A method and system for predicting a grout flow of a grouting project

Abstract

The application provides a grouting engineering gushing prediction method and system, and belongs to the technical field of tunnel and underground engineering construction safety monitoring. The method comprises the following steps: collecting grouting pressure, flow, grout viscosity and microseismic event data; calculating grout resistance index and microseismic event aggregation density; inputting the same into a pre-trained machine learning model to obtain an initial gushing risk index and a confidence score; selecting direct early warning or entering a parameter correction path according to whether the confidence score is lower than a preset threshold; when the confidence score is low, generating a dynamic correction factor based on the score and risk change, adjusting the calculation weight of the grout resistance index, obtaining the corrected parameters and performing secondary prediction; and obtaining a final risk index by comprehensively considering the initial and secondary prediction results, and executing hierarchical early warning. The application improves the accuracy and reliability of gushing risk prediction through the confidence-driven double-path decision and online adaptive correction of parameters, and constructs an early warning mechanism.

Description

Technical Field

[0001] This invention relates to the field of tunnel and underground engineering construction safety monitoring technology, and in particular to a method and system for predicting grouting in grouting projects. Background Technology

[0002] Grouting technology is a key construction technique used in tunnel engineering, mining, and underground structure reinforcement to seal high-pressure water inrushes, reinforce fractured rock masses, and improve geological properties. However, during grouting operations in high-pressure, water-rich strata, karst areas, or large geological structures, the grout, driven by high pressure, may erupt uncontrollably and rapidly along potential channels or weak surfaces, resulting in a geyser accident. Such accidents are characterized by their suddenness and destructive power, seriously threatening the safety of workers, damaging construction equipment, delaying project schedules, and potentially triggering secondary disasters.

[0003] Currently, in grouting engineering practice, the monitoring and early warning of gushing risks mainly rely on the following methods. Among them, the empirical threshold method sets fixed safety thresholds for a few parameters such as pressure and flow rate based on historical experience. This method is too simplistic and crude, making it difficult to adapt to complex and ever-changing geological and construction dynamics, resulting in high false alarm and false negative rates; essentially, it is a passive, post-event alarm. The multi-parameter linear monitoring method makes a comprehensive judgment by observing the trends of more parameters such as pressure, flow rate, and grout density. While this represents progress, it still relies on manual interpretation based on experience, lacks a quantitative risk fusion model, and cannot effectively perceive the dynamic characteristics of the interaction between grout and the formation. With technological advancements, prediction methods based on simple machine learning models have begun to be applied. Historical data is input into regression models or shallow neural networks to predict parameters or identify anomalies. However, these models are often black boxes, only outputting a single risk probability, lacking a quantitative evaluation mechanism for the model's own prediction confidence. When the input data is noisy or in a state that the model has not learned from, the system cannot recognize that its judgment may be unreliable, potentially leading to incorrect decisions. Furthermore, the static rule judgment method uses a preset combination of fixed rules for judgment, which is difficult to adjust online and cannot adapt to the nonlinear changes in the rheological properties of the grout and the formation response during the grouting process, resulting in poor system adaptability.

[0004] In summary, existing technologies generally suffer from drawbacks such as limited perception dimensions, rigid judgment models, and fixed decision-making paths. They are mostly open-loop, static monitoring or simple predictions, and cannot build an early warning system that can assess its own reliability in real time and dynamically adjust parameters and decision-making logic. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for predicting gushing in grouting projects, so as to overcome the shortcomings of existing early warning methods, such as poor adaptability and inability to assess the reliability of prediction.

[0006] To achieve the above objectives, the present invention provides a method for predicting gushing in grouting projects, comprising the following steps:

[0007] Step S1: Collect data from multiple sources;

[0008] Step S2: Calculate the first slurry resistance index and the microseismic event cluster density based on multi-source data;

[0009] Step S3: Input the first slurry resistance index and the microseismic event cluster density into the pre-trained machine learning model to generate the initial gushing risk index, and simultaneously obtain the model's confidence score for the current prediction.

[0010] Step S4: Determine whether the confidence score is lower than the preset threshold;

[0011] If the value is below the preset threshold, proceed to step S5;

[0012] If it is not lower than the preset threshold, skip step S5, use the initial surge risk index directly as the effective risk index to be synthesized, and proceed to step S6.

[0013] Step S5: Based on the change characteristics of the confidence score and the initial gushing risk index, a dynamic correction factor is generated, and the first slurry resistance index used in step S2 is adjusted to obtain the second slurry resistance index.

[0014] Step S6: When step S5 has been executed, the second slurry resistance index and the microseismic event cluster density are input into the machine learning model for prediction again to obtain the secondary risk index. The initial gushing risk index and the secondary risk index are then combined to obtain the final risk index. When step S5 has not been executed, the effective risk index to be synthesized is directly used as the final risk index.

[0015] Step S7: Execute tiered early warning based on the final risk index.

[0016] Preferably, in step S1, the multi-source data includes real-time acquisition of grouting pressure, grouting flow rate, grout viscosity, and the occurrence time, three-dimensional coordinates, and energy of micro-seismic events.

[0017] Preferably, step S2 specifically includes:

[0018] Step S21, based on grouting pressure Grouting flow rate and slurry viscosity Calculate the first slurry resistance index :

[0019] ;

[0020] in, To be taken from The instantaneous pressure value, Based on the current sliding time window The calculated pressure standard deviation, and These are preset weighting coefficients;

[0021] Step S22: Based on microseismic event energy Calculate the cluster density of microseismic events :

[0022] ;

[0023] in, In the time window Internal and spatial unit volume The sum of the energies of all microseismic events.

[0024] Preferably, the specific process of step S3 includes:

[0025] Step S31: Set the first slurry resistance index at the current moment. and microseismic event cluster density Each with their own past Combining historical values ​​from each time step, a feature vector with temporal context information is constructed. ;

[0026] Step S32: Convert the feature vector Input a pre-trained machine learning model, which is a bidirectional long short-term memory network containing a self-attention mechanism;

[0027] Step S33: The machine learning model outputs two scalars: an initial surge risk index normalized by the sigmoid function. and confidence score Confidence score Based on the model for the same input conduct Output during the second forward propagation variance The calculation is as follows:

[0028] ;

[0029] in, It is an integer greater than 1.

[0030] Preferably, the specific process of step S4 includes:

[0031] Step S41: Set the preset threshold ;

[0032] Step S42: Compare confidence scores With preset threshold ;

[0033] Step S43, if If the result is low confidence, the parameter correction path is triggered, and a path identifier is generated. Then proceed to step S5;

[0034] Step S44, if If the result is deemed high confidence, a fast decision-making path is triggered, and a path identifier is generated. The initial surge risk index Assign a value to the effective risk index to be synthesized Then, proceed directly to step S6.

[0035] Preferably, step S5 only applies to the path identifier. The execution process includes:

[0036] Step S51: Calculate the initial surge risk index In recent Gradient of change within each time step ;

[0037] Step S52: Based on confidence score and gradient of change Calculate the dynamic correction factor :

[0038] ;

[0039] in, For adjustment coefficients, For symbolic functions, This is the gradient limit value;

[0040] Step S53: Utilize dynamic correction factor Adjust the calculation of the first slurry resistance index Weighting coefficients and The formula is adjusted as follows:

[0041] ;

[0042] ;

[0043] in, and The preset coupling coefficient, and These are the adjusted weighting coefficients;

[0044] Step S54: Use the adjusted weighting coefficients and and the real-time grouting pressure collected in step S11. and grouting flow rate The real-time slurry viscosity collected in step S12 ,according to The calculation formula was recalculated to obtain the second slurry resistance index. .

[0045] Preferably, the specific process of step S6 includes:

[0046] Step S61: Determine the path identifier The value;

[0047] Step S62, if Then execute:

[0048] Step S621: Adjust the second slurry resistance index Compared with the microseismic event cluster density calculated based on step S22 at the current time Combine to construct new feature vectors ;

[0049] Step S622, to Input into a machine learning model to obtain a secondary risk index. ;

[0050] Step S623: Calculate the final risk index :

[0051] ;

[0052] Among them, the composite weight Confidence score The decision is made using the following formula:

[0053] ;

[0054] Step S63, if Then the effective risk index to be synthesized will be directly used. As the final risk index ,Right now .

[0055] Preferably, the specific process of step S7 includes:

[0056] Step S71: Preset at least three risk level thresholds. , , ,and The risk index range is divided into low-risk, medium-risk, and high-risk zones.

[0057] Step S72: Final Risk Index The current risk level is determined by comparing it with the risk level threshold.

[0058] Step S73: Trigger the audible and visual alarm signal corresponding to the risk level.

[0059] The present invention also provides a gushing prediction system for grouting projects, comprising:

[0060] The data acquisition unit is used to collect grouting pressure, grouting flow rate, grout viscosity and micro-vibration event data in real time;

[0061] The data processing unit is connected in communication with the data acquisition unit and is used to calculate the first slurry resistance index and the microseismic event cluster density based on the acquired multi-source data.

[0062] The prediction unit, which is connected in communication with the data processing unit, is used to input the first slurry resistance index and the microseismic event cluster density into the pre-trained machine learning model to generate an initial gushing risk index and simultaneously obtain the model's confidence score for the current prediction.

[0063] The decision control unit communicates with the prediction unit to determine whether the confidence score is lower than a preset threshold, and selects either the execution parameter correction path or the fast decision path based on the judgment result.

[0064] The parameter correction unit is connected in communication with the decision control unit and the data processing unit. When the confidence score is lower than the preset threshold, it generates a dynamic correction factor based on the change characteristics of the confidence score and the initial gushing risk index, and adjusts the slurry resistance index calculation logic to obtain the second slurry resistance index.

[0065] The risk synthesis unit, which is connected in communication with the decision control unit and the prediction unit, is used to re-input the second slurry resistance index and the microseismic event cluster density into the machine learning model to obtain the secondary risk index when executing the parameter correction path, and to synthesize the initial surge risk index and the secondary risk index to obtain the final risk index; when executing the fast decision path, the initial surge risk index is directly used as the final risk index.

[0066] The early warning output unit is connected in communication with the risk synthesis unit and is used to execute graded early warnings based on the final risk index.

[0067] Therefore, the present invention employs the above-mentioned method and system for predicting grouting in grouting projects, and the beneficial technical effects are as follows:

[0068] (1) To address the problems of existing methods having a single perception dimension, rigid judgment models, and an inability to effectively perceive the dynamic characteristics of the interaction between grout and formation, this invention integrates multi-source data on grouting pressure, flow rate, grout viscosity, and microseismic events, and designs grout resistance index and microseismic event aggregation density as core fusion features. This achieves collaborative perception and intelligent feature extraction of grout rheological properties and formation mechanical response during the grouting process. Based on this, a bidirectional long short-term memory network model with a self-attention mechanism is used to deeply mine temporal features and predict risks, improving the ability to identify and warn of complex nonlinear gushing precursors.

[0069] (2) To address the problems in existing technologies, such as fixed decision paths, lack of self-assessment mechanisms for predictive reliability, and susceptibility to erroneous decisions under noisy or unknown conditions, this invention innovatively introduces a model confidence scoring mechanism and constructs a confidence-driven dual-path decision logic based on this. When the confidence level is high, a rapid early warning path is executed to ensure timely response; when the confidence level is low, a parameter correction path is automatically triggered, generating dynamic correction factors based on risk change trends, adaptively adjusting the weights of input features online, and initiating secondary prediction and result synthesis. This mechanism enables the system to have self-assessment and calibration capabilities, improving the reliability of early warnings under uncertain data or abnormal operating conditions.

[0070] (3) To address the problem that existing methods are mostly static, open-loop systems that cannot adapt to the nonlinear dynamic changes in the grouting process, this invention constructs an early warning system by integrating online parameter correction, dual-path decision-making, and risk synthesis feedback. This system not only monitors and provides feedback on the real-time reliability of the prediction model, but also iteratively optimizes the feature calculation logic through dynamic correction factors. This enables the early warning model to adapt to the continuous changes in grout characteristics and formation conditions during the grouting process, thereby overcoming the limitations of static rules or fixed threshold methods and achieving a simultaneous improvement in early warning accuracy and adaptability. Attached Figure Description

[0071] Figure 1 This is a flowchart illustrating the flow prediction method for grouting engineering according to the present invention.

[0072] Figure 2 Data acquisition layout diagram (partial schematic diagram);

[0073] Figure 3 This is a schematic diagram illustrating the tiered early warning execution and feedback process.

[0074] Figure 4 This is an architectural diagram of a grouting engineering gushing prediction system according to the present invention;

[0075] Figure 5This is a time series diagram comparing the predicted risk index of the experimental group and the control group before and after a typical gushing event in the simulation test of this invention;

[0076] Figure 6 This is a time series diagram showing the relationship between model confidence and decision path in the simulation experiment of this invention. Detailed Implementation

[0077] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0078] 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.

[0079] Example 1

[0080] This embodiment details the implementation process of the method proposed in this invention.

[0081] System initialization and parameter preset.

[0082] Before executing the method, perform the following preparatory work:

[0083] Deploy data acquisition hardware: Install pressure sensors and flow meters in the grouting pipeline, install online viscometers in the slurry circulation pipeline, and deploy a micro-vibration monitoring sensor array around the grouting area.

[0084] A pre-trained machine learning model is loaded into the computing unit. This model is a bidirectional long short-term memory network (BiLSTM-Attention) with a self-attention mechanism. Its training data comes from historical grouting project datasets and can correlate grout dynamics with formation response characteristics.

[0085] The parameters and thresholds required for the preset method are as follows:

[0086] Confidence threshold .

[0087] Weighting coefficients in the formula for calculating slurry resistance index and .

[0088] Time window length for microseismic event statistics With spatial unit volume Definition.

[0089] The historical time step length used to calculate the gradient change. and the historical time step length used to construct the feature vector. .

[0090] The adjustment coefficient required for dynamic correction factor calculation Gradient limit value Coupling coefficient and .

[0091] Risk level threshold , , .

[0092] The specific work process is as follows: Figure 1 As shown.

[0093] Step S1: Synchronous acquisition of multi-source data.

[0094] After the method is started, it enters the real-time cyclic acquisition phase (e.g. Figure 2 ):

[0095] The pressure sensor and flow meter operate synchronously at a set sampling frequency (e.g., 1Hz) to continuously acquire real-time raw data sequences of grouting pressure. and grouting flow rate instantaneous value .

[0096] Online viscometer continuously measures and outputs real-time slurry viscosity values. .

[0097] The microseismic monitoring array operates continuously, outputting characteristic data for each detected microseismic event, including at least the event's timestamp, three-dimensional spatial coordinates, and released energy value. .

[0098] Step S2: Calculation of primary risk parameters.

[0099] In each processing cycle (e.g., per second), the collected raw data is processed to generate two core primary risk parameters:

[0100] (1) Calculate the first slurry resistance index :

[0101] ;

[0102] in, To be taken from The instantaneous pressure value, Based on the current sliding time window The calculated pressure standard deviation, and These are preset weighting coefficients;

[0103] This parameter comprehensively reflects the instantaneous resistance state and its fluctuations during the slurry pumping process.

[0104] (2) Calculate the cluster density of microseismic events :

[0105] ;

[0106] in, In the time window Internal and spatial unit volume The sum of the energies of all microseismic events.

[0107] This parameter quantifies the spatial and temporal concentration of microfracture activity in the rock mass under the influence of grouting.

[0108] Step S3: Initial risk prediction and confidence assessment.

[0109] Input the real-time parameters obtained from the above calculations into the prediction model:

[0110] Feature vector construction: [The text abruptly ends here, likely due to an incomplete sentence or a formatting error.] and The values, respectively, are compared with their respective predecessors. The combination of historical values ​​at each time step forms a feature vector containing temporal context information. This allows the model to perceive trends in parameter changes.

[0111] Model prediction: using feature vectors Input a pre-trained BiLSTM-Attention model. After processing, the model outputs two key scalars:

[0112] Initial Gust Risk Index : A value normalized to the [0, 1] interval by the Sigmoid function, which directly represents the model's probability estimate of the current risk of a gushing event.

[0113] Confidence score The rating is based on multiple (e.g.) =10) Forward propagation and using stochasticity such as Dropout to estimate the model's impact on the current prediction. The uncertainty. The specific calculation formula is:

[0114] ;

[0115] in, yes The variance of the predicted values. The higher the value, the more confident the model is in its prediction.

[0116] Step S4: Confidence Decision and Path Selection.

[0117] The system makes key decisions based on confidence scores:

[0118] The calculated confidence score With the preset threshold Compare them.

[0119] Decision 1 (High Confidence Path): If The system determines that the prediction result is reliable. Path identifiers are generated. and will The value is directly assigned to an intermediate variable—the effective risk index to be synthesized. Then, skip step S5 (parameter correction) and proceed directly to step S6 to synthesize the results.

[0120] Decision 2 (Low Confidence Path): If The system determines that this prediction has high uncertainty and the result may be unreliable. Generate path identifiers. This triggers step S5, initiating the parameter correction process.

[0121] Step S5: Dynamic parameter correction and recalculation.

[0122] This step is only applicable to... This is executed periodically, aiming to improve the reliability of the next prediction by refining the input parameters.

[0123] Calculate the risk change gradient :analyze In recent Calculate the gradient of change within each time step and the trend of change. .

[0124] ;

[0125] in, For the current moment The initial surge risk index, for The initial surge risk index prior to the time step. This refers to the data processing cycle or sampling interval. The preset number of time steps is used to determine the length of the historical window for calculating the gradient.

[0126] Generate dynamic correction factor : Based on confidence score and gradient of change Calculate the dynamic correction factor :

[0127] ;

[0128] in, For adjustment coefficients, For symbolic functions, This is the gradient limiting value. To ensure greater correction strength at low confidence levels. This links the direction of the correction to the trend of risk changes.

[0129] Using dynamic correction factors Adjust the calculation of the first slurry resistance index Weighting coefficients and The formula is adjusted as follows:

[0130] ;

[0131] ;

[0132] in, and The preset coupling coefficient, and These are the adjusted weighting coefficients;

[0133] Step S54: Use the adjusted weighting coefficients and and the collected real-time grouting pressure Grouting flow rate Real-time slurry viscosity ,according to The calculation formula was recalculated to obtain the second slurry resistance index. .

[0134] Step S6: Secondary prediction and risk index synthesis.

[0135] This step generates the final risk index for early warning using different methods depending on the different decision-making paths:

[0136] Scenario 1 (from a low-confidence path) ):

[0137] Secondary prediction: the result after correction The density of microseismic events at the current moment Combine to form new feature vectors The same machine learning model is input again for prediction to obtain a secondary risk index. .

[0138] Exponential synthesis: the initial prediction With secondary prediction The final risk index is obtained by weighted synthesis. .

[0139] ;

[0140] Among them, the composite weight Confidence score The decision is made using the following formula: This reflects the level of trust in the initial prediction results.

[0141] Scenario 2 (from a high-confidence path) Since no parameter correction was performed, the data saved in step S4 was directly used. As the final risk index .

[0142] Step S7, Execute tiered early warning (e.g.) Figure 3 ).

[0143] Synthesized Compare with the preset risk level threshold:

[0144] At least three risk level thresholds are pre-set (e.g.) , , The risk index range of [0, 1] is divided into corresponding low, medium and high risk zones.

[0145] according to The current risk level is determined by the range in which the risk falls.

[0146] Trigger the audible and visual alarm signal corresponding to the risk level to complete the early warning output.

[0147] Step S8, adaptive update of model parameters, is executed periodically and includes:

[0148] Step S81: Collect the confidence score for each prediction within a statistical period. Path selection results And feedback on the comparison between early warning results and actual working conditions;

[0149] Step S82, if in When the frequency of false alarms or missed alarms exceeds the frequency threshold. Then the preset threshold will be automatically lowered. ;

[0150] Step S83, if in At that time, the accuracy of the early warning after parameter correction The accuracy rate of early warnings remains higher than before the revision. And the difference Exceeding the accuracy threshold Then adjust according to the preset amount. Increase the adjustment coefficient .

[0151] The invention will be further illustrated below with specific examples.

[0152] Simulation test setup.

[0153] Experimental Environment and Data Source: To verify the effectiveness of the method of this invention, a simulation test environment based on historical engineering data was constructed. The test dataset comes from three completed high-risk tunnel grouting projects, containing more than 500 hours of grouting process data, including 17 manually confirmed gushing events (or clear precursors) records.

[0154] Data preprocessing:

[0155] Desensitization processing: Sensitive information in the original data, such as the specific location of the project and the name of the unit, is replaced or deleted, and only pure technical parameters related to the grouting process and geological response are retained.

[0156] Standardization processing: Unifying the processing of raw data from different sources and with different dimensions.

[0157] Time alignment: Based on the absolute timestamp, the data streams of pressure, flow rate, viscosity (sampling frequency 1Hz) and microseismic events (discrete occurrence) are interpolated and aligned to form a unified 1Hz time series.

[0158] Outlier handling: Employing statistically based methods (such as...) The method (principle) identifies and removes obviously erroneous collected values ​​(such as pressure dropping to zero instantaneously), and fills them in using linear interpolation of valid values ​​before and after the time interval.

[0159] Normalization: For numerical characteristics such as grouting pressure, flow rate, viscosity and micro-vibration energy, the Min-Max normalization method is used to linearly scale them to the [0,1] interval to eliminate the influence of dimensions and accelerate model convergence.

[0160] Serialization: The processed continuous data is sliced ​​according to a preset time window length (e.g., 300 seconds) and step size (e.g., 1 second) to generate sample sequences for model training and testing.

[0161] Control group setup:

[0162] Control Group A (Single Threshold Method): The empirical threshold method commonly used in engineering is adopted. When the grouting pressure exceeds the safety threshold set according to the 95th percentile of historical data distribution, an alarm is triggered.

[0163] Control Group B (Traditional Machine Learning Method): This method uses the same BiLSTM-Attention model as the present invention for risk prediction, but does not include confidence assessment and parameter correction loops. The model directly outputs a risk index and provides an early warning by comparing it with a fixed threshold optimized through a validation set.

[0164] Experimental group (method of this invention): The method described in this invention is fully implemented, including confidence assessment, dual-path decision-making, and dynamic parameter adjustment. The preset key parameters are: =0.75, =0.7, =0.3, =0.5, =0.1, =0.05, =0.1.

[0165] Evaluation metrics: Accuracy, recall, precision, and F1-Score are used as the main evaluation metrics. The average lead time of the warnings (i.e., the time interval from the first triggering of a medium-to-high risk warning to the actual occurrence of the gushing event) is also recorded.

[0166] Processing procedure.

[0167] Play back the test dataset on the simulation platform, and let the three methods process the same data stream in parallel:

[0168] Data input: mimicking the real-time system, pre-processed, timestamp-aligned, standardized data is injected into the three systems at a frequency of 1Hz.

[0169] Process execution:

[0170] For the experimental group (the method of this invention), the system operates according to the steps described above. The following is a specific numerical example demonstrating the execution process of the low-confidence path:

[0171] During simulation time The system collects and calculates the following: =0.65, =0.08, =0.70, =0.20, calculated as follows . =0.15.

[0172] Feature vector After inputting the model, the output is: =0.62, =0.68 (variance) The calculated value is approximately 0.47.

[0173] Decision: due to Entering the low-confidence correction path ( ).

[0174] Parameter correction:

[0175] Calculate gradient :set up =0.58, =1 second, =5, then .

[0176] Calculate the dynamic correction factor:

[0177] .

[0178] Adjusting weights: ; .

[0179] calculate Using the same real-time data with new weights, Although the numerical changes are minor, the direction of the correction has been determined.

[0180] Secondary prediction and synthesis: =3.361 and =0.15 input model, get =0.64. (According to...) =0.68 Synthesis: .

[0181] Warning: =0.626, higher than =0.6, triggering a medium-level warning.

[0182] For control group B, under the same conditions... At that time, using the same model, we obtain =0.62. Since there is no confidence level assessment, it is directly compared with a fixed threshold (e.g., 0.60), which also triggers a medium-level warning. However, its decision-making process cannot assess its own uncertainty. =0.68).

[0183] For control group A, only standardized pressure was assessed. Does 0.65 exceed the preset threshold (e.g., 0.70)? If not, no warning will be issued.

[0184] Repeat the above process on the entire test set, record the warning trigger time and warning level for each method, and compare them with the actual gushing event time marked in the dataset to calculate various evaluation indicators.

[0185] Experimental results.

[0186] The simulation test results were statistically analyzed, and the results are shown in Table 1.

[0187] Table 1 Comparison Results

[0188]

[0189] like Figure 5 The paper presents a time-series graph comparing the risk index predictions of the method of this invention and the control group before and after a typical gushing event. The results show that the risk index curve output by the method of this invention can capture the upward trend of risk earlier and more smoothly than the traditional method, and triggers a medium-level warning about 4.5 minutes before the event occurs, which is significantly earlier than the control group. Figure 6 A time-series graph showing the relationship between model confidence and decision path is presented. The results show that the system's decision path dynamically switches as the model confidence score changes in real time. When the confidence score falls below a preset threshold, the system automatically switches to a parameter correction path to improve the reliability of subsequent predictions. Figure 5 , Figure 6 Together, they demonstrate that the present invention achieves advanced risk warning in terms of time by introducing a confidence-driven adaptive decision-making mechanism.

[0190] Example 2

[0191] like Figure 4 As shown, a grouting project blowout prediction system includes:

[0192] The data acquisition unit is used to collect grouting pressure, grouting flow rate, grout viscosity and micro-vibration event data in real time;

[0193] The data processing unit is connected in communication with the data acquisition unit and is used to calculate the first slurry resistance index and the microseismic event cluster density based on the acquired multi-source data.

[0194] The prediction unit, which is connected in communication with the data processing unit, is used to input the first slurry resistance index and the microseismic event cluster density into the pre-trained machine learning model to generate an initial gushing risk index and simultaneously obtain the model's confidence score for the current prediction.

[0195] The decision control unit communicates with the prediction unit to determine whether the confidence score is lower than a preset threshold, and selects either the execution parameter correction path or the fast decision path based on the judgment result.

[0196] The parameter correction unit is connected in communication with the decision control unit and the data processing unit. When the confidence score is lower than the preset threshold, it generates a dynamic correction factor based on the change characteristics of the confidence score and the initial gushing risk index, and adjusts the slurry resistance index calculation logic to obtain the second slurry resistance index.

[0197] The risk synthesis unit, which is connected in communication with the decision control unit and the prediction unit, is used to re-input the second slurry resistance index and the microseismic event cluster density into the machine learning model to obtain the secondary risk index when executing the parameter correction path, and to synthesize the initial surge risk index and the secondary risk index to obtain the final risk index; when executing the fast decision path, the initial surge risk index is directly used as the final risk index.

[0198] The early warning output unit is connected in communication with the risk synthesis unit and is used to execute graded early warnings based on the final risk index.

[0199] The warning levels are divided into three levels, each corresponding to a different color, to visually reflect the degree of risk:

[0200] Green alert (low risk): indicates that the current risk of gushing is low and construction can proceed normally.

[0201] Yellow Alert (Medium Risk): Audible and visual alarms indicate a potential risk. It is recommended to strengthen monitoring and prepare response measures (adjust parameters).

[0202] Red alert (high risk): Indicates a high risk of gushing, and immediate measures such as work stoppage and evacuation should be taken.

[0203] It is worth noting that all contents not described in detail in this invention are existing technologies and are well known to those skilled in the art.

[0204] Therefore, this invention employs the aforementioned method and system for predicting grouting surges in grouting projects. Based on a machine learning model confidence assessment and online dynamic parameter correction scheme, it achieves adaptive and accurate prediction of the grouting surge risk early warning system. The system autonomously selects between direct early warning and secondary prediction after parameter optimization based on the confidence score output by the model. Furthermore, it performs real-time calibration of key feature parameters when the confidence level is low, thereby improving the accuracy, predictability, and reliability of the early warning.

[0205] 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 predicting gushing during grouting, characterized in that, Includes the following steps: Step S1: Collect data from multiple sources; Step S2: Calculate the first slurry resistance index and the microseismic event cluster density based on multi-source data; First slurry resistance index Based on grouting pressure Grouting flow rate and slurry viscosity calculate: ； in, To be taken from The instantaneous pressure value, Based on the current sliding time window The calculated pressure standard deviation, and These are preset weighting coefficients; Step S3: Input the first slurry resistance index and the microseismic event cluster density into the pre-trained machine learning model to generate the initial gushing risk index, and simultaneously obtain the model's confidence score for the current prediction. Step S4: Determine whether the confidence score is lower than the preset threshold; If the value is below the preset threshold, proceed to step S5; If it is not lower than the preset threshold, skip step S5, use the initial surge risk index directly as the effective risk index to be synthesized, and proceed to step S6. Step S5: Based on the change characteristics of the confidence score and the initial gushing risk index, a dynamic correction factor is generated, and the first slurry resistance index used in step S2 is adjusted to obtain the second slurry resistance index. Step S6: When step S5 has been executed, the second slurry resistance index and the microseismic event cluster density are input into the machine learning model for prediction again to obtain the secondary risk index. The initial gushing risk index and the secondary risk index are then combined to obtain the final risk index. When step S5 has not been executed, the effective risk index to be synthesized is directly used as the final risk index. Step S7: Execute tiered early warning based on the final risk index.

2. The method for predicting gushing during grouting as described in claim 1, characterized in that, In step S1, the multi-source data includes real-time acquisition of grouting pressure, grouting flow rate, grout viscosity, and the occurrence time, three-dimensional coordinates, and energy of micro-seismic events.

3. The method for predicting gushing during grouting as described in claim 2, characterized in that, Step S2 also includes: Based on microseismic event energy Calculate the cluster density of microseismic events : ； in, In the time window Internal and spatial unit volume The sum of the energies of all microseismic events.

4. The method for predicting gushing during grouting as described in claim 3, characterized in that, The specific process of step S3 includes: Step S31: Set the first slurry resistance index at the current moment. and microseismic event cluster density Each with their own past Combining historical values ​​from each time step, a feature vector with temporal context information is constructed. ; Step S32: Transfer the feature vector Input a pre-trained machine learning model, which is a bidirectional long short-term memory network containing a self-attention mechanism; Step S33: The machine learning model outputs two scalars: an initial surge risk index normalized by the sigmoid function. and confidence score Confidence score Based on the model for the same input conduct Output during the second forward propagation variance The calculation is as follows: ； in, It is an integer greater than 1.

5. The method for predicting gushing during grouting as described in claim 4, characterized in that, The specific process of step S4 includes: Step S41: Set the preset threshold ; Step S42: Compare confidence scores With preset threshold ; Step S43, if If the result is low confidence, the parameter correction path is triggered, and a path identifier is generated. Then proceed to step S5; Step S44, if If the result is deemed high confidence, a fast decision-making path is triggered, and a path identifier is generated. The initial surge risk index Assign a value to the effective risk index to be synthesized Then, proceed directly to step S6.

6. The method for predicting gushing during grouting as described in claim 5, characterized in that, Step S5 only applies to the path identifier. The execution process includes: Step S51: Calculate the initial surge risk index In recent Gradient of change within each time step ; Step S52: Based on confidence score and gradient of change Calculate the dynamic correction factor : ； in, For adjustment coefficients, For symbolic functions, This is the gradient limit value; Step S53: Utilize dynamic correction factor Adjust the calculation of the first slurry resistance index Weighting coefficients and The formula is adjusted as follows: ； ； in, and The preset coupling coefficient, and These are the adjusted weighting coefficients; Step S54: Use the adjusted weighting coefficients and and the real-time grouting pressure collected in step S1 Grouting flow rate Real-time slurry viscosity ,according to The calculation formula was recalculated to obtain the second slurry resistance index. .

7. The method for predicting gushing during grouting as described in claim 6, characterized in that, The specific process of step S6 includes: Step S61: Determine the path identifier The value; Step S62, if Then execute: Step S621: Adjust the second slurry resistance index Compared with the microseismic event cluster density calculated based on step S2 at the current time Combine to construct new feature vectors ; Step S622, to Input into a machine learning model to obtain a secondary risk index. ; Step S623: Calculate the final risk index : ； Among them, the composite weight Confidence score The decision is made using the following formula: ； Step S63, if Then the effective risk index to be synthesized will be directly used. As the final risk index ,Right now .

8. The method for predicting gushing during grouting as described in claim 7, characterized in that, The specific process of step S7 includes: Step S71: Preset at least three risk level thresholds. , , ,and The risk index range is divided into low-risk, medium-risk, and high-risk zones. Step S72: Final Risk Index The current risk level is determined by comparing it with the risk level threshold. Step S73: Trigger the audible and visual alarm signal corresponding to the risk level.

9. A gushing prediction system for grouting projects, characterized in that, A method for predicting gushing during grouting as described in any one of claims 1-8, comprising: The data acquisition unit is used to collect grouting pressure, grouting flow rate, grout viscosity and micro-vibration event data in real time; The data processing unit is connected in communication with the data acquisition unit and is used to calculate the first slurry resistance index and the microseismic event cluster density based on the acquired multi-source data. The prediction unit, which is connected in communication with the data processing unit, is used to input the first slurry resistance index and the microseismic event cluster density into the pre-trained machine learning model to generate an initial gushing risk index and simultaneously obtain the model's confidence score for the current prediction. The decision control unit communicates with the prediction unit to determine whether the confidence score is lower than a preset threshold, and selects either the execution parameter correction path or the fast decision path based on the judgment result. The parameter correction unit is connected in communication with the decision control unit and the data processing unit. When the confidence score is lower than the preset threshold, it generates a dynamic correction factor based on the change characteristics of the confidence score and the initial gushing risk index, and adjusts the slurry resistance index calculation logic to obtain the second slurry resistance index. The risk synthesis unit, which is connected in communication with the decision control unit and the prediction unit, is used to re-input the second slurry resistance index and the microseismic event cluster density into the machine learning model to obtain the secondary risk index when executing the parameter correction path, and to synthesize the initial surge risk index and the secondary risk index to obtain the final risk index; when executing the fast decision path, the initial surge risk index is directly used as the final risk index. The early warning output unit is connected in communication with the risk synthesis unit and is used to execute graded early warnings based on the final risk index.

Images