correction propagation times into multiple equidistant intervals; statistically analyze the probability distribution of correction propagation times falling into each interval.

7. The intelligent control method for a mine ventilation system based on deep learning according to claim 1, characterized in that, The topological dynamic complexity index is obtained based on the probability distribution value of the corrected propagation time, including: In the formula, This is a metric for topological dynamic complexity. It is the hyperbolic tangent function; This represents the total number of equally spaced intervals. For interval numbering; These are probability distribution values; This is the operation for the natural logarithm.

8. The intelligent control method for a mine ventilation system based on deep learning according to claim 1, characterized in that, Obtaining the real-time hidden layer capacity based on the aforementioned topological dynamic complexity index includes: Calculate the sum of 1 and the topology dynamic complexity index, calculate the product of the sum and the preset initial hidden layer capacity, round the product to the nearest integer, and use the rounded result as the real-time hidden layer capacity.

9. The intelligent control method for a mine ventilation system based on deep learning according to claim 1, characterized in that, The harmful gas concentration data and the wind speed data are input into a hidden layer configured with the real-time hidden layer capacity, and the required air volume prediction value and the harmful gas concentration prediction value are output, including: A sliding time window is constructed to segment the time series data, generating multiple sets of historical operating condition input feature vectors. The actual air demand at future moments is used as the actual air demand label, and the actual harmful gas concentration at future moments is used as the actual harmful gas concentration label. Initialize the network weights of the Long Short-Term Memory network, input the input feature vector into the hidden layer configured with real-time hidden layer capacity, and output the required air volume prediction value and the harmful gas concentration prediction value.

10. The intelligent control method for a mine ventilation system based on deep learning according to claim 1, characterized in that, The alarm device is triggered based on the predicted value of the harmful gas concentration, and the fan speed control command and damper opening control command are generated based on the predicted air volume demand, including: During the mine operation phase, real-time monitoring data streams are input into the model that has completed model training, and the model outputs predicted values ​​of harmful gas concentrations for future moments. The alarm device is triggered when the predicted concentration of harmful gases exceeds a preset safety threshold. Based on the predicted air demand at future times, blower speed control commands and damper opening control commands are generated and sent to the downhole actuators.

Images