A gas inspection task dispatching and assessment method

Abstract

This invention discloses a method for assigning and assessing gas inspection tasks, belonging to the field of coal mine safety information technology. The method includes: real-time acquisition of multi-source heterogeneous data underground through multi-source sensors and a UWB positioning system; preprocessing, feature extraction, and spatiotemporal fusion of the data before inputting it into an improved LSTM model to predict gas concentration distribution and risk level; constructing a multi-dimensional dynamic capability profile of gas inspectors and calculating the task-person matching degree matrix; constructing and solving a multi-objective optimization function including assessment feedback items to generate the optimal task allocation and path planning scheme; issuing tasks and collecting execution data in real time; using a dynamic weighting algorithm to calculate assessment scores in real time and update the capability profile; generating graded assessment early warning signals based on real-time scores compared with historical baselines, and feeding them back to the assignment stage to trigger dynamic adjustments to the assignment strategy. This invention achieves closed-loop linkage between task assignment and assessment, improving the timeliness and accuracy of gas risk prevention and control.

Description

Technical Field

[0001] This invention relates to the field of coal mine safety information technology, specifically a method for assigning and assessing gas inspection tasks. Background Technology

[0002] Coal mine gas inspection is a key link in ensuring safe production in the mine. The rational allocation of gas inspection tasks and the scientific assessment of the work quality of gas inspectors are of vital importance to preventing gas accidents and protecting the lives of miners.

[0003] Currently, the assignment of gas inspection tasks and the performance evaluation of personnel generally suffer from the following technical deficiencies:

[0004] 1. Task assignment relies on static planning and lacks dynamic risk response capabilities. Existing inspection plans are based on fixed schedules and roadmaps, and mainly rely on spatial topology simulation analysis based on geological structure distribution data and real-time gas concentration data. They cannot adaptively adjust according to real-time changes in underground gas concentration and dynamic evolution of environmental risks. When the gas concentration in a certain area underground rises abnormally or the mining progress changes, the existing system cannot promptly add or adjust inspection tasks, resulting in insufficient inspection frequency in high-risk areas.

[0005] 2. Task assignment and personnel assessment are independent of each other and lack a two-way linkage mechanism at the technical level. In the existing system, the task assignment module only allocates tasks based on static inspection plans or risk data, and the assessment module only performs statistical scoring on historical execution data at the end of the month or quarter. There is no real-time data interaction and feedback loop between the two. The assessment results cannot immediately affect the subsequent task assignment strategy, and the assignment strategy cannot be dynamically adjusted according to personnel performance, resulting in a break in the closed-loop link of "assignment, execution, assessment and optimization".

[0006] 3. The performance evaluation indicators lack timeliness, dynamic adjustment capability, and real-time early warning mechanism. Existing evaluations mostly adopt fixed indicator systems, which cannot dynamically adjust the evaluation weights according to factors such as task difficulty, risk level, and unexpected situations. Moreover, they are all post-event evaluations, which cannot dynamically score and warn based on the real-time performance of gas inspectors during task execution, and cannot intervene and correct possible false inspections, missed inspections, and other behaviors in advance.

[0007] Therefore, a method for assigning and assessing gas inspection tasks is proposed. Summary of the Invention

[0008] The purpose of this invention is to provide a method for assigning and assessing gas inspection tasks, so as to solve the problems mentioned in the background art.

[0009] To achieve the above objectives, the present invention provides the following technical solution: a method for assigning and assessing gas inspection tasks, comprising the following steps:

[0010] S1: Through multi-parameter gas sensors deployed at various inspection points underground, UWB positioning base stations deployed underground, intrinsically safe intelligent terminals for gas inspectors, and explosion-proof inspection robots deployed in roadways, multi-source heterogeneous data from underground coal mines are collected in real time, and the collected data is transmitted to the ground data center in real time through the mine industrial Ethernet ring network or 5G private network.

[0011] S2: Preprocess, extract features, and perform spatiotemporal fusion of the multi-source heterogeneous data;

[0012] S3: Input the fused feature data output from step S2 into the improved LSTM gas concentration prediction model to predict the gas concentration distribution and risk level of each inspection area within multiple future time steps; the improved LSTM gas concentration prediction model adopts a dual-layer LSTM architecture combined with an attention mechanism.

[0013] S4: Based on the inspectors' historical assessment records, skill level certifications, execution efficiency statistics, and violation records, construct a dynamic personnel capability vector containing multiple dimensions, and update the values ​​of each dimension of the capability vector in real time;

[0014] S5: Based on the requirement feature vector of each inspection task output in step S3 and the ability vector of each gas inspector output in step S4, calculate the matching degree matrix between each gas inspector and each inspection task.

[0015] S6: Based on the matching degree matrix output in step S5, the risk prediction results output in step S3, and the real-time UWB positioning and working status of each inspector, construct and solve a multi-objective optimization function that includes assessment feedback items to generate the optimal task allocation scheme and inspection path planning scheme.

[0016] S7: The task allocation scheme generated in step S6 is sent to the corresponding intrinsically safe intelligent terminal or inspection robot in the mine through the underground 5G private network, and the UWB positioning data, time data and detection data during the execution process are collected in real time.

[0017] S8: Based on the execution data collected in step S7, combined with the risk level and difficulty coefficient of the task, a dynamic weighting algorithm is used to calculate the current task assessment score of each inspector in real time, and update the corresponding dimension data in the personnel capability vector mentioned in step S4.

[0018] S9: During task execution, the real-time scoring results from step S8 are compared with the historical baseline of the inspector's capabilities. When the real-time score deviates from the historical baseline by more than a preset threshold, an assessment warning signal is generated and fed back to step S6, triggering an adjustment to the subsequent task assignment strategy.

[0019] Preferably, the multi-source heterogeneous data in step S1 includes: static basic data, dynamic environmental data, and inspection resource data;

[0020] The static basic data includes: geological structure distribution data of coal mine shafts, roadway topology data—including spatial coordinates, connectivity, roadway length and slope information, and spatial layout data of fixed sensors and inspection points.

[0021] The dynamic environmental data includes: real-time gas concentration data at each monitoring point—including ventilation parameters such as CH4 and CO concentrations, wind speed, wind direction, temperature, and air pressure; time-series data of mining machinery positions; and microseismic monitoring event sequences.

[0022] The inspection resource data includes: real-time location information of the gas inspector, skill level information of the gas inspector, historical assessment score information of the gas inspector, current working status information of the gas inspector, real-time location information of the inspection robot, remaining battery power information of the inspection robot, equipment status information and availability information of the inspection robot, all obtained in real time through the UWB positioning system.

[0023] Preferably, the preprocessing in step S2 includes: removing abnormal jump values ​​caused by sensor faults, interpolating missing values ​​using linear interpolation or KNN interpolation, and resampling data from different sampling frequencies to the same time granularity to achieve spatiotemporal alignment of multi-source data.

[0024] The feature extraction includes constructing the following enhanced features: the rate of change of gas concentration is a first-order difference, the gas concentration acceleration is a second-order difference, the spatiotemporal gradient of gas concentration is the concentration difference between adjacent monitoring points, the statistical features of gas concentration in the same historical period include mean, variance and peak value, the cross features of ventilation parameters and gas concentration, and the correlation features between mining progress and gas outburst.

[0025] Preferably, the improved LSTM gas concentration prediction model in step S3 specifically includes:

[0026] Input layer, used to receive past data. A multi-dimensional feature vector for each time step—the multi-dimensional feature vector includes time-series data of gas concentration, ventilation parameters, and mining parameters;

[0027] The first LSTM layer contains 64 to 128 hidden units, which are used to extract the short-term dependency features of the multidimensional feature vector in the time dimension.

[0028] The second LSTM layer contains 32 to 64 hidden units, which are used to extract the long-term dependency features of the multidimensional feature vector in the time dimension.

[0029] The attention layer is used to assign different attention weights to the output of the second LSTM at each time step to enhance the feature representation of key time points;

[0030] The fully connected output layer is used to output the predicted gas concentration values ​​of each inspection point in the next K time steps.

[0031] Preferably, the risk level classification method in step S3 is as follows:

[0032] For predicted concentration The inspection points were designated as Level I low-risk areas, and regular inspection frequencies were maintained.

[0033] For predicted concentration The inspection points were designated as Level II medium-risk areas, and the inspection frequency was increased to twice the normal frequency.

[0034] For predicted concentration The inspection points were designated as Level III high-risk areas, and the inspection frequency was increased to three times the normal frequency.

[0035] For predicted concentration The inspection point was designated as a Level IV danger zone, triggering an immediate response mechanism and simultaneously initiating continuous monitoring.

[0036] Wherein, C is the predicted percentage of gas concentration, and the threshold is dynamically adjusted according to the mine safety regulations.

[0037] Preferably, the dynamic personnel capability vector in step S4 includes the following dimensions:

[0038] Ability Level Dimension The value is determined based on the national vocational skills assessment level passed by the inspector, and is an integer ranging from 1 to 5.

[0039] Historical performance dimension It is calculated based on the weighted moving average of the assessment scores from the past N tasks, using the following formula:

[0040] ;

[0041] in, For the first The weighting coefficient of the sub-task For the first The assessment score for this task;

[0042] Execution efficiency dimension The formula is calculated based on the ratio of the average completion time to the standard completion time of historical tasks by gas inspectors:

[0043] ;

[0044] Risk management capability dimension The formula is calculated based on the ratio of the number of high-risk hazards successfully identified and handled by gas inspectors in their historical missions to the total number of high-risk missions:

[0045] ;

[0046] Integrity and Compliance Dimensions The ratio of the number of tasks in which gas inspectors did not commit violations such as false inspections, missed inspections, or data tampering to the total number of tasks is calculated using the following formula:

[0047] ;

[0048] The matching degree matrix in step S5 Calculated as follows: For the... The inspection task and the first Individual inspectors, matching degree The calculation formula is:

[0049] ;

[0050] in: For the first The capability feature vector of each task; For the first The capability vector of each tile inspector; This is a function for calculating vector similarity. For the first Location feature vectors of each task; For the first Historical service area preference vector of each inspector; For the first The current location of the inspector is the first Proximity score for each task location; For the first The current remaining workload of the tile inspector is similar to that of the first... A score indicating the match between the workload required for each task and the target workload.

[0051] Preferably, the multi-objective optimization function in step S6 is:

[0052] ;

[0053] Among them, all those with superscript All variables are derived from... The normalized dimensionless value is obtained using the following normalization formula:

[0054] ;

[0055] Mapped to respectively interval;

[0056] in, This is the normalized system response time; For nonlinear mapping functions of response time, This represents the normalized total length of the inspection path; To normalize the overall energy consumption of the system; To normalize the coverage of high-risk areas and adopt The form is introduced to reflect the diminishing marginal utility effect; This is a normalized prediction of the task execution success rate. To normalize the personnel matching degree in the task; For binary decision variables, taking the value 1 indicates the first... The inspection task was assigned to the first... One inspector, taking 0 indicates no allocation; For the first The performance evaluation incentives / penalties for each tile inspector are signed real values. to The weight coefficients for each objective term and satisfying ;

[0057] The aforementioned performance incentives / penalties Calculate as follows:

[0058] ;

[0059] in, For the first The real-time cumulative assessment score of each tile inspector for the current shift. For the first Historical baseline values ​​of the inspector's capabilities; and These are the incentive coefficient and the penalty coefficient, respectively.

[0060] Preferably, the constraints of the multi-objective optimization function in step S6 include:

[0061] Task skill matching constraints: For the first Each task, and the skill level of the assigned tile inspector. The skill level must be no lower than the minimum skill level required for the task. ;

[0062] Task-person matching lower limit constraint: For the first task... The degree of match between the assigned tile inspector and the task. It must be no less than the preset minimum matching threshold. ;

[0063] Assessment status constraint: If a tile inspector is currently in an assessment warning state, the system will not assign inspection tasks for risk areas of level III or above to him within the preset cooldown time.

[0064] Workload constraints: The continuous working time of any tile inspector in a single session shall not exceed the prescribed limit;

[0065] Time window constraint: Each inspection point must complete the inspection within the specified time window;

[0066] Equipment status constraints: Inspection robots must not be assigned new tasks when their battery level is below a threshold.

[0067] The multi-objective optimization function mentioned in step S6 is solved using an improved NSGA-II multi-objective genetic algorithm, or switched to an online decision-making algorithm based on multi-agent reinforcement learning after the system has stabilized.

[0068] Preferably, the formula for calculating the current task assessment score using the dynamic weighting algorithm in step S8 is:

[0069] ;

[0070] in, The scores for each assessment dimension; The dynamic weights for each assessment dimension are automatically adjusted based on the risk level and difficulty coefficient of the task.

[0071] For high-risk tasks in Level III and Level IV risk areas, the weight of the task execution authenticity dimension is automatically increased. to The weight of the work quality dimension will automatically increase. to ;

[0072] For urgent inspection tasks triggered by the system, the weights of task response and execution efficiency are automatically increased. to ;

[0073] For review tasks that are simultaneously assigned to both human and robot personnel, the weight of the data consistency dimension is automatically increased. to .

[0074] The assessment warning signal mentioned in step S9 is generated in the following manner: (Setting up a tile inspector) In the The real-time assessment score for this task is Its historical capability baseline value is The standard deviation of the historical ability baseline is ;

[0075] when At that time, a preliminary assessment warning signal is generated;

[0076] when At that time, an intermediate assessment warning signal will be generated, and no inspection tasks of level III or above risk areas will be assigned to the inspector within the preset 2-hour cooldown period;

[0077] when When the time comes, an advanced assessment warning signal will be generated, and no inspection tasks will be assigned to the inspector within a preset 24-hour cooldown period. A personnel replacement plan will also be automatically generated.

[0078] Preferably, the dispatch strategy in step S6 is automatically switched according to the global system state:

[0079] In the initial state, a performance-based assignment strategy is adopted, that is, tasks are assigned to the inspectors with the highest comprehensive scores in the personnel capability profile.

[0080] When a fixed periodic risk is detected in a certain area, i.e., the same area repeatedly shows a pattern of increased gas concentration within the same time period, the strategy switches to a specialty matching assignment strategy. This means that tasks in the area are preferentially assigned to the dimension with the highest risk handling capability in that area. The tile inspector who scores the tiles;

[0081] When a global task backlog is detected, meaning the number of tasks to be executed exceeds 1.5 times the sum of the number of on-duty inspectors and the number of robots, the "efficiency-first" dispatch strategy is switched to – that is, tasks are prioritized for allocation based on execution efficiency. The highest-rated tile inspectors and inspection robots;

[0082] The task allocation scheme described in step S7 is issued to the execution terminal in the following ways: For gas inspectors, task instructions are received via text and image through an intrinsically safe intelligent terminal for mining; for inspection robots, task instructions are received via data instructions through an underground wireless network; the task instructions include a list of inspection points, the optimal route, a specified time window, key inspection items, and risk warning information; during execution, the location is verified by comparing the real-time UWB location of the execution terminal with the planned route, and any deviation is automatically recorded; after arriving at the inspection point, the gas inspector confirms arrival by swiping an NFC card or scanning a QR code; gas concentration data is automatically collected and uploaded by the intelligent gas detector; when a hidden danger is discovered, it is reported immediately through a mobile APP, and a rectification task is automatically generated and pushed to the responsible person.

[0083] Compared with the prior art, the beneficial effects of the present invention are:

[0084] 1. This invention addresses the technical shortcomings of existing technologies that rely on static planning for task assignment and cannot adaptively adjust to real-time changes in underground gas concentration and dynamic evolution of environmental risks. By using multi-source sensor fusion perception and a deep learning prediction model, this invention achieves advanced prediction of gas concentration. It can dynamically generate and adjust inspection tasks based on the prediction results, enabling high-risk areas to obtain sufficient inspection resources in a timely manner. This fundamentally solves the technical problem of insufficient inspection frequency in high-risk areas under the static planning model, and significantly improves the timeliness and accuracy of gas risk prevention and control.

[0085] 2. This invention addresses the technical shortcomings of existing technologies where task assignment and assessment modules are independent and lack real-time data interaction and feedback loops. By using dynamic personnel competency profiling and multi-objective optimization linked to assessment, the invention ensures that the execution results of each task are instantly fed back to the personnel competency model through assessment scores. The competency model directly participates in the optimization decisions for subsequent task assignments, forming a tightly coupled closed-loop mechanism where assessment guides assignment and assignment drives assessment. This fundamentally solves the technical problems that assessment results cannot immediately influence assignment strategies and that assignment strategies cannot be dynamically adjusted based on personnel performance, thus achieving a technological upgrade in human resource management from static evaluation to dynamic optimization.

[0086] 3. This invention addresses the technical shortcomings of existing technologies, such as the lack of dynamic adjustment capabilities and real-time early warning mechanisms for assessment indicators, and the inability to intervene in advance against false or missed inspections during task execution. This invention uses a dynamic weighting algorithm to automatically adjust assessment weights based on the task's risk level, and performs real-time scoring and tiered early warnings during task execution. It shifts the assessment point from post-event to in-event, enabling violations to be detected and corrected promptly during the process. This significantly improves the timeliness and binding force of assessments, effectively ensures the authenticity and reliability of inspection data, and promotes the technological advancement of coal mine gas inspection management from static assessment to dynamic intelligent control. Attached Figure Description

[0087] Figure 1 This is a schematic diagram of the process of the present invention. Detailed Implementation

[0088] 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.

[0089] Please see Figure 1 This invention provides a technical solution for a method of assigning and assessing gas inspection tasks:

[0090] A method for assigning and assessing gas inspection tasks includes the following steps:

[0091] S1: Multi-parameter gas sensors deployed at various inspection points underground, UWB positioning base stations deployed underground, intrinsically safe intelligent terminals for gas inspectors, and explosion-proof inspection robots deployed in roadways are used to collect multi-source heterogeneous data in the coal mine in real time. The collected data is then transmitted to the ground data center in real time via the mine industrial Ethernet ring network or 5G private network. The multi-source heterogeneous data in step S1 includes: static basic data, dynamic environmental data, and inspection resource data.

[0092] The static basic data includes: geological structure distribution data of coal mine shafts, roadway topology data—including spatial coordinates, connectivity, roadway length and slope information, and spatial layout data of fixed sensors and inspection points.

[0093] The dynamic environmental data includes: real-time gas concentration data at each monitoring point—including ventilation parameters such as CH4 and CO concentrations, wind speed, wind direction, temperature, and air pressure; time-series data of mining machinery positions; and microseismic monitoring event sequences.

[0094] The inspection resource data includes: real-time location information of the gas inspector, skill level information of the gas inspector, historical assessment score information of the gas inspector, current working status information of the gas inspector, real-time location information of the inspection robot, remaining battery power information of the inspection robot, equipment status information and availability information of the inspection robot, all obtained in real time through the UWB positioning system.

[0095] S2: Preprocessing, feature extraction, and spatiotemporal fusion of the multi-source heterogeneous data; the preprocessing in step S2 includes: removing abnormal jump values ​​caused by sensor faults, interpolating missing values ​​using linear interpolation or KNN interpolation, and resampling data with different sampling frequencies to the same time granularity to achieve spatiotemporal alignment of multi-source data.

[0096] The feature extraction includes constructing the following enhanced features: the rate of change of gas concentration is a first-order difference, the gas concentration acceleration is a second-order difference, the spatiotemporal gradient of gas concentration is the concentration difference between adjacent monitoring points, the statistical features of gas concentration in the same historical period include mean, variance and peak value, the cross features of ventilation parameters and gas concentration, and the correlation features between mining progress and gas outburst.

[0097] S3: Input the fused feature data output from step S2 into the improved LSTM gas concentration prediction model to predict the gas concentration distribution and risk level of each inspection area within multiple future time steps; the improved LSTM gas concentration prediction model adopts a dual-layer LSTM architecture combined with an attention mechanism.

[0098] The improved LSTM gas concentration prediction model in step S3 specifically includes:

[0099] Input layer, used to receive past data. A multi-dimensional feature vector for each time step—the multi-dimensional feature vector includes time-series data of gas concentration, ventilation parameters, and mining parameters;

[0100] The first LSTM layer contains 64 to 128 hidden units, which are used to extract the short-term dependency features of the multidimensional feature vector in the time dimension.

[0101] The second LSTM layer contains 32 to 64 hidden units, which are used to extract the long-term dependency features of the multidimensional feature vector in the time dimension.

[0102] The attention layer is used to assign different attention weights to the output of the second LSTM at each time step to enhance the feature representation of key time points;

[0103] A fully connected output layer is used to output the predicted gas concentration values ​​of each inspection point in the next K time steps;

[0104] The risk level classification method in step S3 is as follows:

[0105] For predicted concentration The inspection points were designated as Level I low-risk areas, and regular inspection frequencies were maintained.

[0106] For predicted concentration The inspection points were designated as Level II medium-risk areas, and the inspection frequency was increased to twice the normal frequency.

[0107] For predicted concentration The inspection points were designated as Level III high-risk areas, and the inspection frequency was increased to three times the normal frequency.

[0108] For predicted concentration The inspection point was designated as a Level IV danger zone, triggering an immediate response mechanism and simultaneously initiating continuous monitoring.

[0109] Wherein, C is the predicted percentage of gas concentration, and the threshold is dynamically adjusted according to the mine safety regulations.

[0110] S4: Based on the inspectors' historical assessment records, skill level certifications, execution efficiency statistics, and violation records, construct a dynamic personnel capability vector containing multiple dimensions, and update the values ​​of each dimension of the capability vector in real time;

[0111] The dynamic personnel capability vector in step S4 includes the following dimensions:

[0112] Ability Level Dimension The value is determined based on the national vocational skills assessment level passed by the inspector, and is an integer ranging from 1 to 5.

[0113] Historical performance dimension It is calculated based on the weighted moving average of the assessment scores from the past N tasks, using the following formula:

[0114] ;

[0115] in, For the first The weighting coefficient of the sub-task For the first The assessment score for this task;

[0116] Execution efficiency dimension The formula is calculated based on the ratio of the average completion time to the standard completion time of historical tasks by gas inspectors:

[0117] ;

[0118] Risk management capability dimension The formula is calculated based on the ratio of the number of high-risk hazards successfully identified and handled by gas inspectors in their historical missions to the total number of high-risk missions:

[0119] ;

[0120] Integrity and Compliance Dimensions The ratio of the number of tasks in which gas inspectors did not commit violations such as false inspections, missed inspections, or data tampering to the total number of tasks is calculated using the following formula:

[0121] .

[0122] S5: Based on the requirement feature vector of each inspection task output in step S3 and the ability vector of each gas inspector output in step S4, calculate the matching degree matrix between each gas inspector and each inspection task.

[0123] The matching degree matrix in step S5 Calculated as follows: For the... The inspection task and the first Individual inspectors, matching degree The calculation formula is:

[0124] ;

[0125] in: For the first The capability feature vector of each task; For the first The capability vector of each tile inspector; This is a function for calculating vector similarity. For the first Location feature vectors of each task; For the first Historical service area preference vector of each inspector; For the first The current location of the inspector is the first Proximity score for each task location; For the first The current remaining workload of the tile inspector is similar to that of the first... A score indicating the match between the workload required for each task and the target workload.

[0126] S6: Based on the matching degree matrix output in step S5, the risk prediction results output in step S3, and the real-time UWB positioning and working status of each inspector, construct and solve a multi-objective optimization function that includes assessment feedback items to generate the optimal task allocation scheme and inspection path planning scheme.

[0127] The multi-objective optimization function in step S6 is:

[0128] ;

[0129] Among them, all those with superscript All variables are derived from... The normalized dimensionless value is obtained using the following normalization formula:

[0130] ;

[0131] Mapped to respectively interval;

[0132] in, This is the normalized system response time; For nonlinear mapping functions of response time, This represents the normalized total length of the inspection path; To normalize the overall energy consumption of the system; To normalize the coverage of high-risk areas and adopt The form is introduced to reflect the diminishing marginal utility effect; This is a normalized prediction of the task execution success rate. To normalize the personnel matching degree in the task; For binary decision variables, taking the value 1 indicates the first... The inspection task was assigned to the first... One inspector, taking 0 indicates no allocation; For the first The performance evaluation incentives / penalties for each tile inspector are signed real values. to The weight coefficients for each objective term and satisfying ;

[0133] The aforementioned performance incentives / penalties Calculate as follows:

[0134] ;

[0135] in, For the first The real-time cumulative assessment score of each tile inspector for the current shift. For the first Historical baseline values ​​of the inspector's capabilities; and These are the incentive coefficient and the penalty coefficient, respectively.

[0136] The constraints of the multi-objective optimization function in step S6 include:

[0137] Task skill matching constraints: For the first Each task, and the skill level of the assigned tile inspector. The skill level must be no lower than the minimum skill level required for the task. ;

[0138] Task-person matching lower limit constraint: For the first task... The degree of match between the assigned tile inspector and the task. It must be no less than the preset minimum matching threshold. ;

[0139] Assessment status constraint: If a tile inspector is currently in an assessment warning state, the system will not assign inspection tasks for risk areas of level III or above to him within the preset cooldown time.

[0140] Workload constraints: The continuous working time of any tile inspector in a single session shall not exceed the prescribed limit;

[0141] Time window constraint: Each inspection point must complete the inspection within the specified time window;

[0142] Equipment status constraints: Inspection robots must not be assigned new tasks when their battery level is below a threshold.

[0143] The multi-objective optimization function mentioned in step S6 is solved using an improved NSGA-II multi-objective genetic algorithm, or switched to an online decision-making algorithm based on multi-agent reinforcement learning after the system has stabilized.

[0144] The distribution strategy in step S6 is automatically switched according to the global system state:

[0145] In the initial state, a performance-based assignment strategy is adopted, that is, tasks are assigned to the inspectors with the highest comprehensive scores in the personnel capability profile.

[0146] When a fixed periodic risk is detected in a certain area, i.e., the same area repeatedly shows a pattern of increased gas concentration within the same time period, the strategy switches to a specialty matching assignment strategy. This means that tasks in the area are preferentially assigned to the dimension with the highest risk handling capability in that area. The tile inspector who scores the tiles;

[0147] When a global task backlog is detected, meaning the number of tasks to be executed exceeds 1.5 times the sum of the number of on-duty inspectors and the number of robots, the "efficiency-first" dispatch strategy is switched to – that is, tasks are prioritized for allocation based on execution efficiency. The highest-rated tile inspectors and inspection robots.

[0148] S7: The task allocation scheme generated in step S6 is sent to the corresponding intrinsically safe intelligent terminal or inspection robot in the mine through the underground 5G private network, and the UWB positioning data, time data and detection data during the execution process are collected in real time.

[0149] The task allocation scheme described in step S7 is issued to the execution terminal in the following ways: For gas inspectors, task instructions are received via text and image through an intrinsically safe intelligent terminal for mining; for inspection robots, task instructions are received via data instructions through an underground wireless network; the task instructions include a list of inspection points, the optimal route, a specified time window, key inspection items, and risk warning information; during execution, the location is verified by comparing the real-time UWB location of the execution terminal with the planned route, and any deviation is automatically recorded; after arriving at the inspection point, the gas inspector confirms arrival by swiping an NFC card or scanning a QR code; gas concentration data is automatically collected and uploaded by the intelligent gas detector; when a hidden danger is discovered, it is reported immediately through a mobile APP, and a rectification task is automatically generated and pushed to the responsible person.

[0150] S8: Based on the execution data collected in step S7, combined with the risk level and difficulty coefficient of the task, a dynamic weighting algorithm is used to calculate the current task assessment score of each inspector in real time, and update the corresponding dimension data in the personnel capability vector mentioned in step S4.

[0151] The formula for calculating the assessment score of the current task using the dynamic weighting algorithm described in step S8 is as follows:

[0152] ;

[0153] in, The scores for each assessment dimension; The dynamic weights for each assessment dimension are automatically adjusted based on the risk level and difficulty coefficient of the task.

[0154] For high-risk tasks in Level III and Level IV risk areas, the weight of the task execution authenticity dimension is automatically increased. to The weight of the work quality dimension will automatically increase. to ;

[0155] For urgent inspection tasks triggered by the system, the weights of task response and execution efficiency are automatically increased. to ;

[0156] For review tasks that are simultaneously assigned to both human and robot personnel, the weight of the data consistency dimension is automatically increased. to .

[0157] S9: During the task execution process, the real-time scoring results of step S8 are compared with the historical ability baseline of the inspector. When the real-time score deviates from the historical baseline by more than a preset threshold, an assessment warning signal is generated and fed back to step S6, triggering the adjustment of the subsequent task assignment strategy.

[0158] The assessment warning signal mentioned in step S9 is generated in the following manner: (Setting up a tile inspector) In the The real-time assessment score for this task is Its historical capability baseline value is The standard deviation of the historical ability baseline is ;

[0159] when At that time, a preliminary assessment warning signal is generated;

[0160] when At that time, an intermediate assessment warning signal will be generated, and no inspection tasks of level III or above risk areas will be assigned to the inspector within the preset 2-hour cooldown period;

[0161] when When the time comes, an advanced assessment warning signal will be generated, and no inspection tasks will be assigned to the inspector within a preset 24-hour cooldown period. A personnel replacement plan will also be automatically generated.

[0162] As one embodiment of the present invention, multi-parameter gas sensors for mining are deployed at 312 inspection points underground in a coal mine, with a detection range of... for , for Measurement accuracy for Data is collected at a frequency of 1 minute per session for real-time acquisition. Concentration and Concentration data;

[0163] Deployed in the main underground roadways, intersections, and around the mining faces. The system has 42 positioning base stations and adopts a TDOA+TOF hybrid ranging algorithm. The unobstructed wireless transmission distance between the base station and the positioning card is no less than 400 meters. The positioning update frequency is once per second, enabling centimeter-level real-time positioning of personnel and inspection robots in the well. After on-site calibration and testing, the static positioning error under unobstructed conditions is no more than 0.30 meters, and the dynamic positioning error is no more than 0.50 meters.

[0164] Equipped with 45 intrinsically safe intelligent terminals for mining, running an Android operating system and a dedicated inspection APP, supporting NFC reading, QR code scanning, and 4G / 5G communication, these terminals are used by gas inspectors to receive inspection tasks, clock in at designated times, input detection data, and report potential hazards. Each gas inspector and each inspection robot wears / installs one UWB positioning tag.

[0165] Four explosion-proof rail-mounted inspection robots were deployed, each equipped with a multi-parameter gas detection module. , , Equipped with temperature and humidity control, a starlight-level high-definition visible light camera, an infrared thermal imager, a UWB positioning module, and an underground WiFi 6 communication module, the robot runs along an I-beam track with an adjustable speed of 0-1.5 m / s, a climbing ability of ≥15°, a single charge endurance of ≥8 hours, and supports automatic charging.

[0166] Two high-performance computing servers are deployed on the ground for training and inference operations of the LSTM prediction model; three database servers are deployed for storing multi-source historical data and real-time data; and two web application servers are deployed for the intelligent dispatch and assessment platform for inspections.

[0167] All hardware devices are interconnected via the mine's 5G private network, and data is transmitted to the ground data center in real time.

[0168] As one embodiment of the present invention, the system collects three types of data in real time through a data perception layer:

[0169] The first category is static basic data, including pre-entered geological structure distribution data of coal mines, roadway topology data containing spatial coordinates and connectivity relationships, and spatial layout data of 312 inspection points.

[0170] The second category is dynamic environmental data, including data collected every minute from each monitoring point. Concentration and Concentration data, ventilation parameters collected every minute by anemometers, positional and temporal data of mining machinery, and event sequence data from microseismic monitoring instruments;

[0171] The third category is inspection resource data, including the real-time location coordinates of the inspectors and inspection robots obtained every second through the UWB positioning system, the pre-entered skill level and historical assessment scores of the inspectors, and the real-time battery level and equipment status of the inspection robots.

[0172] The data analysis layer performs the following preprocessing on the collected multi-source heterogeneous data:

[0173] First, abnormal jump values ​​caused by sensor malfunctions are eliminated—the abnormal jump values ​​are defined as data points that exceed three standard deviations of the historical normal range;

[0174] Secondly, linear interpolation is used to impute the missing values: for the first... missing data points The interpolation is performed using the formula shown below:

[0175] ;

[0176] in, Represents a timestamp.

[0177] Then, data from different sampling frequencies are uniformly resampled to a time granularity of 1 minute to achieve spatiotemporal alignment of multi-source data.

[0178] Finally, enhanced features are constructed: the 1-minute rate of change (first-order difference), 5-minute acceleration (second-order difference), spatiotemporal gradient of gas concentration at adjacent monitoring points, and the statistical mean and variance of gas concentration at the same time in the past 24 hours are calculated.

[0179] As an embodiment of the present invention, the improved LSTM gas concentration prediction model is constructed and trained in the following manner:

[0180] Model building:

[0181] The model employs a two-layer LSTM architecture combined with an attention mechanism;

[0182] The input layer receives the past 60 time steps (multi-dimensional feature vectors of the past 60 minutes). The feature vector of each time step has a dimension of 18 and includes the gas concentration value at the current moment, the gas concentration change rate over 1 minute, the gas concentration acceleration over 5 minutes, the spatiotemporal gradient values ​​of gas concentration at 3 adjacent monitoring points, wind speed, wind direction angle, temperature, air pressure, the pose parameters of the mining machinery (x, y, z coordinates and direction of travel), the frequency of microseismic events, and the mean and variance of the gas concentration at the same time in the previous shift.

[0183] The first LSTM layer contains 128 hidden units, which receive the 18-dimensional feature vector as input and output a 128-dimensional short-term dependency feature sequence.

[0184] The second LSTM layer contains 64 hidden units, which receive the 128-dimensional short-term dependency feature sequence as input and output a 64-dimensional long-term dependency feature sequence.

[0185] The attention layer calculates attention weights for the 64-dimensional feature vectors at 60 time steps of the second LSTM output: for the feature vector at time step t... Its attention weight is Normalized weights are The weighted summation yields the context vector as follows: ,in, and This represents the trainable parameters.

[0186] The fully connected output layer receives the context vector. As input, the output is the predicted gas concentration value of each inspection point for the next 30 minutes—that is, the output layer contains 312 neurons, corresponding to the predicted values ​​of 312 inspection points.

[0187] Model training:

[0188] Historical monitoring data from the coal mine over the past 18 months was collected as the training dataset, totaling approximately 946,080 records—divided into training, validation, and test sets in a 7:2:1 ratio.

[0189] Mean square error Add an L2 regularization term as the loss function, where, Represents the actual value. Indicates the predicted value. Indicates the number of samples.

[0190] The differential evolution algorithm was used to automatically optimize the hyperparameters of the LSTM: the search range of the hyperparameters was set as follows: learning rate between 0.0001 and 0.01, number of hidden units in the first LSTM layer between 64 and 128, number of hidden units in the second LSTM layer between 32 and 64, and dropout rate between 0.1 and 0.5; the population size of the differential evolution algorithm was set to 50, and the maximum number of iterations was set to 200; the optimal hyperparameters determined by the optimization were: learning rate 0.0012, number of hidden units in the first LSTM layer 128, number of hidden units in the second LSTM layer 64, dropout rate 0.25, and L2 regularization coefficient 0.0005.

[0191] With root mean square error Mean absolute error and coefficient of determination As a performance evaluation metric for the model. After training, the model's performance on the test set... achieve , achieve , It reached 0.975.

[0192] Online update mechanism:

[0193] The model employs a rolling window online learning strategy, with the window size set to 7 days—that is, data from the most recent 7 days. Whenever new measured data arrives, the system automatically adds the latest data to the training window and removes the oldest data from 7 days ago. The model is incrementally trained every 24 hours, achieving daily online updates.

[0194] As an embodiment of the present invention, the intelligent prediction layer calls the trained LSTM model every 5 minutes, inputs the multidimensional feature data of the past 60 minutes, and outputs the predicted gas concentration values ​​of each inspection point for the next 30 minutes.

[0195] Based on the prediction results, the system classifies the risk level of each inspection area according to the following rules:

[0196] For predicted concentration The inspection points are classified as Level I low-risk areas, and the regular inspection frequency of every 2 hours is maintained.

[0197] For predicted concentration The inspection points were designated as Level II medium-risk areas, and the inspection frequency was increased to once every hour.

[0198] For predicted concentration The inspection points were designated as Level III high-risk areas, and the inspection frequency was increased to once every 40 minutes.

[0199] For predicted concentration The inspection point was designated as a Level IV danger zone, triggering an immediate response mechanism. The system automatically generated an urgent inspection task and dispatched it immediately, while simultaneously initiating continuous monitoring.

[0200] In a practical operation of this embodiment, the system predicted that the methane concentration at three inspection points near the mining face would rise to a certain level within the next 30 minutes. to The system automatically marked the three inspection points as Level III high-risk areas and generated corresponding encrypted inspection tasks. On-site measurements showed that the actual gas concentration in the area reached [a certain level] after 30 minutes. to Prediction error is in Within.

[0201] As an embodiment of the present invention, taking the 38 gas inspectors on duty at the coal mine as the object, the personnel capability profiling module constructs a dynamic personnel capability vector for each gas inspector in the following manner:

[0202] Taking gas inspector Zhang (senior technician, 18 years of service) as an example, the initial values ​​of each dimension of his capability vector are set as follows:

[0203] Ability Level Dimension This corresponds to the highest level in the national vocational skills assessment system.

[0204] Historical performance dimension The initial value is set at 87 points, based on the weighted moving average of the monthly performance scores over the past 6 months, with a decay rate parameter. ;

[0205] Execution efficiency dimension The initial value was set to 1.12, based on the fact that its historical average task completion time was 12% faster than the standard time;

[0206] Risk management capability dimension The initial value was set at 0.93, based on a 93% success rate in handling high-risk hazards in its historical missions.

[0207] Integrity and Compliance Dimensions The initial value was set to 0.99, based on the proportion of violations in its historical tasks. .

[0208] After inspector Zhang completes an inspection task, the dynamic assessment and scoring module calculates a score of 92 points for this task, and the personnel capability profile module immediately updates his historical performance dimensions.

[0209] ;

[0210] That is, the current rating has a weight of approximately 63.2%, while the historical rating has a weight of approximately 36.8%, achieving smooth rating updates while highlighting recent performance.

[0211] At the same time, the risk management capability will be updated accordingly based on whether potential hazards were successfully addressed and whether any violations occurred during this mission. and integrity and compliance dimensions .

[0212] In one embodiment of the present invention, the system generates 10 inspection tasks, of which 4 are Level III high-risk area tasks, 3 are Level II medium-risk area tasks, and 3 are Level I low-risk area tasks, with one of the Level III high-risk area tasks being selected. For example, the task-person matching degree calculation module calculates its matching degree with each gas inspector in the following way:

[0213] task demand capability feature vector is: (required skill level=4, risk tolerance=0.9, efficiency requirement=0.8);

[0214] the capability vector of gas inspector Zhang is: (capability level , historical performance , execution efficiency , risk handling capability , integrity and compliance );

[0215] calculate and the cosine similarity:

[0216] ;

[0217] the task has position coordinates of (125, 84, -380) — representing underground coordinates, the real-time position of gas inspector Zhang is (118, 79, -380), and the proximity score is obtained by normalizing after calculating the three-dimensional Euclidean distance .

[0218] The current remaining workload of gas inspector Zhang is that 2 inspection points have been completed, and the remaining available working time is 3.5 hours, the workload required for the task is 1.5 hours, and the workload matching degree score .

[0219] set weight coefficients , , and calculate the matching degree:

[0220] ;

[0221] Similarly, the matching degrees of the other 37 gas inspectors with the task are calculated. The system takes gas inspector Zhang with the highest matching degree as the preferred candidate dispatch object for the task .

[0222] As an embodiment of the present invention, the assessment linkage dispatch module constructs and solves the multi-objective optimization function in the following manner:

[0223] The system currently has 10 inspection tasks to be dispatched, of which 4 are Class III high-risk tasks, 3 are Class II medium-risk tasks, and 3 are Class I low-risk tasks. Among the 38 gas inspectors, 22 are currently on duty, and all 4 inspection robots are available

[0224] Normalization process:

[0225] Each original index is mapped to after min-max normalization. The range of values ​​for each indicator is determined by statistical analysis of historical data from the 30 days prior to system launch: System response time minutes, total path length kilometers, system energy consumption High-risk area coverage Task execution success rate prediction .

[0226] With system response time For example, the system response time of a certain solution is 12 minutes, after normalization Similarly, the total path length, system energy consumption, high-risk area coverage, task execution success rate, and task-person matching degree are all normalized in the same way.

[0227] The weight coefficients for each objective are set as follows: For system response time, For the total path length, For system energy consumption, To increase coverage in high-risk areas, For predicting the success rate of task execution, The total score for task-person matching. This is for performance evaluation incentives / penalties.

[0228] Set constraints: Task to For Level III high-risk missions, the minimum skill level requirement is... (For technicians and above only); Minimum matching threshold (Inspectors with a matching score below 0.55 should not be assigned this task); the continuous working time of an inspector should not exceed 4 hours at a time; the inspection time window for Level III and above risk areas is ±5 minutes; the robot's battery level should be below... New tasks may not be assigned at this time.

[0229] Calculation of performance incentives / penalties:

[0230] Inspector Zhang has completed 3 tasks in his current shift, with scores of 91, 94, and 92 respectively. His historical baseline score is 87, and his cumulative score is higher than the baseline.

[0231] ;

[0232] Inspector Li completed two tasks, scoring 76 and 73 respectively, with a baseline score of 78.

[0233] ;

[0234] Solution process:

[0235] The assessment-linked distribution module is executed according to the following steps:

[0236] The first step is to screen candidates based on the constraints: for the four Level III high-risk tasks, only 12 of the 22 gas inspectors on duty meet the skill level requirements (skill level ≥ 4).

[0237] The second step is to read the assessment warning status: There are currently no inspectors in the warning status in the system, and all inspectors are qualified to distribute tiles.

[0238] The third step is to use the improved NSGA-II multi-objective genetic algorithm to solve for the optimal allocation scheme: the population size is set to 150, the number of iterations is set to 500, the crossover probability is 0.85, and the mutation probability is 0.12.

[0239] After calculation, the optimal dispatch scheme output by the system is: Level III high-risk task. and The tiles were assigned to inspector Zhang, with matching scores of 0.85 and 0.87 respectively. Assigned to inspector Zhao – matching degree 0.83. Gas inspector Qian was assigned a matching score of 0.79; Level II medium-risk tasks were assigned to gas inspectors with moderate matching scores; Level I low-risk tasks were assigned to gas inspectors with low matching scores or those in training period; and four inspection robots were assigned to routine inspection tasks in areas such as the main roadway and the central substation.

[0240] As one embodiment of the present invention, inspector Zhang was assigned a Level III high-risk task. and Then execution begins. The dynamic assessment and scoring module collects data in real time and performs dynamic scoring during the task execution by inspector Zhang.

[0241] Inspector Zhang received the task at 8:00 and arrived at the first inspection point at 8:04—the response time was 4 minutes, while the standard response time is 8 minutes, and this item scores 98 points.

[0242] Inspector Zhang completed the arrival confirmation by NFC check-in. The system recorded the arrival timestamp as 8:04, while the planned arrival time was 8:03—a deviation of 1 minute, which is within the specified ±5 minute range. This item scores 95 points.

[0243] Gas inspector Zhang used a smart gas detector to automatically collect the gas concentration. , and the task Fixed sensor synchronous data The deviation is The data accuracy rate is This item scores 98 points;

[0244] Inspector Zhang discovered a ventilation facility that was not properly sealed during an inspection and reported it immediately. He will receive an extra 10 points for this.

[0245] The dynamic assessment and scoring module is based on the task. The Level III high-risk rating automatically adjusts the weight of the task execution authenticity dimension from... Rise to The weight of the work quality dimension from Rise to Other dimensions are adjusted accordingly, and the score for this task is calculated as follows:

[0246] ;

[0247] The assessment and early warning module compares inspector Zhang's current real-time score of 96.45 with his historical baseline of 87 and standard deviation. Comparison:

[0248] ;

[0249] The current performance is excellent and does not trigger any warnings.

[0250] Meanwhile, inspector Wang was assigned a Level II medium-risk task. During the execution, the following situation occurred: Inspector Wang failed to arrive at the inspection point 18 minutes after the scheduled arrival time. The system sent two reminders but received no response. UWB location verification showed that his location deviated from the planned route by more than 25 meters for 10 minutes. The dynamic assessment and scoring module gave him a real-time score of 48 points in this task, including 0 points for the attendance rate, 25 points for the route matching, and no points deducted for the data collection stage.

[0251] The assessment and early warning module compares inspector Wang's real-time score of 48 with his historical baseline score of 65 and the standard deviation. Comparison:

[0252] ;

[0253] The intermediate assessment warning signal was triggered, and inspector Wang was marked as being in a "risk state".

[0254] As an embodiment of the present invention, when inspector Wang triggers the intermediate assessment warning signal, the assessment linkage distribution module immediately performs the following dynamic adjustments:

[0255] First, activate the assessment status constraint: During the preset 2-hour cooldown period, the system will not assign any inspection tasks for Level III or higher risk areas to inspector Wang. The system's current task pool for assignment will contain one Level III high-risk task. Originally planned to be assigned to tile inspector Wang, the system automatically switched the candidate for assignment to tile inspector Sun, who had the second highest matching degree, because Wang was the closest to the area.

[0256] Second, a warning message was sent to the smart terminal of inspector Wang.

[0257] Third, send intervention prompts to the on-duty dispatcher;

[0258] Fourth, the system automatically reassigns the two Level III high-risk tasks following Wang's current shift to the more suitable Wang and Zhao and Qian, while keeping Wang's remaining three Level II medium-risk tasks unchanged to give him an opportunity to correct his mistakes.

[0259] If inspector Wang's assessment score rises above the baseline level in the next three consecutive tasks, the system will automatically remove his "risk status" marker and restore his normal distribution qualification.

[0260] As an embodiment of the present invention, after the system has been running for a full quarter, the personnel capability profiling module has accumulated the complete dynamic capability evolution trajectory of each gas inspector:

[0261] Inspector Zhang's overall score rose from an initial 87 points to 94 points, mainly due to his historical performance. The score rose from 87 to 93 points, reflecting the risk management capability dimension. From 0.93 to 0.97, as the competency profile score improved, the proportion of Level III high-risk tasks obtained by inspector Zhang in subsequent months increased from [previous figure]. Rise to This creates a positive cycle of "good performance → obtaining excellent tasks → improved capabilities → obtaining even better tasks";

[0262] Inspector Zhou's overall score rose from an initial 59 points to 76 points—mainly due to his performance in the integrity and compliance dimension. The execution efficiency dimension rose from 0.68 to 0.96 (after six consecutive months without any violations). The percentage of high-risk (Level III) tasks obtained by inspector Zhou increased from 0.78 to 0.95, reflecting the improvement in his competency profile. Rise to This indicates that the system's distribution strategy successfully guided the positive behavioral shift of low-scoring individuals.

[0263] The system performs a quarterly comparative analysis of the actual detection data from each gas inspector and the predicted data from the LSTM prediction model. It was found that the data from the area under the responsibility of gas inspector Zhang showed the highest consistency with the model predictions (average deviation rate was...). Inspector Zhou's area had the lowest consistency (average deviation rate was ). The system automatically determined that inspector Zhou might have had issues with data collection, and therefore reduced the number of high-risk tasks in his assigned area in the next month's distribution strategy. At the same time, Zhou's training assignment was increased to involve shadowing five-star gas inspector Zhang. After two months of shadowing training, gas inspector Zhou's data consistency deviation rate decreased from... Descending to .

[0264] Through the above-mentioned close linkage mechanism between assessment and distribution, the system has achieved continuous evolution of personnel capabilities and continuous optimization of distribution strategies during continuous operation, and has fully constructed a deep closed-loop coupling system of "distribution, execution, assessment, feedback, distribution adjustment, re-execution and re-assessment".

[0265] As an embodiment of the present invention, the system was officially launched on March 1, 2025. The following data comes from the actual operation records of 12 consecutive months from March 2025 to February 2026.

[0266] The task dispatch effect data is shown in Table 1:

[0267] Table 1

[0268] index Before the system went live After the system goes live change Number of inspectors in Gangwa 38 people 28 people Reduced by 26.3% Daily average number of inspection tasks completed 276 312 Increased by 13.0% Average response time for missions in high-risk areas 23.5 minutes 8.7 minutes Shortened by 63.0% Average completion time of inspection tasks 6.8 hours 4.1 hours Shortened by 39.7% Task-person matching 0.51 0.78 An increase of 52.9%

[0269] As shown in Table 1, inspection tasks in high-risk areas (Level III and IV) were 100% performed by personnel with skill level 4 or above, compared to only 67% before the system went live. Regarding human-machine collaboration, four inspection robots handled 62% of the routine inspection tasks in main roadways and central substations.

[0270] Table 2 shows the comparison data of prediction model accuracy:

[0271] Table 2

[0272] Evaluation indicators BP neural network LSTM This invention model Increase RMSE 0.067% 0.045% 0.028% 37.8% MAE 0.051% 0.032% 0.019% 40.6% 0.892 0.938 0.975 3.9%

[0273] As shown in Table 2, on a 30-minute prediction scale, the model achieved a 94.2% recall rate and a 6.8% false alarm rate for high-risk events with gas concentrations exceeding 0.5%. In actual operation, the system provided early warnings for 47 abnormal gas events, of which 43 were verified on-site, resulting in an effective early warning accuracy rate of 91.5%.

[0274] Assessment and early warning data

[0275] During the 12 months of system operation:

[0276] Automatic assessment and scoring: A total of 102,648 single-task assessments and scores have been completed automatically with zero human intervention;

[0277] Assessment warnings were triggered in a total of 186 times (involving 21 people), 37 times (involving 9 people), and 3 times (involving 2 people).

[0278] Improvement rate after warning: Among the 9 gas inspectors who triggered the intermediate warning, 8 of them saw their task assessment scores rise above the baseline for three consecutive times after the cooling-off period, with an improvement rate of 88.9%;

[0279] False detection and missed detection incidents: 14 incidents were recorded in the 12 months before the system went live; 1 incident was recorded in the 12 months after the system went live, a reduction of 92.9%.

[0280] Comparison of the effects of multi-objective optimization algorithms

[0281] The improved NSGA-II multi-objective optimization algorithm used in this invention is compared with the single-objective greedy algorithm and the traditional weighted summation method. The comparison results are shown in Table 3.

[0282] Table 3

[0283] Optimization Objective Single-objective greedy algorithm Traditional weighted summation method This invention System response time (minutes) 15.2 12.8 8.7 Total length of inspection route (km) 8.6 7.1 5.8 High-risk area coverage rate (%) 78.3 86.5 97.2 Task-person matching 0.62 0.71 0.78

[0284] As shown in Table 3, the method of the present invention achieves optimal or near-optimal results for each optimization objective, especially in the coverage rate of high-risk areas, which reaches 97.2%, an improvement of 10.7 percentage points compared with the traditional weighted summation method.

[0285] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A gas inspection task assignment and evaluation method, characterized in that, Includes the following steps: S1: Through multi-parameter gas sensors deployed at various inspection points underground, UWB positioning base stations deployed underground, intrinsically safe intelligent terminals for gas inspectors, and explosion-proof inspection robots deployed in roadways, multi-source heterogeneous data from underground coal mines are collected in real time, and the collected data is transmitted to the ground data center in real time through the mine industrial Ethernet ring network or 5G private network. S2: Preprocess, extract features, and perform spatiotemporal fusion of the multi-source heterogeneous data; S3: Input the fused feature data output from step S2 into the improved LSTM gas concentration prediction model to predict the gas concentration distribution and risk level of each inspection area within multiple future time steps; the improved LSTM gas concentration prediction model adopts a dual-layer LSTM architecture combined with an attention mechanism. S4: Based on the inspectors' historical assessment records, skill level certifications, execution efficiency statistics, and violation records, construct a dynamic personnel capability vector containing multiple dimensions, and update the values ​​of each dimension of the capability vector in real time; S5: Based on the requirement feature vector of each inspection task output in step S3 and the ability vector of each gas inspector output in step S4, calculate the matching degree matrix between each gas inspector and each inspection task. S6: Based on the matching degree matrix output in step S5, the risk prediction results output in step S3, and the real-time UWB positioning and working status of each inspector, construct and solve a multi-objective optimization function that includes assessment feedback items to generate the optimal task allocation scheme and inspection path planning scheme. S7: The task allocation scheme generated in step S6 is sent to the corresponding intrinsically safe intelligent terminal or inspection robot in the mine through the underground 5G private network, and the UWB positioning data, time data and detection data during the execution process are collected in real time. S8: Based on the execution data collected in step S7, combined with the risk level and difficulty coefficient of the task, a dynamic weighting algorithm is used to calculate the current task assessment score of each inspector in real time, and update the corresponding dimension data in the personnel capability vector mentioned in step S4. S9: During task execution, the real-time scoring results from step S8 are compared with the historical baseline of the inspector's capabilities. When the real-time score deviates from the historical baseline by more than a preset threshold, an assessment warning signal is generated and fed back to step S6, triggering an adjustment to the subsequent task assignment strategy.

2. The method for assigning and assessing gas inspection tasks according to claim 1, characterized in that: The multi-source heterogeneous data in step S1 includes: static basic data, dynamic environmental data, and inspection resource data; The static basic data includes: geological structure distribution data of coal mine shafts, roadway topology data—including spatial coordinates, connectivity, roadway length and slope information, and spatial layout data of fixed sensors and inspection points. The dynamic environmental data includes: real-time gas concentration data at each monitoring point—including ventilation parameters such as CH4 and CO concentrations, wind speed, wind direction, temperature, and air pressure; time-series data of mining machinery positions; and microseismic monitoring event sequences. The inspection resource data includes: real-time location information of the gas inspector, skill level information of the gas inspector, historical assessment score information of the gas inspector, current working status information of the gas inspector, real-time location information of the inspection robot, remaining battery power information of the inspection robot, equipment status information and availability information of the inspection robot, all obtained in real time through the UWB positioning system.

3. The method for assigning and assessing gas inspection tasks according to claim 1, characterized in that: The preprocessing in step S2 includes: removing abnormal jump values ​​caused by sensor faults, interpolating missing values ​​using linear interpolation or KNN interpolation, and resampling data from different sampling frequencies to the same time granularity to achieve spatiotemporal alignment of multi-source data. The feature extraction includes constructing the following enhanced features: the rate of change of gas concentration is a first-order difference, the gas concentration acceleration is a second-order difference, the spatiotemporal gradient of gas concentration is the concentration difference between adjacent monitoring points, the statistical features of gas concentration in the same historical period include mean, variance and peak value, the cross features of ventilation parameters and gas concentration, and the correlation features between mining progress and gas outburst.

4. The method for assigning and assessing gas inspection tasks according to claim 1, characterized in that: The improved LSTM gas concentration prediction model in step S3 specifically includes: Input layer, used to receive past data. A multi-dimensional feature vector for each time step—the multi-dimensional feature vector includes time-series data of gas concentration, ventilation parameters, and mining parameters; The first LSTM layer contains 64 to 128 hidden units, which are used to extract the short-term dependency features of the multidimensional feature vector in the time dimension. The second LSTM layer contains 32 to 64 hidden units, which are used to extract the long-term dependency features of the multidimensional feature vector in the time dimension. The attention layer is used to assign different attention weights to the output of the second LSTM at each time step to enhance the feature representation of key time points; The fully connected output layer is used to output the predicted gas concentration values ​​of each inspection point in the next K time steps.

5. The method for assigning and assessing gas inspection tasks according to claim 1, characterized in that: The risk level classification method in step S3 is as follows: For predicted concentration The inspection points were designated as Level I low-risk areas, and regular inspections were maintained. For predicted concentration The inspection points were designated as Level II medium-risk areas, and the inspection frequency was increased to twice the normal frequency. For predicted concentration The inspection points were designated as Level III high-risk areas, and the inspection frequency was increased to three times the normal frequency. For predicted concentration The inspection point was designated as a Level IV danger zone, triggering an immediate response mechanism and simultaneously initiating continuous monitoring. Wherein, C is the predicted percentage of gas concentration, and the threshold is dynamically adjusted according to the mine safety regulations.

6. The method for assigning and assessing gas inspection tasks according to claim 1, characterized in that: The dynamic personnel capability vector in step S4 includes the following dimensions: Ability Level Dimension The value is determined based on the national vocational skills assessment level passed by the inspector, and is an integer ranging from 1 to 5. Historical performance dimension It is calculated based on the weighted moving average of the scores from the past N tasks, using the following formula: ； in, For the first The weighting coefficient of the sub-task For the first The assessment score for this task; Execution efficiency dimension The formula is calculated based on the ratio of the average completion time to the standard completion time of historical tasks by gas inspectors: ； Risk management capability dimension The ratio of the number of high-risk hazards successfully identified and handled by gas inspectors in their historical missions to the total number of high-risk missions is calculated using the following formula: ； Integrity and Compliance Dimensions The ratio of the number of tasks in which gas inspectors did not commit violations such as false inspections, missed inspections, or data tampering to the total number of tasks is calculated using the following formula: ； The matching degree matrix in step S5 Calculated as follows: For the... The inspection task and the first Individual inspector, matching degree The calculation formula is: ； in: For the first The feature vector of the required capabilities for each task; For the first The capability vector of each tile inspector; This is a function for calculating vector similarity. For the first Location feature vectors of each task; For the first Historical service area preference vector of each inspector; For the first The current location of the inspector is the first Proximity score for each task location; For the first The current remaining workload of the tile inspector is similar to that of the first... A score indicating the match between the workload required for each task and the target workload.

7. The method for assigning and assessing gas inspection tasks according to claim 1, characterized in that: The multi-objective optimization function in step S6 is: ； Among them, all those with superscript All variables are derived from... The normalized dimensionless value is obtained using the following normalization formula: ； Mapped to respectively interval; in, This is the normalized system response time; For nonlinear mapping functions of response time, This represents the normalized total length of the inspection path; To normalize the overall energy consumption of the system; To normalize the coverage of high-risk areas and adopt The form is introduced to reflect the diminishing marginal utility effect; This is a normalized prediction of the task execution success rate. To normalize the personnel matching degree in the task; For binary decision variables, taking the value 1 indicates the first... The inspection task was assigned to the first... One inspector, taking 0 indicates no allocation; For the first The performance evaluation incentives / penalties for each tile inspector are signed real values. to The weight coefficients for each objective term and satisfying ; The aforementioned performance incentives / penalties Calculate as follows: ； in, For the first The real-time cumulative assessment score of each tile inspector for the current shift. For the first Historical baseline values ​​of the inspector's capabilities; and These are the incentive coefficient and the penalty coefficient, respectively.

8. The method for assigning and assessing gas inspection tasks according to claim 1, characterized in that: The constraints of the multi-objective optimization function in step S6 include: Task skill matching constraints: For the first Each task, and the skill level of the assigned tile inspector. The skill level must be no lower than the minimum skill level required for the task. ; Task-person matching lower limit constraint: For the first task... The degree of match between the assigned tile inspector and the task. It must be no less than the preset minimum matching threshold. ; Assessment status constraint: If a tile inspector is currently in an assessment warning state, the system will not assign inspection tasks for risk areas of level III or above to him within the preset cooldown time. Workload constraints: The continuous working time of any tile inspector in a single session shall not exceed the prescribed limit; Time window constraint: Each inspection point must complete the inspection within the specified time window; Equipment status constraints: Inspection robots must not be assigned new tasks when their battery level is below a threshold; The multi-objective optimization function mentioned in step S6 is solved using an improved NSGA-II multi-objective genetic algorithm, or switched to an online decision-making algorithm based on multi-agent reinforcement learning after the system has stabilized.

9. The method for assigning and assessing gas inspection tasks according to claim 1, characterized in that: The formula for calculating the current task assessment score using the dynamic weighting algorithm in step S8 is as follows: ； in, The scores for each assessment dimension; The dynamic weights for each assessment dimension are automatically adjusted based on the risk level and difficulty coefficient of the task. For high-risk tasks in Level III and Level IV risk areas, the weight of the task execution authenticity dimension is automatically increased. to The weight of the work quality dimension will automatically increase. to ; For urgent inspection tasks triggered by the system, the weights of task response and execution efficiency are automatically increased. to ; For review tasks that are simultaneously assigned to both human and robot personnel, the weight of the data consistency dimension is automatically increased. to ; The assessment warning signal mentioned in step S9 is generated in the following manner: [Setting up a tile inspector] In the The real-time assessment score for this task is Its historical capability baseline value is The standard deviation of the historical ability baseline is ; when At that time, a preliminary assessment warning signal is generated; when At that time, an intermediate assessment warning signal will be generated, and no inspection tasks of level III or above risk areas will be assigned to the inspector within the preset 2-hour cooldown period; when When the time comes, an advanced assessment warning signal will be generated, and no inspection tasks will be assigned to the inspector within a preset 24-hour cooldown period. A personnel replacement plan will also be automatically generated.

10. A method for assigning and assessing gas inspection tasks according to claim 1, characterized in that: The distribution strategy in step S6 is automatically switched according to the global system state: In the initial state, a performance-based assignment strategy is adopted, that is, tasks are assigned to the inspectors with the highest comprehensive scores in the personnel capability profile. When a fixed periodic risk is detected in a certain area, i.e., the same area repeatedly shows a pattern of increased gas concentration within the same time period, the strategy switches to a specialty matching assignment strategy. This means that tasks in the area are preferentially assigned to the dimension with the highest risk handling capability in that area. The tile inspector who scores the tiles; When a global task backlog is detected, meaning the number of tasks to be executed exceeds 1.5 times the sum of the number of inspectors and robots on duty, the "efficiency-first" dispatch strategy is switched to – that is, tasks are prioritized for allocation based on execution efficiency. The highest-rated tile inspectors and inspection robots; The task allocation scheme in step S7 is issued to the execution terminal in the following ways: for gas inspectors, the task instructions are received in a graphic and textual manner through an intrinsically safe intelligent terminal for mining; for inspection robots, the task instructions are received in a data instruction manner through an underground wireless network; the task instructions include a list of inspection points, the optimal path, the specified time window, key inspection items, and risk warning information. During the execution process, the location is verified by comparing the real-time UWB location of the execution terminal with the planned route, and any deviation is automatically recorded; after the gas inspector arrives at the inspection point, he / she completes the arrival confirmation by swiping an NFC card or scanning a QR code; gas concentration data is automatically collected and uploaded by the intelligent gas detector; when a hidden danger is found, it is reported in real time through the mobile APP, and a rectification task is automatically generated and pushed to the responsible person.

Images