A method and system for dynamically evaluating the grouting effect of a coal mine goaf

Abstract

The present application provides a kind of coal mine goaf grouting effect dynamic evaluation method and system, belong to mine safety and geotechnical engineering disaster prevention technical field.The method includes: collecting the multi-source monitoring data of grouting whole cycle and intelligently dividing grouting stage;Intelligent evolution model is constructed, and the sequence of initial indicators, mid-term indicators, late indicators to maintenance period indicators is calculated by the first to fourth neural networks;The contribution weight of historical indicators is dynamically determined using attention mechanism;Grouting process parameters and geophysical data are fused to constrain the inversion process, and a grout diffusion model is generated;Based on the dynamic evaluation results and the characteristics of the three-dimensional model, a comprehensive rating is performed, and if the indicators do not meet the standards, grouting parameter control instructions are generated to realize closed-loop feedback.The present application solves the problem of static isolation, dependence on experience and inability to real-time control in the prior art, and realizes the whole process adaptive evaluation and control of grouting effect.

Description

Technical Field

[0001] This invention relates to the field of mine safety and geotechnical engineering disaster prevention technology, and in particular to a method and system for dynamic evaluation of grouting effect in coal mine goaf areas. Background Technology

[0002] Grouting is a key technical means to control coal mine goaf and surface subsidence. However, grouting projects are hidden projects, and their effects are difficult to observe directly. Therefore, establishing a scientific, accurate, and reliable method for evaluating grouting effects is of paramount importance for ensuring the quality of remediation projects, optimizing grouting parameters, and saving project costs.

[0003] Currently, evaluation methods for the effectiveness of grouting in coal mine goaf areas have significant limitations: Core drilling, while intuitive, is a point-based detection method with limited representativeness, high cost, and significant destructiveness; geophysical exploration methods (such as high-density electrical resistivity tomography and seismic wave methods) have broad coverage but their interpretation results are ambiguous, lacking accuracy and reliability; surface subsidence monitoring is directly related to remediation targets but suffers from severe lag, serving as post-event verification and unable to guide process control; experience-based judgment methods based on grouting parameters rely too heavily on personal experience, lacking quantification and objectivity. Furthermore, while some existing comprehensive evaluation methods attempt to integrate multiple indicators, their indicator weights are often fixed, failing to adapt to the dynamic characteristics of different grouting stages, and are merely simple data aggregations lacking in-depth integration and intelligent analysis, thus failing to form a closed-loop feedback loop.

[0004] In summary, existing technologies generally suffer from prominent problems such as incomplete point-based evaluation, inaccurate indirect detection, untimely post-evaluation, subjective experience-based judgment, and unintelligent comprehensive evaluation. Therefore, there is an urgent need for a new dynamic evaluation method for grouting effects that can span the entire grouting process, integrate multi-source information, possess adaptive learning capabilities, and achieve intelligent feedback decision-making, in order to fill the gaps in existing technologies. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for dynamic evaluation of grouting effect in coal mine goaf areas, so as to solve the problems of static isolation, reliance on experience, and inability to provide real-time feedback and control in the existing evaluation process.

[0006] To achieve the above objectives, this invention provides a method for dynamically evaluating the grouting effect in coal mine goaf areas, comprising the following steps: Step S1: Collect grouting process parameters, geophysical data, soil and rock mechanical response data, and surface deformation data throughout the entire grouting cycle; Step S2: Based on the grouting process parameters, the initial, middle, and late stages of grouting and the curing period are automatically divided through time sequence analysis; Step S3: Construct an intelligent evolution model. The input of this model includes the monitoring data of the current stage and the evaluation results of at least one historical stage. The output is the grouting effect index of different stages. Step S4: Based on the current stage effect indicators output by the intelligent evolution model, determine whether to enter the next stage or generate specific grouting parameter control instructions, and feed them back to the grouting execution system.

[0007] Preferably, in step S1, the grouting process parameters include real-time grouting pressure. Real-time grouting flow rate Cumulative grouting volume Geophysical data includes apparent resistivity data of mined-out areas obtained through high-density electrical resistivity tomography; geotechnical response data includes changes in formation dip angle obtained through pre-embedded sensors. With stress change Surface deformation data includes the cumulative settlement at surface subsidence monitoring points. .

[0008] Preferably, step S2 specifically includes: Step S21: Based on cumulative grouting volume Plot the time-grouting volume curve and calculate its first derivative to obtain the grouting rate curve; Step S22: Identify the key inflection points of the grouting rate curve, and divide the rapid rise of the grouting rate into the initial stage of grouting, the stable rate stage into the middle stage of grouting, and the rate decay stage into the later stage of grouting. Step S23: After grouting is completed, the cumulative surface settlement will be used as the basis for calculation. The period is defined as the maintenance period, starting from the point where the rate of change stabilizes within a threshold.

[0009] Preferably, step S3 specifically includes: Step S31: Constructing the initial feature vector Its dimensions are The calculation formula is: ; in, This indicates that the parameters have been standardized. This represents the rate of change of apparent resistivity before and after grouting; Input is fed into the first neural network, and the initial performance metrics are output. The calculation formula is: ; in, , These are the weights of the first neural network. , These are the bias parameters of the first neural network. For activation function, For the Sigmoid function; Step S32: Construct the intermediate feature vector Its dimensions are The calculation formula is: ; in, This represents the characteristics of slurry diffusion morphology obtained by inverting apparent resistivity data. Will and Concatenate into extended feature vectors ;Will The input is fed into a second neural network, where it is calculated through an attention mechanism. Contribution weight The calculation formula is: ; in, For the query vector, by The last hidden state after being processed by the gated recurrent unit (GRU) is passed through the weight matrix. Mapped to obtain, The key vector is formed by passing the hidden states of the GRU at all time steps through the weight matrix. Mapped to obtain, Let be the dimension of the key vector. Indicates the output of the gated loop unit; based on Calculate the weighted hidden state ,in The value vector is formed by passing the hidden states of the GRU at all time steps through the weight matrix. Mapped to obtain; Final output mid-term performance metrics: ; in, and The weight matrix and bias vector of the output layer; Step S33: Constructing the later feature vector Its dimensions are The calculation formula is: ; Will and , Concatenate into extended feature vectors ; Will The input is fed into a third neural network, where it is computed through a multi-head attention mechanism. and The contribution weight for each head The calculation formula is: ; in, , , Each by Obtained through different linear transformations Indicates transpose; The outputs from multiple heads are concatenated and then fed through a feedforward neural network to output the post-processing performance metrics. ; Step S34: Construct the feature vector for the maintenance period Its dimensions are The calculation formula is: ; in, and These represent the long-term trends of changes in stratum dip angle and surface subsidence, respectively; and , , Concatenate into extended feature vectors ;Will The input is fed into the fourth neural network, where it is calculated through an attention mechanism. , , Contribution weight The calculation formula is: ; in, For query vector, For key vectors, , Represents the weight matrix; based on Calculate the weighted historical index vector , Indicates a historical index; Will and After fusion, the comprehensive curing period effect index is output through the fully connected layer. .

[0010] Preferably, the structure of the first neural network is a five-layer fully connected network, consisting of: an input layer, a layer with a number of neurons and a dimension of the initial feature vector. The same; three hidden layers with 128, 64, and 32 neurons respectively, using the ReLU activation function; and an output layer containing one neuron, using the Sigmoid activation function to output initial performance metrics. ; Loss function of the first neural network A contrastive learning loss based on Euclidean distance is adopted, the expression of which is: ; in, This is the initial feature vector of the current sample. To and Initial feature vectors of successful grouting cases belonging to the same geological conditions. To and Initial feature vectors of failed grouting cases belonging to the same geological conditions; This represents the feature mapping function of the first neural network; This is the function for calculating Euclidean distance; As a label, when the sample When it is a success story =1, otherwise =0; These are preset boundary value hyperparameters.

[0011] Preferably, the structure of the second neural network includes, in sequence: a feature concatenation layer, used to combine the intermediate feature vector with the expected performance index. The concatenation is performed; a gated recurrent unit layer with 64 hidden units is used to process the concatenated serialized features; an attention layer is used to query the vector. The key vector originates from the last hidden state of the gated recurrent unit layer. Sum value vector Derived from the hidden states of all time steps in the GRU; a fully connected output layer that outputs mid-term performance metrics based on weighted hidden states. ; Loss function of the second neural network Let the mean squared error loss function be expressed as follows: ; in, These are labels for mid-term effects determined based on historical data.

[0012] Preferably, the structure of the third neural network includes, in sequence: a feature concatenation layer, used to concatenate the later feature vectors with... , Perform splicing; layer normalization layer; multi-head self-attention layer, number of heads =4, the query, key, and value vector dimensions for each header. = =32; Feedforward network layer, consisting of two fully connected layers, with 128 neurons in the middle layer, and the activation function is GELU; Residual connections are applied between the input and output of the multi-head self-attention layer and the feedforward network layer; Fully connected output layer, outputting the post-performance index. ; Loss function of the third neural network Let the mean squared error loss function be expressed as follows: ; in, These are labels for post-production effects determined based on historical data.

[0013] Preferably, the fourth neural network adopts a multi-task learning architecture based on an encoder-decoder: a shared encoder, which is a three-layer fully connected network with 256, 128, and 64 neurons respectively, is used to extract shared features from the input vector; the main task decoder receives the shared features and outputs the comprehensive maintenance period effect index through a fully connected layer with 32 neurons. The auxiliary task decoder, running in parallel with the main task decoder, receives shared features and outputs future time points through a fully connected layer with 32 neurons. Predicted value of surface subsidence ; Its overall loss function for: ; in, Let the mean squared error loss function be . Let the mean absolute error loss function be . A comprehensive effect label based on expert experience. For at a certain point in time Actual monitored surface subsidence The weighting coefficient is used to balance the losses of the two tasks.

[0014] Preferably, step S4 specifically includes: Step S41: During the middle and later stages of grouting, the effect indicators will be... or With preset threshold Comparison; among which, comparison during the middle stage of grouting With threshold In the later stage of grouting With threshold ;like or If so, then regulation will be initiated; Step S42: The control strategy generator, based on the type of currently unmet indicator and the changes in apparent resistivity and formation stress, Commonly located defect areas generate control commands; Grouting pressure adjustment amount Grouting flow rate adjustment The calculation formula is: ; ; in, As a current performance indicator, The corresponding threshold; and The gain coefficient related to the grouting system; and This is a mapping function related to the characteristics of the defect region, whose input is the average stress change in the defect region. and apparent resistivity anomaly ; Step S43: After implementing the control measures, input the new round of monitoring data into the model to obtain new performance indicators. Calculate the degree of improvement Based on this improvement, the attention weight matrix of the second or third neural network is fine-tuned online. When fine-tuning the second neural network, the fine-tuning formula is: ; When fine-tuning the third neural network, the fine-tuning formula is: ; in, For online learning rate, The loss function of the second neural network Attention weight matrix gradient, The loss function of the third neural network Attention weight matrix gradient, , This is the fine-tuned attention weight matrix.

[0015] This invention also provides a dynamic evaluation system for grouting effect in coal mine goaf areas, comprising: The multi-source data acquisition module is used to collect grouting process parameters, geophysical data, rock and soil mechanical response data and surface deformation data throughout the entire grouting cycle; The phase intelligent division module communicates with the multi-source data acquisition module to draw a time-grouting volume curve based on the cumulative grouting volume and calculate its first derivative to obtain the grouting rate curve. It identifies the key inflection points of the grouting rate curve, divides the rapid rise of the grouting rate into the initial stage of grouting, the stable rate into the middle stage of grouting, and the rate decay into the late stage of grouting. It also takes the change rate of the cumulative settlement of the ground surface after grouting being stabilized within the threshold as the starting point of the curing period. The intelligent evolution model module communicates and connects with the multi-source data acquisition module and the stage intelligent division module. The intelligent evolution model module includes: The first neural network unit is used to receive the initial feature vector and output the initial effect index. The initial feature vector is constructed based on the grouting pressure, grouting flow rate, cumulative grouting volume and the rate of change of apparent resistivity before and after grouting. The second neural network unit is used to receive the concatenated vector of the intermediate feature vector and the initial effect index, and outputs the intermediate effect index through the gated recurrent unit and the attention mechanism. The intermediate feature vector includes the slurry diffusion morphology features. The third neural network unit is used to receive the concatenated vector of the later feature vector and the initial and intermediate effect indicators, and outputs the later effect indicators through layer normalization, multi-head self-attention mechanism and feedforward network. The fourth neural network unit is used to receive the concatenated vector of the maintenance period feature vector and the effect indicators of the initial, middle and late stages, and output the comprehensive effect indicators of the maintenance period and the predicted value of future land subsidence through the encoder-decoder multi-task learning architecture. The closed-loop feedback control module communicates with the intelligent evolution model module and is used to compare the mid-term or late-term effect indicators with the preset thresholds during the mid-term and late-term of grouting. If the target is not met, the control instructions for adjusting the grouting pressure and grouting flow rate are generated and fed back to the grouting execution system based on the type of the target indicator and the defect area located by the apparent resistivity and the change in formation stress. The closed-loop feedback control module is also used to calculate the degree of improvement after the control is performed, and to fine-tune the attention weight matrix of the second or third neural network module online based on the degree of improvement.

[0016] Therefore, the present invention employs the above-mentioned method and system for dynamic evaluation of grouting effect in coal mine goaf areas, and the beneficial technical effects are as follows: (1) This invention constructs a phased dynamic weighting system and uses the evaluation results of the previous stage as the input for the evaluation of the next stage, so that the evaluation model can automatically adapt to the core contradictions and internal logical connections of different stages of grouting, including the initial, middle, late and curing stages. This overcomes the shortcomings of the existing comprehensive evaluation method, which has fixed weights and cannot reflect the characteristics of the engineering stages, making the evaluation results more in line with the actual evolution of the engineering, and significantly improving its scientificity and accuracy.

[0017] (2) This invention uses grouting process parameters (such as grouting pressure and grouting volume) as physical constraints and integrates them with geophysical exploration data for analysis, effectively overcoming the industry problem of multiple solutions in geophysical exploration methods and generating a more accurate grout diffusion model. At the same time, by systematically integrating multi-source data such as process parameters, geotechnical mechanical response, and surface deformation, and conducting in-depth correlation analysis using an intelligent evolution model, a comprehensive and multi-dimensional evaluation system is formed, solving the problems of incomplete point-based evaluation and inaccurate indirect detection in traditional methods.

[0018] (3) By setting evaluation thresholds and linking them with the control strategy generator at key stages of grouting (mid-term and late-term), this invention can judge the grouting effect in real time and immediately generate targeted parameter adjustment instructions (such as grouting pressure and flow rate), which are then fed back to the execution system. This not only solves the problem of untimely post-event evaluation in the surface settlement monitoring method, but also directly uses the evaluation results to guide construction, forming an intelligent closed loop of monitoring-evaluation-decision-control, realizing proactive intervention and optimized control of the entire grouting project process, and effectively ensuring project quality and safety. Attached Figure Description

[0019] Figure 1 This is a flowchart of a method for dynamic evaluation of grouting effect in coal mine goaf according to the present invention; Figure 2 This is a structural diagram of the intelligent evolution model. Detailed Implementation

[0020] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0021] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0022] Example 1 like Figure 1 As shown, a dynamic evaluation method for grouting effect in coal mine goaf includes the following steps: Step S1: Multi-source data acquisition and preprocessing.

[0023] Collect grouting process parameters, geophysical data, geotechnical response data, and surface deformation data throughout the entire grouting cycle.

[0024] Real-time grouting pressure is collected via a sensor network in the grouting pipeline. Real-time grouting flow rate Cumulative grouting volume .

[0025] Apparent resistivity data of the goaf were obtained using a high-density electrical resistivity meter before, during and after grouting.

[0026] The change in the dip angle of the formation is collected by dip angle sensors and stress sensors pre-embedded in the formation. With stress change .

[0027] Cumulative settlement was collected by setting up settlement monitoring points on the ground surface. .

[0028] This step systematically collects multi-source data throughout the grouting process, providing a comprehensive and continuous data foundation for subsequent intelligent evaluation. It overcomes the shortcomings of traditional methods, such as single data sources and limited coverage, and ensures the integrity and reliability of the evaluation basis.

[0029] Step S2: Intelligent division of grouting process stages.

[0030] Based on grouting process parameters, the initial, middle, late and curing periods of grouting are automatically divided through time sequence analysis.

[0031] Step S21: Based on cumulative grouting volume Plot the time-grouting volume curve and calculate its first derivative to obtain the grouting rate curve.

[0032] Step S22: The rapid rise of the grouting rate is divided into the initial stage of grouting (the grouting rate rises from 0 to the peak value). The stable grouting rate phase is divided into the mid-grouting phase (grouting rate is maintained at 80%-100% of the peak value). The grouting rate decay stage is divided into the late grouting stage (the grouting rate drops to less than 30% of the peak value).

[0033] Step S23: After grouting is completed, the cumulative surface settlement will be used as the basis for calculation. The period is defined as the maintenance period starting from the point where the rate of change stabilizes within the threshold (0.1 mm / d).

[0034] This step achieves objective and accurate stage division of the grouting process by automatically identifying the characteristic inflection points of the grouting rate curve, avoiding the subjectivity of traditional methods that rely on human experience, and laying the foundation for subsequent stage-specific evaluation.

[0035] Step S3: Evaluation of serialization effect based on intelligent evolution model.

[0036] like Figure 2 A smart evolution model is constructed. The input of the model includes the monitoring data of the current stage and the evaluation results of at least one historical stage, and the output is the grouting effect index of different stages.

[0037] Step S31: Calculate initial performance indicators : Constructing the initial feature vector Its dimensions are The calculation formula is: ; in, This indicates that the parameters have been standardized. This represents the rate of change of apparent resistivity before and after grouting; Input is fed into the first neural network, and the initial performance metrics are output. The calculation formula is: ; in, , These are the weights of the first neural network. , These are the bias parameters of the first neural network. For activation function, This is the Sigmoid function.

[0038] The first neural network has a five-layer fully connected structure, consisting of: an input layer, a layer with a number of neurons, and a layer with the initial feature vector dimension... The same; three hidden layers with 128, 64, and 32 neurons respectively, using the ReLU activation function; and an output layer containing one neuron, using the Sigmoid activation function to output initial performance metrics. .

[0039] Loss function of the first neural network A contrastive learning loss based on Euclidean distance is adopted, the expression of which is: ; in, This is the initial feature vector of the current sample. To and Initial feature vectors of successful grouting cases belonging to the same geological conditions. To and Initial feature vectors of failed grouting cases belonging to the same geological conditions; This represents the feature mapping function of the first neural network; This is the function for calculating Euclidean distance; As a label, when the sample When it is a success story =1, otherwise =0; These are preset boundary value hyperparameters.

[0040] Step S32: Calculate the mid-term performance indicators Constructing intermediate feature vectors Its dimensions are The calculation formula is: ; in, This indicates the characteristics of slurry diffusion morphology obtained by inverting apparent resistivity data.

[0041] Will and Concatenate into extended feature vectors ;Will The input is fed into a second neural network, where it is calculated through an attention mechanism. Contribution weight The calculation formula is: ; in, For the query vector, by The last hidden state after being processed by the gated recurrent unit (GRU) is passed through the weight matrix. Mapped to obtain, The key vector is formed by passing the hidden states of the GRU at all time steps through the weight matrix. Mapped to obtain, Let be the dimension of the key vector. This indicates the output of the gated loop unit.

[0042] based on Calculate the weighted hidden state ,in The value vector is formed by passing the hidden states of the GRU at all time steps through the weight matrix. The mapping is obtained.

[0043] Final output mid-term performance metrics: ; in, and These are the weight matrix and bias vector for the output layer.

[0044] The second neural network structure includes, in sequence: a feature concatenation layer, used to combine the intermediate feature vector with the expected performance index. The concatenation is performed; a gated recurrent unit layer with 64 hidden units is used to process the concatenated serialized features; an attention layer is used to query the vector. The key vector originates from the last hidden state of the gated recurrent unit layer. Sum value vector Derived from the hidden states of all time steps in the GRU; a fully connected output layer that outputs mid-term performance metrics based on weighted hidden states. .

[0045] Loss function of the second neural network Let the mean squared error loss function be expressed as follows: ; in, These are labels for mid-term effects determined based on historical data.

[0046] Step S33: Calculate post-performance indicators : Constructing late-stage feature vectors Its dimensions are The calculation formula is: ; Will and , Concatenate into extended feature vectors .

[0047] Will The input is fed into a third neural network, where it is computed through a multi-head attention mechanism. and The contribution weight for each head The calculation formula is: ; in, , , Each by Obtained through different linear transformations This indicates transpose.

[0048] The outputs from multiple heads are concatenated and then fed through a feedforward neural network to output the post-processing performance metrics. .

[0049] The structure of the third neural network includes, in sequence: a feature concatenation layer, used to combine the later feature vectors with... , Perform splicing; layer normalization layer; multi-head self-attention layer, number of heads =4, the query, key, and value vector dimensions for each header. = =32; Feedforward network layer, consisting of two fully connected layers, with 128 neurons in the middle layer, and the activation function is GELU; Residual connections are applied between the input and output of the multi-head self-attention layer and the feedforward network layer; Fully connected output layer, outputting the post-performance index. .

[0050] Loss function of the third neural network Let the mean squared error loss function be expressed as follows: ; in, These are labels for post-production effects determined based on historical data.

[0051] Step S34: Calculate the comprehensive effect index of the maintenance period. Constructing the feature vector of the maintenance period Its dimensions are The calculation formula is: ; in, and These represent the long-term trends of changes in stratum dip angle and surface subsidence, respectively; and , , Concatenate into extended feature vectors ;Will The input is fed into the fourth neural network, where it is calculated through an attention mechanism. , , Contribution weight The calculation formula is: ; in, For query vector, For key vectors, , This represents the weight matrix.

[0052] based on Calculate the weighted historical index vector , This represents a historical index.

[0053] Will and After fusion, the comprehensive curing period effect index is output through the fully connected layer. .

[0054] The fourth neural network adopts a multi-task learning architecture based on an encoder-decoder: a shared encoder, which is a three-layer fully connected network with 256, 128, and 64 neurons respectively, used to extract shared features from the input vector; and a main task decoder, which receives the shared features and outputs the comprehensive maintenance period effect index through a fully connected layer with 32 neurons. The auxiliary task decoder, running in parallel with the main task decoder, receives shared features and outputs future time points through a fully connected layer with 32 neurons. Predicted value of surface subsidence ; Overall loss function for: ; in, Let the mean squared error loss function be . Let the mean absolute error loss function be . A comprehensive effect label based on expert experience. For at a certain point in time Actual monitored surface subsidence The weighting coefficient is used to balance the losses of the two tasks.

[0055] Step S4: Closed-loop feedback and decision-making.

[0056] Based on the current stage effect indicators output by the intelligent evolution model, it is determined whether to enter the next stage or to generate specific grouting parameter control instructions, and then fed back to the grouting execution system.

[0057] Step S41: During the middle and later stages of grouting, the effect indicators will be... or With preset threshold Comparison; among which, comparison during the middle stage of grouting With threshold (0.7), compared in the later stage of grouting With threshold (0.8); if or If so, then regulation will be initiated; initial effect indicators Used for initial risk warning and comprehensive performance indicators during the maintenance period. Used for final effect verification, neither of which serves as a direct triggering condition for the control actions during the process.

[0058] Step S42: The control strategy generator, based on the type of currently unmet indicator and the changes in apparent resistivity and formation stress, Commonly identified defective areas generate control commands.

[0059] Grouting pressure adjustment amount and grouting flow rate adjustment The calculation formula is: ; ; in, For the current performance indicators ( or ), For the corresponding threshold ( or ); and The gain coefficient related to the grouting system; and The mapping functions related to the characteristics of the defect region are all constructed using a three-layer fully connected neural network, with the input being the average stress change in the defect region. and apparent resistivity anomaly .

[0060] Step S43: After implementing the control measures, input the new round of monitoring data into the model to obtain new performance indicators. Calculate the degree of improvement Based on this improvement, the attention weight matrix of the second or third neural network is fine-tuned online.

[0061] When fine-tuning the second neural network, the fine-tuning formula is: ; When fine-tuning the third neural network, the fine-tuning formula is: ; in, For online learning rate, The loss function of the second neural network Attention weight matrix gradient, The loss function of the third neural network Attention weight matrix gradient, , This is the fine-tuned attention weight matrix.

[0062] Example 2 A dynamic evaluation system for grouting effect in coal mine goaf areas includes: The multi-source data acquisition module is used to collect grouting process parameters, geophysical data, rock and soil mechanical response data and surface deformation data throughout the entire grouting cycle; The phase intelligent division module communicates with the multi-source data acquisition module to draw a time-grouting volume curve based on the cumulative grouting volume and calculate its first derivative to obtain the grouting rate curve. It identifies the key inflection points of the grouting rate curve, divides the rapid rise of the grouting rate into the initial stage of grouting, the stable rate into the middle stage of grouting, and the rate decay into the late stage of grouting. It also takes the change rate of the cumulative settlement of the ground surface after grouting being stabilized within the threshold as the starting point of the curing period. The intelligent evolution model module communicates and connects with the multi-source data acquisition module and the stage intelligent division module. The intelligent evolution model module includes: The first neural network unit is used to receive the initial feature vector and output the initial effect index. The initial feature vector is constructed based on the grouting pressure, grouting flow rate, cumulative grouting volume and the rate of change of apparent resistivity before and after grouting. The second neural network unit is used to receive the concatenated vector of the intermediate feature vector and the initial effect index, and outputs the intermediate effect index through the gated recurrent unit and the attention mechanism. The intermediate feature vector includes the slurry diffusion morphology features. The third neural network unit is used to receive the concatenated vector of the later feature vector and the initial and intermediate effect indicators, and outputs the later effect indicators through layer normalization, multi-head self-attention mechanism and feedforward network. The fourth neural network unit is used to receive the concatenated vector of the maintenance period feature vector and the effect indicators of the initial, middle and late stages, and output the comprehensive effect indicators of the maintenance period and the predicted value of future land subsidence through the encoder-decoder multi-task learning architecture. The closed-loop feedback control module communicates with the intelligent evolution model module and is used to compare the mid-term or late-term effect indicators with the preset thresholds during the mid-term and late-term of grouting. If the target is not met, the control instructions for adjusting the grouting pressure and grouting flow rate are generated and fed back to the grouting execution system based on the type of the target indicator and the defect area located by the apparent resistivity and the change in formation stress. The closed-loop feedback control module is also used to calculate the degree of improvement after the control is performed, and to fine-tune the attention weight matrix of the second or third neural network module online based on the degree of improvement.

[0063] Therefore, this invention adopts the above-mentioned dynamic evaluation method and system for grouting effect in coal mine goaf areas. By constructing a sequential evaluation system based on an intelligent evolution model, the entire grouting process is divided into interrelated stages, and multiple neural networks are used to realize the cumulative influence of the early evaluation results on the later evaluation. At the same time, grouting process parameters and geophysical and other multi-source data are integrated to constrain the intelligent algorithm with physical mechanisms and improve the evaluation accuracy. Finally, based on the dynamic evaluation results, closed-loop feedback and intelligent decision-making of the grouting process are realized, thereby achieving full-process, accurate, adaptive evaluation and active control of the grouting effect.

[0064] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for dynamic evaluation of grouting effect in coal mine goaf areas, characterized in that, Includes the following steps: Step S1: Collect grouting process parameters, geophysical data, soil and rock mechanical response data, and surface deformation data throughout the entire grouting cycle; Step S2: Based on the grouting process parameters, the initial, middle, and late stages of grouting and the curing period are automatically divided through time sequence analysis; Step S3: Construct an intelligent evolution model. The input of this model includes the monitoring data of the current stage and the evaluation results of at least one historical stage. The output is the grouting effect index of different stages. Step S4: Based on the current stage effect indicators output by the intelligent evolution model, determine whether to enter the next stage or generate specific grouting parameter control instructions, and feed them back to the grouting execution system.

2. The method for dynamic evaluation of grouting effect in coal mine goaf according to claim 1, characterized in that, In step S1, the grouting process parameters include real-time grouting pressure. Real-time grouting flow rate Cumulative grouting volume Geophysical data includes apparent resistivity data of mined-out areas obtained through high-density electrical resistivity tomography; geotechnical response data includes changes in formation dip angle obtained through pre-embedded sensors. With stress change Surface deformation data includes the cumulative settlement at surface subsidence monitoring points. .

3. The method for dynamic evaluation of grouting effect in coal mine goaf according to claim 2, characterized in that, Step S2 specifically includes: Step S21: Based on cumulative grouting volume Plot the time-grouting volume curve and calculate its first derivative to obtain the grouting rate curve; Step S22: Identify the key inflection points of the grouting rate curve, and divide the rapid rise of the grouting rate into the initial stage of grouting, the stable rate stage into the middle stage of grouting, and the rate decay stage into the later stage of grouting. Step S23: After grouting is completed, the cumulative surface settlement will be used as the basis for calculation. The period is defined as the maintenance period, starting from the point where the rate of change stabilizes within a threshold.

4. The method for dynamic evaluation of grouting effect in coal mine goaf according to claim 3, characterized in that, Step S3 specifically includes: Step S31: Constructing the initial feature vector Its dimensions are The calculation formula is: ； in, This indicates that the parameters have been standardized. This represents the rate of change of apparent resistivity before and after grouting; Input is fed into the first neural network, and the initial performance metrics are output. The calculation formula is: ； in, , These are the weights of the first neural network. , These are the bias parameters of the first neural network. For activation function, For the Sigmoid function; Step S32: Construct the intermediate feature vector Its dimensions are The calculation formula is: ； in, This represents the characteristics of slurry diffusion morphology obtained by inverting apparent resistivity data. Will and Concatenate into extended feature vectors ;Will The input is fed into a second neural network, where it is calculated through an attention mechanism. Contribution weight The calculation formula is: ； in, For the query vector, by The last hidden state after being processed by the gated recurrent unit (GRU) is passed through the weight matrix. Mapped to obtain, The key vector is formed by passing the hidden states of the GRU at all time steps through the weight matrix. Mapped to obtain, Let be the dimension of the key vector. Indicates the output of the gated loop unit; based on Calculate the weighted hidden state ,in The value vector is formed by passing the hidden states of the GRU at all time steps through the weight matrix. Mapped to obtain; Final output mid-term performance metrics: ； in, and The weight matrix and bias vector of the output layer; Step S33: Constructing the later feature vector Its dimensions are The calculation formula is: ； Will and , Concatenate into extended feature vectors ; Will The input is fed into a third neural network, where it is computed through a multi-head attention mechanism. and The contribution weight for each head The calculation formula is: ； in, , , Each by Obtained through different linear transformations Indicates transpose; The outputs from multiple heads are concatenated and then fed into a feedforward neural network to output the post-processing performance metrics. ; Step S34: Construct the feature vector for the maintenance period Its dimensions are The calculation formula is: ； in, and These represent the long-term trends of changes in stratum dip angle and surface subsidence, respectively; and , , Concatenate into extended feature vectors ;Will The input is fed into the fourth neural network, where it is calculated through an attention mechanism. , , Contribution weight The calculation formula is: ； in, For query vector, For key vectors, , Represents the weight matrix; based on Calculate the weighted historical index vector , Indicates a historical index; Will and After fusion, the comprehensive curing period effect index is output through the fully connected layer. .

5. The method for dynamic evaluation of grouting effect in coal mine goaf according to claim 4, characterized in that, The first neural network has a five-layer fully connected structure, consisting of: an input layer, a layer with a number of neurons, and a layer with the initial feature vector dimension... The same; three hidden layers with 128, 64, and 32 neurons respectively, using the ReLU activation function; and an output layer containing one neuron, using the Sigmoid activation function to output initial performance metrics. ; Loss function of the first neural network A contrastive learning loss based on Euclidean distance is adopted, the expression of which is: ； in, This is the initial feature vector of the current sample. To and Initial feature vectors of successful grouting cases belonging to the same geological conditions. To and Initial feature vectors of failed grouting cases belonging to the same geological conditions; This represents the feature mapping function of the first neural network; This is the function for calculating Euclidean distance; As a label, when the sample When it is a success story =1, otherwise =0; These are preset boundary value hyperparameters.

6. The method for dynamic evaluation of grouting effect in coal mine goaf according to claim 5, characterized in that, The second neural network structure includes, in sequence: a feature concatenation layer, used to combine the intermediate feature vector with the expected performance index. The concatenation is performed; a gated recurrent unit layer with 64 hidden units is used to process the concatenated serialized features; an attention layer is used to query the vector. The key vector originates from the last hidden state of the gated recurrent unit layer. Sum value vector Derived from the hidden states of all time steps in the GRU; a fully connected output layer that outputs mid-term performance metrics based on weighted hidden states. ; Loss function of the second neural network Let the mean squared error loss function be expressed as follows: ； in, These are labels for mid-term effects determined based on historical data.

7. The method for dynamic evaluation of grouting effect in coal mine goaf according to claim 6, characterized in that, The structure of the third neural network includes, in sequence: a feature concatenation layer, used to combine the later feature vectors with... , Perform splicing; layer normalization layer; multi-head self-attention layer, number of heads =4, the query, key, and value vector dimensions for each header. = =32; Feedforward network layer, consisting of two fully connected layers, with 128 neurons in the middle layer, and the activation function is GELU; Residual connections are applied between the input and output of the multi-head self-attention layer and the feedforward network layer; Fully connected output layer, outputting the post-performance index. ; Loss function of the third neural network Let the mean squared error loss function be expressed as follows: ； in, These are labels for post-production effects determined based on historical data.

8. The method for dynamic evaluation of grouting effect in coal mine goaf according to claim 7, characterized in that, The fourth neural network adopts a multi-task learning architecture based on an encoder-decoder: a shared encoder, which is a three-layer fully connected network with 256, 128, and 64 neurons respectively, used to extract shared features from the input vector; the main task... The decoder receives shared features and outputs a comprehensive maintenance period effect index through a fully connected layer with 32 neurons. The auxiliary task decoder, running in parallel with the main task decoder, receives shared features and outputs future time points through a fully connected layer with 32 neurons. Predicted value of surface subsidence ; Its overall loss function for: ； in, Let the mean squared error loss function be . Let the mean absolute error loss function be . A comprehensive effect label based on expert experience. For at a certain point in time Actual monitored surface subsidence The weighting coefficient is used to balance the losses of the two tasks.

9. A method for dynamic evaluation of grouting effect in coal mine goaf according to claim 8, characterized in that, Step S4 specifically includes: Step S41: During the middle and later stages of grouting, the effect indicators will be... or With preset threshold Comparison; among which, comparison during the middle stage of grouting With threshold In the later stage of grouting With threshold ;like or If so, then regulation will be initiated; Step S42: The control strategy generator, based on the type of currently unmet indicator and the changes in apparent resistivity and formation stress, Commonly located defect areas generate control commands; Grouting pressure adjustment amount and grouting flow rate adjustment The calculation formula is: ； ； in, As a current performance indicator, The corresponding threshold; and The gain coefficient related to the grouting system; and This is a mapping function related to the characteristics of the defect region, whose input is the average stress change in the defect region. and apparent resistivity anomaly ; Step S43: After implementing the control measures, input the new round of monitoring data into the model to obtain new performance indicators. Calculate the degree of improvement Based on this improvement, the attention weight matrix of the second or third neural network is fine-tuned online. When fine-tuning the second neural network, the fine-tuning formula is: ； When fine-tuning the third neural network, the fine-tuning formula is: ； in, For online learning rate, The loss function of the second neural network Attention weight matrix gradient, The loss function of the third neural network Attention weight matrix gradient, , This is the fine-tuned attention weight matrix.

10. A dynamic evaluation system for grouting effect in coal mine goaf areas, characterized in that, include: The multi-source data acquisition module is used to collect grouting process parameters, geophysical data, rock and soil mechanical response data and surface deformation data throughout the entire grouting cycle; The phase intelligent division module communicates with the multi-source data acquisition module to draw a time-grouting volume curve based on the cumulative grouting volume and calculate its first derivative to obtain the grouting rate curve. It identifies the key inflection points of the grouting rate curve, divides the rapid rise of the grouting rate into the initial stage of grouting, the stable rate into the middle stage of grouting, and the rate decay into the late stage of grouting. It also takes the change rate of the cumulative settlement of the ground surface after grouting being stabilized within the threshold as the starting point of the curing period. The intelligent evolution model module communicates and connects with the multi-source data acquisition module and the stage intelligent division module. The intelligent evolution model module includes: The first neural network unit is used to receive the initial feature vector and output the initial effect index. The initial feature vector is constructed based on the grouting pressure, grouting flow rate, cumulative grouting volume and the rate of change of apparent resistivity before and after grouting. The second neural network unit is used to receive the concatenated vector of the intermediate feature vector and the initial effect index, and outputs the intermediate effect index through the gated recurrent unit and the attention mechanism. The intermediate feature vector contains the slurry diffusion morphology features. The third neural network unit is used to receive the concatenated vector of the later feature vector and the initial and intermediate effect indicators, and outputs the later effect indicators through layer normalization, multi-head self-attention mechanism and feedforward network. The fourth neural network unit is used to receive the concatenated vector of the maintenance period feature vector and the effect indicators of the initial, middle and late stages, and output the comprehensive effect indicators of the maintenance period and the predicted value of future land subsidence through the encoder-decoder multi-task learning architecture. The closed-loop feedback control module communicates with the intelligent evolution model module and is used to compare the mid-term or late-term effect indicators with the preset thresholds during the mid-term and late-term of grouting. If the target is not met, the control instructions for adjusting the grouting pressure and grouting flow rate are generated and fed back to the grouting execution system based on the type of the target indicator and the defect area located by the apparent resistivity and the change in formation stress. The closed-loop feedback control module is also used to calculate the degree of improvement after the control is performed, and to fine-tune the attention weight matrix of the second or third neural network module online based on the degree of improvement.

Images