A digital twin driven mine ventilation network reinforcement learning regulation system

Abstract

The application relates to the technical field of intelligent decision-making of mine ventilation networks, and discloses a digital twin driven mine ventilation network reinforcement learning regulation system. The safety benefits of the historical ventilation action, the twin state uncertainty and the reinforcement learning action value correction are limited to the same training chain for processing under the action response divergence verification of preventive ventilation segment identification, non-intervention counterfactual risk deduction and observation equivalent candidate twin state, the problems that the preventive ventilation action is misjudged as low value and the unidentifiable twin return is miswritten into the strategy in the historical sample are solved, the mine ventilation regulation strategy has more reliable training access, online execution verification and feedback closed update constraints while retaining the preposed risk response capability.

Description

Technical Field

[0001] This application relates to the field of intelligent decision-making technology for mine ventilation networks, and in particular to a digital twin-driven reinforcement learning control system for mine ventilation networks. Background Technology

[0002] Mine ventilation networks are responsible for tasks such as gas dilution, dust dispersion, thermal and humidity environment regulation, and ensuring air supply to working faces. Their operation is continuously affected by factors such as roadway extension, mining continuity, door opening and closing, local ventilation fan connection, transportation equipment movement, and adjustments to temporary ventilation structures. In actual production, dispatchers typically take preventative ventilation actions such as increasing air volume, adjusting doors, starting and stopping local ventilation fans, or correcting the main fan frequency based on trends in gas concentration, air volume decrease, abnormal pressure differences, increased working face load, and on-site maintenance arrangements. This allows potential risks to be suppressed before alarms are triggered. Existing digital twin mine ventilation systems can establish virtual ventilation networks based on roadway topology, air resistance parameters, measuring point air volume, air pressure, and gas concentration. They can also generate control strategies for fans and doors by combining simulation pre-playing, predictive models, or deep reinforcement learning. Graph neural networks can also be used to express the branch relationships of ventilation networks, while reinforcement learning can learn the value of control actions based on energy consumption, safety constraints, and ventilation needs. These technologies provide effective support for mine ventilation status perception, strategy deduction, and automated control.

[0003] In mines with relatively standardized safety management, strong experience in manual dispatching, and a small number of historical accident samples, historical ventilation data often contains a large number of stable operation records after preventive ventilation adjustments. These records can reflect the operational status after ventilation adjustments, but may not directly reflect the risk evolution process when the corresponding ventilation adjustment action was not executed. If the offline reinforcement learning model evaluates the value of the action based solely on alarm results, current measurement point compliance status, and energy consumption changes, it may be difficult to accurately distinguish the implicit safety benefits corresponding to preventive ventilation adjustment actions. At the same time, the mine ventilation resistance coefficient, leakage path, and local execution response are affected by roadway conditions, structure status, and measurement point coverage. Multiple candidate digital twin states may all meet the error requirements at the current measurement point, but after inputting the same control action, they will form different downstream air volume, pressure difference, and gas dilution responses. Therefore, when using digital twin results to train reinforcement learning control strategies, it is still necessary to further address the problem that insufficient expression of the safety benefits of historical preventive actions and insufficient consistency of action responses in candidate twin states jointly affect the strategy value assessment, in order to improve the safety and reliability of mine ventilation control strategies in real-world execution scenarios. Summary of the Invention

[0004] This application proposes a digital twin-driven reinforcement learning control system for mine ventilation networks to address the problems mentioned in the background art.

[0005] To achieve the above objectives, this application adopts the following technical solution: a digital twin-driven mine ventilation network reinforcement learning control system, comprising: a segment recognition module, which acquires mine ventilation operation records and digital twin ventilation network status, adopts event anchor alignment and influence branch constraint method to convert the segment-level risk before and after historical ventilation adjustment actions into a homogeneous risk representation, and corrects the comparison process with interference coverage results to form a preventive ventilation adjustment segment confidence result;

[0006] The counterfactual deduction module uses the confidence results of the preventive ventilation adjustment segment to locate the initial counterfactual state, removes the corresponding historical ventilation adjustment actions, reconstructs the sequence of uninterrupted counterfactual actions, and generates uninterrupted counterfactual risk results and implicit safety benefit results through ventilation network balancing playback;

[0007] The response identification module takes the non-intervention counterfactual risk result, constructs a set of observation equivalent candidate twin states that meet the measurement point error allowance and ventilation network physical constraints, performs action response divergence calculation between the observation equivalent candidate twin states to perform the same candidate control action, and forms action response divergence result and action response identifiable allowance result.

[0008] The strategy correction module uses an action value correction permission control method to process implicit safety benefit results, action response divergence results, and action response identifiable permission results, generate action value correction permission results, correct reinforcement learning action values ​​when the action value correction permission result is in the permission update state, and output mine ventilation control strategy.

[0009] The beneficial effects of this invention are as follows:

[0010] This invention solves the problems of preventive ventilation adjustment actions being misjudged as low-value in historical samples and unidentifiable twin rewards being mistakenly written into the strategy by identifying preventive ventilation adjustment segments, extrapolating uninterrupted counterfactual risks, and verifying action response divergences under observed equivalent candidate twin states. It confines the safety benefits of historical ventilation adjustment actions that are obscured, the uncertainty of twin states, and the correction of the value of reinforcement learning actions within the same training chain. This enables the mine ventilation control strategy to retain its proactive risk response capabilities while possessing more reliable training access, online execution verification, and feedback closed-loop update constraints. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort:

[0012] Figure 1 This is a system framework diagram of the present invention. Detailed Implementation

[0013] 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 embodiments of the present invention, and not all embodiments. 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.

[0014] like Figure 1 As shown, this invention discloses a digital twin-driven reinforcement learning control system for mine ventilation networks, comprising: a fragment recognition module, a counterfactual inference module, a response identification module, and a strategy correction module.

[0015] The segment recognition module acquires mine ventilation operation records and digital twin ventilation network status. The mine ventilation operation records serve as the time and operation input for segment recognition, providing historical ventilation adjustment actions, scheduling direction information, measuring point operation status, equipment operation status, sensor maintenance status, and operation interference records. The digital twin ventilation network status serves as the spatial and topological input for segment recognition, providing ventilation branch connection relationships, airflow direction, branch identification, location of ventilation adjustment execution objects, and mapping relationship between measuring points and ventilation branches. The acquisition of mine ventilation operation records, timestamp calibration, sensor calibration, and digital twin ventilation network topology maintenance are completed using existing methods in the mine safety monitoring system and ventilation digital twin platform. Its output provides basic data for subsequent event anchor point alignment, influence branch constraints, and segment-level risk conversion.

[0016] The segment recognition module first performs event anchor point alignment. Using historical ventilation adjustment actions from the mine ventilation operation record as input, the module sets the occurrence time of these actions as the event anchor point. Based on this anchor point, the module forms a pre-adjustment window by looking forward from the warning lead time in the mine ventilation operation record, and a post-adjustment verification window by looking backward from the ventilation propagation delay. The warning lead time is determined by the time interval between the occurrence of risk precursors and the alarm occurrence time in the historical pre-alarm samples, measured in minutes. The ventilation propagation delay is determined by the ventilation path length from the ventilation adjustment target to the affected branch, the airflow direction, and historical similar ventilation adjustment action response records, also measured in minutes. The pre-adjustment window covers at least three consecutive valid sampling periods, and in continuous production scenarios, it ranges from 5 to 60 minutes. When the sampling period is greater than 1 minute, the pre-adjustment window uses at least three consecutive valid sampling periods as the lower limit. The post-adjustment verification window covers the affected area. The maximum ventilation propagation delay in the affected branch set is taken as 5 to 90 minutes in a typical local ventilation adjustment scenario. When the ventilation propagation delay from the ventilation adjustment target to the affected branch exceeds 90 minutes, the post-adjustment verification window is extended to the response arrival time corresponding to the maximum ventilation propagation delay. When the number of verified preventive ventilation adjustment segments is less than 30, the pre-adjustment window and the post-adjustment verification window adopt a conservative window jointly determined by the mine ventilation regulations, scheduling review records, and ventilation engineer calibration records. After the number of verified preventive ventilation adjustment segments reaches 30, the segment identification module corrects the pre-adjustment window and the post-adjustment verification window according to the response arrival time distribution in the verified preventive ventilation adjustment segments. The 30 samples are used to form the initial stable basis for quartile statistics and cross-validation, and are not used as fixed business thresholds for mine operation. The output of event anchor point alignment is the pre-adjustment window and the post-adjustment verification window, which are used to limit the time range of subsequent risk characterization and risk fallback verification.

[0017] The segment recognition module then executes the influence branch constraint. Taking the air-regulating execution object of historical air-regulating actions and the state of the digital twin ventilation network as input, the segment recognition module maps the air-regulating execution object to branches in the digital twin ventilation network. During mapping, the segment recognition module reads the location description, branch identifier, associated measurement points, and historical action object association records of the air-regulating execution object, matching the air-regulating execution object with the branch identifier in the digital twin ventilation network. If the match fails, the segment recognition module calls the air-regulating execution object association records for a secondary match, and marks the historical air-regulating actions that failed the secondary match as segments with insufficient object affiliation, not entering the high-confidence segment output. After mapping, the segment recognition module performs reachable path expansion along the airflow direction in the digital twin ventilation network state corresponding to the event anchor point, including ventilation branches with airflow connectivity to the air-regulating execution object in the candidate range. For branches adjacent to the air-regulating execution object and with pressure... For ventilation branches with synchronous changes, the segment identification module reads the direction and magnitude of differential pressure changes within the window before air adjustment, as well as historical records of similar air adjustment responses. When the direction of differential pressure change of adjacent branches satisfies the same response relationship as the direction of differential pressure change of the branch where the air adjustment target is located, and the magnitude of differential pressure change exceeds the upper limit of historical stable differential pressure fluctuations, the corresponding ventilation branch is identified as satisfying the differential pressure coupling constraint and included in the candidate range. If the digital twin ventilation network status shows that there are closed dampers, sealed walls, disconnected paths, and inaccessible airflow paths between the target ventilation branch and the air adjustment target, the segment identification module performs exclusion processing on the target ventilation branch. The output of the influence branch constraint is the influence branch set. This influence branch set is used to limit the calculation range of subsequent homogenization risk characterization, actual air adjustment observation risk results, and interference coverage results, thereby avoiding the inclusion of measurement point changes without ventilation response association into the same preventive air adjustment segment.

[0018] After the influence branch set is formed, the segment identification module converts the segment-level risks before and after historical air adjustment actions into a homogeneous risk representation. The segment identification module takes the ventilation risk change process of the window before air adjustment within the influence branch set as input, and performs homogeneous conversion in a way that ensures the risk direction is consistent. The homogeneous conversion uses the historical stability boundary as the correspondence between 0 and the normalized upper bound as the correspondence between 1. Risk changes exceeding the normalized upper bound are truncated to 1, and changes opposite to the risk direction are set to 0. Gas concentration changes are used to form the gas risk representation quantity. The segment identification module forms the gas risk based on the slope of the gas concentration change within the window before air adjustment and the position of the gas concentration relative to the warning limit. Characteristic quantities include airflow gap changes, which are used to form airflow risk characterization quantities. The segment identification module forms airflow risk characterization quantities based on the gap between the ventilation branch airflow and the corresponding working face airflow demand and the gap change trend. Pressure difference deviation changes are used to form pressure difference risk characterization quantities. The segment identification module forms pressure difference risk characterization quantities based on the degree of deviation of the pressure difference from the historical stable pressure difference baseline. Working face air demand load changes are used to form load risk characterization quantities. The segment identification module forms load risk characterization quantities based on the air demand changes corresponding to the work intensity, equipment operating status, and working face load changes. All of the above risk characterization quantities are intermediate results used to constitute the segment-level risk characterization results before air adjustment.

[0019] Before fusion, the gas risk characterization quantity, air volume risk characterization quantity, pressure difference risk characterization quantity, and load risk characterization quantity were all converted to dimensionless results ranging from 0 to 1, maintaining a consistent numerical direction. The normalization boundaries were jointly determined by historical stable samples, historical pre-alarm samples, and mine safety regulations. The upper normalization bound for the gas risk characterization quantity was taken as the upper quartile of the slope of gas concentration change in the historical pre-alarm samples, and it was not higher than the risk boundary corresponding to the formal alarm limit. The normalization boundary for the air volume risk characterization quantity was determined by the historical abnormal sample distribution of the required air volume gap at the working face. The normalization boundary for the pressure difference risk characterization quantity was determined by the allowable deviation range of the historical stable pressure difference baseline. The normalization boundary for the load risk characterization quantity was determined by the historical abnormal sample distribution of the required air volume gap at the working face. The upper limit of the wind load and the scheduling verification record are determined. If the affected branch has communication interruption and sensor maintenance status within the pre-adjustment window, the segment identification module corrects the corresponding measurement point confidence weight and writes the measurement point quality status into the segment quality mark. The measurement point confidence weight is a dimensionless weight from 0 to 1. The weight correction range corresponding to communication interruption and maintenance status is determined by the historical error record of similar measurement points and sensor maintenance record. When there is no historical error record of similar measurement points, the segment identification module marks the corresponding measurement point as a low confidence measurement point and corrects the contribution of the corresponding affected branch in the subsequent fusion. Through this processing, gas, air volume, pressure difference and load changes with different original dimensions and directions are unified to the same evaluation scale.

[0020] The segment identification module then generates segment-level risk characterization results before airflow adjustment. Based on the ventilation path correlation strength from the airflow adjustment target to each affected branch, the segment identification module generates branch influence weights. The ventilation path correlation strength is jointly determined by path distance, airflow direction consistency, pressure difference coupling strength, and historical similar airflow adjustment response strength. Path distance reflects the propagation distance between the airflow adjustment target and the affected branch; airflow direction consistency determines whether the response can reach the affected branch along the airflow path corresponding to the event anchor point; pressure difference coupling strength compensates for the ventilation correlation between adjacent coupled branches; and historical similar airflow adjustment response strength calibrates the actual response degree of similar airflow adjustment actions to the same branch. Each branch influence weight is a dimensionless weight ranging from 0 to 1, and after normalization, it participates in the fusion process. To ensure that the sum of the influence weights of each branch is 1, when the number of historical similar wind adjustment response samples is less than 30, the branch influence weights are formed primarily based on path distance, airflow direction consistency, and pressure difference coupling strength. After the number of historical similar wind adjustment response samples reaches 30, the branch influence weights are further calibrated by incorporating the intensity of historical similar wind adjustment responses. The segment identification module integrates the homogenized risk characterization of each influencing branch according to the branch influence weights to obtain the segment-level risk characterization results before wind adjustment. The segment-level risk characterization results before wind adjustment are dimensionless results ranging from 0 to 1, used to describe the degree of risk convergence within the same influencing branch set before the occurrence of historical wind adjustment actions, and used as input for comparing the observed risk results after subsequent actual wind adjustment and forming the confidence results of preventive wind adjustment segments.

[0021] The segment identification module uses the same conversion caliber to form the actual observed risk results after wind adjustment within the post-wind adjustment verification window. The effective starting point of the post-wind adjustment verification window is jointly determined by the completion time of historical wind adjustment actions and the ventilation propagation delay corresponding to the set of influence branches. When the completion time of the actuator action is missing, the segment identification module uses the upper quartile of the completion time of similar historical wind adjustment actions as the effective starting point correction. The segment identification module takes the same set of influence branches as the processing object and uses the same variable caliber, window statistical method, homogenization rule, normalization boundary, and branch influence weight as the segment-level risk characterization results before wind adjustment to form the actual risk. The observed risk results after the actual wind adjustment are as follows: if the segment-level risk characterization results before the wind adjustment adopt the highest risk caliber within the window, the observed risk results after the actual wind adjustment also adopt the highest risk caliber within the window; if the segment-level risk characterization results before the wind adjustment adopt the caliber of the warning duration, the observed risk results after the actual wind adjustment adopt the caliber of the warning duration. The observed risk results after the actual wind adjustment are dimensionless results ranging from 0 to 1, which are used to compare with the segment-level risk characterization results before the wind adjustment using the same caliber. This comparison using the same caliber is used to determine whether there is an observable correlation between the risk decline before and after the historical wind adjustment action, and to avoid mixing risk results with different evaluation calibers.

[0022] After forming the segment-level risk characterization results before ventilation adjustment and the observed risk results after actual ventilation adjustment, the segment identification module reads the operational interference records from the mine ventilation operation records and generates interference coverage results. The operational interference records serve as the basis for judging the interference coverage results, used to determine whether the risk decline is covered by independent external interference. The segment identification module determines the source relationship based on the time of occurrence of operational interference, the ventilation branch to which the interference object belongs, the ventilation path relationship between the interference object and the ventilation adjustment execution object, and the scheduling direction information. When the interference object is located within the set of affected branches, the time of occurrence falls within the historical ventilation adjustment action response propagation window, and the scheduling direction information is consistent with the historical ventilation adjustment action, the corresponding changes are retained in the segment evidence chain; when the interference object does not belong to the set of affected branches and the time of occurrence is within the historical ventilation adjustment action response propagation window, the corresponding changes are retained in the segment evidence chain. When a change is independent of the historical wind control response propagation window and can independently explain the risk decline, the corresponding change is identified as an independent external disturbance. Changes whose source cannot be confirmed are marked as sources to be confirmed. The disturbance coverage result is a dimensionless result from 0 to 1. The closer the value is to 1, the more likely the risk decline is to be explained by an independent external disturbance. The initial value of the disturbance coverage result is formed based on the proportion of risk decline explained by independent external disturbances in the historical review segments. When the number of historical review segments is less than 30, the segment identification module uses the review records of the ventilation engineer and the dispatching manager on the historical segments to determine the initial value of the disturbance coverage result. The disturbance coverage result is used as deduction evidence in the formation process of the confidence result of the preventive wind control segment to reduce the misattribution caused by independent external disturbances.

[0023] The segment identification module uses the interference coverage results to subtract the confidence strength of the risk comparison before and after wind adjustment, generating a preventative wind adjustment segment confidence result. The module uses the segment-level risk characterization results before wind adjustment as risk precursor evidence, the decline in the observed risk results after actual wind adjustment relative to the segment-level risk characterization results before wind adjustment as response evidence, the event anchor point being located before the formal alarm as preliminary evidence, and the interference coverage results as subtraction evidence. Risk precursor evidence, response evidence, and preliminary evidence are used to support the preventative characteristics of historical wind adjustment actions. The interference coverage results are used to correct misattributions caused by independent external interference. Each piece of evidence has a dimensionless weight from 0 to 1, derived from historically verified... The cross-validation results of preventive air adjustment segments, abnormal handling segments, and ordinary air adjustment segments are determined and normalized before being used to form confidence strength. When the number of validated segments is less than 30, the deduction weight corresponding to the interference coverage result is not less than the weight of any positive evidence, so as to avoid over-confirmation of preventive air adjustment segments when there is insufficient evidence. The confidence result of preventive air adjustment segments is limited to the range of 0 to 1. The closer the value is to 1, the more the corresponding segment matches the characteristics of preventive air adjustment segments. The confidence result of preventive air adjustment segments is used as the output of the segment identification module and enters the counterfactual inference module to determine the initial credible basis of the counterfactual initial state, the historical air adjustment actions to be removed, and the subsequent implicit safety benefits.

[0024] The segment identification module also performs reverse retention processing to prevent early preventative ventilation adjustment segments from being filtered by the threshold. If the segment-level risk characterization result before ventilation adjustment does not reach the preventative observation threshold, but the scheduling information in the mine ventilation operation record points to at least one risk state such as approaching gas, insufficient air supply, abnormal pressure difference, and high working face air demand load, and there are at least two weak risk changes within the affected branch set that deviate in the same direction from historical stable samples, last for no less than two consecutive effective sampling periods, and do not reach the preventative observation threshold, the segment identification module does not directly exclude the corresponding candidate segment. Instead, it outputs the corresponding candidate segment to the segment set to be checked against the facts. The preventative observation threshold is a dimensionless threshold of 0 to 1, determined by the lower quartile of the segment-level risk characterization result before ventilation adjustment in the historical pre-alarm samples. When the number of historical pre-alarm samples is less than 30, the preventative observation threshold is set between 0.4 and 0.6 and calibrated through scheduling review records. 0.4 to 0.6 corresponds to weak risk changes before the alarm under the same direction risk characterization. The intermediate risk range is not considered a formal alarm threshold in mine safety regulations. A change in risk in the same direction refers to at least two risk characteristics within the affected branch set changing in the same risk direction relative to historical stable samples, and the corresponding risk characteristics not reaching the preventative observation threshold. When the set of segments to be counterfactually verified enters subsequent counterfactual deduction, only the initial verification confidence weight is used as the initial confidence basis for the counterfactual deduction module; it does not directly enter the high-confidence training path. The initial verification confidence weight is formed by multiplying the segment confidence value of the corresponding candidate segment by the verification reduction coefficient. The verification reduction coefficient is a dimensionless coefficient of 0.5 to 0.8, determined by the proportion of historical segments to be counterfactually verified that have been confirmed by subsequent counterfactual deduction. When the number of historical segments to be counterfactually verified is less than 30, the verification reduction coefficient is 0.5 to 0.6. The output of the reverse retention processing is the set of segments to be counterfactually verified, used for counterfactual deduction verification in the counterfactual deduction module, preventing early weak-risk segments from being permanently excluded due to failure to meet the preventative observation threshold.

[0025] The final segment identification module outputs the confidence results of preventive wind adjustment segments. These results include segment start and end times, historical wind adjustment actions, influencing branch set, segment-level risk characterization results before wind adjustment, observed risk results after actual wind adjustment, interference coverage results, segment confidence value, and source label. The segment start and end times and historical wind adjustment actions determine the historical wind adjustment actions that need to be removed in the subsequent counterfactual inference module. The segment-level risk characterization results before wind adjustment are used to determine the initial state of subsequent counterfactual events. The influencing branch set limits the calculation scope of subsequent uninterrupted counterfactual risk evolution. The interference coverage results and segment confidence value are used to determine the initial credible basis for subsequent implicit safety benefits. The source label is used for subsequent verification and threshold calibration.

[0026] The counterfactual deduction module receives the confidence results of preventive wind adjustment segments and the set of segments to be counterfactually verified from the segment identification module. The confidence results of preventive wind adjustment segments serve as the main input for this step, providing the segment start and end times, historical wind adjustment actions, set of influencing branches, segment-level risk characterization results before wind adjustment, observed risk results after actual wind adjustment, interference coverage results, segment confidence values, and source labels. The set of segments to be counterfactually verified serves as supplementary input, allowing segments that do not meet the preventive segment admission criteria but have early weak risk changes to enter the counterfactual deduction verification. Historically verifiable segments serve as parameter calibration samples, used to determine the historical playback error allowable range, the counterfactual credibility threshold, and the evidence weight in the counterfactual deduction credibility. Historically verifiable segments refer to historical segments that have undergone scheduling verification, actual observation playback, and source label verification. This step only processes the preventive wind adjustment segments and segments to be counterfactually verified output by the segment identification module. The counterfactual deduction object is consistent with the event anchor point, set of influencing branches, and segment-level risk caliber in the segment identification module.

[0027] The counterfactual inference module locates the initial counterfactual state. It reads the start and end times of the preventative ventilation adjustment segment from the confidence results, historical ventilation adjustment actions, the set of influencing branches, and the segment-level risk characterization results before the adjustment. Based on the segment start and end times, it extracts the stable ventilation state before the historical ventilation adjustment actions. The stable ventilation state includes at least the gas concentration, air volume, pressure difference, temperature and humidity, airflow direction, required air load at the work surface, and actuator status within the influencing branch set. It also retains the ventilation branch connection relationships, the mapping relationship between measuring points and ventilation branches, and the location of the ventilation adjustment execution object from the digital twin ventilation network state. The extraction window for the stable ventilation state covers at least three consecutive valid sampling periods. If the sampling period is greater than one minute, the extraction window for the stable ventilation state still uses at least three consecutive valid sampling periods as the lower limit. If historical ventilation adjustments... Before the action is executed, there may be single-point jumps, communication interruptions, and sensor maintenance states. The counterfactual inference module does not use abnormal sampling points, but instead uses the state results within a continuous and stable sampling window before the action is executed. For gas concentration, air volume, pressure difference, temperature and humidity, and the required air load of the work surface, the counterfactual inference module forms a weighted average state value based on the credibility weight of the measurement points. For airflow direction and actuator state, the counterfactual inference module uses the state with the highest frequency of occurrence and closest to the event anchor point within the continuous and stable sampling window. If a stable ventilation state of no less than 3 consecutive effective sampling cycles cannot be obtained before the event anchor point, the counterfactual inference module writes a counterfactual initial state insufficient label in the counterfactual evidence state. The corresponding segment does not form a high credibility counterfactual initial state. This processing outputs the counterfactual initial state, which is used as the starting point for the playback of the uninterrupted counterfactual action sequence.

[0028] The formation of the counterfactual initial state maintains consistency in terms of object, time, and spatial boundaries. The counterfactual inference module only uses the ventilation state before the event anchor point, and does not use the observed state after the historical ventilation adjustment action as the counterfactual starting point. The set of influence branches follows the set of influence branches formed in the fragment identification module, and is not expanded to the entire mine ventilation network in the counterfactual inference module. If there is a topology update record in the digital twin ventilation network state near the event anchor point, the counterfactual inference module uses the digital twin ventilation network state at the time corresponding to the event anchor point as the topological basis of the counterfactual initial state. Through this processing, the counterfactual initial state and the preventive ventilation adjustment fragment of the fragment identification module maintain the same event anchor point, the same set of influence branches, and the same ventilation network state, avoiding writing the changes after ventilation adjustment into the uninterrupted path.

[0029] The counterfactual deduction module reconstructs the sequence of uninterrupted counterfactual actions. It reads the actual action sequence from the mine ventilation operation records, locates the historical ventilation adjustment action corresponding to the preventative ventilation adjustment segment within the actual action sequence, and deletes that historical ventilation adjustment action from the actual action sequence. For other ventilation adjustment actions that occur after the historical ventilation adjustment action, the counterfactual deduction module identifies derived ventilation adjustment actions based on scheduling instruction relationships, ventilation path relationships, and action occurrence time relationships. Specifically, when a subsequent ventilation adjustment action belongs to the same scheduling instruction chain as the deleted historical ventilation adjustment action, the execution object's scope and the set of affected branches overlap, and the occurrence time of the subsequent ventilation adjustment action falls within the response propagation window of the historical ventilation adjustment action, the counterfactual deduction module will assign the subsequent ventilation adjustment action to the historical ventilation adjustment action. Subsequent air adjustment actions are identified as derived air adjustment actions. The response propagation window is jointly determined by the completion time of historical air adjustment actions, the ventilation propagation delay corresponding to the affected branch set, and the response records of similar historical air adjustment actions. When the number of response records of similar historical air adjustment actions is less than 30, the response propagation window is jointly determined by the ventilation propagation delay, the upper limit of the execution time of the actuator, and the scheduling review record. The completion time of historical air adjustment actions is only used to determine the time alignment of the response propagation window and the counterfactual inference window, and is not considered as an execution action in the non-intervention counterfactual action sequence. Subsequent air adjustment actions that meet the conditions for derived air adjustment actions are deleted simultaneously with historical air adjustment actions to form a non-intervention counterfactual action sequence. Subsequent air adjustment actions that do not simultaneously meet the conditions for derived air adjustment actions shall not be deleted simultaneously.

[0030] For subsequent actions with independent sources, the counterfactual deduction module does not perform synchronous deletion. Subsequent actions with independent sources include production plan adjustment actions, mandatory safety procedure actions, actions that are not associated with ventilation paths in the affected branch set, and actions whose occurrence time is independent of the historical ventilation adjustment action response propagation window. Subsequent actions with independent sources are retained in the uninterrupted counterfactual action sequence to maintain the true production boundary during counterfactual deduction. If the source of a subsequent action cannot be confirmed, the counterfactual deduction module writes an action source pending confirmation label into the counterfactual evidence status. Segments with the action source pending confirmation label are not allowed to enter the high-confidence training sample set, but are only used as subsequent review objects to enter the action response identification module for action response identification verification.

[0031] After the non-interventional counterfactual action sequence is formed, the counterfactual deduction module uses the initial counterfactual state as the playback starting point, applies the non-interventional counterfactual action sequence to the ventilation network structure, and performs ventilation network balance playback. The ventilation network balance playback is achieved by using the air volume balance and air pressure balance methods in the mine ventilation network solution. Its inputs are the initial counterfactual state, the non-interventional counterfactual action sequence, the digital twin ventilation network state, and the set of influencing branches. Its processing purpose is to obtain the ventilation response process when the historical ventilation adjustment action was not executed. Its output is the branch air volume, branch pressure difference, airflow direction, and risk variable change process under the non-interventional state.

[0032] During the ventilation network balancing replay process, the counterfactual inference module applies physical constraints to the ventilation network. These physical constraints include at least node airflow conservation, loop air pressure balance, fan operation boundaries, damper opening boundaries, local fan connection boundaries, and airflow path connectivity constraints. If the replay result violates the ventilation network physical constraints, the counterfactual inference module does not treat the corresponding replay result as a highly credible non-interventional counterfactual risk result. Instead, it writes a "physical constraint not satisfied" label into the counterfactual evidence state. Fragments with the "physical constraint not satisfied" label do not enter the highly credible formation process of the implicit safety benefit result, and the "physical constraint not satisfied" label is retained in the counterfactual evidence state. This processing is used to avoid using the original observation state after deleting historical wind adjustment actions from the action sequence, which would cause the non-interventional path to be physically invalid.

[0033] For weak observation areas, the counterfactual inference module uses a weak observation compensation method to correct playback bias. Weak observation areas refer to ventilation branches where the influencing branch set lacks direct measurement points, the measurement point confidence weight is lower than the measurement point availability threshold, or the measurement point is in a communication interruption or maintenance state within the counterfactual inference window. Ventilation branches with measurement point confidence weights lower than the measurement point availability threshold enter weak observation compensation and are not considered true observation branches for high-confidence judgment. The measurement point availability threshold is a dimensionless threshold between 0 and 1, jointly determined by the upper bound of historical errors for similar measurement points, sensor maintenance records, and mine monitoring system calibration records. When there are fewer than 30 historical error records, the measurement point availability threshold is set... Between 0.6 and 0.8, and verified and calibrated by ventilation engineers, the available threshold for measuring points is between 0.7 and 0.8 in high-gas areas, long-distance ventilation areas, and areas with a high proportion of weak observations; in areas with sufficient measuring point coverage and stable operation, the available threshold for measuring points is between 0.6 and 0.7. The weak observation compensation method uses the observation status of adjacent ventilation branches, airflow direction, historical similar operating status, and digital twin ventilation network status as inputs to complete the changes in air volume, pressure difference, and risk variables in the weak observation area. The weak observation compensation results are included as a component of the uninterrupted counterfactual risk results in subsequent comparisons, and at the same time form the uncertainty results of weak observation completion.

[0034] The uncertainty result for weak observation completion is a dimensionless result ranging from 0 to 1. Its value is jointly determined by the proportion of weak observation branches, the consistency of adjacent branch observations, and the support level of historical similar operating states. The proportion of weak observation branches is the ratio of the number of branches corresponding to the weak observation area in the affected branch set to the total number of branches in the affected branch set. The consistency of adjacent branch observations is formed by the proportion of adjacent branch risk change direction consistency and the normalized deviation of adjacent branch risk change. If the weak observation area lacks effective adjacent branch observations, the consistency of adjacent branch observations is treated as low consistency. The lower the proportion of consistent risk change direction, the lower the consistency of adjacent branch observations; the greater the normalized deviation of adjacent branch risk change, the lower the consistency of adjacent branch observations. The support level for the operational status is determined by the number of similar operational status segments and the playback error of those segments. Fewer similar operational status segments result in lower support for historical similar operational statuses, and greater playback error further reduces support. When historical similar operational statuses are lacking, their support is treated as low. A higher proportion of weak observation branches leads to higher uncertainty in weak observation completion results. Lower consistency between adjacent branch observations results in higher uncertainty in weak observation completion results. Lower support for historical similar operational statuses leads to higher uncertainty in weak observation completion results. These uncertainties are subsequently used to correct the credibility of counterfactual inferences, preventing weak observation areas from excessively influencing the implicit safety benefit results.

[0035] The counterfactual inference module generates uninterrupted counterfactual risk results. Within the counterfactual inference window, the module reads the uninterrupted ventilation response process from the set of influencing branches and, using the same risk caliber as the segment identification module, converts the gas risk, air volume risk, pressure difference risk, and load risk under the uninterrupted state into homogeneous risk characterization quantities ranging from 0 to 1. The counterfactual inference window adopts the window length and risk statistical caliber of the post-air adjustment verification window in the segment identification module, and determines the effective calculation interval based on the completion time of historical air adjustment actions and the ventilation propagation delay corresponding to the set of influencing branches. If the segment identification module uses the highest risk caliber within the window, the counterfactual inference module uses the highest risk caliber within the window to generate uninterrupted counterfactual risk results; if the segment identification module uses the proportion of warning duration, the counterfactual inference module uses the same proportion of duration to generate uninterrupted counterfactual risk results. Through this processing, the uninterrupted counterfactual risk results and the observed risk results after actual air adjustment have the same set of influencing branches, the same risk variables, the same homogeneity rules, the same normalization boundaries, and the same window statistical caliber.

[0036] The counterfactual inference module compares the non-interventional counterfactual risk result with the observed risk result after the actual wind adjustment using the same risk caliber. When the non-interventional counterfactual risk result is higher than the observed risk result after the actual wind adjustment, the counterfactual inference module determines the difference between the two as the counterfactual risk difference result. When the non-interventional counterfactual risk result is not higher than the observed risk result after the actual wind adjustment, the counterfactual risk difference result is determined to be 0. The counterfactual risk difference result is a dimensionless intermediate result between 0 and 1, used to represent the possible increase in risk under counterfactual conditions without the implementation of historical wind adjustment actions, compared to the observed state after the actual wind adjustment. The counterfactual risk difference result is not directly used as a reinforcement learning reward, but rather enters the counterfactual inference credibility control and implicit security benefit result formation process.

[0037] The counterfactual extrapolation module further generates counterfactual extrapolation confidence levels. Counterfactual extrapolation confidence levels are dimensionless results ranging from 0 to 1. Inputs include confidence results for preventative ventilation adjustment segments, consistency results from counterfactual model replay, consistency results from disturbances, interference coverage results, and uncertainty results from weak observation completion. The consistency results from counterfactual model replay are obtained through the replay of real action sequences. Specifically, the counterfactual extrapolation module applies real action sequences to the same ventilation network structure and compares the risk results obtained from the replay with the observed risk results after actual ventilation adjustment using the same caliber. When the deviation between the two is within the permissible range of historical replay errors, the counterfactual extrapolation is considered successful. The model playback consistency result is converted into a dimensionless result of 0 to 1 according to the relative position of the deviation within the historical playback error allowable range. When the deviation is close to 0, the value is close to 1, and when the deviation is close to the upper limit of the historical playback error allowable range, the value is close to 0. When the deviation exceeds the historical playback error allowable range, the counterfactual model playback consistency result is 0, and a playback consistency insufficiency label is written in the counterfactual evidence status. The historical playback error allowable range is determined by the playback error distribution of historically verifiable segments. When the number of historically verifiable segments is less than 30, the ventilation engineer determines the conservative error range based on the ventilation network measurement point error and scheduling verification records.

[0038] The perturbation consistency result is used to indicate whether the production load, ventilation topology state, and external perturbation inputs are consistent with the actual segments during the counterfactual simulation process. The counterfactual simulation module matches the production load records, ventilation topology state records, and external perturbation records within the counterfactual simulation window with the corresponding records of the actual segments using the same timestamp, and converts the completeness and deviation of the match into dimensionless results from 0 to 1. The more complete the match and the smaller the deviation, the higher the perturbation consistency result. When key perturbation records are missing or the perturbation records are inconsistent with the actual segments, the perturbation consistency result is not used as positive evidence in the formation of counterfactual simulation credibility, and a perturbation missing label is written into the counterfactual evidence status. The confidence results of the preventive ventilation adjustment segment, the consistency results of the counterfactual model playback, and the perturbation consistency result are used as positive evidence to improve the credibility of the counterfactual simulation; the interference coverage result and The uncertainty results of weak observation completion are used as deduction evidence to reduce the credibility of counterfactual inference. The weight of each piece of evidence is a dimensionless weight between 0 and 1, and it is normalized before participating in credibility fusion. When the number of historically verifiable segments is less than 30, the weight of the deduction evidence corresponding to the interference coverage result and the uncertainty results of weak observation completion is not less than the weight of any positive evidence. After the number of historically verifiable segments reaches 30, the counterfactual inference module corrects the weight of each piece of evidence according to the historical playback error distribution and the verification results. The counterfactual inference module first forms a positive evidence weighted result based on the confidence result of the preventive wind adjustment segment, the consistency result of the counterfactual model playback, and the consistency result of the disturbance. Then, it forms a deduction evidence weighted result based on the interference coverage result and the uncertainty results of weak observation completion. The weight of the deduction evidence weighted result is then reduced by the positive evidence weighted result and limited to the range of 0 to 1 to obtain the credibility of counterfactual inference.

[0039] After establishing the counterfactual inference credibility, the counterfactual inference module generates implicit security benefit results. The module takes the counterfactual risk difference results and the counterfactual inference credibility as input. When the counterfactual risk difference results are greater than 0 and the counterfactual inference credibility reaches the counterfactual credibility threshold, the module retains the credible portion of the counterfactual risk difference results according to the corresponding proportion of the counterfactual inference credibility, forming implicit security benefit results. The implicit security benefit results are not higher than the counterfactual risk difference results. The higher the counterfactual inference credibility, the greater the contribution of the counterfactual risk difference results retained as implicit security benefit results. The counterfactual credibility threshold is a dimensionless threshold between 0 and 1, determined by the lower quartile of the counterfactual inference credibility in historically verifiable segments. The number of historically verifiable segments... When the number of counterfactual data points is less than 30, the counterfactual credibility threshold is set between 0.50 and 0.70 and is verified and calibrated by ventilation engineers and dispatch managers. In high-gas areas, long-distance ventilation areas, and areas with a high proportion of weak observations, the counterfactual credibility threshold is set between 0.6 and 0.7. In areas with sufficient measurement point coverage and stable playback errors, the counterfactual credibility threshold is set between 0.5 and 0.6. The range of 0.5 to 0.7 corresponds to the range where the counterfactual inference results have a medium to high credibility basis and is not used as the alarm threshold for safety procedures. The counterfactual credibility threshold is only used to control the access of implicit safety benefit results to subsequent action response identification and action value correction permission. Implicit safety benefit results are dimensionless results between 0 and 1, used to represent the safety benefits that are masked by the actual execution of historical ventilation actions in historical records.

[0040] If the uninterrupted counterfactual risk result is not higher than the observed risk result after the actual wind adjustment, the counterfactual inference module does not generate an implicit safety benefit result and writes a "no implicit safety benefit" label into the counterfactual evidence status. If the uninterrupted counterfactual risk result is higher than the observed risk result after the actual wind adjustment, but the counterfactual inference credibility does not reach the counterfactual credibility threshold, the counterfactual inference module writes a "insufficient counterfactual evidence" label into the counterfactual evidence status. Segments with "no implicit safety benefit" and "insufficient counterfactual evidence" labels do not enter the positive update process of action value, nor are they used as negative samples to reduce the action value of historical wind adjustment actions. The counterfactual evidence status is used by the response identification module and the strategy correction module to determine the sample role. Segments with "action source to be confirmed," "physical constraint not met," "disturbance missing," "no implicit safety benefit," and "insufficient counterfactual evidence" labels do not enter the positive update process of action value. Among them, segments with "action source to be confirmed" and "disturbance missing" labels only enter the subsequent review path. This processing is used to avoid erroneously weakening the value of preventive wind adjustment actions due to unreliable model playback, too many weak observation areas, and independent external interference coverage.

[0041] The final counterfactual inference module outputs the non-interventional counterfactual risk result, the observed risk result after actual wind adjustment, the counterfactual risk difference result, the counterfactual inference credibility, the implicit safety benefit result, and the counterfactual evidence status. The non-interventional counterfactual risk result enters the response identification module to determine the response comparison object of the candidate control action in the observation equivalent candidate twin state. The implicit safety benefit result enters the response identification module and the strategy correction module to determine whether the corresponding segment has the benefit basis for action value correction. The counterfactual inference credibility enters the response identification module to limit the influence weight of the action response divergence result on the admission of training samples. The counterfactual evidence status is used to prevent segments with insufficient counterfactual evidence from being mistakenly regarded as negative samples. Through the above processing, the counterfactual inference module converts the confidence result of the preventive wind adjustment segment into non-interventional counterfactual risk evidence with physical constraints, weak observation compensation, and credibility control, providing a benefit basis for subsequent action response divergence calculation and action value correction permission.

[0042] In this embodiment, the response identification module receives the non-interventional counterfactual risk result, the observation risk result after actual wind adjustment, the counterfactual risk difference result, the counterfactual inference credibility, the implicit safety benefit result, and the counterfactual evidence status output by the counterfactual inference module. The non-interventional counterfactual risk result is used to determine the response comparison object of the candidate control action in the observation equivalent candidate twin state. The implicit safety benefit result is used to determine whether the corresponding segment has the benefit basis for entering the action value correction permission. The counterfactual inference credibility is used to limit the influence weight of the action response divergence result on the admission of training samples. The counterfactual evidence status is used to identify segments with unconfirmed action source, unmet physical constraints, missing disturbances, and insufficient counterfactual evidence. The response identification module only performs observation equivalent candidate twin state construction on segments for which the counterfactual inference module has formed implicit safety benefit results or entered the path to be reviewed. Segments marked as not being able to enter the action value update in the counterfactual evidence status are only retained as review objects.

[0043] The response identification module first constructs a set of observation-equivalent candidate twin states that satisfy the allowable measurement point error and the physical constraints of the ventilation network. The module then reads the counterfactual initial state, the set of influencing branches, and the digital twin ventilation network state. Based on the digital twin ventilation network state and historical playback errors, it performs boundary calibration on the uncertain parameters of the observation-equivalent candidate twin states. These uncertain parameters describe the ventilation network state that cannot be uniquely determined by measurement point observations under stable ventilation conditions before the event anchor point. They involve at least the ventilation resistance coefficient within the set of influencing branches, leakage path parameters coupled with pressure differentials to the set of influencing branches, weak observation branch completion parameters, and actuator response parameters. The boundary of the ventilation resistance coefficient is jointly determined by historical ventilation measurement records, roadway type, support status, and historical playback errors. When there are fewer than 30 historical ventilation measurement records, the boundary of the ventilation resistance coefficient is determined by the roadway... The boundary of the leakage path parameter is determined by the type, support status, and scheduling review records. The insufficient resistance boundary calibration is written into the candidate state fitting error result. The boundary of the leakage path parameter is determined by the air door sealing status, the pressure difference coupling relationship of adjacent branches, and historical leakage verification records. When there are fewer than 30 historical leakage verification records, the boundary of the leakage path parameter is mainly based on the air door sealing status and pressure difference coupling relationship. The insufficient leakage boundary calibration is written into the candidate state fitting error result. The boundary of the weak observation branch completion parameter is determined by the adjacent branch observation status, the uncertainty result of weak observation completion, and the support level of historical similar operating status. The boundary of the actuator response parameter is determined by the completion time of historical ventilation adjustment actions, actuator feedback records, and scheduling review records. The above uncertain parameters are used to form the boundary of the observation equivalent candidate twin state and are not directly used as the output of the mine ventilation control strategy.

[0044] After completing the boundary calibration of uncertain parameters, the response identification module performs constraint sampling within the calibrated value range to generate two or more initial candidate twin states. After the initial candidate twin states are verified by the measurement point error allowance and the ventilation network physical constraint, they form observation equivalent candidate twin states. The input of constraint sampling is the calibrated value range of each uncertain parameter. The processing action is to form candidate parameter combinations within the value range of each uncertain parameter and write the candidate parameter combinations into the digital twin ventilation network state to obtain the initial candidate twin states. For uncertain parameters with wide boundaries, the response identification module forms initial candidate parameter combinations at the lower boundary, middle position, and upper boundary of the value range. For uncertain parameters with concentrated historical playback error distribution, the response identification module prioritizes forming candidate parameter combinations in the high-frequency interval corresponding to the historical playback error distribution. The initial candidate twin states output by this processing must continue to be verified by the measurement point error allowance and the ventilation network physical constraint, and cannot be used for action response comparison based solely on the sampling results.

[0045] The number of observation-equivalent candidate twin states should be no less than two, preferably two to twenty. When the proportion of weak observations in the digital twin ventilation network state is high, the historical playback error is large, or there are multiple differential pressure coupling branches in the influencing branch set, the number of observation-equivalent candidate twin states should be 10 to 20. When the measurement point coverage is sufficient and the historical playback error is stable, the number of observation-equivalent candidate twin states should be 2 to 10. This range is used to cover multiple possible ventilation network states that meet the stable ventilation state before the event anchor point under controllable computational load. If no less than two observation-equivalent candidate twin states that simultaneously meet the measurement point error allowance and ventilation network physical constraints cannot be formed within the 20 initial candidate twin states, the response identification module writes a candidate state support insufficient label in the counterfactual evidence state. The implicit security benefit result of the corresponding segment shall not enter the action value positive update process, and the corresponding segment shall be transferred to the low confidence verification sample set. When it is necessary to supplement the scheduling review material, the subsequent review mark shall be retained in the counterfactual evidence state.

[0046] The response identification module performs measurement point error verification on each observation equivalent candidate twin state. Measurement point error verification uses the observed values ​​of measurement points in the stable ventilation state before the event anchor point corresponding to the counterfactual initial state and the simulated values ​​of measurement points in the observation equivalent candidate twin states as input. The processing action involves converting the measurement point deviations of airflow, pressure difference, and gas concentration into dimensionless errors of 0 to 1, and fusing them according to the measurement point confidence weights. The normalized reference value of the measurement point deviation is jointly determined by the median value of measurement points in historical stable samples, the lower limit of sensor resolution, and the corresponding reference boundary in the mine safety regulations. When there are fewer than 30 historical stable samples, the normalized reference value is preferentially determined by the lower limit of sensor resolution and the boundary of the mine safety regulations. The permissible threshold for measurement point error is a dimensionless threshold of 0 to 1, determined by the error distribution of the digital twin playback of historically verifiable segments. The sensor allowable error and scheduling verification records are used to jointly determine the error. When the number of historically verifiable segments is less than 30, the allowable threshold for measurement point error is preferably 0.10 to 0.25, 0.1 to 0.15 for high-gas areas and key return air paths, and 0.15 to 0.25 for areas with insufficient measurement point coverage. The label of insufficient measurement point coverage is written into the candidate state fitting error result. The observation equivalent candidate twin state with the label of insufficient measurement point coverage is not used to improve the credibility of the action response identifiable permission result. It only participates in the action response divergence calculation as a low measurement point support state in the state competition range. The observation equivalent candidate twin state that passes the measurement point error verification indicates that the measurement point fitting error of the observation equivalent candidate twin state under the stable ventilation state before the event anchor point is acceptable, but it still needs to be verified by the physical constraints of the ventilation network.

[0047] The response identification module further performs ventilation network physical constraint verification on the observed equivalent candidate twin states. This verification takes the observed equivalent candidate twin states, the digital twin ventilation network state, and the set of influencing branches as input. The processing steps involve verifying whether the ventilation resistance coefficient in the observed equivalent candidate twin state is positive, whether the leakage path has structural connection and pressure differential coupling basis, whether the fan operating point is within the permissible range of the fan characteristic curve, whether the damper opening is within the mechanically reachable range, whether each node satisfies airflow balance, and whether the airflow path satisfies the connectivity relationship of the digital twin ventilation network. If the ventilation resistance coefficient is not positive, the leakage path lacks structural connection basis, the fan operating point exceeds the permissible range, the damper opening exceeds the mechanically reachable range, the node airflow is unbalanced, or the airflow path does not satisfy the connectivity relationship, the response identification module excludes the corresponding observed equivalent candidate twin state. When selecting a twin state, if the observed equivalent candidate twin state satisfies both the measurement point error allowance and the ventilation network physical constraints, the response identification module includes the observed equivalent candidate twin state in the observed equivalent candidate twin state set. If the observed equivalent candidate twin state only satisfies the measurement point error allowance but violates the ventilation network physical constraints, the response identification module excludes the observed equivalent candidate twin state and writes a label indicating that the physical constraints of the candidate state are not satisfied in the counterfactual evidence state. The candidate state fitting error result is formed by the measurement point error verification result and the ventilation network physical constraint verification result, which is used to record the fitting reliability of the observed equivalent candidate twin state before entering the action response comparison. Through this processing, the observed equivalent candidate twin state set retains the observation equivalence under the stable ventilation state before the event anchor point, and does not allow states that violate the physical process of mine ventilation to enter the action response comparison.

[0048] After forming the observation-equivalent candidate twin state set, the response identification module determines the same candidate control action for action response divergence calculation. For segments with implicit safety benefit results generated by the counterfactual inference module, the candidate control action adopts the historical ventilation adjustment action deleted from the real action sequence in the counterfactual inference module. If the counterfactual inference module has identified a derived ventilation adjustment action that has a causal relationship with the historical ventilation adjustment action, the candidate control action adopts the combined action boundary of the historical ventilation adjustment action and the derived ventilation adjustment action. The combined action boundary includes the ventilation adjustment execution object, action direction, action amplitude, and action duration. Among them, the ventilation adjustment execution object is obtained by mapping the ventilation adjustment action object in the mine ventilation operation record and the branch identifier in the digital twin ventilation network state; the action direction is determined by the air volume, air pressure, or opening adjustment direction in the control command; and the action amplitude is determined by the control command value in the mine ventilation operation record. The control command value and the actuator feedback value are determined; the duration of the action is determined by the start time and completion time of the action. When the control command value and the actuator feedback value are inconsistent, the action amplitude is based on the actuator feedback value within the reachable range of the actuator, and an execution feedback deviation label is written in the counterfactual evidence state. When the actuator feedback value is missing, the delay exceeds one sampling period in the unified response observation window, or exceeds the reachable range of the actuator, the response identification module does not use the corresponding segment of the candidate control action as a high-confidence training sample, and an execution feedback insufficiency label is written in the counterfactual evidence state. The boundary of this combined action follows the derived wind adjustment action identification result in the counterfactual deduction module, and the candidate control action range is not expanded again in the response identification module. Through this processing, the response identification module compares the response differences of the same candidate control action in multiple observation equivalent candidate twin states.

[0049] The response identification module inputs the same candidate control action into each observation equivalent candidate twin state in the observation equivalent candidate twin state set, and replays the candidate baseline risk process without inputting the candidate control action and the candidate action risk process after inputting the candidate control action in the unified response observation window. The unified response observation window is jointly determined by the ventilation propagation time of the affected branch set, the actuator response time, and the sensor sampling period, and its unit is minutes. The unified response observation window covers no less than 3 consecutive effective sampling periods and covers the maximum ventilation propagation time in the affected branch set. Under normal local ventilation adjustment scenarios, the unified response observation window is preferably set to 5 minutes to 90 minutes. When the ventilation propagation time corresponding to the affected branch set exceeds 90 minutes, the unified response observation window is extended to the response arrival time corresponding to the maximum ventilation propagation time. When the number of historical similar ventilation adjustment action response records is less than 30, the unified response observation window is jointly determined by the ventilation propagation time, the upper bound of the actuator action completion time, and the scheduling review record. This window is only used to compare the response differences in the observation equivalent candidate twin states and is not used as an online execution confirmation window.

[0050] For each observed equivalent candidate twin state, the response identification module first replays the candidate baseline risk process without inputting candidate control actions, then replays the candidate action risk process after inputting candidate control actions. The difference between the candidate baseline risk process and the candidate action risk process forms the risk change quantity. The risk change quantity includes the risk change direction result and the risk change magnitude result. The risk change direction result is used to distinguish between risk reduction, risk maintenance, and risk increase. The risk change magnitude result is a dimensionless intermediate result between 0 and 1. When the risk after the candidate control action is lower than the candidate baseline risk process, the risk change direction result is risk reduction; when the risk after the candidate control action is higher than the candidate baseline risk process, the risk change direction result is risk increase. When the risk change direction is determined, the risk increases; when the difference between the two is lower than the minimum effective response threshold, the risk change direction is determined, the risk remains unchanged, and the risk change magnitude is treated as no effective response. The minimum effective response threshold is a dimensionless threshold of 0 to 1, which is jointly determined by the upper bound of the response noise of similar historical wind regulation actions and the sensor resolution. When the number of historical similar wind regulation action response records is less than 30, the minimum effective response threshold is preferably 0.03 to 0.1. When the noise at the measuring point is high, it is 0.07 to 0.1. When the measuring point is stable and the sampling frequency is high, it is 0.03 to 0.07. This threshold is used to filter random fluctuations at the measuring point and avoid identifying noise as a candidate control action response.

[0051] The response identification module then extracts the response direction, response location, and response time corresponding to the candidate control actions along the unified response observation window. It converts the risk change, response direction, response location, and response time into a response description vector. The response direction is primarily determined by the risk change direction result, with the airflow change direction, pressure difference change direction, and gas dilution trend serving as consistency verification criteria. When the risk change direction result is inconsistent with the verification results of the airflow change direction, pressure difference change direction, and gas dilution trend, the response identification module writes a direction conflict label into the response description vector. The observation equivalent candidate twin state corresponding to the direction conflict label is included in the conflict candidate state in the divergence conversion of the response direction dimension. The higher the proportion of conflict candidate states, the closer the divergence in the response direction dimension is to 1. The response location is jointly determined by the branch set where the risk exceeds the early warning observation threshold and the branch set where the risk decrease reaches the minimum effective response threshold. The response description vector is used to indicate whether the risk distribution location is stable after the candidate control action is taken. The early warning observation threshold is preferentially adopted from the preventive observation threshold calibrated by the scheduling review record in the segment recognition module. When the number of historical safe execution segments reaches 30 and the observation boundary before the alarm is stable, the response identification module uses the observation boundary before the alarm in the historical safe execution segments to correct the early warning observation threshold. The response time is determined by the time when the minimum effective response threshold is first reached after the candidate control action is completed. If the observation equivalent candidate twin state does not reach the minimum effective response threshold within the unified response observation window, the response identification module marks the observation equivalent candidate twin state as having no effective response and determines the response time according to the upper limit of the unified response observation window plus a sampling period. Through this processing, the response description vector unifies the candidate control action response under different observation equivalent candidate twin states to four comparable dimensions: direction, amplitude, position, and time.

[0052] The response identification module performs hierarchical divergence transformation on the response description vectors of each observation equivalent candidate twin state, obtaining dimensionless divergence values ​​for the response direction dimension, response magnitude dimension, response location dimension, and response time dimension. The divergence value for the response direction dimension is formed based on the consistency of the risk change direction results among the observation equivalent candidate twin states. When multiple observation equivalent candidate twin states simultaneously have response directions of risk reduction, risk maintenance, and risk increase, the divergence value for the response direction dimension is set to 1. When multiple observation equivalent candidate twin states have only two types of response directions, the divergence value for the response direction dimension is determined according to the lesser of the two types of response directions. The proportion of candidate states corresponding to the response direction is converted to a dimensionless result ranging from 0 to 1; when the response directions of all equivalent candidate twin states are consistent, the divergence in the response direction dimension is set to 0, and the divergence in the response amplitude dimension is formed based on the dispersion of the risk change amplitude results under each equivalent candidate twin state, and the upper bound of the dispersion of the response amplitude of historically verified wind adjustment actions under stable candidate states is used as the normalization boundary; when the number of historically verified wind adjustment actions is less than 30, the upper bound of the dispersion of the response amplitude is jointly determined by the minimum effective response threshold, sensor resolution, and scheduling verification records; when the risk change amplitude results are consistent, the upper bound of the response amplitude is set to 0. When the dispersion of the result exceeds the upper bound of the response amplitude dispersion, the discrepancy in the response amplitude dimension is truncated to 1. The discrepancy in the response location dimension is formed based on the degree of overlap of response locations in each equivalent candidate twin state. When response locations completely overlap, the discrepancy in the response location dimension is 0; when response locations do not overlap, the discrepancy in the response location dimension is 1; when partially overlapping, it is converted to a dimensionless result of 0 to 1 according to the overlap ratio. The discrepancy in the response time dimension is formed based on the degree of dispersion of response time in each equivalent candidate twin state, and the upper bound of the allowable response time dispersion is used as the normalization boundary; the upper bound of the allowable response time dispersion... The boundaries are derived from the ventilation propagation time, actuator response error, and sensor sampling cycle that affect the branch set. When the number of historical records of similar air adjustment actions is less than 30, the upper limit of the allowable response time dispersion is determined by 20% to 40% of the length of the unified response observation window. For high-gas areas and critical return air paths, the upper limit is 20% to 30%, and for ordinary ventilation areas, the upper limit is 30% to 40%. When the degree of response time dispersion exceeds the upper limit of the allowable response time dispersion, the divergence of the response time dimension is truncated to 1. The divergence of the four dimensions is a dimensionless result of 0 to 1. The closer the value is to 1, the more inconsistent the corresponding response dimension is.

[0053] The response identification module sets fusion weights for each dimension of divergence based on response misjudgment records in historically verified wind transfer actions. Historically verified wind transfer actions refer to those that have been reviewed by scheduling, execution result playback, and source tag verification, and whose response direction, response location, response amplitude, and response time have all been reviewed. The fusion weights for each dimension are dimensionless weights ranging from 0 to 1, and are normalized before participating in the formation of action response divergence results. When the number of historically verified wind transfer actions is less than 30, the fusion weights of the response direction and response location dimensions are not lower than the fusion weights of the response amplitude and response time dimensions. When the number of historically verified wind transfer actions reaches 30, the response identification module adjusts the fusion weights according to the contribution of direction misjudgment, amplitude misjudgment, location misjudgment, and response delay misjudgment to subsequent risks. This weight setting is used to prioritize and constrain response divergences that may cause risk direction misjudgment and risk location transfer.

[0054] After the aforementioned divergence quantities are formed, the response identification module generates an action response divergence result, one preferred expression of which is: ;

[0055] in, The output of the segment recognition module represents the first segment. A preventive air conditioning segment, which is derived from the segment start and end time and source label in the preventive air conditioning segment confidence results; Indicates the first The divergence results of the action response to the candidate control action for each preventive wind adjustment segment range from 0 to 1. The closer the value is to 1, the more inconsistent the response of the same candidate control action is under the observed equivalent candidate twin state. This represents the divergence in the response direction dimension, which is derived from the consistency of the risk change direction results and the direction conflict labels. This represents the divergence in the response magnitude dimension, which is derived from the response description vector and originates from the degree of dispersion of the risk change magnitude results under each observation equivalent candidate twin state. This represents the divergence of the response location dimension, which is derived from the response description vector and originates from the degree of overlap of response locations in the equivalent candidate twin states of each observation. This represents the divergence in the response time dimension, which is derived from the response description vector and originates from the degree of dispersion of response time in each observation-equivalent candidate twin state. These represent the fusion weights for the response direction, response amplitude, response location, and response time dimensions, respectively. All are dimensionless weights ranging from 0 to 1, and the sum of the four fusion weights is 1. All the aforementioned divergence quantities are dimensionless results ranging from 0 to 1. Each fusion weight is determined from historically verified wind adjustment actions based on response misjudgment records and review records. When the number of historically verified wind adjustment actions is less than 30, the basic weights for the response direction and response location dimensions are prioritized, with the fusion weights for both dimensions not less than 0.25. The fusion weights for the response amplitude and response time dimensions are allocated from the remaining weights and then distributed during the final calculation. After being processed, the formula uses a weighted fusion method because directional divergence is used to constrain whether candidate control actions will cause misjudgment of risk direction, positional divergence is used to constrain whether the risk transfer position is stable, amplitude divergence is used to constrain whether the degree of risk correction is consistent, and time divergence is used to constrain whether the response arrival is reproducible. The four types of divergence jointly determine whether candidate control actions have identifiable action responses in the observation equivalent candidate twin state. The action response divergence results enter the formation process of subsequent action response identifiable permission results and are compared with the response divergence permission threshold to restrict unidentifiable response fragments from entering the action value correction permission path.

[0056] The response identification module also performs boundary state processing. When all observed equivalent candidate twin states have not formed response positions and the risk change magnitude results are all below the minimum effective response threshold, the divergence of the response position dimension is set to 0, and a "no effective response" label is written in the counterfactual evidence state. When some observed equivalent candidate twin states form response positions and others do not, the divergence of the response position dimension is set to 1. If the observed equivalent candidate twin states do not reach the minimum effective response threshold within the unified response observation window, the response time is determined by adding one sampling period to the upper limit of the unified response observation window. This boundary processing is used to avoid the irreproducibility of action response divergence results caused by empty sets, no effective responses, and undefined response times.

[0057] After generating action response divergence results, the response identification module generates action response identifiable permission results. These results are jointly determined by the action response divergence results, implicit security benefit results, and counterfactual inference credibility. When the action response divergence results do not exceed the response divergence permission threshold, the implicit security benefit results reach the security benefit access threshold, and the counterfactual inference credibility reaches the counterfactual credibility threshold, the response identification module generates action response identifiable permission results. Historical security execution segments refer to historical segments that have undergone scheduling review, actual execution result playback, and source tag verification, and have not triggered the prohibited state of the security procedure. The response divergence permission threshold is a dimensionless threshold between 0 and 1, determined by the upper quartile of the action response divergence results in the historical security execution segments. When the number of historical security execution segments is less than 30, the response divergence permission threshold is preferably set to 0. The threshold is 0.25 to 0.45. For high-gas areas, critical return air paths, and areas with a high proportion of weak observations, the threshold is 0.25 to 0.35. For areas with sufficient measurement point coverage and stable playback errors, the threshold is 0.35 to 0.45. Historically verified preventive ventilation segments refer to historical segments that have been confirmed to have implicit safety benefits after scheduling review, counterfactual deduction review, and source label verification. The safety benefit admission threshold is a dimensionless threshold of 0 to 1, determined by the lower quartile of the implicit safety benefit results in the historically verified preventive ventilation segments. When the number of historically verified preventive ventilation segments is less than 30, the safety benefit admission threshold is preferably 0.1 to 0.3. For high-gas areas and critical return air paths, the threshold is 0.1 to 0.2. For ordinary ventilation areas, the threshold is 0.2 to 0.3. The above thresholds are used for training sample admission and action value correction permission, and are not used as alarm thresholds in mine safety regulations.

[0058] When the action response divergence result exceeds the response divergence permission threshold, the response identification module transfers the corresponding fragment to the low-confidence verification sample set and prohibits the corresponding fragment from reducing the action value of historical wind-adjustment actions. When the implicit security benefit result does not reach the security benefit admission threshold, or the counterfactual inference credibility does not reach the counterfactual credibility threshold, the response identification module does not form an action response identifiable permission result and retains the corresponding source label in the counterfactual evidence state. The low-confidence verification sample set is only used for policy risk verification in the policy correction module and does not participate in the positive increase or negative decrease of action value. This processing is used to avoid fragments with inconsistent responses between observation equivalent candidate twin states being directly written into the reinforcement learning action value.

[0059] The final response identification module outputs the set of observed equivalent candidate twin states, the candidate state fitting error result, the response description vector, the action-response divergence result, and the action-response identifiable permission result. The set of observed equivalent candidate twin states is used to record the state competition range on which the action-response comparison is based. The candidate state fitting error result is used to record the fitting reliability of the observed equivalent candidate twin states before entering the action-response comparison. The action-response divergence result is used to gating the implicit security gain result. The action-response identifiable permission result is used to determine which fragments enter the high-confidence training sample set, the low-confidence verification sample set, and the fragments that do not enter the action value update in the policy correction module. Through the above processing, the response identification module further limits the implicit security gain result generated by the counterfactual inference module to a credible gain that can be used for reinforcement learning action value correction, avoiding the writing of unstable rewards into the policy network due to the inconsistency of responses between observed equivalent candidate twin states.

[0060] The strategy correction module receives the implicit security gain result, counterfactual inference credibility, and counterfactual evidence state output by the counterfactual inference module, and receives the observation equivalent candidate twin state set, candidate state fitting error result, response description vector, action response divergence result, and action response identifiable permission result output by the response identification module. The implicit security gain result is used to represent the security gain that was hidden after the historical wind control action was executed in the historical record. The counterfactual inference credibility is used to represent the reliability of the counterfactual evidence of the security gain. The action response divergence result is used to represent the degree of inconsistency in the response of the same candidate control action in the observation equivalent candidate twin state. The action response identifiable permission result is used to determine whether the corresponding segment is allowed to enter the action value correction path. The counterfactual evidence state and candidate state fitting error result are used to restrict low-confidence segments from entering the action value update. The offline reinforcement learning action value update is used as a supporting technology. Its input is the training sample set. Its processing purpose is to correct the value mapping between the state and the action. Its output is the reinforcement learning action value. The core processing of this implementation method is to perform permission control, reward correction, sample contribution control, online pre-execution verification, and closed-loop write-back restriction on the training samples.

[0061] The strategy correction module first generates action value correction permission results. Taking implicit security benefit results, counterfactual deduction credibility, action response divergence results, action response identifiable permission results, counterfactual evidence state, and candidate state fitting error results as input, the module sequentially performs benefit admission verification, credibility admission verification, response divergence permission verification, and identifiable state verification. Benefit admission verification determines whether the implicit security benefit results reach the security benefit admission threshold determined in the response identification module. Credibility admission verification determines whether the counterfactual deduction credibility reaches the counterfactual credibility threshold determined in the counterfactual deduction module. Response divergence permission verification determines whether the action response divergence results do not exceed the response identification threshold. The response divergence permission thresholds determined in the block; all of the above thresholds are dimensionless thresholds from 0 to 1, used only for action value correction admission, and not as alarm thresholds in mine safety regulations. The identifiable state verification is used to determine whether the action response identifiable permission result is a permission state, and at the same time to determine whether the counterfactual evidence state does not contain the action source to be confirmed label, physical constraint not met label, disturbance missing label, counterfactual evidence insufficient label, and execution feedback insufficient label, and to determine whether the candidate state fitting error result does not contain the candidate state support insufficient label. When all of the above verifications pass, the action value correction permission result is determined to be a permission update state; when any verification fails, the action value correction permission result is determined to be a prohibited positive update state.

[0062] The strategy correction module forms training sample roles based on the action value correction permission results. Segments with action value correction permission results in a permission update state enter the high-confidence training sample set for positive action value correction in reinforcement learning. Segments with implicit security benefit results that do not reach the security benefit admission threshold, action response divergence results that exceed the response divergence permission threshold, action response identifiable permission results that do not form a permission state, insufficient candidate state support, or counterfactual evidence states containing the label of insufficient execution feedback enter the low-confidence verification sample set. These are only used for strategy risk verification and do not participate in positive or negative action value updates. Segments with labels of no implicit security benefit, unmet physical constraints, or insufficient counterfactual evidence are determined not to enter action value updates, and their source labels are retained for subsequent parameter calibration. Through this process, trainable segments, verification segments, and segments that do not enter action value updates complete role separation before entering action value correction.

[0063] For segments entering the high-confidence training sample set, the policy correction module constructs a reinforcement learning training tuple. The reinforcement learning training tuple includes the state representation before the wind adjustment action is executed, the candidate adjustment action, the corrected reward, the state representation after the candidate adjustment action is executed, and the sample update weights. The state representation before the wind adjustment action is executed comes from the segment-level risk representation result before the wind adjustment in the segment recognition module, the counterfactual initial state in the counterfactual inference module, and the observation equivalent candidate twin state set in the response identification module. The candidate adjustment action comes from the same candidate adjustment action determined in the response identification module. The state representation after the candidate adjustment action is executed is based on the observed risk result after the actual wind adjustment. The candidate action risk process is only used as the basis for consistency verification and the basis for sample update weight correction, and does not replace the observed risk result after the actual wind adjustment. The above processing is used to prevent the inference result in the candidate twin state from being mistakenly written into the training tuple as the real next state.

[0064] The strategy correction module generates a positive reward component, a deduction component, and a corrected reward. The positive reward component is a dimensionless result ranging from 0 to 1, and its inputs include the implicit security gain result, the counterfactual inference confidence level, and the confidence result of the preventive wind adjustment segment. The strategy correction module first reads the implicit security gain result as the basic security gain, and then performs a confidence correction on the basic security gain according to the counterfactual inference confidence level and the preventive wind adjustment segment confidence result to obtain the positive reward component. When the number of historical high-confidence training samples is less than 30, the positive reward component adopts a conservative value, not exceeding the lower value between the implicit security gain result and the counterfactual inference confidence level. The deduction component is a dimensionless result ranging from 0 to 1, and its inputs include the action response divergence result, the candidate state fitting error result, and the counterfactual evidence state. The strategy correction module uses the action response score... The divergence result serves as the basic deduction basis, and a label deduction basis is formed based on the insufficient measurement point coverage label, insufficient resistance boundary calibration label, insufficient air leakage boundary calibration label, insufficient candidate state support label, and execution feedback deviation label in the candidate state fitting error result. When the number of historical high-confidence training samples is less than 30, the deduction contribution of the label deduction basis is not less than the deduction contribution of the action response divergence result. The corrected reward is formed by deducting the contribution corresponding to the deduction component from the positive reward component and is limited to the range of 0 to 1. When the contribution corresponding to the deduction component reaches or exceeds the positive reward component, the corrected reward is determined to be 0. When the action value correction permission result is not a permission update state, or the action response divergence result exceeds the response divergence permission threshold, the strategy correction module does not form a reward contribution for positive update.

[0065] The policy correction module further generates sample update weights, which are dimensionless weights ranging from 0 to 1. Their inputs include the action value correction permission result, counterfactual inference credibility, action response divergence result, and candidate state fitting error result. When the action value correction permission result is in a permitted update state, the policy correction module uses the counterfactual inference credibility as the positive source of the sample update weights and the action response divergence result and candidate state fitting error result as the deduction source to form the sample update weights. When the action value correction permission result is in a prohibited positive update state, the sample update weights are reset to 0. When the number of historically high-credibility training samples is less than 30, the improvement contribution corresponding to the counterfactual inference credibility does not exceed the combined deduction contribution of the action response divergence result and candidate state fitting error result. After the number of historically high-credibility training samples reaches 30, the policy correction module corrects the sample update weights based on the historical policy risk verification results and action execution review results. The sample update weights are used to limit the contribution of a single training tuple to the value of reinforcement learning actions.

[0066] After forming the reinforcement learning training tuples, the policy correction module uses the reinforcement learning training tuples in the high-confidence training sample set as training input and corrects the reinforcement learning action value using the offline reinforcement learning action value update method. During the correction process, the positive reward component is used to compensate for the implicit safety benefits of preventive wind adjustment actions, the deduction component is used to suppress the erroneous writing of rewards caused by the indistinguishability of action responses, and the sample update weight is used to limit the contribution ratio of each training tuple to the reinforcement learning action value. The low-confidence verification sample set does not participate in the positive and negative updates of reinforcement learning action value and is only used for subsequent policy risk verification. After the offline reinforcement learning action value update is completed, the policy correction module outputs the updated reinforcement learning action value as the basis for the action value generation of online candidate control actions.

[0067] During online operation, the strategy correction module generates the current mine ventilation status based on the current digital twin ventilation network status and real-time ventilation monitoring data, and performs a status validity confirmation on the current mine ventilation status. The status validity confirmation takes the real-time measurement point communication status, maintenance status, sensor calibration status, measurement point missing ratio, and digital twin ventilation network status update time as inputs. The processing action is to determine whether the real-time ventilation monitoring data and the digital twin ventilation network status are in the same online confirmation window, and output a valid result or an invalid result for the current mine ventilation status. After a valid result for the current mine ventilation status is formed, the strategy correction module generates candidate control actions based on the current mine ventilation status and the value of reinforcement learning actions. When an invalid result for the current mine ventilation status is formed, the strategy correction module outputs a review prompt and does not generate automatically issued candidate control actions.

[0068] The strategy correction module sequentially performs action response identification judgment, procedure safety constraint judgment, and actuator reachability judgment on candidate control actions. The action response identification judgment is implemented using the observation-equivalent candidate twin state construction and action response divergence calculation methods in the response identification module. Starting from the current mine ventilation state, it regenerates the current observation-equivalent candidate twin state set and the current action response divergence result to determine whether the candidate control action has a risk of unidentifiable response under the current mine ventilation state. The procedure safety constraint judgment uses the mine safety regulations, ventilation operation limits, and the current mine ventilation state as the basis. The input determines whether the candidate control action will cause the gas concentration, air volume, pressure difference, thermal humidity state, and return air path to exceed the safety allowable boundary. The actuator reachability determination takes the candidate control action, actuator status, and actuator operating boundary as input to determine whether the action direction, action amplitude, and action duration of the candidate control action are within the executable range of the fan, damper, and local ventilation fan. The actuator operating boundary is determined by the fan frequency range, damper opening range, local ventilation fan operating status, and equipment interlock status. When the equipment interlock status prohibits execution, the actuator reachability result is determined to be unreachable.

[0069] When a candidate control action simultaneously passes the action response identification judgment, procedure safety constraint judgment, and actuator reachability judgment, the strategy correction module outputs a permitted execution result and incorporates the candidate control action into the mine ventilation control strategy. The permitted execution result includes the candidate control action, action value result, action response divergence result, procedure safety constraint result, actuator reachability result, and strategy credibility result. The action value result is determined by reinforcement learning action value and converted to a dimensionless action value confidence level of 0 to 1 according to the action value distribution in the historical high-confidence training sample set. When the number of historical online execution samples is less than 30, the action value confidence level is conservatively converted, and its value is not higher than the historical review pass rate of similar candidate control actions, and it is not used as the sole basis for the validity of the permitted execution result. The strategy credibility result is a dimensionless result of 0 to 1. The inputs include action value results, action response divergence results, candidate state fitting error results, and real-time measurement point credibility. Real-time measurement point credibility is formed by the current measurement point communication status, maintenance status, sensor calibration status, and measurement point missing ratio. The strategy correction module first generates a positive strategy credibility result based on the action value results and real-time measurement point credibility, and then generates a strategy credibility deduction result based on the action response divergence results and candidate state fitting error results. The strategy credibility result is then limited to the range of 0 to 1 after deducting the strategy credibility deduction result from the positive strategy credibility result. When the number of historical online execution samples is less than 30, the contribution of the strategy credibility deduction result is no less than the contribution of the positive strategy credibility result. The strategy credibility result is used to characterize the credibility of the candidate control action as the output of the mine ventilation control strategy under the current mine ventilation state.

[0070] When a candidate control action fails to pass the action response identification, procedure safety constraint judgment, and actuator reachability judgment simultaneously, and the current mine ventilation status meets the risk warning conditions of a preventive ventilation adjustment segment, the strategy correction module generates a downgraded candidate action. The risk warning conditions are determined by the segment-level risk characterization results before ventilation adjustment, the preventive observation threshold, and the source label in the segment identification module. When the current mine ventilation status exhibits risk warning conditions such as approaching methane levels, insufficient air supply, abnormal pressure difference, increasing air demand at the working face, or deteriorating thermal and humidity conditions, the strategy correction module generates a restricted downgraded candidate action. The action amplitude of the downgraded candidate action is reduced step by step according to the minimum adjustable step size of the corresponding actuator, and must not exceed the amplitude of the original candidate control action. Its scope of action is limited to the current set of affected branches or the ventilation branch corresponding to the current risk warning. The current mine ventilation status, the target of the candidate control action, and the current digital twin ventilation network status are re-formed according to the influence branch constraint method in the segment recognition module. The duration of the action does not exceed the duration of the original candidate control action and is jointly constrained by the unified response observation window and the mine safety regulations. The minimum adjustable step size is determined by the technical parameters of the executing agency and the scheduling review record. When the number of historical downgrade execution samples is less than 30, the downgrade candidate action is preferentially reduced by one minimum adjustable step size. The downgrade candidate action shall not exceed the minimum value among the mine safety regulations, the upper limit of the executing agency's mechanical permit, and the upper limit of the scheduling review permit. The upper limit of the executing agency's mechanical permit is determined by the fan frequency range, the air door opening range, and the local ventilation fan operating status. The upper limit of the scheduling review permit is determined by the mine ventilation scheduling record and safety management requirements.

[0071] The strategy correction module continues to perform action response identification judgment, procedure safety constraint judgment, and actuator reachability judgment for downgraded candidate actions. When a downgraded candidate action passes the above judgments, the strategy correction module outputs the downgrade execution result; when a downgraded candidate action fails the above judgments, the strategy correction module outputs a review prompt and does not issue automatic control instructions to the actuator. The review prompt includes the candidate control action, the downgraded candidate action, the action response divergence result, the reason for the procedure safety constraint failure, the reason for the actuator unreachability, the strategy credibility result, and the source label. Through this processing, when there are risk warnings but the candidate control action does not meet the execution conditions, the system still retains the constrained preventive response path and blocks actions that fail the safety constraints from entering the automatic control chain.

[0072] The strategy correction module records and updates the final execution results. The system writes the permitted execution result, downgraded execution result, review prompts, and post-execution data such as air volume, pressure difference, gas concentration, thermal and humidity conditions, actuator response time, and risk reduction results into the mine ventilation operation record. This record is not directly used as a new high-confidence training sample but is used for subsequent review and source label verification. If the risk reduction direction, response location, response time, and measurement point deviation after execution are consistent with the expected results in the permitted or downgraded execution results, and the deviation does not exceed the error allowable range determined by the historical safe execution segment, and no low-confidence label in the counterfactual evidence state is triggered, the strategy correction module uses the corresponding record as a candidate high-confidence training sample. When the number of execution segments is less than 30, the error tolerance range is jointly determined by the mine safety regulations, sensor allowable errors, and scheduling review records. Candidate high-confidence training samples are only allowed to enter the subsequent high-confidence training sample set after the source label verification and execution result review. The benchmark update chain only accepts records that have been reviewed and have not triggered low-confidence labels. The policy execution record itself does not directly update the action value benchmark. When the observation results after execution are inconsistent with the expectations, the policy correction module writes a policy execution deviation label and prohibits the record from entering the benchmark update chain. Through the above processing, the result writeback of the policy correction module distinguishes between policy execution records, review records, and subsequent training candidate records, avoiding low-confidence control results from entering the next round of action value correction.

[0073] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A digital twin-driven reinforcement learning control system for mine ventilation networks, characterized in that, include: The segment identification module acquires mine ventilation operation records and digital twin ventilation network status. Using event anchor alignment and influence branch constraint, it converts the segment-level risks before and after historical ventilation adjustment actions into homogeneous risk representations. The comparison process is corrected with interference coverage results to form confidence results for preventive ventilation adjustment segments. The counterfactual deduction module uses the confidence results of the preventive ventilation adjustment segment to locate the initial counterfactual state, removes the corresponding historical ventilation adjustment actions, reconstructs the sequence of uninterrupted counterfactual actions, and generates uninterrupted counterfactual risk results and implicit safety benefit results through ventilation network balancing playback; The response identification module takes the non-intervention counterfactual risk result, constructs a set of observation equivalent candidate twin states that meet the measurement point error allowance and ventilation network physical constraints, performs action response divergence calculation between the observation equivalent candidate twin states to perform the same candidate control action, and forms action response divergence result and action response identifiable allowance result. The strategy correction module uses an action value correction permission control method to process implicit safety benefit results, action response divergence results, and action response identifiable permission results, generate action value correction permission results, correct reinforcement learning action values ​​when the action value correction permission result is in the permission update state, and output mine ventilation control strategy.

2. The digital twin-driven mine ventilation network reinforcement learning control system according to claim 1, characterized in that, The specific operations of event anchor point alignment and influence branch constraint are as follows: read the historical ventilation adjustment actions in the mine ventilation operation record, and establish event anchor points based on the time when the historical ventilation adjustment actions occurred; based on the event anchor points, form a pre-adjustment window and a post-adjustment verification window according to the early warning lead and ventilation propagation delay in the mine ventilation operation record; map the ventilation adjustment execution objects of the historical ventilation adjustment actions to the branches of the digital twin ventilation network; expand the ventilation branches associated with the ventilation adjustment execution objects using airflow reachability constraints and pressure difference coupling constraints; perform exclusion processing on ventilation branches that do not meet the connectivity relationship of the digital twin ventilation network to form an influence branch set.

3. The digital twin-driven mine ventilation network reinforcement learning control system according to claim 1, characterized in that, The specific steps for converting the segment-level risks before and after historical ventilation adjustments into a uniform risk representation and correcting the comparison with interference coverage results are as follows: Within the set of influencing branches, read the ventilation risk change process of the window before ventilation adjustment; use a uniform conversion method to convert the ventilation risk change process into a risk representation quantity with a consistent numerical direction; form branch influence weights based on the correlation strength of ventilation paths, and merge the risk representation quantities of each branch into the segment-level risk representation result before ventilation adjustment according to the branch influence weights; within the verification window after ventilation adjustment, use the same conversion caliber to form the observed risk result after actual ventilation adjustment; generate interference coverage results using operational interference records from mine ventilation operation records; and use the interference coverage results to subtract the confidence strength of the risk comparison before and after ventilation adjustment to generate a preventative ventilation adjustment segment confidence result.

4. The digital twin-driven mine ventilation network reinforcement learning control system according to claim 1, characterized in that, The specific steps for locating the initial counterfactual state and reconstructing the sequence of uninterrupted counterfactual actions based on the confidence results of preventive air adjustment segments are as follows: Read the segment start and end times, historical air adjustment actions, influencing branch sets, and segment-level risk characterization results before air adjustment from the confidence results of the preventive air adjustment segments; extract the stable ventilation state before the execution of historical air adjustment actions based on the segment start and end times, and convert the stable ventilation state into the initial counterfactual state; delete the historical air adjustment actions corresponding to the preventive air adjustment segments from the actual action sequence; identify derived air adjustment actions that have a causal relationship with historical air adjustment actions using the air adjustment action association identification method, and simultaneously delete the derived air adjustment actions to form the sequence of uninterrupted counterfactual actions.

5. The digital twin-driven mine ventilation network reinforcement learning control system according to claim 1, characterized in that, The specific operation for generating uninterrupted counterfactual risk results and implicit safety benefit results through ventilation network balancing playback is as follows: taking the initial counterfactual state as the playback starting point, applying the uninterrupted counterfactual action sequence to the ventilation network structure; and using the ventilation network balancing solution method to replay the ventilation response process under the uninterrupted state. During playback, physical constraints on the ventilation network are applied, and a weak observation compensation method is used to correct playback biases caused by weak observation areas, resulting in non-interventional counterfactual risk outcomes. The counterfactual risk results without intervention are compared with the observed risk results after actual wind adjustment, using the same risk caliber. The credibility of the counterfactual inference is formed by combining the consistency results of the counterfactual model playback and the consistency results of the disturbance. When the counterfactual risk outcome without intervention is higher than the observed risk outcome after the actual wind adjustment, and the credibility of the counterfactual inference reaches the counterfactual credibility threshold, the risk difference is converted into an implicit security benefit outcome.

6. The digital twin-driven mine ventilation network reinforcement learning control system according to claim 1, characterized in that, The specific operations for constructing a set of observation-equivalent candidate twin states that satisfy the measurement point error allowance and ventilation network physical constraints are as follows: read the counterfactual initial state, the set of influence branches, and the digital twin ventilation network state; perform boundary calibration on the uncertain parameters of the observation-equivalent candidate twin states based on the digital twin ventilation network state and historical playback errors; perform constraint sampling within the calibrated value range to generate two or more initial candidate twin states; perform measurement point error verification and ventilation network physical constraint verification on each initial candidate twin state; determine the initial candidate twin state that simultaneously satisfies the measurement point error allowance and ventilation network physical constraints as the observation-equivalent candidate twin state and include it in the set of observation-equivalent candidate twin states.

7. The digital twin-driven mine ventilation network reinforcement learning control system according to claim 1, characterized in that, The specific operation for calculating the action response divergence between observed equivalent candidate twin states is as follows: input the same candidate control action into each observed equivalent candidate twin state in the observed equivalent candidate twin state set; and replay the candidate baseline risk process without inputting the candidate control action and the candidate action risk process after inputting the candidate control action in the unified response observation window. The risk change is calculated based on the difference between the candidate baseline risk process and the candidate action risk process. Extract the response direction, response location, and response time corresponding to candidate control actions along the unified response observation window, and convert the risk change, response direction, response location, and response time into response description vectors.

8. A digital twin-driven mine ventilation network reinforcement learning control system according to claim 7, characterized in that, The specific operations for forming action response divergence results and action response identifiable permit results are as follows: A hierarchical divergence transformation is performed on the response description vectors of each observation equivalent candidate twin state to obtain dimensionless divergence quantities in the response direction dimension, response amplitude dimension, response location dimension, and response time dimension; fusion weights are set for each dimension of divergence quantities based on historically verified response misjudgment records in wind regulation actions; the divergence quantities of each dimension are fused into action response divergence results according to the fusion weights; when the action response divergence results do not exceed the response divergence permit threshold, the implicit security benefit results reach the security benefit admission threshold, and the counterfactual inference credibility reaches the counterfactual credibility threshold, an action response identifiable permit result is formed; when the action response divergence results exceed the response divergence permit threshold, the corresponding segment is transferred to the low-credibility verification sample set, and the corresponding segment is prohibited from entering the negative update of action value.

9. A digital twin-driven mine ventilation network reinforcement learning control system according to claim 1, characterized in that, The specific operation of handling implicit security benefit results, action response divergence results, and action response identifiable permission results using the action value correction permission control method is as follows: Read the implicit security benefit result, the counterfactual inference credibility generated by the counterfactual inference module, the action response divergence result, and the action response identifiable permission result; sequentially execute the benefit admission verification, credibility admission verification, response divergence permission verification, and identifiable state verification; when all three verifications pass, the action value correction permission result is determined as a permission update state; when any one of these verifications fails, the action value correction permission result is determined as a prohibited forward update state; the segment corresponding to the permission update state is included in the high-credibility training sample set, and the segment corresponding to the unidentifiable response state is included in the low-credibility verification sample set.

10. A digital twin-driven mine ventilation network reinforcement learning control system according to claim 1, characterized in that, The specific operation of correcting the value of reinforcement learning actions and outputting the mine ventilation control strategy is as follows: constructing fragments in the high-confidence training sample set as reinforcement learning training tuples; using the implicit safety benefit results to form positive reward components, using the action response divergence results to form deduction components, and using the sample update weights formed by the action value correction permission results and the action response divergence results to control the contribution of the reinforcement learning training tuples to the correction of the value of reinforcement learning actions. During online operation, the system performs action response identification, procedure safety constraint judgment, and actuator reachability judgment on candidate control actions. If all three judgments pass, a permitted execution result is output. If a candidate control action fails all three judgments simultaneously, and the current mine ventilation status meets the risk warning conditions for a preventative ventilation adjustment segment, a downgraded candidate action is generated, and these judgments are applied. If the downgraded candidate action passes all three judgments, a downgrade execution result is output. If the downgraded candidate action fails all three judgments, a review prompt is output.

Images