Gas extraction coupling anomaly monitoring grading method and system fusing AI time series analysis

Abstract

This invention discloses a method and system for classifying gas drainage coupled anomaly monitoring by integrating AI time-series analysis, belonging to the field of coal mine safety monitoring technology. The method includes: collecting time-series data of coupled parameters to generate a multi-time-resolution original time-series dataset; performing step-wise time-series processing on the original time-series dataset to generate regularized time-series data; constructing a multi-dimensional time-series hazard index sequence and mapping the regularized time-series data to this sequence; adaptively identifying coupled mutation trends using a good-point set-improved adaptive compression factor particle swarm optimization algorithm and outputting mutation identification results; weighted fusion of the multi-dimensional time-series hazard indicators to generate a dynamic hazard index, and dynamically adjusting the update frequency of the dynamic hazard index according to the mutation identification results; comparing the latest dynamic hazard index with a preset classification threshold to output the current risk level of the target monitoring area. This improves the real-time performance, accuracy, and stability of gas drainage anomaly monitoring.

Description

Technical Field

[0001] The embodiments of the present invention relate to the field of coal mine safety monitoring technology, and in particular to a gas extraction coupled anomaly monitoring and classification method and system that integrates AI time series analysis. Background Technology

[0002] Gas drainage is a core technical means to prevent gas outburst accidents in coal mine safety production. Especially in small tectonic areas affected by mining, the physical properties of the coal body-gas pressure-coal seam stress coupling system exhibit significant time variability. Coupled abrupt changes such as instantaneous increase in gas emission and reverse stress rise occur frequently, posing a huge challenge to monitoring and early warning.

[0003] Existing gas extraction anomaly monitoring technologies largely rely on static threshold determination and offline analysis of single parameters, which have several shortcomings: First, the monitoring data lacks a systematic time-series processing flow, and problems such as environmental noise interference and sensor drift lead to signal distortion, making it difficult to extract effective time-series evolution patterns. Second, the identification of abrupt change signals mainly relies on manually setting fixed thresholds, which cannot adaptively adapt to the time-varying characteristics of the coupled system, easily resulting in response lag, missed judgments, and false judgments. Third, the relationship between the monitoring terminal and the analysis model is one-way, lacking a dynamic feedback iteration mechanism, and the model parameters cannot be adjusted in real time with changes in operating conditions, causing the early warning accuracy to continuously decline over time. Fourth, although some monitoring systems have introduced intelligent algorithms, they suffer from poor algorithm scenario adaptability, coarse data preprocessing, and a lack of stability constraints in the model update mechanism, making it difficult to meet the accurate early warning requirements under complex mining conditions.

[0004] Therefore, there is an urgent need to build an intelligent monitoring system that focuses on time-series data processing, sudden change trend capture, and dynamic feedback calibration to break through existing technical bottlenecks and improve the real-time performance, accuracy, and stability of gas extraction anomaly monitoring. Summary of the Invention

[0005] The core objective of this invention is to address the technical pain points of existing gas extraction monitoring technologies, such as coarse time-series signal processing, reliance on manually fixed thresholds for identifying abrupt changes, and the lack of dynamic iteration mechanisms for monitoring data and analysis models. This invention provides a gas extraction coupled anomaly monitoring classification method and system that integrates AI time-series analysis, thereby breaking through existing technical bottlenecks and improving the real-time performance, accuracy, and stability of gas extraction anomaly monitoring.

[0006] In a first aspect, embodiments of the present invention provide a gas extraction coupled with anomaly monitoring and classification method that integrates AI time-series analysis, including:

[0007] The system collects coupled parameter time series data based on multiple types of monitoring sensors, and dynamically adjusts the sampling frequency according to the real-time fluctuation amplitude of the coupled parameter time series data to generate a multi-time resolution original time series dataset.

[0008] The original time series dataset is subjected to step-by-step time series processing to generate regularized time series data;

[0009] Construct a multidimensional time-series hazard index sequence, and map the regularized time-series data to the multidimensional time-series hazard index sequence;

[0010] Based on the multidimensional time-series hazard index sequence, the coupled mutation trend is adaptively identified by the good point set improved adaptive compression factor particle swarm optimization algorithm, and the mutation identification result is output.

[0011] The multidimensional time-series risk indicators are weighted and fused to generate a dynamic risk index, and the update frequency of the dynamic risk index is dynamically adjusted according to the mutation identification results.

[0012] The latest dynamic risk index is compared with the preset classification threshold to output the current risk level of the target monitoring area.

[0013] As a preferred embodiment, dynamically adjusting the sampling frequency based on the real-time fluctuation amplitude of the coupled parameter time-series data includes:

[0014] The current sampled value of each parameter in the coupled parameter time series data is compared with the corresponding static reference value to calculate the real-time fluctuation amplitude.

[0015] When the fluctuation amplitude of any parameter exceeds the preset fluctuation threshold, the sampling frequency of that parameter and its associated parameters will be automatically switched from the reference sampling interval to the high-frequency sampling interval.

[0016] When the fluctuation range of all parameters falls back to within the preset fluctuation threshold, the baseline sampling interval is restored.

[0017] In a preferred embodiment, the original time-series dataset is subjected to step-wise time-series processing to generate regularized time-series data, including:

[0018] The original time-series dataset is subjected to combined noise reduction using a combination of time-frequency domain analysis and exponential smoothing.

[0019] The isolated forest algorithm is used to automatically detect outliers in the denoised original time series dataset, identify the types of outliers, and perform hierarchical repair based on the types of outliers.

[0020] The dataset after hierarchical repair is aligned at multiple scales, and the width of the time series analysis window is adaptively adjusted according to the current working conditions to generate regular time series data with continuity and consistency.

[0021] As a preferred embodiment, a multidimensional time-series hazard index sequence is constructed, and the regularized time-series data is mapped to the multidimensional time-series hazard index sequence, including:

[0022] The parameters in the regularized time series data are normalized to obtain the standard values ​​of each parameter.

[0023] Based on the physical coupling correlation between each parameter and the dangerous state of the coupled system, a corresponding fusion weight is assigned to each parameter.

[0024] The standard values ​​of each parameter are weighted and fused according to their corresponding fusion weights to obtain the time-series risk index values ​​of each dimension.

[0025] The multidimensional time-series hazard index sequence is constructed based on the time-series hazard index values ​​of each dimension.

[0026] In a preferred embodiment, based on the multidimensional time-series hazard index sequence, a good-point set improved adaptive compression factor particle swarm optimization algorithm is used to adaptively identify coupled mutation trends and output mutation identification results, including:

[0027] An optimization objective function is constructed with mutation identification accuracy and response delay as dual optimization objectives.

[0028] The particle population is initialized using the good point set improved adaptive compression factor particle swarm optimization algorithm to make the particles uniformly distributed in a preset search space, wherein the search space includes at least the temporal sliding window width, the mutation discrimination threshold and the index correction coefficient.

[0029] An adaptive compression factor and an iteration stagnation judgment mechanism are introduced to dynamically adjust the particle velocity and position update rules according to the iteration process.

[0030] The search is iteratively optimized within the search space until the convergence condition is met, and the mutation identification result that makes the objective function optimal is output. The mutation identification result includes the mutation time, mutation trend, mutation type and mutation confidence.

[0031] In a preferred embodiment, the multidimensional time-series hazard indicators are weighted and fused to generate a dynamic hazard index, and the update frequency of the dynamic hazard index is dynamically adjusted according to the mutation identification results, including:

[0032] The contribution weight coefficients of time-series risk indicators in each dimension are determined by combining subjective experience with historical monitoring data.

[0033] The time-series hazard index values ​​of each dimension at the current time step are weighted and fused according to their corresponding contribution weight coefficients to obtain the dynamic hazard index of the target monitoring area.

[0034] When the mutation confidence in the mutation identification result exceeds the preset confidence threshold, the update frequency of the dynamic risk index is switched to a finer time step that is higher than the preset baseline time step, until the dynamic risk index falls back to within the preset safety threshold range, and no new mutation trend is identified for a consecutive preset number of time steps, then the preset baseline time step update is resumed.

[0035] In a preferred implementation, the latest dynamic hazard index is compared with a preset classification threshold to output the current risk level of the target monitoring area, including:

[0036] The current dynamic risk index is matched with the preset classification threshold to determine and output the risk level of the target monitoring area at the current moment;

[0037] The mutation type and mutation confidence level in the mutation identification results are correlated with the risk level to generate composite early warning information containing mutation type and mutation confidence information.

[0038] In a preferred embodiment, the method further includes a bidirectional dynamic feedback calibration step, specifically comprising:

[0039] The regularized time series data drives a pre-built time series prediction model to generate dynamic risk index prediction values ​​in real time.

[0040] Determine the deviation between the predicted value of the dynamic hazard index and the latest dynamic hazard index;

[0041] When the deviation meets the preset triggering condition, the parameter inversion of the pre-constructed physical numerical simulation model is driven by the regularized time series data to obtain the corrected physical parameters used to characterize the coupling characteristics of coal and rock media.

[0042] Based on the corrected physical parameters, the parameters of the time series prediction model and the physical numerical simulation model are optimized.

[0043] In a preferred embodiment, the preset triggering conditions include:

[0044] The deviations of the consecutive preset number of time steps all exceed the preset deviation threshold, the mutation trend deviates from the mutation trend predicted by the time series prediction model, and the number of times the bidirectional dynamic feedback calibration step is triggered per unit time meets the preset frequency constraint condition.

[0045] Secondly, embodiments of the present invention also provide a gas extraction coupled anomaly monitoring and classification system that integrates AI time-series analysis, comprising:

[0046] The multi-source data acquisition and adaptive sampling module is used to acquire coupled parameter time series data based on multiple types of monitoring sensors, and dynamically adjust the sampling frequency according to the real-time fluctuation amplitude of the coupled parameter time series data to generate a multi-time resolution original time series dataset.

[0047] The time series data purification module is used to perform step-by-step time series processing on the original time series dataset to generate regularized time series data.

[0048] A multidimensional hazard index construction module is used to construct a multidimensional time-series hazard index sequence and map the regularized time-series data to the multidimensional time-series hazard index sequence;

[0049] The coupled mutation trend identification module is used to adaptively identify coupled mutation trends based on the multidimensional time-series hazard index sequence using a good point set improved adaptive compression factor particle swarm optimization algorithm, and output mutation identification results.

[0050] The dynamic risk index generation and update module is used to perform weighted fusion of the multidimensional time-series risk indicators to generate a dynamic risk index, and dynamically adjust the update frequency of the dynamic risk index according to the mutation identification results.

[0051] The risk grading and early warning output module is used to compare the latest dynamic risk index with the preset grading threshold and output the current risk level of the target monitoring area.

[0052] Thirdly, embodiments of the present invention also provide an electronic device, the electronic device comprising:

[0053] One or more processors;

[0054] Storage device for storing one or more programs;

[0055] When the one or more programs are executed by the one or more processors, the one or more processors implement the gas extraction coupling anomaly monitoring and classification method that integrates AI time-series analysis as described in any embodiment of the present invention.

[0056] Fourthly, embodiments of the present invention also provide a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the gas extraction coupling anomaly monitoring and classification method integrating AI time-series analysis as described in any embodiment of the present invention.

[0057] Compared with existing technologies, the present invention achieves the following beneficial effects:

[0058] (1) Accurate and efficient coupling mutation capture: This invention completes the exclusive reconstruction of the coupling mutation trend capture scenario of the GPSCF-PSO algorithm, realizes the uniform initialization of the particle population through the good point set, and optimizes the algorithm's search characteristics by combining the adaptive compression factor and the iterative stagnation judgment mechanism. It avoids the technical defects of traditional optimization algorithms that are prone to getting trapped in local optima from the root. It can capture coal-gas-stress coupling mutation trends such as instantaneous increase in gas outburst and reverse stress recovery in an unsupervised and adaptive manner, and compresses the mutation identification response delay to the hour level. Compared with the traditional manual fixed threshold identification method, it greatly improves the accuracy and timeliness of mutation trend identification under complex mining conditions.

[0059] (2) The early warning accuracy has the characteristic of self-iterative improvement: The present invention designs a monitoring-model bidirectional dynamic feedback calibration closed loop with stability constraints. By strictly controlling the model update frequency through multi-level triggering conditions, it effectively solves the problems of system operation fluctuation and false alarm rate caused by frequent model iteration. At the same time, it completes the inversion correction of core coupled physical parameters based on real-time monitoring time series data and feeds it back to the prediction model and numerical simulation model in a synchronous manner, realizing the adaptive optimization of model parameters. This makes the early warning accuracy continuously improve with the continuous accumulation of field monitoring data, perfectly adapting to the time variability characteristics of the physical properties of coupled systems in small tectonic areas under the influence of mining.

[0060] (3) Strong horizontal expansion capability of the technical system: The integrated technical system of time-series purification-unsupervised mutation capture-dynamic feedback calibration constructed by this invention is designed around the common needs of engineering safety early warning with multi-parameter coupling and significant time-series evolution characteristics. It is not developed exclusively for gas extraction scenarios, but can be directly transferred and adapted to similar engineering scenarios such as slope stability monitoring and tunnel surrounding rock instability early warning. It provides a general technical solution for intelligent monitoring and early warning of engineering safety in multiple fields and has extremely strong scenario expansion and reuse value.

[0061] (4) The invention is highly practical and economical for engineering implementation: It abandons conventional improvements in hardware deployment and system architecture throughout the entire process, and focuses on algorithm optimization and data processing mechanism innovation. It does not require large-scale modification of the existing gas drainage monitoring hardware in coal mines and has good compatibility with the existing monitoring system. At the same time, the deployment process does not require a large amount of additional hardware costs, the modification cost is low and the implementation difficulty is small. It can be quickly implemented in various coal mine mining small structure areas for gas drainage monitoring scenarios, and the engineering promotion and application prospects are broad. Attached Figure Description

[0062] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings. The drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0063] Figure 1 This is a flowchart of the gas extraction coupled anomaly monitoring and classification method provided in the embodiments of the present invention;

[0064] Figure 2 This is a system architecture diagram of the gas extraction coupled anomaly monitoring and classification method that integrates AI time-series analysis provided in this embodiment of the invention;

[0065] Figure 3 This is a flowchart of the GPSCF-PSO algorithm mutation trend identification provided in this embodiment of the invention;

[0066] Figure 4 This is a schematic diagram of the working principle of the bidirectional dynamic feedback calibration closed loop provided in the embodiment of the present invention;

[0067] Figure 5 This is a schematic diagram comparing the performance of different algorithms provided in the embodiments of the present invention;

[0068] Figure 6 This is a schematic diagram of the structure of the gas extraction coupled anomaly monitoring and hierarchical system that integrates AI time-series analysis provided in an embodiment of the present invention;

[0069] Figure 7 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0070] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, the accompanying drawings show only the parts relevant to the present invention, and not all of the structures.

[0071] Before discussing the exemplary embodiments in more detail, it should be mentioned that some exemplary embodiments are described as processes or methods depicted as flowcharts. Although the flowcharts describe operations (or steps) as being processed sequentially, many of these operations (or steps) may be performed in parallel, concurrently, or simultaneously. Furthermore, the order of the operations may be rearranged. The process may be terminated when its operation is completed, but may also have additional steps not included in the figures. The process may correspond to a method, function, procedure, subroutine, subroutine, etc.

[0072] Example 1

[0073] like Figure 1 The diagram shows a flowchart of a gas extraction coupled anomaly monitoring and classification method 100 based on AI time-series analysis provided in Embodiment 1 of the present invention. The method 100 specifically includes the following steps:

[0074] S110: Based on the acquisition of coupling parameter time series data by multiple types of monitoring sensors, the sampling frequency is dynamically adjusted according to the real-time fluctuation amplitude of the coupling parameter time series data to generate a multi-time resolution original time series dataset.

[0075] This step only establishes the basic data acquisition framework required for time series analysis and does not involve any improvement or innovation in hardware structure. The core purpose is to provide a raw data source with a unified time dimension and adaptability to disturbance scenarios for subsequent full-process time series processing.

[0076] As a preferred embodiment, multiple types of monitoring sensors are deployed to collect the following three types of core coupling parameters, based on the coupling mechanism of coal body-gas pressure-coal seam stress:

[0077] (1) Gas-related parameters: Time-series data of pore pressure, extraction flow rate and gas concentration were collected by pore pressure sensor and gas extraction flow sensor;

[0078] (2) Stress-deformation related parameters: Time series data of principal stress, shear stress increment and displacement of surrounding rock were collected by surrounding rock stress sensor and roadway displacement convergence meter;

[0079] (3) Microseismic energy related parameters: The frequency of microseismic events and the time series data of energy released by microseismic events are collected by microseismic monitoring probes.

[0080] As a preferred embodiment, all the data collected by the above sensors are bound to the Universal Time (UTC) timestamp to achieve rigid alignment of timestamps of multi-source data; an adaptive sampling mechanism is adopted, with a 5-minute sampling interval set under normal stable operating conditions.

[0081] Furthermore, the current sampled value of each parameter is compared with the corresponding static baseline value to calculate the real-time fluctuation amplitude. When the fluctuation amplitude of a single parameter exceeds the preset fluctuation threshold (such as 15% of the static baseline value), the sampling frequency of the parameter and its associated parameters is automatically switched from the baseline sampling interval to a 1-minute high-frequency sampling interval. When the fluctuation amplitude of all parameters falls back to within the preset fluctuation threshold, the 5-minute baseline sampling interval is restored, and finally, a multi-time resolution original time series dataset is generated to adapt to the time series analysis needs under different disturbance intensities.

[0082] S120: Perform step-by-step time series processing on the original time series dataset to generate regular time series data.

[0083] In a preferred embodiment, this step completes signal denoising, outlier repair, and time-series scale normalization through a stepped time-series processing flow. This addresses the issues of noise interference, data distortion, and inconsistent multi-source scales in the original monitoring signal, providing high-quality time-series input for subsequent abrupt trend capture. Specifically, it includes the following steps:

[0084] Step S121: Temporal noise suppression based on wavelet transform

[0085] (1) For non-stationary time-series signals in gas extraction monitoring, the Daubechies 4th order (db4) wavelet basis function is selected. This basis function has linear phase and compact support characteristics, which are suitable for the characteristics of gradual and abrupt changes in geological engineering signals. The number of wavelet decomposition layers is set. =4, this value has been verified by frequency band decomposition, and can effectively separate the effective signal of the coupling response between environmental high-frequency noise and coal and rock mass. The filtering calculation formula is:

[0086] ;

[0087] in, : The noise-reduced signal of the gas extraction monitoring time series signal after wavelet filtering; The wavelet decomposition level is set to 4. After frequency band decomposition verification, it can effectively separate the effective signal of the coupling response between environmental high-frequency noise and coal and rock mass. Wavelet decomposition layer number, with a value ranging from 1 to... ; : The displacement coefficient index of the wavelet transform; : No. In the wavelet decomposition layer, the first Wavelet coefficients; : Daubechies 4th order (db4) wavelet basis functions, with linear phase and compact support characteristics, suitable for the characteristics of gradual and abrupt changes in signals in gas extraction geological engineering; : Timestamp of timing signals.

[0088] (2) For discrete microseismic signals, exponential smoothing is used to assist in noise reduction. The calculation formula is as follows:

[0089] ;

[0090] in, Microseismic discrete signals in The signal value after exponential smoothing and noise reduction at any given moment; The smoothing coefficient of the exponential smoothing method has a range of values. It can balance the signal smoothing effect with the preservation of the integrity of abrupt features; Microseismic discrete signals in The original sampled value at time; Microseismic discrete signals in The signal value after exponential smoothing and noise reduction at any given moment.

[0091] In summary, through the above combined noise reduction processing, the system's timing noise suppression rate is improved. The above methods preserve the effective timing characteristics of the signal to the greatest extent possible.

[0092] Step S122: Temporal outlier identification and repair based on isolated forest

[0093] The isolated forest algorithm is used to achieve automated detection of time-series outliers. Key operating parameters are set as follows: number of base estimators n_estimators=100, maximum number of samples max_samples=256, and prior anomaly proportion contamination=0.03. This parameter configuration matches the statistical proportion of anomaly data in the historical monitoring on site to ensure the generalization ability of anomaly identification.

[0094] The hierarchical repair logic is implemented: progressive anomalies such as sensor drift are repaired using linear interpolation, while intermittent anomalies such as instantaneous power outages and signal interruptions are repaired by cross-validation with data from adjacent sensors of the same type and then by mean filling. This process removes invalid interference while retaining true mutation characteristics, and the anomaly identification accuracy can reach 96%.

[0095] Step S123: Multi-scale temporal alignment and normalization

[0096] This embodiment uses rolling average combined with timestamp rigid alignment technology to perform multi-scale alignment on the hierarchical repaired dataset and fills in short-term missing data through linear interpolation.

[0097] Simultaneously, an adaptive sliding window T is set, with a window length of 24h under normal operating conditions and 1h under severe mining disturbance conditions. This completes the time series dimension normalization of multi-source parameters, ensuring the consistency of data format for subsequent catastrophe analysis and generating regular time series data with continuity and consistency.

[0098] S130: Construct a multidimensional time-series hazard index sequence and map the regularized time-series data to the multidimensional time-series hazard index sequence.

[0099] As a preferred embodiment, the multidimensional time-series hazard index sequence includes the following steps:

[0100] (1) Index normalization: All basic parameters are normalized using extreme value normalization (min-max) to eliminate dimensional differences and obtain the standard values ​​of each parameter. The calculation formula is as follows:

[0101] ;

[0102] in, The basic parameters for gas extraction monitoring, after extreme value normalization (min-max), have a range of values. ; : Raw values ​​of basic parameters for gas extraction monitoring; The minimum values ​​of the basic parameters for gas drainage monitoring in the target mining area over the past three years; The maximum value of the basic parameters of gas extraction monitoring in the target mining area over the past three years is statistically analyzed.

[0103] For permeability change compared to analogous value indicators, logarithmic normalization is used, and the calculation formula is as follows:

[0104] ;

[0105] in, The change in penetration rate is the log-normalized value of the indicator. : The ratio of changes in coal and rock permeability during gas extraction; The historical maximum value of the change ratio of gas extraction permeability in the target mining area.

[0106] (2) Determine the fusion weight of each sub-parameter: Based on the physical coupling correlation between each basic parameter and the dangerous state of the coupled system, assign a corresponding fusion weight to each sub-parameter in the subsequent four-dimensional index; wherein, the physical coupling correlation is used to quantify the strength of the physical interaction mechanism between each monitoring parameter (such as gas pressure, surrounding rock stress, microseismic energy, etc.) and dangerous states such as coal and rock mass instability and gas outburst.

[0107] Specific acquisition methods include: determining the response sensitivity of parameter changes and critical failure states based on coal and rock mechanics tests; conducting grey relational analysis using historical outburst accident case data to calculate the correlation coefficient between each parameter and outburst risk; and obtaining a normalized correlation degree value after expert experience correction, which serves as the basis for allocating fusion weights; for example, the weight coefficients of each sub-parameter in the structural sensitivity index. , , .

[0108] (3) Construction of four-dimensional indicators: Based on the normalized parameters and the determined weights of the sub-parameters, the normalized indicator values ​​of each parameter are weighted and fused according to their corresponding fusion weights to construct a four-dimensional time series risk indicator. The specific formula is as follows:

[0109] Constructing sensitivity indicators: ;

[0110] Residual gas index: ;

[0111] Dynamic stress index: ;

[0112] Decoupling criteria: .

[0113] in, , , The weighting coefficients of each sub-parameter in the constructed sensitivity index are determined jointly by the sensitivity analysis of historical prominent cases and the requirements of coal mine safety regulations. : The index value after normalizing the elevation difference; The index value of the fracture zone width after extreme value normalization; : The joint development index after extreme value normalization; , The weighting coefficients of each sub-parameter in the residual gas index are determined jointly by the sensitivity analysis of historical outburst cases and the requirements of coal mine safety regulations. The residual gas content is the index value after extreme value normalization; The residual pore pressure is the index value after extreme value normalization; , , The weighting coefficients of each sub-parameter in the dynamic stress index are determined jointly by the sensitivity analysis of historical prominent cases and the requirements of coal mine safety regulations. The index value of the principal stress increment after extreme value normalization; The index value of shear stress increment after extreme value normalization; The index value of the equivalent rate of change after extreme value normalization; , The weighting coefficients of each sub-parameter in the decoupling criterion index are determined jointly by the sensitivity analysis of prominent historical cases and the requirements of coal mine safety regulations. The index value of the pore pressure-stress imbalance coefficient after extreme value normalization.

[0114] A multidimensional time-series hazard index sequence is constructed based on the time-series hazard index values ​​of each dimension.

[0115] S140: Based on the multidimensional time-series risk index sequence, the coupling mutation trend is adaptively identified by the good point set improved adaptive compression factor particle swarm optimization algorithm, and the mutation identification result is output.

[0116] As a preferred embodiment, this invention proposes a particle swarm optimization based on a good-point set and adaptive compression factor (GPSCF-PSO) algorithm. This algorithm is based on the traditional particle swarm optimization (PSO) algorithm, introducing a good-point set to achieve uniform initialization of the particle population, avoiding local optima problems caused by random initialization. Simultaneously, it embeds an adaptive compression factor and an iteration stagnation judgment mechanism to balance the algorithm's global search and local optimization capabilities, overcoming the shortcomings of traditional algorithms such as slow convergence speed and low optimization accuracy. This invention reconstructs and adapts this algorithm from hyperparameter optimization scenarios to temporal abrupt trend capture scenarios, redesigning the optimization objective, search space, and execution logic to achieve unsupervised and adaptive identification of coupled abrupt trends.

[0117] Combination Figure 3 As shown, the GPSCF-PSO algorithm adaptation design includes the following:

[0118] (1) Optimize the design of the objective function

[0119] An optimization objective function is constructed with mutation identification accuracy and response delay as the dual optimization objectives. :

[0120] ;

[0121] in, The optimization objective function of the GPSCF-PSO algorithm is used to balance the accuracy of cooperative mutation identification and the identification response delay. The harmonic mean of precision and recall in mutation trend detection represents the mutation identification accuracy of the algorithm. The single-window mutation identification delay of the GPSCF-PSO algorithm characterizes the algorithm's response speed. : Hourly reference delay, which serves as the baseline reference value for mutation identification delay.

[0122] The objective function aims to simultaneously maximize mutation identification accuracy and minimize response latency, and uses the formula... The problem of maximizing accuracy is transformed into a problem of minimization, which facilitates co-optimization with the delay term. The weight coefficients of 0.7 and 0.3 reflect the priority given to ensuring recognition accuracy.

[0123] (2) Definition of search space

[0124] Using key parameters for identifying time-series abrupt changes as the optimization target, and combining on-site measured data to verify and determine the parameter boundaries: the width of the time-series sliding window. (Unit: standard time step), threshold for discriminating abrupt changes in four-dimensional indicators Correction coefficient for contribution of indicator mutation .

[0125] (3) Core Iteration Formula

[0126] Population initialization uses a good point set to generate uniformly distributed particles, calculated using the following formula:

[0127] ;

[0128] in, The first step in the GPS CF-PSO algorithm The first good point set generated Initial positions of the particles; The first step in the GPS CF-PSO algorithm The upper limit of the value of each optimization parameter; The first step in the GPS CF-PSO algorithm The lower limit of the value of each optimization parameter; : Good point set sequence, used to achieve uniform initialization of the particle population in the GPSCF-PSO algorithm.

[0129] Simultaneously, an adaptive compression factor and an iteration stagnation judgment mechanism are introduced. The formula for calculating the compression factor is:

[0130] ;

[0131] ;

[0132] in, The adaptive compression factor of the GPSCF-PSO algorithm is used to balance the algorithm's global search and local optimization capabilities. The acceleration constant of the GPSCF-PSO algorithm is fixed at 4.1. The individual learning factor in the GPSCF-PSO algorithm is a component of the speedup constant. The group learning factor of the GPSCF-PSO algorithm is a component of the speedup constant.

[0133] The particle velocity update formula is:

[0134] ;

[0135] in, The first step in the GPS CF-PSO algorithm The velocity of each particle; The inertial weight of the GPSCF-PSO algorithm is linearly decreased from 0.9 to 0.3 to adjust the search range of particles. A random number between 0 and 1, used to increase the randomness of the particle search; The first step in the GPS CF-PSO algorithm The historical optimal solution for each particle; : The global optimal solution of the particle population in the GPSCF-PSO algorithm; flagnum: The number of stagnation iterations of the global optimal value in the GPSCF-PSO algorithm, used to trigger the switching of the particle velocity update rule.

[0136] As a preferred embodiment, the time-normalized four-dimensional time-series index sequence is used as the algorithm input. The GPSCF-PSO algorithm is used to find the optimal combination of identification parameters and output the mutation identification results, including information such as mutation time, mutation trend, mutation type and mutation confidence. It can automatically determine typical coupled mutation trends such as instantaneous increase in gas outburst and reverse stress recovery. The identification response delay is compressed to the hour level, which overcomes the adaptability defects of traditional fixed threshold identification methods and improves the accuracy of mutation trend identification under complex working conditions.

[0137] S150: The multidimensional time-series risk indicators are weighted and fused to generate a dynamic risk index, and the update frequency of the dynamic risk index is dynamically adjusted according to the mutation identification results.

[0138] In a preferred embodiment, the time-series hazard index values ​​of each dimension at the current time step are weighted and fused according to their corresponding contribution weight coefficients to obtain the dynamic hazard index of the target monitoring area. The calculation formula is:

[0139]

[0140] in, The dynamic risk index is a comprehensive risk assessment indicator that couples gas extraction with abnormal monitoring, characterizing the degree of gas outburst risk in the monitored area. : Structural sensitivity index, one of the four-dimensional time-series hazard indicators, characterizes the degree of influence of geological structure in the gas drainage area on outburst risk; The residual gas index is one of the four-dimensional time-series hazard indicators, which characterizes the degree of influence of residual gas in the gas drainage area on the risk of outburst. Dynamic stress index, one of the four-dimensional time-series hazard indicators, characterizes the degree of influence of dynamic stress in the surrounding rock of the gas drainage area on the risk of outburst. The decoupling criterion index is one of the four-dimensional time-series hazard indicators, which characterizes the impact of the degree of pore pressure-stress imbalance in the gas drainage area on the risk of outburst.

[0141] The contribution weight coefficients of each dimension of time-series hazard indicators were determined by combining the analytic hierarchy process (AHP) with historical accident case data, taking into account both the coupling physical mechanism of coal and rock mass and the actual situation of on-site engineering.

[0142] In a preferred embodiment, dynamically adjusting the update frequency of the dynamic risk index based on the mutation identification results in step S140 includes:

[0143] Establish a multi-scale time-step adaptive update mechanism that is linked to the mutation trend, and use daily time-step updates for the dynamic hazard index under normal stable operating conditions;

[0144] When the GPSCF-PSO algorithm detects a coupled mutation trend, such as when the mutation confidence in the mutation identification result exceeds a preset confidence threshold (e.g., 0.8), it automatically switches the update frequency to hourly fine-grained updates (updating once every hour) to enhance real-time tracking of risk evolution.

[0145] Once the risk level corresponding to the dynamic risk index falls back to the low to medium level or below, and no new mutation trend is identified within two or more consecutive monitoring windows, the system automatically resumes to the regular daily update step size, achieving a dynamic balance between computing resources and early warning timeliness.

[0146] S160: Compare the latest dynamic risk index with the preset classification threshold and output the current risk level of the target monitoring area.

[0147] As a preferred embodiment, based on cluster analysis of historical on-site data, the present invention classifies the risk level into four levels:

[0148] Low risk: Low to medium risk: ;

[0149] Medium to high risk: High risk: .

[0150] Field verification showed that this risk-level classification strategy can achieve dynamic and accurate risk assessment, with the system achieving a high accuracy rate in identifying high-risk levels. The above methods can effectively reduce the probability of missed or false alarms. By matching the current dynamic hazard index with the above risk levels, the risk level of the target monitoring area at the current moment is determined and output.

[0151] In addition, the mutation type and mutation confidence level in the mutation identification results are associated with the risk level to generate composite early warning information containing mutation type and mutation confidence level information. The association method is as follows: based on the identified mutation type and mutation confidence level, combined with the risk level corresponding to the current dynamic risk index, a structured early warning message is generated through a preset mapping rule and output to the monitoring and early warning interface or the upper-level control system. The specific generation method of the structured early warning message is as follows: the risk level, mutation type, mutation confidence level and suggested response measures are spliced ​​together according to a predetermined format.

[0152] As a preferred embodiment, combined with Figure 2 and Figure 4 As shown, the method further constructs a two-way dynamic feedback calibration closed loop driven by the measured-prediction deviation. Its core logic is as follows: using real-time monitoring time-series data as a benchmark, a gated recurrent unit (GRU) model is driven to predict the dynamic hazard index; the deviation between the predicted and measured values ​​is used as a trigger signal to initiate parameter inversion of the Fast Lagrange Analysis 3D model (FLAC3D); the key physical parameters obtained from the inversion are synchronously transmitted back to the GRU model and the FLAC3D model, completing one two-way calibration iteration. By introducing an update mechanism with stability constraints, the system fluctuations and increased false alarms caused by frequent model iterations are effectively solved, achieving adaptive iterative improvement of early warning accuracy as field data accumulates.

[0153] Training data system and basic model configuration:

[0154] The training data consists of two types of data sources: three years of field-measured time-series data from the target mining area, covering 12 mining faces, with 17,200 valid samples after removing invalid values; and simulated data generated using a coupled numerical model constructed with FLAC3D, with 6,400 samples. The simulated data was calibrated against field standards, and its relative error with the measured values ​​is less than 8%, ensuring the validity and representativeness of the data. The total dataset is divided into training and testing sets in an 8:2 ratio, and 5-fold cross-validation is used to improve the model's generalization ability. A single-hidden-layer GRU time-series prediction model is constructed, with the input layer being preprocessed multi-parameter time-series window data and the output layer being the predicted value of the dynamic hazard index. The mean squared error (MSE) loss function and the Adam optimizer are used.

[0155] The time-series processed measured data is input into the GRU model, which outputs a real-time predicted value of the dynamic risk index. The relative error between this predicted value and the measured value after mutation identification correction is calculated using the following formula:

[0156]

[0157] in, The relative error of the dynamic risk index prediction is the percentage of the deviation between the model prediction and the measured value. The measured value of the dynamic risk index after mutation identification correction is calculated from the on-site monitoring data of gas extraction. The predicted dynamic hazard index output by the gated cyclic unit (GRU) model is obtained by inputting the time-series processed measured data into the model.

[0158] Furthermore, to ensure the stability of the real-time monitoring system, this embodiment sets multi-level constraint triggering conditions, and initiates the model calibration process only when all conditions are met, specifically including:

[0159] (1) Relative error of prediction over three consecutive standard time steps ;

[0160] (2) The results of the mutation trend identification deviate from the trend predicted by the model;

[0161] (3) Meet the frequency constraints: the number of model updates per day shall not exceed 1, and the cumulative number of updates per week shall not exceed 3.

[0162] Furthermore, after the update is triggered, the real-time monitoring time-series data is used as the target, and iterative comparison is performed with the FLAC3D numerical simulation results to inversely correct core coupled physical parameters such as permeability-strain sensitivity coefficient and coal body damage threshold. A model version management and gray-scale release strategy is adopted: before the update, the historical optimal model weights are retained; after the update, it is first tested on edge computing nodes, and the full system is deployed only after the early warning accuracy does not decrease for 12 consecutive hours to ensure the continuity of system operation; the corrected optimal parameter combination is entered into the mining area parameter sensitivity database, and the gated cyclic unit (GRU) model weights and numerical simulation model parameters are optimized simultaneously to form a two-way dynamic feedback closed loop in which monitoring data drives model optimization and model output feeds back into monitoring and early warning, and the early warning accuracy continues to improve with the accumulation of field data.

[0163] In summary, the "time-series purification-unsupervised mutation capture-dynamic feedback calibration" technical system constructed in this invention has core methods (scene reconstruction of the good point set improved adaptive compression factor particle swarm optimization (GPSCF-PSO) algorithm and closed-loop update mechanism with stability constraints). These methods are not only applicable to gas extraction anomaly monitoring, but can also be transferred to engineering safety early warning scenarios with multi-parameter coupling and significant time-series evolution characteristics, such as slope stability monitoring and tunnel surrounding rock instability early warning. It has good potential for horizontal expansion.

[0164] To verify the performance advantages of this invention, four algorithms—GRU, Long Short-Term Memory Network (LSTM), Particle Swarm Optimization-GRU (PSO-GRU), and the GPSCF-PSO-GRU proposed in this invention—were compared in an experiment. The results are as follows: Figure 5 As shown:

[0165] The convergence comparison curves of the loss function (mean squared error, MSE) clearly show that the four algorithms—Gated Recurrent Unit (GRU), Long Short-Term Memory (LSTM), Particle Swarm Optimization-Gated Recurrent Unit (PSO-GRU), and Good Point Set Improved Adaptive Compression Factor Particle Swarm Optimization-Gated Recurrent Unit (GPSCF-PSO-GRU)—exhibit significant performance differences during the iteration process. LSTM and GRU, as the basic time-series models, have the slowest convergence speed, with final MSE values ​​remaining at 0.30 and 0.28 respectively, reflecting the difficulty of quickly capturing the coupled abrupt changes in gas extraction time-series data without the introduction of optimization algorithms. PSO-GRU, by optimizing the GRU hyperparameters using the Particle Swarm Optimization (PSO) algorithm, significantly improves the convergence speed; the MSE decreases from 0.75 to 0.30 in the first 30 iterations and reaches 0.05 after 100 iterations, demonstrating… The intelligent optimization algorithm improves the performance of the time series model; however, the GPSCF-PSO-GRU algorithm proposed in this invention exhibits the best convergence characteristics. In the first 30 iterations, the MSE drops rapidly from 0.70 to 0.20, with a convergence speed significantly better than PSO-GRU. Moreover, it maintains a rapid decreasing trend with the number of iterations, and the MSE drops precisely to 0.01 after 100 iterations. This result verifies that the GPSCF-PSO algorithm is more efficient in optimizing the hyperparameters of the GRU model. Through the improved strategy of good point set initialization and adaptive compression factor, it effectively avoids the problem of traditional PSO algorithm being prone to getting trapped in local optima. It can more accurately adapt to the needs of capturing the abrupt change features of four-dimensional coupled time series data of gas extraction, and finally achieves rapid convergence of the model loss function and extremely low final error, which fully demonstrates the superior performance of the algorithm in the gas extraction anomaly monitoring scenario.

[0166] Based on the above embodiments, the core beneficial effects of the present invention are as follows:

[0167] (1) Accurate and efficient coupling mutation capture: This invention completes the exclusive reconstruction of the coupling mutation trend capture scenario of the GPSCF-PSO algorithm, realizes the uniform initialization of the particle population through the good point set, and optimizes the algorithm's search characteristics by combining the adaptive compression factor and the iterative stagnation judgment mechanism. It avoids the technical defects of traditional optimization algorithms that are prone to getting trapped in local optima from the root. It can capture coal-gas-stress coupling mutation trends such as instantaneous increase in gas outburst and reverse stress recovery in an unsupervised and adaptive manner, and compresses the mutation identification response delay to the hour level. Compared with the traditional manual fixed threshold identification method, it greatly improves the accuracy and timeliness of mutation trend identification under complex mining conditions.

[0168] (2) The early warning accuracy has the characteristic of self-iterative improvement: The present invention designs a monitoring-model bidirectional dynamic feedback calibration closed loop with stability constraints. By strictly controlling the model update frequency through multi-level triggering conditions, it effectively solves the problems of system operation fluctuation and false alarm rate caused by frequent model iteration. At the same time, it completes the inversion correction of core coupled physical parameters based on real-time monitoring time series data and feeds it back to the prediction model and numerical simulation model in a synchronous manner, realizing the adaptive optimization of model parameters. This makes the early warning accuracy continuously improve with the continuous accumulation of field monitoring data, perfectly adapting to the time variability characteristics of the physical properties of coupled systems in small tectonic areas under the influence of mining.

[0169] (3) Strong horizontal expansion capability of the technical system: The integrated technical system of time-series purification-unsupervised mutation capture-dynamic feedback calibration constructed by this invention is designed around the common needs of engineering safety early warning with multi-parameter coupling and significant time-series evolution characteristics. It is not developed exclusively for gas extraction scenarios, but can be directly transferred and adapted to similar engineering scenarios such as slope stability monitoring and tunnel surrounding rock instability early warning. It provides a general technical solution for intelligent monitoring and early warning of engineering safety in multiple fields and has extremely strong scenario expansion and reuse value.

[0170] (4) The invention is highly practical and economical for engineering implementation: It abandons conventional improvements in hardware deployment and system architecture throughout the entire process, and focuses on algorithm optimization and data processing mechanism innovation. It does not require large-scale modification of the existing gas drainage monitoring hardware in coal mines and has good compatibility with the existing monitoring system. At the same time, the deployment process does not require a large amount of additional hardware costs, the modification cost is low and the implementation difficulty is small. It can be quickly implemented in various coal mine mining small structure areas for gas drainage monitoring scenarios, and the engineering promotion and application prospects are broad.

[0171] Example 2

[0172] Figure 6 This is a schematic diagram of a gas extraction coupled with anomaly monitoring and grading system that integrates AI time-series analysis, as provided in Embodiment 2 of the present invention. Figure 6 As shown, the system includes:

[0173] The multi-source data acquisition and adaptive sampling module 610 is used to acquire coupled parameter time series data based on multiple types of monitoring sensors, and dynamically adjust the sampling frequency according to the real-time fluctuation amplitude of the coupled parameter time series data to generate a multi-time resolution original time series dataset.

[0174] The time series data purification module 620 is used to perform step-by-step time series processing on the original time series dataset to generate regular time series data.

[0175] A multidimensional hazard index construction module 630 is used to construct a multidimensional time-series hazard index sequence and map the regularized time-series data to the multidimensional time-series hazard index sequence;

[0176] The coupling mutation trend identification module 640 is used to adaptively identify coupling mutation trends based on the multidimensional time-series hazard index sequence using a good point set improved adaptive compression factor particle swarm optimization algorithm, and output mutation identification results.

[0177] The dynamic risk index generation and update module 650 is used to perform weighted fusion of the multidimensional time-series risk indicators to generate a dynamic risk index, and dynamically adjust the update frequency of the dynamic risk index according to the mutation identification results.

[0178] The risk grading and early warning output module 660 is used to compare the latest dynamic risk index with the preset grading threshold and output the current risk level of the target monitoring area.

[0179] The gas extraction coupling anomaly monitoring and classification system fused with AI time-series analysis provided in this embodiment of the invention can execute the gas extraction coupling anomaly monitoring and classification method fused with AI time-series analysis provided in any of the embodiments of the invention above. It has the corresponding functions and beneficial effects of executing the gas extraction coupling anomaly monitoring and classification method fused with AI time-series analysis. For detailed process, please refer to the relevant operations of the gas extraction coupling anomaly monitoring and classification method fused with AI time-series analysis in the foregoing embodiments.

[0180] Example 3

[0181] Figure 7 This is a schematic diagram of the structure of an electronic device provided in Embodiment 3 of the present invention. The electronic device 10 is intended to represent various forms of digital computers, and may also represent various forms of mobile devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the invention described and / or claimed herein.

[0182] like Figure 7 As shown, the electronic device 10 includes at least one processor 11 and a memory, such as a read-only memory (ROM) 12 or a random access memory (RAM) 13, communicatively connected to the at least one processor 11. The memory stores computer programs executable by the at least one processor. The processor 11 can perform various appropriate actions and processes based on the computer program stored in the ROM 12 or loaded into the RAM 13 from storage unit 18. The RAM 13 can also store various programs and data required for the operation of the electronic device 10. The processor 11, ROM 12, and RAM 13 are interconnected via a bus 14. An input / output (I / O) interface 15 is also connected to the bus 14.

[0183] Multiple components in electronic device 10 are connected to I / O interface 15, including: input unit 16, such as keyboard, mouse, etc.; output unit 17, such as various types of displays, speakers, etc.; storage unit 18, such as disk, optical disk, etc.; and communication unit 19, such as network card, modem, wireless transceiver, etc. Communication unit 19 allows electronic device 10 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0184] Processor 11 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of processor 11 include, but are not limited to, central processing unit (CPU), graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various processors running machine learning model algorithms, digital signal processors (DSPs), and any suitable processor, controller, microcontroller, etc. Processor 11 executes the gas extraction coupled anomaly monitoring classification method with fused AI time-series analysis described above.

[0185] It should be understood that the various forms of processes shown above can be used, with steps reordered, added, or deleted. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution of this invention can be achieved, and this is not limited herein.

[0186] The above embodiments are merely illustrative examples and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A gas extraction coupling anomaly monitoring hierarchical method fused with AI time series analysis, characterized in that, include: The system collects coupled parameter time series data based on multiple types of monitoring sensors, and dynamically adjusts the sampling frequency according to the real-time fluctuation amplitude of the coupled parameter time series data to generate a multi-time resolution original time series dataset. The original time series dataset is subjected to step-by-step time series processing to generate regularized time series data; Construct a multidimensional time-series hazard index sequence, and map the regularized time-series data to the multidimensional time-series hazard index sequence; Based on the multidimensional time-series hazard index sequence, the coupled mutation trend is adaptively identified by the good point set improved adaptive compression factor particle swarm optimization algorithm, and the mutation identification result is output. The multidimensional time-series risk indicators are weighted and fused to generate a dynamic risk index, and the update frequency of the dynamic risk index is dynamically adjusted according to the mutation identification results. The latest dynamic risk index is compared with the preset classification threshold to output the current risk level of the target monitoring area.

2. The method of claim 1, wherein, Dynamically adjusting the sampling frequency based on the real-time fluctuation amplitude of the coupled parameter time series data includes: The current sampled value of each parameter in the coupled parameter time series data is compared with the corresponding static reference value to calculate the real-time fluctuation amplitude. When the fluctuation amplitude of any parameter exceeds the preset fluctuation threshold, the sampling frequency of that parameter and its associated parameters will be automatically switched from the reference sampling interval to the high-frequency sampling interval. When the fluctuation range of all parameters falls back to within the preset fluctuation threshold, the baseline sampling interval is restored.

3. The method of claim 1, wherein, The original time series dataset is subjected to step-wise time series processing to generate regularized time series data, including: The original time-series dataset is subjected to combined noise reduction using a combination of time-frequency domain analysis and exponential smoothing. The isolated forest algorithm is used to automatically detect outliers in the denoised original time series dataset, identify the types of outliers, and perform hierarchical repair based on the types of outliers. The dataset after hierarchical repair is aligned at multiple scales, and the width of the time series analysis window is adaptively adjusted according to the current working conditions to generate regular time series data with continuity and consistency.

4. The method of claim 3, wherein, Constructing a multidimensional time-series hazard index sequence and mapping the regularized time-series data to the multidimensional time-series hazard index sequence includes: The parameters in the regularized time series data are normalized to obtain the standard values ​​of each parameter. Based on the physical coupling correlation between each parameter and the dangerous state of the coupled system, a corresponding fusion weight is assigned to each parameter. The standard values ​​of each parameter are weighted and fused according to their corresponding fusion weights to obtain the time-series risk index values ​​of each dimension. The multidimensional time-series hazard index sequence is constructed based on the time-series hazard index values ​​of each dimension.

5. The method of claim 4, wherein, Based on the multidimensional time-series hazard index sequence, a good-point set improved adaptive compression factor particle swarm optimization algorithm is used to adaptively identify coupled mutation trends, and the mutation identification results are output, including: An optimization objective function is constructed with mutation identification accuracy and response delay as dual optimization objectives. The particle population is initialized using the good point set improved adaptive compression factor particle swarm optimization algorithm to make the particles uniformly distributed in a preset search space, wherein the search space includes at least the temporal sliding window width, the mutation discrimination threshold and the index correction coefficient. An adaptive compression factor and an iteration stagnation judgment mechanism are introduced to dynamically adjust the particle velocity and position update rules according to the iteration process. The search is iteratively optimized within the search space until the convergence condition is met, and the mutation identification result that makes the objective function optimal is output. The mutation identification result includes the mutation time, mutation trend, mutation type and mutation confidence.

6. The method of claim 5, wherein, The multidimensional time-series hazard indicators are weighted and fused to generate a dynamic hazard index, and the update frequency of the dynamic hazard index is dynamically adjusted based on the mutation identification results, including: The contribution weight coefficients of time-series risk indicators in each dimension are determined by combining subjective experience with historical monitoring data. The time-series hazard index values ​​of each dimension at the current time step are weighted and fused according to their corresponding contribution weight coefficients to obtain the dynamic hazard index of the target monitoring area. When the mutation confidence in the mutation identification result exceeds the preset confidence threshold, the update frequency of the dynamic risk index is switched to a finer time step that is higher than the preset baseline time step, until the dynamic risk index falls back to within the preset safety threshold range, and no new mutation trend is identified for a consecutive preset number of time steps, then the preset baseline time step update is resumed.

7. The method of claim 5, wherein, The latest dynamic hazard index is compared with the preset classification threshold to output the current risk level of the target monitoring area, including: The current dynamic risk index is matched with the preset classification threshold to determine and output the risk level of the target monitoring area at the current moment; The mutation type and mutation confidence level in the mutation identification results are correlated with the risk level to generate composite early warning information containing mutation type and mutation confidence information.

8. The method of claim 1, wherein, The method further includes a bidirectional dynamic feedback calibration step, specifically including: The regularized time series data drives a pre-built time series prediction model to generate dynamic risk index prediction values ​​in real time. Determine the deviation between the predicted value of the dynamic hazard index and the latest dynamic hazard index; When the deviation meets the preset triggering condition, the parameter inversion of the pre-constructed physical numerical simulation model is driven by the regularized time series data to obtain the corrected physical parameters used to characterize the coupling characteristics of coal and rock media. Based on the corrected physical parameters, the parameters of the time series prediction model and the physical numerical simulation model are optimized.

9. The method of claim 8, wherein, The preset triggering conditions include: The deviations of the consecutive preset number of time steps all exceed the preset deviation threshold, the mutation trend deviates from the mutation trend predicted by the time series prediction model, and the number of times the bidirectional dynamic feedback calibration step is triggered per unit time meets the preset frequency constraint condition.

10. A gas extraction coupling anomaly monitoring hierarchical system fused with AI time series analysis, characterized in that, include: The multi-source data acquisition and adaptive sampling module is used to acquire coupled parameter time series data based on multiple types of monitoring sensors, and dynamically adjust the sampling frequency according to the real-time fluctuation amplitude of the coupled parameter time series data to generate a multi-time resolution original time series dataset. The time series data purification module is used to perform step-by-step time series processing on the original time series dataset to generate regularized time series data. A multidimensional hazard index construction module is used to construct a multidimensional time-series hazard index sequence and map the regularized time-series data to the multidimensional time-series hazard index sequence; The coupled mutation trend identification module is used to adaptively identify coupled mutation trends based on the multidimensional time-series hazard index sequence using a good point set improved adaptive compression factor particle swarm optimization algorithm, and output mutation identification results. The dynamic risk index generation and update module is used to perform weighted fusion of the multidimensional time-series risk indicators to generate a dynamic risk index, and dynamically adjust the update frequency of the dynamic risk index according to the mutation identification results. The risk grading and early warning output module is used to compare the latest dynamic risk index with the preset grading threshold and output the current risk level of the target monitoring area.

Images