Coal mine gas prevention and control extraction effect dynamic evaluation and closed-loop management and control method

Abstract

The present application relates to the field of coal mine safety engineering, and provides a coal mine gas prevention and control extraction effect dynamic evaluation and closed-loop management and control method, comprising: S1, based on the basic threshold in the extraction design scheme, periodically dynamically calibrating the threshold through a multiple linear regression model to generate a dynamic threshold matrix of core indexes; S2, collecting multi-dimensional data to construct a fusion data matrix; S3, constructing a dynamic multi-scale convolutional neural network model and training it, the model containing a variable-length filter generator for adaptively generating a variable-length filter according to input data to extract multi-scale time series features; and outputting a compliance state judgment result of the current extraction effect and a core index prediction value of a future preset time period in parallel; S4, according to the compliance state judgment result and the prediction value output by S3, starting a three-level linkage feedback management and control process coordinated by an AI model and an expert rule base. The present application realizes accurate evaluation and short-term prediction of core indexes such as extraction flow, gas concentration and attenuation coefficient.

Description

Technical Field

[0001] This invention relates to the field of coal mine safety engineering technology, specifically to a method for dynamic evaluation and closed-loop management of coal mine gas control and extraction effects. Background Technology

[0002] Coal mine gas drainage is a core component of ensuring safe underground production. Real-time evaluation and closed-loop management of drainage effectiveness directly impact gas control efficiency, construction costs, and operational safety. Current technologies for controlling gas drainage effectiveness have the following significant shortcomings:

[0003] 1. Poor adaptability of evaluation models: Traditional evaluations often rely on fixed threshold comparisons or simple time series analysis, which makes it difficult to capture the complex time series characteristics of gas extraction, which are characterized by "slow decay + sudden fluctuations". Furthermore, they cannot adapt to the dynamic changes in different extraction stages (initial, stable, and final stages), resulting in low evaluation accuracy and a high misjudgment rate.

[0004] 2. Insufficient feature extraction capability: Existing convolutional neural networks (CNNs) use fixed-length filters, which cannot adaptively mine multi-scale key information (such as short-term flow fluctuations and long-term decay trends) in extracted time-series data, and have insufficient sensitivity to identify abnormal operating conditions (sudden drop in flow, concentration rebound).

[0005] 3. Lack of coordination in control: There is a lack of effective coordination between data-driven models and expert experience. Either the reliance on pure rule bases leads to insufficient adaptability, or the reliance on "black box" models leads to a lack of interpretability in control instructions. Furthermore, the division of labor and cooperation mechanism between models and rule bases is not clear, making it difficult to achieve a combination of accurate diagnosis and efficient control.

[0006] 4. Incomplete closed-loop management: Existing solutions mostly remain at the level of single early warning or partial adjustment, failing to form a complete closed-loop process of "dynamic target setting → multi-dimensional data fusion → intelligent evaluation and prediction → multi-level feedback management → effect verification and iteration". Furthermore, equipment adjustments lack safety verification and manual confirmation mechanisms, posing safety risks.

[0007] Although preliminary applications of time-series data evaluation models and simple closed-loop control have emerged in existing technologies, they have not designed adaptive feature extraction mechanisms for the multi-scale characteristics of gas extraction time-series data, nor have they constructed a collaborative management and control system of models and rule bases. As a result, they cannot effectively solve problems such as inaccurate evaluation of extraction effects, delayed management and control response, and high safety risks under complex working conditions. There is an urgent need to propose an intelligent management and control solution that combines high adaptability, strong feature extraction capabilities, and safety and reliability. Summary of the Invention

[0008] To address the shortcomings of existing technologies, this invention proposes a dynamic evaluation and closed-loop management method for the gas control and drainage effect in coal mines. It proposes a closed-loop management scheme for the gas control and drainage effect covering all time and space and the entire process. By integrating dynamic multi-scale temporal feature extraction, intelligent dynamic evaluation and multi-level closed-loop feedback mechanism, it takes into account adaptability, safety and practicality.

[0009] To achieve the above objectives, the following technical solution is adopted: A method for dynamic evaluation and closed-loop management of coal mine gas control and drainage effects, comprising the following steps: S1, based on the basic thresholds in the drainage design scheme and combined with coal seam geological parameters, the thresholds are periodically and dynamically calibrated using a multiple linear regression model to generate a dynamic threshold matrix containing core indicators such as drainage flow rate, gas concentration, and attenuation coefficient, which serves as the evaluation benchmark; S2, multi-dimensional data including drainage monitoring data, geological parameters, and construction parameters are collected, cleaned, aligned, and fused to construct a spatiotemporally aligned fused data matrix; S3, a dynamic multi-scale convolutional neural network (DMS-CNN) model is constructed and trained. The Dynamic Multi-Scale Convolutional Neural Network (DMS-CNN) model includes a variable-length filter generator, which adaptively generates variable-length filters based on the input data to extract multi-scale temporal features. The spatiotemporally aligned fused data matrix is ​​input into the trained Dynamic Multi-Scale Convolutional Neural Network (DMS-CNN) model, which outputs in parallel the current sampling effect's compliance status judgment result and the predicted value of core indicators for a future preset period. S4: Based on the compliance status judgment result and predicted value output in step S3, and based on preset risk judgment rules, different levels of early warning are triggered. Based on the early warning level, a three-level linkage feedback control process coordinated by the AI ​​model and expert rule base is initiated.

[0010] Furthermore, in step S1, dynamic thresholds adapted to specific geological conditions are generated for various core indicators using the following formula:

[0011]

[0012] The dynamic calibration threshold of the nth core indicator among the core indicators of extraction flow rate, gas concentration and attenuation coefficient; The basic threshold weight of the nth type of indicator; The basic threshold for the nth type of indicator; The geological parameter influence coefficient of the nth type of index; Key geological parameters of coal seams related to the nth type of index, including gas content, coal seam permeability, coal seam thickness, and fault density. : Correction for threshold calibration.

[0013] Furthermore, in step S2, the construction of the spatiotemporally aligned fused data matrix specifically includes: using the timestamps of the monitoring data as a benchmark, globally associating the static geological model parameters with all timestamps, and interpolating the dynamic construction parameters to the same time interval as the monitoring data to achieve time alignment; establishing a mapping relationship based on the spatial coordinates of the extraction boreholes and the grid division of the geological model, and associating the geological parameters of the corresponding grids with the monitoring data of the boreholes to achieve spatial alignment; and fusing the aligned monitoring data, geological parameters, and construction parameters to construct a two-dimensional data matrix containing multiple types of key parameters.

[0014] Furthermore, in step S3, the Dynamic Multi-Scale Convolutional Neural Network (DMS-CNN) model includes at least an input layer, a variable-length filter generator, a variable-length filter convolutional layer, a temporal max-pooling layer, and an output layer connected in sequence; wherein, the variable-length filter generator is used to dynamically generate a set of filters with variable lengths based on the input temporal data; the output layer contains two parallel branches, the first branch is used to output the judgment result of the current sampling effect meeting the target status, and the second branch is used to output the continuous prediction result of the core indicators within a preset future time period.

[0015] Furthermore, in step S3, the construction process of the variable-length filter generator includes: sliding window slicing: slicing the input time-series data with a window length l to obtain a sub-sequence matrix S; embedding representation generation: convolving the sub-sequence matrix S with K fixed-length base filters to obtain the embedding representation corresponding to each base filter. This forms the embedding matrix E; mask generation: each embedding representation in the embedding matrix E is... The corresponding basic filter parameters are fused with features and processed by a neural network to generate a soft mask matrix M; where each soft mask is used to control the effective range of the weights of the corresponding basic filter in the time dimension; the soft mask matrix M is multiplied by the basic filter matrix to generate a variable-length filter that can adaptively extract features at different time scales.

[0016] Furthermore, in the variable-length filter generator, the length of the sliding window is dynamically adjusted proportionally according to the length L of the input sequence of the model configured in the current sampling stage; the soft mask matrix is ​​used to control the effective length of each variable-length filter, so that it presents a smooth weight transition in the time dimension.

[0017] Furthermore, the input layer of the dynamic multi-scale convolutional neural network model dynamically configures its input sequence length L according to the current sampling stage to adapt to the different needs of extracting long-term trends or recent change features at different stages; wherein, it is configured with a first preset duration in the initial sampling stage, a second preset duration less than the first preset duration in the stable sampling stage, and a third preset duration less than the second preset duration in the final sampling stage.

[0018] Furthermore, the neural network is a fully connected network, which calculates an intermediate value for calibrating the effective length of the filter based on the vector after feature fusion; each element value in the soft mask matrix is ​​generated by comparing the effective length obtained by converting the intermediate value with the position index parameter j, and then mapping it via the Sigmoid function.

[0019] Furthermore, the dynamic multi-scale convolutional neural network model is trained through the following steps: A training dataset is prepared, comprising real mining time-series data from multiple coal mines and virtual data covering abnormal working conditions generated through sampling simulation, and the data is expanded to increase data diversity; a hybrid loss function is used to train the model, the hybrid loss function being a weighted sum of a classification loss function and a regression loss function; wherein the classification loss function uses focus loss to handle class imbalance in the mining effect status classification task, and the regression loss function uses Hubel loss to handle the core indicator regression prediction task and is robust to predicting outliers; during training, training is terminated when the model's loss value on the validation set meets a preset convergence condition or reaches a preset early stopping condition.

[0020] Furthermore, in step S4, the three-level linkage feedback control process, which is coordinated by the AI ​​model and the expert rule base, specifically includes: First-level feedback: if the warning level is a low-risk warning, automatic warning notification and manual response tracking are executed; Second-level feedback: if the warning level is a medium-risk warning or the first-level feedback is not responded to in a timely manner, equipment adjustment instructions based on the expert rule base are generated, safety verification is performed, and manual confirmation is conducted, and the effect of the adjustment is verified using the dynamic multi-scale convolutional neural network model; Third-level feedback: if the warning level is a high-risk warning, the second-level feedback adjustment is ineffective, or the rule base cannot match the instructions, intelligent generation of supplementary measures is performed, expert review and decision-making is conducted, and the solution is iterated; wherein, the AI ​​model is used for problem diagnosis and risk prediction, the expert rule base is used to generate specific equipment control instructions in the second-level feedback, and the expert review and decision-making is used to review and determine a systematic supplementary governance solution in the third-level feedback.

[0021] Compared with the prior art, the present invention achieves the following beneficial effects:

[0022] 1. High evaluation and prediction accuracy: To address the shortcomings of traditional models in adapting to the complex features of extraction time-series data, this invention proposes a Dynamic Multi-Scale Convolutional Neural Network (DMS-CNN) model. This model constructs a dual-function model integrating "real-time achievement judgment + future situation prediction". By adaptively extracting multi-scale time-series features through a variable-length filter generator and combining dynamic input sequence length configuration, it adapts to the feature extraction requirements of different extraction stages, achieving accurate evaluation and short-term prediction of core indicators such as extraction flow rate, gas concentration, and attenuation coefficient. This results in an extraction status classification accuracy of ≥92% and a prediction deviation of core indicators ≤5% for the next 120 minutes, significantly improving the evaluation accuracy compared to traditional models and effectively reducing the false positive and false negative rates.

[0023] 2. Strong Feature Extraction Adaptability: To address the problem that fixed-length filters struggle to capture multi-scale temporal features, a variable-length filter generation mechanism based on soft masks is designed. The variable-length filter generator can dynamically adjust the filter length according to the input data, capturing both short-term sudden fluctuations (such as sudden drops in traffic caused by equipment failure) and long-term decay trends. Through sliding window slicing, embedded representation generation, and dynamic mask calibration, a variable-length filter adapted to the input data is automatically generated, accurately mining key features at different time scales (such as short-term abrupt changes and long-term decay), improving feature extraction capabilities under complex conditions, adapting to different sampling stages and abnormal conditions, and improving feature extraction capabilities by 40% compared to the fixed-length filter model, with significantly enhanced sensitivity to extreme scenarios.

[0024] 3. Highly efficient and secure collaborative management: The collaborative mechanism of "AI model diagnosis + rule base instructions" leverages the precise analytical advantages of data-driven models while retaining the security and reliability of expert rules. The accuracy of secondary feedback instructions is high, and security verification and a 120-second manual confirmation window eliminate security risks caused by model misjudgment. The response time for equipment adjustment is shortened to within 2 minutes, solving the pain points of poor interpretability of pure models and insufficient adaptability of pure rules, and achieving an organic combination of accurate diagnosis and secure management.

[0025] In summary, this invention constructs a closed-loop control mechanism for the entire process, forming a complete closed loop of dynamic target setting → multi-dimensional data fusion preprocessing → intelligent evaluation and prediction → three-level linkage feedback control → effect verification and iteration. It supplements the safety verification and manual confirmation links for equipment adjustment, ensuring the safety, effectiveness and continuity of control, and replacing the traditional fragmented control mode.

[0026] It should be understood that the description in the Summary of the Invention is not intended to limit the key or essential features of the embodiments of the present invention, nor is it intended to restrict the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0027] The above and other features, advantages, and aspects of the various embodiments of the present invention will become more apparent from the accompanying drawings and the following detailed description. The drawings are provided for a better understanding of the invention and are not intended to limit the invention. In the drawings, the same or similar reference numerals denote the same or similar elements, wherein:

[0028] Figure 1 This is a flowchart illustrating a method for dynamic evaluation and closed-loop management of coal mine gas control and extraction effectiveness according to an embodiment of the present invention.

[0029] Figure 2 This is a schematic diagram of the structure of the Dynamic Multi-Scale Convolutional Neural Network (DMS-CNN) model in an embodiment of the present invention;

[0030] Figure 3 This is a schematic diagram of the workflow of the variable-length filter generator in an embodiment of the present invention;

[0031] Figure 4 This is a schematic diagram of the three-level linkage feedback control process in an embodiment of the present invention;

[0032] Figure 5 This is a comparison curve showing the change in prediction error (RMSE) of the DMS-CNN model of this invention with RNN, LSTM, GRU, and TCN models as a function of training epochs. Detailed Implementation

[0033] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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.

[0034] Furthermore, the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0035] Figure 1 This is a flowchart illustrating a method for dynamic evaluation and closed-loop management of coal mine gas control and extraction effectiveness according to an embodiment of the present invention. Figure 1 As shown, the present invention provides a method for dynamic evaluation and closed-loop management of coal mine gas control and extraction effects, comprising the following steps:

[0036] S1. Based on the basic threshold in the extraction design scheme and combined with the geological parameters of the coal seam, the threshold is periodically and dynamically calibrated through a multiple linear regression model to generate a dynamic threshold matrix containing core indicators such as extraction flow rate, gas concentration and attenuation coefficient, which serves as the evaluation benchmark.

[0037] Step S1: To achieve the expected target setting, the system interfaces with the intelligent design scheme for gas extraction, constructing a structured target matrix based on the "basic threshold + dynamic calibration" mode to avoid the limitations of fixed thresholds.

[0038] S1.1: Core Indicator System

[0039] Extraction flow rate indicators: Basic expected extraction flow rate ( Dynamic fluctuation threshold (initial stage) Stable phase );

[0040] Gas concentration indicators: Gas concentration decrease curve (unit: %), stage target value (initial stage) Stable phase Late stage );

[0041] Attenuation-related indicators: Basic expected attenuation coefficient ( ), minimum decay rate threshold ( Where d represents the time unit "day"; 0.05 This indicates that the average daily decrease in extraction volume or concentration is approximately 5% of the remaining amount. A higher value indicates faster decay and a more significant extraction effect; conversely, a lower value indicates slower decay, requiring adjustment of the extraction strategy. The minimum decay rate threshold set in this embodiment of the invention (…) The attenuation rate (AFCR) is a critical control point for engineering safety and techno-economic efficiency. It requires that the actual attenuation rate of the extraction system not fall below this value to determine the continued effectiveness of the extraction. If the measured AFCR is lower than this threshold, an early warning is triggered, indicating insufficient extraction efficiency and a risk of gas accumulation, requiring intervention through a closed-loop management process.

[0042] S1.2: Dynamic initialization logic

[0043] For various indicators such as extraction flow rate, gas concentration, and attenuation coefficient, the basic thresholds are derived from industry safety regulations and theoretical values ​​from design schemes. An adaptation model is constructed by combining coal seam geological parameters (gas content, permeability, coal seam thickness, fault density), and the thresholds are periodically and dynamically calibrated using a multiple linear regression model.

[0044]

[0045] The dynamically calibrated threshold of the nth core indicator (extraction flow rate, gas concentration, and attenuation) is the benchmark value for subsequent extraction effect evaluation; n=1,2,3, corresponding to the core indicators of extraction flow rate, gas concentration, and attenuation respectively. The basic threshold weight of the nth type of indicator, with a range of values. This is used to balance the core function of the basic threshold and adapt and adjust geological parameters; The basic threshold of the nth type of index is derived from the theoretical target value in the safety regulations and drainage design scheme of the coal mine gas drainage industry. The geological parameter influence coefficient of the nth type of index is obtained by training the correlation between historical sampling data and corresponding geological conditions, and is used to quantify the degree of influence of geological factors on the threshold. : Coal seam geological parameters related to the nth type of index, including key geological characteristic parameters such as gas content, coal seam permeability, coal seam thickness, and fault density; : A correction term for threshold calibration, used to offset calibration deviations caused by model simplification or data noise, thereby improving threshold accuracy.

[0046] Threshold after dynamic calibration of the nth type of core indicator A dynamic threshold matrix containing core indicators such as extraction flow rate, gas concentration, and attenuation coefficient is generated as an evaluation benchmark. Furthermore, the thresholds are recalibrated every 24 hours during the extraction process to adapt to changes in operating conditions.

[0047] S2. Collect multi-dimensional data including extraction monitoring data, geological parameters and construction parameters, and clean, align and fuse them to construct a spatiotemporally aligned fused data matrix;

[0048] Step S2 is used to achieve multi-dimensional data acquisition and preprocessing. Specifically, it includes the following steps:

[0049] S2.1 Monitoring Deployment:

[0050] High-precision flow meters and gas concentration sensors are deployed at key nodes of the extraction pipeline (drill outlet, main pipe, and branch pipe), and pressure sensors and frequency sensors are deployed at the extraction pump station. The monitoring frequency is set to 1 time / minute (to take into account both high-frequency capture of sudden changes and data redundancy control).

[0051] S2.2: Data Cleaning

[0052] S2.2.1: Outlier Removal

[0053] use Criteria: Calculate the mean of each indicator. with standard deviation It will exceed Data within a certain range are marked as outliers and replaced with the median of the five consecutive valid data points for that metric.

[0054] The mean value of a certain extraction monitoring indicator (such as extraction flow rate or gas concentration) is calculated by statistically analyzing the historical valid monitoring data of that indicator. The standard deviation of a certain sampling monitoring indicator is used to reflect the dispersion of the monitoring data for that indicator. The core parameters for identifying outliers; This refers to the normal data range for a certain sampling and monitoring indicator. Data exceeding this range are considered outliers and need to be replaced.

[0055] S2.2.2: Missing value completion

[0056] The piecewise interpolation method based on the trend of adjacent time periods is adopted, and the formula is as follows:

[0057]

[0058] : The sampling monitoring index value at time t after missing values ​​are filled in, ensuring the integrity of the data sequence; t: The missing time of the sampling monitoring data, that is, the specific time point when no valid data was collected; k: The continuous missing duration of the monitoring data, in minutes, which is the same as the data sampling interval set by the system. For example, if the data sampling interval is 1 minute, the missing duration k=5 means that 5 sampling points are missing consecutively, that is, 5 minutes of data are actually missing. The most recent effective monitoring indicator value before the missing segment is one of the basic data for interpolation calculation; The most recent effective monitoring indicator value after the missing segment, and Together, we construct the trend characteristics of the missing time period.

[0059] If there is insufficient valid data before and after the missing segment ( If the sample size is less than 1, the historical average of the sampling phase will be used to fill the data to ensure data integrity.

[0060] S2.2.3: Standardization Processing

[0061] Z-Score standardization is performed on the cleaned data to generate a standardized measured data sequence (timestamp + index value).

[0062] Z-Score standardization: A data preprocessing method that converts sampled monitoring data into standardized data with a mean of 0 and a standard deviation of 1, eliminating differences in the dimensions of different indicators and providing a data foundation with a unified scale for subsequent model input.

[0063] S2.3: Data Fusion

[0064] Step S2.3 involves constructing a spatiotemporally aligned fused data matrix, including: using the timestamps of the monitoring data as a baseline, globally associating static geological model parameters with all timestamps, and interpolating dynamic construction parameters to the same time interval as the monitoring data to achieve time alignment; establishing a mapping relationship based on the spatial coordinates of the extraction boreholes and the grid division of the geological model, associating the geological parameters of the corresponding grid with the monitoring data of the boreholes to achieve spatial alignment; and fusing the aligned monitoring data, geological parameters, and construction parameters to construct a two-dimensional data matrix containing multiple types of key parameters. Specifically, this includes the following steps:

[0065] S2.3.1: Spatiotemporal Alignment Logic

[0066] Based on the timestamps of the monitoring data, accurate alignment of the three types of data is achieved:

[0067] Time alignment: 3D geological model parameters (static parameters, such as coal seam thickness and fault density) are bound to all timestamps in a "global association" manner; construction parameters (dynamic parameters, such as fracturing pressure and drilling progress) are interpolated to 1-minute intervals according to the time nodes of the construction log and matched with the monitoring data timestamps;

[0068] Spatial alignment: Based on the spatial coordinates of the extraction borehole and the grid division of the three-dimensional geological model, a "borehole-grid" mapping relationship is established, and the geological parameters of the corresponding grid (such as gas content and permeability within the grid) are associated with the monitoring data of the borehole.

[0069] S2.3.2: Matrix Construction

[0070] After alignment, a fused data matrix with dimensions T×Q is formed, where T is the monitoring duration and Q (e.g., 15) is the total number of key parameters. Key parameters mainly include, but are not limited to: 6 monitoring indicators (such as extraction flow rate, gas concentration, extraction pressure, pump station frequency, concentration change rate, and flow fluctuation value), 5 geological parameters (such as gas content, permeability, coal seam thickness, fault density, and burial depth), and 4 construction parameters (such as borehole diameter, fracturing pressure, sealing length, and extraction time). Those skilled in the art can adjust the specific types and number of parameters Q according to actual conditions.

[0071] S2.4: Sequence Segmentation

[0072] Data segments are divided according to the sampling stage (initial stage, stable stage, and final stage), and the model feature extraction logic is adapted accordingly.

[0073] S3. Construct and train a Dynamic Multi-Scale Convolutional Neural Network (DMS-CNN) model. The DMS-CNN model includes a variable-length filter generator, which is used to adaptively generate variable-length filters according to the input data to extract multi-scale temporal features. Input the spatiotemporally aligned fused data matrix into the trained DMS-CNN model and output the current sampling effect's compliance status judgment result and the core indicator prediction value for the future preset period in parallel.

[0074] Step S3 is used to construct a dynamic evaluation model. This invention proposes a Dynamic Multi-Scale Convolutional Neural Network (DMS-CNN) as the dynamic evaluation model, which is used to optimize and adapt to the characteristics of gas extraction time series data. The structure is optimized for the characteristics of "slow decay + sudden fluctuation" to achieve the dual functions of "real-time compliance judgment + future situation prediction".

[0075] S3.1: Design Model Structure and Key Parameters

[0076] like Figure 2 The diagram shown is a structural schematic of the Dynamic Multi-Scale Convolutional Neural Network (DMS-CNN) model in an embodiment of the present invention, illustrating the connection relationships between the input layer, variable-length filter generator, convolutional layer, pooling layer, and dual-branch output layer.

[0077] S3.1.1: Constructing the Input Layer

[0078] The input layer receives a subset of time-series features from the fused data matrix, including 6 core time-series indicators (extraction flow rate, gas concentration, extraction pressure, pump station frequency, concentration change rate, and flow fluctuation value); the length L of the input sequence is dynamically configured: L=1440 (24 hours) in the initial stage, L=720 (12 hours) in the stable stage, and L=360 (6 hours) in the final stage, taking into account both trend capture and calculation efficiency.

[0079] L: Input sequence length, dynamically configured according to the current sampling stage to adapt to different needs for extracting long-term trends or recent changes in different stages. It is configured as a first preset duration in the initial sampling stage (e.g., 1440, or 24 hours in the initial stage), a second preset duration less than the first preset duration in the stable sampling stage (e.g., 720, or 12 hours in the stable stage), and a third preset duration less than the second preset duration in the final sampling stage (e.g., 360, or 6 hours in the final stage), balancing long-term trend capture and model calculation efficiency.

[0080] S3.1.2: Constructing a Variable Length Filter Generator

[0081] like Figure 3 The diagram shown is a schematic of the workflow of the variable-length filter generator in an embodiment of the present invention, which specifically includes the following steps:

[0082] (1) Sliding window slicing: The input time series data is sliced ​​using an adaptive window length l=0.2L (dynamically adjusted with the length of the input sequence) with a step size of 1. P=L-l+1 time series subsequences are extracted and concatenated to obtain the subsequence matrix. ;

[0083] l: Adaptive window length, dynamically adjusted proportionally according to the input sequence length L of the model configured in the current extraction stage, with a value of 0.2L, dynamically adjusted with the input sequence length L, used to extract subsequences from the input time series data; P: Number of time series subsequences, calculated by the formula P=L-l+1, which is the total number of subsequences obtained after sliding window slicing; The concatenated temporal subsequence matrix, with dimensions P (number of subsequences) × l (window length), serves as the input to the variable-length filter generator;

[0084] (2) Embedding representation generation: The subsequence matrix S is convolved with K=120 basic filters (or fixed filters) of length ω=5 to obtain the embedding representation corresponding to each basic filter. This forms an embedding matrix. ;

[0085] K: The number of fixed filters, with a value of 120, is used to perform convolution operations on the temporal subsequence matrix S to generate embedded representations; ω: The length of the base filter, with a value of 5, is the core structural parameter of the convolution operation, used to extract local features of the subsequence. The embedding representation matrix has dimensions of 120 (fixed number of filters) × l (window length) and is obtained by convolving S with a fixed number of filters.

[0086] (3) Mask generation: embedding matrix Each embedding representation in Feature fusion is performed on the corresponding basic filter parameters, and then processed by a neural network to generate a soft mask. Each soft mask The effective range of the weights of the corresponding basic filter in the time dimension is controlled by a soft mask. A soft mask matrix M is constructed, which is used to control the effective operating length of each variable-length filter, so that it presents a smooth weight transition in the time dimension.

[0087] Furthermore, the neural network is a fully connected network, which is based on the vector after feature fusion. The intermediate value used to calibrate the effective length of the filter was calculated. By using the intermediate value The effective length obtained from the transformation is compared with the position index parameter j, and then mapped by the Sigmoid function to generate each element value in the soft mask matrix M. Specifically, a two-layer fully connected network is used to generate the soft mask, using the following formula:

[0088]

[0089] The intermediate value corresponding to the i-th fixed filter is obtained by processing the concatenation result of the embedding vector and the fixed filter through two fully connected networks, and is used for subsequent mask length calibration. The weight matrix of the first fully connected layer, with dimension 1. It is used to perform a linear transformation on the concatenated features of the embedded vector and the fixed filter; The weight matrix of the second fully connected network has dimensions of [dimension number missing]. This is used to perform a linear transformation on the output of the first fully connected layer to generate intermediate calculated values. ReLU: Rectified Linear Unit Activation Function, used to introduce nonlinear features into fully connected networks to enhance the model's feature representation capabilities; : The i-th row vector of the embedding matrix E, that is, the embedding representation corresponding to the i-th fixed filter; The i-th randomly initialized fixed-length filter is used to filter the embedding vector. After being concatenated, it participates in the mask generation calculation; : Vector concatenation operator, used to concatenate embedded vectors With fixed filter Concatenate them into a feature vector of uniform dimension; The bias term of the first fully connected layer is used to adjust the linear transformation result and improve the model's fitting ability. The bias term in the second-layer fully connected network has the same effect as... Consistency, optimization of intermediate calculation values The generation accuracy; The effective length of the i-th variable-length filter is calculated from an intermediate value. The actual effective length of the corresponding filter is determined by multiplying it by the adaptive window length l. The element in the i-th row and j-th column of the soft mask matrix M is used to control the fixed filter. The effective weight range; Sigmoid: S-shaped activation function, which maps the input value to the [0,1] interval, so that the soft mask elements present a gradient transition of "effective (close to 1) - ineffective (close to 0)"; λ: the sharpness parameter of the soft mask, with a value of 8, used to balance the smooth transition characteristics of the soft mask with the discriminative power of the effective region; j: the column index of the soft mask matrix, corresponding to the length dimension of the fixed filter, used to locate the specific position of the filter.

[0090] (4) Synthesizing a variable-length filter: The soft mask matrix M is multiplied by the basic filter matrix to generate a variable-length filter that can adaptively extract features at different time scales. .

[0091] The generated variable-length filter consists of a fixed filter matrix. Element-wise product operation with the soft mask matrix M We obtained a total of 120 features, which can adaptively adapt to the input data at multiple scales. : Initially randomly initialized fixed-length filter matrix, which forms the basic structure of variable-length filters; M: Soft mask matrix, with dimensions similar to the fixed filter matrix. Consistent, the effective length of the fixed filter is controlled by element-wise product operations, forming a variable length characteristic.

[0092] S3.1.3: Constructing a Variable-Length Filter Convolutional Layer

[0093] Variable-length filter convolutional layers use the generated variable-length filter to fill the zero-padded input sequence. (Length L+l) convolutions yield a feature map matrix after 120 convolutions. The dimension is 120 (number of variable-length filters) × (Length of the feature map after convolution) (where D is the effective length of the variable-length filter), and contains multi-scale temporal features.

[0094] The zero-padding input sequence has a length of L+l, which is used to avoid the loss of temporal data boundary information during convolution operations;

[0095] S3.1.4: Constructing a Temporal Max Pooling Layer

[0096] The temporal max-pooling layer performs max-over-time pooling on each feature map, selects the most discriminative local patterns, and outputs feature vectors. .

[0097] max-over-time pooling: Temporal max pooling operation, which takes the maximum value of each feature map along the time dimension, selects the most discriminative local patterns, and realizes dimensionality reduction of temporal features and enhancement of key information;

[0098] The feature vector after temporal max pooling, with a dimension of 120 (consistent with the number of variable-length filters), serves as the input feature of the model's output layer.

[0099] S3.1.5: Constructing the output layer

[0100] The output layer contains two parallel fully connected branches:

[0101] (1) Standard-reaching judgment branch: 3 neurons (standard-reaching / warning / exceeding limits), Softmax activation, outputting the current risk state classification probability;

[0102] (2) Situation prediction branch: 3 neurons (predicted values ​​of extraction flow rate, gas concentration and attenuation coefficient in the next 120 minutes), linearly activated, outputting continuous prediction results.

[0103] It should be noted that the three status categories of "compliant", "warning", and "exceeding limits" output by the compliance judgment branch constitute the basis for risk level judgment in the subsequent early warning and feedback control process.

[0104] S3.2: Model Training Steps

[0105] S3.2.1: Training Data Generation

[0106] Historical monitoring data: 3,000 sets of drainage engineering monitoring data from 80 coal mines were collected, covering different geological conditions (high gas / low gas, high permeability / low permeability) and drainage technology. Each set includes complete time-series data, compliance status labels, and expert evaluation results.

[0107] Simulated data generation: Based on Latin hypercube sampling, 2000 sets of virtual data covering abnormal operating conditions (sudden drop in flow, concentration rebound, equipment failure) are generated. The KL divergence test (KL<0.1) is used to ensure consistency with the distribution of real data and to supplement extreme scenario samples.

[0108] Data augmentation: Augmentation operations such as time flipping, amplitude scaling, and Gaussian noise addition are performed on the time-series data, ultimately resulting in 8000 valid training samples. Each training sample contains a time-series fused data matrix within a fixed-length window (its time dimension length is denoted as...). The feature dimension is N), the sampling effect status label corresponding to the window (meeting the standard, warning, exceeding the limit), and the real values ​​of the core indicators in subsequent time periods are used for regression supervision. The dataset is divided into training, validation, and test sets in a 7:2:1 ratio.

[0109] Data partitioning: The dataset was divided into a training set (5600 sets), a validation set (1600 sets), and a test set (800 sets) in a 7:2:1 ratio, using 5-fold cross-validation.

[0110] S3.2.2: Model Training Process

[0111] 1. Initialization: Fixed-length filter W t Xavier uniform initialization is employed to ensure that the initialization parameters follow a uniform distribution, guaranteeing stable gradient propagation in the early stages of model training. The mask generator parameters are initialized using He. Xavier uniform initialization is a parameter initialization method that controls the range of initialization parameter values ​​to make the input and output variances of each network layer as consistent as possible, avoiding gradient vanishing or exploding. It is suitable for fixed-length filters W. t Initialization. He initialization is a parameter initialization method designed for ReLU-type activation functions. The mask generator parameters are initialized in this way, which can adapt to the nonlinear characteristics of the activation function and improve the model convergence speed.

[0112] 2. Optimizer: AdamW optimizer, initial learning rate 0.001, using cosine annealing strategy ( max The AdamW optimizer is a weight decay mechanism based on the Adam optimizer, with an initial learning rate of 0.001. It is a core optimization algorithm for training DMS-CNN models, improving upon the Adam optimizer's weight decay mechanism to effectively alleviate overfitting. The AdamW optimizer's initial learning rate is 0.001, a key hyperparameter controlling the step size of model parameter updates and determining the magnitude of parameter adjustments in the early stages of training. The cosine annealing strategy is a learning rate adjustment strategy that periodically adjusts the learning rate using a cosine function. The period parameter of the cosine annealing strategy... The 'round' indicates that the cosine period is 50 rounds of training, meaning that the learning rate is adjusted once every 50 rounds of training. This allows the learning rate to decrease smoothly during training, balancing rapid convergence in the early stages with fine-grained optimization in the later stages. The weight decay coefficient is 0.0001, which serves as the regularization parameter for the AdamW optimizer. By applying L2 penalty to the model parameters, it reduces the risk of overfitting and improves generalization ability.

[0113] 3. Loss Function: The hybrid loss function is a weighted sum of the classification loss function and the regression loss function. The classification loss function uses Focal Loss to address the imbalance in the number of samples in the "meeting standards," "warning," and "exceeding limits" categories in the sampling effect classification task. By reducing the weight of easily classified samples, it improves the model's ability to identify a few abnormal states. The regression loss function uses Huber Loss to handle the core indicator regression prediction task. It is robust to noise and abrupt outliers that may exist in the monitoring data, and can obtain more stable prediction results. Specifically, the hybrid loss function, balancing the classification and regression tasks, is formulated as follows:

[0114]

[0115] in, : The overall loss function value of the DMS-CNN model, used to measure the degree of deviation between the model's prediction results and the true labels, and is the core objective of model parameter optimization; 0.5: Loss function weight coefficient, which assigns equal weight to the classification loss and the regression loss respectively, balancing the training objectives of "real-time achievement judgment" (classification task) and "future situation prediction" (regression task);

[0116] Classification loss The core function of Focal Loss is to alleviate class imbalance in classification tasks, making the model focus more on hard-to-classify samples (such as gas extraction exceeding limits / early warning samples), thereby improving the accuracy of determining compliance status. Its formula is:

[0117]

[0118] in, It is the model's predicted probability of the true category (corresponding to the compliance, warning, and exceeding limits states of this invention). The category weight factor is set for the three categories mentioned above to balance the category frequency. γ is an adjustable focusing parameter (usually γ≥0) used to reduce the weight of easily classified samples, so that the model pays more attention to difficult-to-classify samples.

[0119] Regression loss Huber Loss is employed for the prediction and regression of core indicators (flow rate, concentration, and decay coefficient) in gas extraction. It combines the advantages of mean squared error and absolute error, reducing the interference of outliers on model training. The Huber Loss function is robust to outliers, effectively mitigating the impact of extreme values ​​on the regression task and improving the accuracy of indicator prediction when there are abrupt changes or noise in the gas extraction time-series data. Its formula is as follows:

[0120]

[0121] in, These are the actual values ​​of the core indicators (flow rate, concentration, attenuation coefficient). Here, δ represents the model's predicted value, and δ is a threshold parameter used to control the sensitivity of the loss function to outliers. When the prediction error is less than δ, mean squared error is used; when it is greater than δ, absolute error is used, thereby reducing the impact of outliers. This hybrid loss function simultaneously optimizes both the model's classification accuracy and the robustness of its regression predictions.

[0122] 4. Convergence condition: Total loss on the validation set (Actual, feasible goals) If the loss does not decrease after 10 consecutive rounds, stop early. The maximum number of iterations is 300 rounds.

[0123] Validation set overall loss: The overall loss value of the model on the validation set. The convergence condition is set to ≤0.05 (a practical and feasible target). It is the core basis for judging whether the model training has achieved the expected accuracy. Early stopping: The termination strategy of model training. If the validation set loss does not decrease for 10 consecutive rounds, early stopping is triggered to avoid model overfitting and ensure training efficiency and model generalization ability. Maximum number of iterations: The maximum number of iterations for model training. The value is 300 rounds to prevent the model from getting stuck in meaningless iterations due to data anomalies and to control training costs.

[0124] This invention utilizes a Dynamic Multi-Scale Convolutional Neural Network (DMS-CNN) model constructed and trained in step S3. This model achieves a dual function of "real-time target assessment + future situation prediction." By dynamically configuring the input sequence length, it adapts to the feature extraction requirements of different extraction stages, enabling accurate evaluation and short-term prediction of core indicators such as extraction flow rate, gas concentration, and attenuation coefficient. This solves the problem that traditional models cannot adapt to the complex features of extraction time-series data. Furthermore, a variable-length filter generation mechanism based on soft masks is designed. A variable-length filter generator is constructed, and through sliding window slicing, embedded representation generation, and dynamic mask calibration, a variable-length filter adapted to the input data is automatically generated. This accurately mines key features at different time scales (such as short-term abrupt changes and long-term attenuation), improving the feature extraction capability under complex operating conditions and solving the problem that fixed-length filters are unable to capture multi-scale time-series features.

[0125] S4. Based on the compliance status judgment results and predicted values ​​output in step S3, trigger different levels of early warnings based on preset risk judgment rules; based on the early warning level, initiate a three-level linkage feedback control process that is coordinated by the AI ​​model and the expert rule base.

[0126] Step S4 is used to achieve closed-loop feedback and proactive control. The compliance status judgment result output in step S3 includes three categories: "Compliant," "Warning," and "Exceeding Limits," corresponding to low-risk, medium-risk, and high-risk levels in the control system, respectively. Based on the warning level, a three-level linkage feedback control process, coordinated by the AI ​​model and expert rule base, is initiated. For example... Figure 4 The diagram shown is a schematic representation of the three-level linkage feedback control process in an embodiment of the present invention, including the following steps:

[0127] S4.1: Design early warning trigger conditions

[0128] (1) Real-time target achievement trigger: Any relative deviation of any indicator exceeds the dynamic threshold of the corresponding stage, and the duration is ≥10 minutes (to avoid false triggering).

[0129] (2) Predicting risk triggers: The DMS-CNN model predicts that an indicator will exceed the limit or the risk level will be medium or above within the next 120 minutes;

[0130] (3) Comprehensive trigger: The product of prediction bias and geological parameter sensitivity is ≥0.15 (sensitivity is obtained by training with historical data).

[0131] S4.2: Constructing a three-level linkage feedback control system

[0132] S4.2.1: Construct a first-level feedback mechanism (instant early warning)

[0133] Triggering condition: Low risk level;

[0134] Execution logic: The system automatically pushes early warning information to the intelligent management platform, including deviation indicators, current status, risk analysis, and preliminary handling suggestions, and simultaneously notifies the responsible person (e.g., via SMS / APP) Figure 4 As shown in the figure, a response and feedback on the handling are required within 30 minutes.

[0135] S4.2.2: Construct a two-level feedback mechanism (equipment adjustment, safety first).

[0136] Triggering conditions: Medium risk level, or failure to respond to Level 1 feedback in a timely manner;

[0137] Execution logic:

[0138] (1) Rule base support: The rule base is built based on the experience of no less than 50 senior experts, contains no less than 120 device adjustment scenario rules, and is updated once a quarter in combination with actual operation data;

[0139] (2) Instruction generation: Automatically generate equipment adjustment instructions based on the rule base (such as increasing the negative pressure of extraction by 5%-8%, increasing the pump station frequency by 3Hz-5Hz, and adjusting the opening of branch pipe valves by ±10%).

[0140] (3) Matching verification: If no valid adjustment instruction is matched in the rule base (such as the occurrence of new deviation patterns or extreme working conditions), the third-level feedback is triggered directly; if an instruction is matched, the safety verification process is entered.

[0141] (4) Safety verification: After the instruction is generated, it is verified by safety interlock logic (such as the negative pressure increase not exceeding 90% of the equipment's rated value and the valve adjustment rate not exceeding 0.5% / s). After the verification is passed, it enters the execution process.

[0142] (5) Manual confirmation window: A 120-second manual confirmation window is set up, in which the person in charge can reject or adjust the instruction; if no action is taken within the time limit, the instruction will be executed automatically.

[0143] (6) Real-time monitoring: During the execution of the command, the equipment operating parameters are collected once per second. If any abnormality occurs (such as pressure exceeding the limit or abnormal noise from the equipment), the execution will be stopped immediately and a first-level warning will be triggered.

[0144] Closed-loop verification: After adjustment, continuous monitoring for 60 minutes, the DMS-CNN model in step S3 dynamically evaluates the recovery of the indicators. If the indicators return to within the threshold, the feedback is terminated; otherwise, a third-level feedback is triggered.

[0145] S4.2.3: Construct a three-level feedback system (solution iteration, decision support)

[0146] Triggering conditions: High risk level, or ineffective adjustment of secondary feedback, or no valid instruction matched in the rule base of secondary feedback;

[0147] Execution logic:

[0148] (1) Data Packaging: The system automatically packages the current monitoring data, deviation analysis results, prediction trends, geological parameters, and records of implemented measures to generate a scheme iteration request package;

[0149] (2) Assisted recommendation: Send a request to the intelligent decision-making system for gas extraction, generate 2-3 supplementary measures based on the feedback data (such as hole spacing optimization, supplementary fracturing parameters, and suggestions for new borehole locations), and mark the expected effects, construction difficulty, and cost estimate of each measure.

[0150] (3) Expert review: The plan is sent to the expert review platform, where experts in geology and mining will evaluate and confirm it, and the optimal plan can be selected or modified and adjusted.

[0151] (4) Implementation of the plan: After the plan is approved, the control system will issue construction instructions, update the expected target matrix in time, and start a new round of dynamic evaluation.

[0152] Special handling: If experts determine that existing technical means cannot solve the problem, the system will automatically trigger the highest level of warning and push it to the coal mine safety management leading group.

[0153] S4.3: Verification of Control Effectiveness

[0154] After each feedback execution, the DMS-CNN model in step S3 is used for continuous monitoring for 4 hours to calculate the evaluation results in three dimensions: the compliance rate of the controlled indicators, the deviation reduction rate, and the equipment operation stability.

[0155] (1) Compliance rate ≥ 90%, deviation reduction rate ≥ 70%, and no abnormal equipment operation: the control is deemed effective;

[0156] (2) If the above standards are not met: Re-trigger the three-level feedback and optimize the supplementary plan;

[0157] (3) Three consecutive three-level feedbacks are invalid: The system locks the construction parameter adjustment authority and forces a comprehensive investigation of geological conditions, equipment status and design scheme.

[0158] According to the embodiments of the present invention, a dynamic evaluation and closed-loop management method for coal mine gas control and drainage effects is provided. By constructing a dynamic multi-scale convolutional neural network (DMS-CNN), the model's unique variable-length filter generator can adaptively extract multi-scale temporal features exhibiting both "slow decay" and "sudden fluctuations," solving the problems of insufficient feature extraction capabilities and poor adaptability of traditional models, and achieving high-precision state evaluation and prediction. By establishing a collaborative management mechanism of "AI model diagnosis + expert rule base instructions," the division of labor is clearly defined: the model is responsible for diagnosis, and the rule base is responsible for generating safety instructions, balancing accurate analysis and reliable management. The three-level linkage feedback management loop automatically triggers progressive responses from early warning and equipment adjustment to scheme iteration based on risk levels, and verifies the effects through the model. This completely changes the traditional fragmented and delayed management mode, significantly improving the safety, efficiency, and intelligence level of gas control, thereby constructing a new paradigm for intelligent management of coal mine gas drainage effects.

[0159] like Figure 5The figure shows a performance comparison between the algorithm of this invention and different existing models. By comparing the root mean square error (RMSE, unit: %) of Dynamic Multi-Scale Convolutional Neural Network (DMS-CNN, the algorithm of this invention) with Recurrent Neural Network (RNN), Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and Temporal Convolutional Network (TCN) at different training times, it can be clearly observed that the prediction error of all algorithms shows a monotonically decreasing trend with the increase of training times. This is consistent with the basic law of deep learning models gradually improving feature learning ability through iterative optimization. However, there are significant differences in the convergence rate and final stable accuracy of each algorithm. Among them, DMS-CNN performed best, with an RMSE of 0.85% after 50 training epochs. As the number of training epochs increased to 250, the RMSE continued to decrease to 0.32%, showing not only the fastest convergence rate but also a significantly lower final error than other baseline algorithms. Temporal Convolutional Network (TCN) was second, with an RMSE of 0.38% after 250 training epochs. It demonstrated some competitiveness in capturing long temporal dependencies due to the structural advantages of causal convolution and dilated convolution, but it still lagged behind DMS-CNN's variable-length filter's adaptive extraction capability for multi-scale features. Gated Recurrent Unit (GRU) and Long Short-Term Memory (LSTM) were also mentioned. The performance of LSTM and LSTM networks is similar, with RMSEs of 0.42% and 0.48% respectively after 250 training epochs. Both alleviate the gradient vanishing problem of traditional recurrent neural networks through gating mechanisms. However, they rely on fixed-length hidden layer states to model temporal features, making it difficult to adapt to the multi-scale characteristics of gas extraction data where short-term fluctuations and long-term decays coexist. Recurrent neural networks (RNNs) perform the worst, with an RMSE as high as 0.65% after 250 training epochs. Their simple temporal information transmission mechanism results in insufficient feature capture ability for long sequence data, a slow error reduction rate, and a significantly higher final stable error than other algorithms. In summary, DMS-CNN effectively balances the extraction efficiency of short-term mutation information and long-term trend information by dynamically adapting to the feature requirements of different time scales through a variable-length filter generator. Its RMSE advantage throughout the entire training cycle fully verifies the algorithm's advancement and effectiveness in gas extraction effect evaluation and prediction tasks.

[0160] It should also be noted that, in the embodiments of this application, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0161] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined in the embodiments of this application may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown in this application, but is to be accorded the widest scope consistent with the principles and novel features disclosed in the embodiments of this application.

Claims

1. A method for dynamic evaluation and closed-loop management of gas control and extraction effectiveness in coal mines, characterized in that, Includes the following steps: S1. Based on the basic threshold in the extraction design scheme and combined with the geological parameters of the coal seam, the threshold is periodically and dynamically calibrated through a multiple linear regression model to generate a dynamic threshold matrix containing core indicators such as extraction flow rate, gas concentration and attenuation coefficient, which serves as the evaluation benchmark. S2. Collect multi-dimensional data including extraction monitoring data, geological parameters and construction parameters, and clean, align and fuse them to construct a spatiotemporally aligned fused data matrix; S3. Construct and train a dynamic multi-scale convolutional neural network model. The dynamic multi-scale convolutional neural network model includes a variable-length filter generator, which is used to adaptively generate variable-length filters according to the input data to extract multi-scale temporal features. Input the spatiotemporally aligned fusion data matrix into the trained dynamic multi-scale convolutional neural network model and output the current sampling effect's compliance status judgment result and the core indicator prediction value for the future preset period in parallel. In step S3, the dynamic multi-scale convolutional neural network model includes at least an input layer, a variable-length filter generator, a variable-length filter convolutional layer, a temporal max pooling layer, and an output layer connected in sequence. The variable-length filter generator is used to dynamically generate a set of filters with variable lengths based on the input time-series data; the output layer contains two parallel branches, the first branch is used to output the current sampling effect's compliance status judgment result, and the second branch is used to output the continuous prediction result of the core indicators within a future preset time period. The input layer of the dynamic multi-scale convolutional neural network model dynamically configures its input sequence length L according to the current sampling stage to adapt to the different needs of extracting long-term trends or recent change features at different stages; wherein, it is configured with a first preset duration in the initial sampling stage, a second preset duration less than the first preset duration in the stable sampling stage, and a third preset duration less than the second preset duration in the final sampling stage. S4. Based on the compliance status judgment results and predicted values ​​output in step S3, trigger different levels of early warnings based on preset risk judgment rules; based on the early warning level, initiate a three-level linkage feedback control process that is coordinated by the AI ​​model and the expert rule base.

2. The method for dynamic evaluation and closed-loop management of coal mine gas control and extraction effectiveness according to claim 1, characterized in that, In step S1, dynamic thresholds adapted to specific geological conditions are generated for various core indicators using the following formula: The dynamic calibration threshold of the nth core indicator among the core indicators of extraction flow rate, gas concentration and attenuation coefficient; The basic threshold weight of the nth type of indicator; The basic threshold for the nth type of indicator; The geological parameter influence coefficient of the nth type of index; Key geological parameters of coal seams related to the nth type of index, including gas content, coal seam permeability, coal seam thickness, and fault density. : Correction for threshold calibration.

3. The method for dynamic evaluation and closed-loop management of coal mine gas control and extraction effectiveness according to claim 1, characterized in that, In step S2, constructing the spatiotemporally aligned fusion data matrix specifically includes: Based on the timestamps of the monitoring data, the static geological model parameters are globally associated with all timestamps, and the dynamic construction parameters are interpolated to the same time interval as the monitoring data to achieve time alignment. Based on the spatial coordinates of the extraction boreholes and the grid division of the geological model, a mapping relationship is established, and the geological parameters of the corresponding grids are associated with the monitoring data of the boreholes to achieve spatial alignment. The aligned monitoring data, geological parameters, and construction parameters are integrated to construct a two-dimensional data matrix containing multiple key parameters.

4. The method for dynamic evaluation and closed-loop management of coal mine gas control and extraction effectiveness according to claim 1, characterized in that, In step S3, the construction process of the variable-length filter generator includes: Sliding window slicing: Slicing the input time series data with a window length l to obtain a subsequence matrix S; Embedding representation generation: Convolve the subsequence matrix S using K fixed-length fundamental filters to obtain the embedding representation corresponding to each fundamental filter. This forms the embedding matrix E; Mask generation: Each embedding representation in the embedding matrix E... The corresponding basic filter parameters are fused with features and processed by a neural network to generate a soft mask matrix M; where each soft mask is used to control the effective range of the weights of the corresponding basic filter in the time dimension. Synthesis: The soft mask matrix M is multiplied by the basic filter matrix to generate a variable-length filter that can adaptively extract features at different time scales.

5. The method for dynamic evaluation and closed-loop management of coal mine gas control and extraction effectiveness according to claim 4, characterized in that, In the variable-length filter generator, the length of the sliding window is dynamically adjusted proportionally according to the length L of the input sequence of the model configured in the current sampling stage; the soft mask matrix is ​​used to control the effective length of each variable-length filter, so that it presents a smooth weight transition in the time dimension.

6. The method for dynamic evaluation and closed-loop management of coal mine gas control and extraction effectiveness according to claim 4, characterized in that, The neural network is a fully connected network, which calculates an intermediate value for calibrating the effective length of the filter based on the vector after feature fusion. Each element value in the soft mask matrix is ​​generated by comparing the effective length obtained by converting the intermediate value with the position index parameter j, and then mapping it via the Sigmoid function.

7. The method for dynamic evaluation and closed-loop management of coal mine gas control and extraction effectiveness according to claim 1, characterized in that, The dynamic multi-scale convolutional neural network model is trained through the following steps: A training dataset is prepared, which includes real mining time-series data from multiple coal mines and virtual data covering abnormal working conditions generated through sampling simulation. The data is then expanded to increase data diversity. The model is trained using a hybrid loss function, which is a weighted sum of a classification loss function and a regression loss function. The classification loss function is a focus loss, which is used to handle the class imbalance problem in the classification task of sampling effect status. The regression loss function is a Hubel loss, which is used to handle the regression prediction task of core indicators and is robust to predicting outliers. During training, training is terminated when the model's loss value on the validation set meets the preset convergence condition or reaches the preset early stopping condition.

8. The method for dynamic evaluation and closed-loop management of coal mine gas control and extraction effectiveness according to claim 1, characterized in that, Step S4, the three-level linkage feedback control process involving the collaboration of the AI ​​model and the expert rule base, specifically includes: For the first level of feedback, if the warning level is a low-risk warning, automatic warning notification and manual response tracking will be executed. The second level of feedback involves generating equipment adjustment instructions based on an expert rule base, performing security checks, and obtaining manual confirmation if the warning level is medium risk or if the first level of feedback is not responded to in a timely manner. The dynamic multi-scale convolutional neural network model is then used to verify the effect of the adjustment. The third level of feedback, if the warning level is a high-risk warning, the second level of feedback adjustment is ineffective, or the rule base cannot match the instruction, will execute the intelligent generation of supplementary measures, expert review and decision-making, and scheme iteration. The AI ​​model is used for problem diagnosis and risk prediction, the expert rule base is used to generate specific equipment control instructions in the second-level feedback, and the expert review decision is used to review and determine systematic supplementary governance solutions in the third-level feedback.

Images