Surge shaft rock mass stability control method capable of supporting while digging

Abstract

The invention discloses a surge shaft rock mass stability control method capable of supporting while digging, and relates to the technical field of crossing of hydraulic engineering and geotechnical engineering. A distributed optical fiber sensor is preset on the periphery of an excavation working face to collect rock mass data, and a rock mass data set is formed; based on a historical data set, a stability judgment model is trained at a central control terminal through a parameter grid optimization method and a random forest algorithm, and a support scheme judgment model is trained through a LightGBM gradient lifting regression algorithm; inputting the rock mass data set into a stability judgment model and outputting a current early warning level; on the basis of the early warning level, combining the stability level and the rock mass data, generating support parameters by a support scheme judgment model; after excavation circulation is completed, data are brought into a sample library to conduct incremental training on the double models, model adaptive learning is achieved, it is ensured that the models are continuously matched along with geological condition changes, high judgment accuracy is maintained in the whole process, and powerful support is provided for safety, high efficiency and economical efficiency of surge shaft excavation construction.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of hydraulic engineering and geotechnical engineering, specifically to a method for controlling the stability of rock masses in surge tanks with on-demand support. Background Technology

[0002] As my country's water conservancy projects expand into areas with high water heads, large flow rates, and complex geological conditions, the demand for surge tanks in large hydropower station water conveyance systems continues to grow. Furthermore, the scale of their excavation is expanding, and the geological environment is becoming increasingly complex, placing higher demands on rock mass stability control during surge tank excavation. As a core structure of the hydropower station's water conveyance system, the surge tank faces a complex geological environment during its excavation: on the one hand, the surrounding rock often exhibits multiple fissures and concentrated ground stress, making excavation disturbances prone to rock mass displacement and fissure propagation; on the other hand, fluctuations in fluid pressure within the well can further exacerbate the deterioration of rock mass mechanical properties, leading to safety accidents such as collapses and water inrushes. However, existing methods for controlling the rock mass of surge tanks still have shortcomings that urgently need to be addressed.

[0003] Chinese patent (publication number CN119146871A) discloses a method for monitoring surrounding rock deformation based on three-dimensional point clouds. This method obtains rock surface deformation data through laser scanning and issues an early warning. However, it does not establish a mapping relationship between monitoring data and support parameters. After the early warning, it is necessary to rely on the experience of engineers to determine the support scheme. The average time from early warning to implementation exceeds 8 hours, which cannot meet the timeliness requirement of "supporting as soon as excavation".

[0004] Chinese patent (publication number CN120260230A) discloses a method for controlling the stability of the surrounding rock of a pressure regulating chamber in a high-seismic zone. The method optimizes the support structure through numerical simulation. However, it uses fixed support parameters and cannot dynamically adjust them according to the rock stress and fracture development rate collected in real time during the excavation process. This can easily lead to insufficient support strength in Class V fractured surrounding rock or material waste in Class II intact surrounding rock.

[0005] Chinese patent (publication number CN118542365A) discloses a method for predicting the surrounding rock grade based on LightGBM. Although it introduces machine learning algorithms, the model training only relies on historical static engineering data and does not include the current excavation cycle data of the project in the sample library for incremental training. As a result, when geological conditions change, the prediction accuracy of the model drops from 92% to 68%, and it cannot continuously guarantee the control accuracy.

[0006] Existing rock mass stability control technologies for pressure regulating wells still cannot achieve integrated control that includes real-time monitoring, dynamic modeling, precise support, and multi-factor coordination. There is an urgent need to propose a technical solution that can overcome these shortcomings. Summary of the Invention

[0007] Based on the above-mentioned technical problems, this application discloses a method for controlling the stability of rock mass in a surge tank with on-demand support, specifically as follows:

[0008] Distributed fiber optic sensors are pre-installed around the excavation face of the surge tank well to synchronously collect stress values, displacement, and fracture development rates of the rock mass in the excavation area, forming a rock mass dataset; historical datasets are constructed using characteristic parameters and stability levels from historical projects.

[0009] Based on historical datasets, a rock mass stability judgment model was trained at the central control terminal by combining the parametric grid optimization method with the random forest algorithm, and a support scheme judgment model was trained by the LightGBM gradient boosting regression algorithm.

[0010] Based on the rock mass dataset, the current rock mass stability level is output through the rock mass stability judgment model, including the stable level, critical level and instability warning level. When the rock mass is judged to be at the critical level or instability warning level, the dynamic control mechanism is activated.

[0011] Based on the current rock mass stability level and rock mass data, the optimal support parameters are generated through the support scheme determination model. The initial support thickness of the quick-setting shotcrete, the spacing of the hollow grouting anchors, and the grouting pressure of the anchor channel are determined. The opening degree of the preset two-stage pressure regulating valve group at the bottom of the pressure regulating well is controlled to regulate the fluid pressure in the well.

[0012] After completing the excavation cycle, the central control terminal will include the data from this cycle into the model training sample library to incrementally train the random forest classification model and gradient boosting regression model until all the surge tank excavation operations are completed.

[0013] Preferably, the preset distributed optical fiber sensor specifically refers to: a preset sensor deployment depth multiple. Multiple rings of distributed fiber optic sensors are deployed at equal angular intervals along the excavation outline of the surge tank. These sensors simultaneously collect stress values, displacement, and fracture development rates of the rock mass in the excavation area. The depth of each sensor ring is equal to the excavation radius. The distributed fiber optic sensor acquires rock mass strain data using BOTDR technology and calculates stress values ​​using the following formula:

[0014]

[0015] in, This represents the rock mass stress value. The elastic modulus of the rock mass. The rock mass strain values ​​are collected by distributed fiber optic sensors; the displacement is calculated by integrating the strain distribution curve of the fiber optic sensor; and the fracture development rate is determined by the ratio of the stress change gradient to the time difference in adjacent acquisition cycles.

[0016] Preferably, the step of training the rock mass stability judgment model by fusing the parametric grid optimization method with the random forest algorithm specifically involves:

[0017] The feature parameters and rock mass stress anisotropy coefficients in the historical dataset are standardized to construct the training feature set, as shown in the formula:

[0018]

[0019] in, Let j be the standardized feature parameter of the i-th sample. Let j be the feature parameter of the original i-th sample. Let be the mean of the j-th feature parameter. Let $\frac{j}{j}$ be the standard deviation of the $j$-th feature parameter.

[0020] The parameter grid optimization method is used to optimize the key parameters of the random forest algorithm, including the number of decision trees, the maximum depth of decision trees, and the minimum number of samples required for node splits. The accuracy of 5-fold cross-validation is used as the evaluation index to select the optimal parameter combination.

[0021] A random forest classification model is constructed based on the optimal parameter combination. The model output is the current rock mass stability level. During model training, core influencing features are selected based on feature importance, using the following formula:

[0022]

[0023] in, Score the importance of the j-th feature. The reduction in the Gini coefficient when splitting the j-th feature in the k-th decision tree. For the first The total number of features.

[0024] Preferably, the rock mass stress anisotropy coefficient is specifically defined as the coefficient obtained through the maximum principal stress of the rock mass. Minimum principal stress of rock mass and the mean principal stress of the rock mass Calculate the rock mass stress anisotropy coefficient The formula is:

[0025]

[0026] During model construction, a bagging method should be used for sample sampling, with the sampling ratio being a preset proportion of the total sample size. The number of features selected for each decision tree is set to the total number of features. ,in The total number of features is controlled by the purity threshold of the leaf nodes.

[0027] Preferably, the LightGBM gradient boosting regression algorithm is used to train the support scheme determination model, specifically by introducing the coupling coefficient of rock mass stress and displacement as a key input feature, as shown in the formula:

[0028]

[0029] in, The coupling coefficient is... This represents the change in rock mass stress between adjacent data collection periods. This represents the change in rock mass displacement between adjacent data collection periods. Density of the rock mass;

[0030] During the training process of the support scheme determination model, the gradient descent method is used to minimize the loss function, and the formula is:

[0031]

[0032] in, For loss function, The number of training samples. Let i be the actual support parameters for the i-th sample. These are the support parameters predicted by the model; and feature binning is used during model training to divide continuous features into multiple preset intervals, thereby improving the model's ability to fit nonlinear relationships.

[0033] Preferably, the dynamic control mechanism includes a three-level response strategy, specifically: based on the operating conditions of the pressure regulating well, the opening degree of the first-level valve of the two-stage pressure regulating valve group is preset. , , Rock mass stress growth rate threshold ;

[0034] When the condition is determined to be critical, the first-level response is initiated, and the opening of the first-level valve of the two-stage pressure regulating valve group is adjusted to... The secondary valve remains closed, and the initial support thickness of the quick-setting shotcrete is increased by a certain value based on the model prediction.

[0035] When it is determined to be at the instability warning level and the rock mass stress growth rate is less than At that time, the secondary response is initiated, and the opening of the primary valve is adjusted to... The opening of the secondary valve is adjusted to... Furthermore, the spacing between hollow grouting anchor rods was reduced, and the grouting pressure was increased;

[0036] When it is determined to be at the instability warning level and the rock mass stress growth rate is not less than At that time, a three-level response is initiated: the opening of the first-level valve is adjusted to... The opening of the secondary valve is adjusted to... Excavation work was suspended and advanced support was implemented, which adopted pipe roof support.

[0037] Preferably, the step of generating optimal support parameters through the support scheme determination model specifically involves determining the initial support thickness of the quick-setting shotcrete based on the rock mass fracture development rate, using the following formula:

[0038]

[0039] in, The corrected initial support thickness, To determine the initial support thickness predicted by the support scheme model, For correction factor, The rate of rock mass fracture development;

[0040] Based on the rock mass pull-out resistance requirements, the spacing of the hollow grouting anchors and the grouting pressure of the anchor channels are determined using the following formula:

[0041]

[0042] in, For the pull-out resistance of the anchor bolt, The diameter of the anchor rod. This refers to the anchorage length of the anchor bolt. The bond strength between the rock mass and the mortar;

[0043] The grouting volume is calculated based on the rock mass porosity, using the following formula:

[0044]

[0045] in, This refers to the grouting volume for a single anchor bolt. This represents the rock mass volume corresponding to the anchorage section of the anchor bolt. The porosity of the rock mass This is the grouting filling coefficient.

[0046] Preferably, the opening adjustment of the two-stage pressure regulating valve group needs to be coordinated with the fluid pressure in the well, and the pressure control formula is:

[0047]

[0048] in, The target fluid pressure inside the well. This represents the initial fluid pressure inside the well. This is the pressure regulation coefficient. This represents the current maximum stress value of the rock mass;

[0049] During the coordination process, the actual pressure inside the well is collected in real time through pressure sensors. ,when Preset well pressure difference threshold At that time, the opening degree of the two-stage pressure regulating valve group is automatically adjusted.

[0050] Preferably, the incremental training is performed by setting a preset excavation cycle start threshold. Each completed After each excavation cycle, the rock mass dataset from this cycle, along with the corresponding support scheme parameters and stability level determination results, are used to form new training samples. Incremental learning algorithms are then used to update the parameters of the random forest classification model and the LightGBM gradient boosting regression model.

[0051] In the random forest model, incremental updates combine adding new decision trees with retaining the optimal decision tree, and preset an importance threshold for each decision tree. Preserve the importance of features in historical training before scoring The number of new decision trees is equal to the number of initial decision trees. ;

[0052] The incremental update of the LightGBM model adopts a hot start mode, which initializes the new model based on the parameters of the historical model and only performs a small number of rounds of iterative training on the newly added samples.

[0053] Compared with the prior art, the technical solution of this application has the following technical effects:

[0054] This invention deploys distributed fiber optic sensors around the excavation face to collect data. When the critical level or instability warning level is reached, a dynamic control mechanism is activated. This solves the problems of existing technologies where monitoring and support are disconnected and response is delayed due to reliance on human experience. It reduces the time from warning to support plan generation from more than 8 hours to within 30 minutes, fully meeting the timeliness requirement of "excavation and support as needed" and effectively avoiding the risk of rock mass instability during the delay window.

[0055] This invention generates optimal support parameters such as the initial thickness of quick-setting shotcrete and the spacing of hollow grouting anchors from the support scheme determination model. The parameters can be dynamically adjusted according to the real-time collected rock mass stress and fracture development rate, avoiding the problems of insufficient support for Class V surrounding rock and waste of materials for Class II surrounding rock due to fixed parameters in the existing technology. This improves the matching degree of support strength and reduces the cost of support materials.

[0056] This invention incorporates data into a sample library for incremental model training after each excavation cycle, overcoming the shortcomings of existing technologies that rely on static data and suffer from accuracy degradation when geological conditions change. Even in scenarios where geological conditions transition from granite to sandstone and shale, it can maintain high model prediction accuracy and ensure the stability control precision throughout the excavation process.

[0057] The technical solution of this invention, which coordinates the regulation of fluid pressure in the well with support parameters, can effectively suppress the problem of fluid pressure seepage along fractures in water-rich pressure regulating wells, and significantly improve the support effect and engineering safety in complex scenarios such as water-rich and fractured surrounding rock.

[0058] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the preferred embodiments of this application are described in detail below with reference to the accompanying drawings.

[0059] The above and other objects, advantages and features of this application will become more apparent to those skilled in the art from the following detailed description of specific embodiments in conjunction with the accompanying drawings. Attached Figure Description

[0060] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In all drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0061] Based on the description of the figures and their corresponding technical content in the document, the titles of the figures are as follows:

[0062] Figure 1 This is a flowchart illustrating the overall process of a rock mass stability control method for a surge tank well that requires ongoing support during excavation.

[0063] Figure 2 This is a general framework diagram of a rock mass stability control method for a surge tank well that requires ongoing support during excavation.

[0064] Figure 3 A neural network diagram for training a rock mass stability assessment model using the parametric grid optimization method combined with the random forest algorithm;

[0065] Figure 4 Neural network diagram for training the support scheme determination model using the LightGBM gradient boosting regression algorithm;

[0066] Figure 5 This is a diagram showing the overall experimental setup of this method in a surge tank project of a pumped storage power station;

[0067] Figure 6 This is a graph showing the rock mass displacement data for each excavation cycle in an experiment at a pumped storage power station surge tank.

[0068] Figure 7 This is a comparison chart of the average rock mass displacement in different months during an experiment at a pumped storage power station surge tank.

[0069] Figure 8 This is a comparison chart of the accuracy rates of judgments in different months during an experiment on a surge tank project at a pumped storage power station. Detailed Implementation

[0070] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. In the following description, specific details such as specific configurations and components are provided merely to help fully understand the embodiments of this application. Therefore, those skilled in the art should understand that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this application. In addition, for clarity and brevity, descriptions of known functions and structures are omitted in the embodiments.

[0071] It should be understood that the phrase "an embodiment" or "this embodiment" throughout the specification means that a specific feature, structure, or characteristic related to the embodiment is included in at least one embodiment of this application. Therefore, "an embodiment" or "this embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments.

[0072] Furthermore, reference numerals and / or letters may be repeated in different examples within this application. Such repetition is for the purpose of simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or settings discussed.

[0073] In this article, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can mean: A exists alone, B exists alone, and A and B exist simultaneously. The term " / and" in this article describes another type of relationship between related objects, indicating that two relationships can exist. For example, A / and B can mean: A exists alone, and A and B exist alone. In addition, the character " / " in this article generally indicates that the related objects before and after it are in an "or" relationship.

[0074] In this article, the term "at least one" is merely a description of the relationship between related objects, indicating that there can be three relationships. For example, "at least one of A and B" can mean: A exists alone, A and B exist simultaneously, or B exists alone.

[0075] It should also be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion.

[0076] Example 1 mainly describes a method for controlling the stability of rock mass in a surge tank well with ongoing support during excavation, such as... Figures 1-2 As shown, it specifically includes:

[0077] Distributed fiber optic sensors are pre-installed around the excavation face of the surge tank well to synchronously collect stress values, displacement, and fracture development rates of the rock mass in the excavation area, forming a rock mass dataset; historical datasets are constructed using characteristic parameters and stability levels from historical projects.

[0078] Based on historical datasets, a rock mass stability judgment model was trained at the central control terminal by combining the parametric grid optimization method with the random forest algorithm, and a support scheme judgment model was trained by the LightGBM gradient boosting regression algorithm.

[0079] Based on the rock mass dataset, the current rock mass stability level is output through the rock mass stability judgment model, including the stable level, critical level and instability warning level. When the rock mass is judged to be at the critical level or instability warning level, the dynamic control mechanism is activated.

[0080] Based on the current rock mass stability level and rock mass data, the optimal support parameters are generated through the support scheme determination model. The initial support thickness of the quick-setting shotcrete, the spacing of the hollow grouting anchors, and the grouting pressure of the anchor channel are determined. The opening degree of the preset two-stage pressure regulating valve group at the bottom of the pressure regulating well is controlled to regulate the fluid pressure in the well.

[0081] After completing the excavation cycle, the central control terminal will include the data from this cycle into the model training sample library to incrementally train the random forest classification model and gradient boosting regression model until all the surge tank excavation operations are completed.

[0082] Furthermore, the preset distributed optical fiber sensor specifically refers to: a preset sensor deployment depth multiple. Multiple rings of distributed fiber optic sensors are deployed at equal angular intervals along the excavation outline of the surge tank. These sensors simultaneously collect stress values, displacement, and fracture development rates of the rock mass in the excavation area. The depth of each sensor ring is equal to the excavation radius. The distributed fiber optic sensor acquires rock mass strain data using BOTDR technology and calculates stress values ​​using the following formula:

[0083]

[0084] in, This represents the rock mass stress value. The elastic modulus of the rock mass. The rock mass strain values ​​are collected by distributed fiber optic sensors; the displacement is calculated by integrating the strain distribution curve of the fiber optic sensor; and the fracture development rate is determined by the ratio of the stress change gradient to the time difference in adjacent acquisition cycles.

[0085] Furthermore, the method of training the rock mass stability judgment model by fusing the parameter grid optimization method with the random forest algorithm specifically involves: standardizing the feature parameters and rock mass stress anisotropy coefficients in the historical dataset to construct a feature set to be trained, as shown in the formula:

[0086]

[0087] in, Let j be the standardized feature parameter of the i-th sample. Let j be the feature parameter of the original i-th sample. Let be the mean of the j-th feature parameter. Let $\frac{j}{j}$ be the standard deviation of the $j$-th feature parameter.

[0088] The key parameters of the random forest algorithm are optimized using the parametric grid optimization method. The parameter range is set as follows: number of decision trees 50-200, maximum depth of decision trees 5-20 layers, minimum number of samples for node splitting 2-10. The accuracy of 5-fold cross-validation is used as the evaluation index. The optimal parameters are selected by traversing the parameter combinations in the grid.

[0089] A random forest classification model is constructed based on the optimal parameter combination. The model output is the current rock mass stability level. During model training, core influencing features are selected based on feature importance, using the following formula:

[0090]

[0091] in, Score the importance of the j-th feature. The reduction in the Gini coefficient when splitting the j-th feature in the k-th decision tree. For the first The total number of features.

[0092] Furthermore, the rock mass stress anisotropy coefficient is specifically defined as: the coefficient obtained through the maximum principal stress of the rock mass. Minimum principal stress of rock mass and the mean principal stress of the rock mass Calculate the rock mass stress anisotropy coefficient The formula is:

[0093]

[0094] During model construction, a bagging method should be used for sample sampling, with the sampling ratio being a preset proportion of the total sample size. The number of features selected for each decision tree is set to the total number of features. ,in The total number of features is controlled by the purity threshold of the leaf nodes.

[0095] Furthermore, the LightGBM gradient boosting regression algorithm is used to train the support scheme determination model, specifically by introducing the coupling coefficient of rock mass stress and displacement as a key input feature, as shown in the formula:

[0096]

[0097] in, The coupling coefficient is... This represents the change in rock mass stress between adjacent data collection periods. This represents the change in rock mass displacement between adjacent data collection periods. Density of the rock mass;

[0098] During the training process of the support scheme determination model, the gradient descent method is used to minimize the loss function, and the formula is:

[0099]

[0100] in, For loss function, The number of training samples. Let i be the actual support parameters for the i-th sample. The support parameters are predicted by the model; and feature binning is used during model training to divide continuous features into multiple preset intervals to improve the model's ability to fit nonlinear relationships. The training iteration rounds are set to 100-150 rounds, and the learning rate is set to 0.05.

[0101] Furthermore, the dynamic control mechanism includes a three-level response strategy, specifically: based on the operating conditions of the pressure regulating well, the opening degree of the first-level valve of the two-stage pressure regulating valve group is preset. , , Rock mass stress growth rate threshold ;

[0102] When the condition is determined to be critical, the first-level response is initiated, and the opening of the first-level valve of the two-stage pressure regulating valve group is adjusted to... The secondary valve remains closed, and the initial support thickness of the quick-setting shotcrete is increased by a certain value based on the model prediction.

[0103] When it is determined to be at the instability warning level and the rock mass stress growth rate is less than At that time, the secondary response is initiated, and the opening of the primary valve is adjusted to... The opening of the secondary valve is adjusted to... Furthermore, the spacing between hollow grouting anchor rods was reduced, and the grouting pressure was increased;

[0104] When it is determined to be at the instability warning level and the rock mass stress growth rate is not less than At that time, a three-level response is initiated: the opening of the first-level valve is adjusted to... The opening of the secondary valve is adjusted to... Excavation work was suspended and advanced support was implemented, which adopted pipe roof support.

[0105] Furthermore, the specific method for generating optimal support parameters through the support scheme determination model is as follows: determining the initial support thickness of the quick-setting shotcrete based on the rock mass fracture development rate, using the following formula:

[0106]

[0107] in, The corrected initial support thickness, To determine the initial support thickness predicted by the support scheme model, For correction factor, The rate of rock mass fracture development;

[0108] Based on the rock mass pull-out resistance requirements, the spacing of the hollow grouting anchors and the grouting pressure of the anchor channels are determined using the following formula:

[0109]

[0110] in, For the pull-out resistance of the anchor bolt, The diameter of the anchor rod. This refers to the anchorage length of the anchor bolt. The bond strength between the rock mass and the mortar;

[0111] The grouting volume is calculated based on the rock mass porosity, using the following formula:

[0112]

[0113] in, This refers to the grouting volume for a single anchor bolt. This represents the rock mass volume corresponding to the anchorage section of the anchor bolt. The porosity of the rock mass This is the grouting filling coefficient.

[0114] Furthermore, the opening adjustment of the two-stage pressure regulating valve group needs to be coordinated with the fluid pressure inside the well. The pressure control formula is as follows:

[0115]

[0116] in, The target fluid pressure inside the well. This represents the initial fluid pressure inside the well. This is the pressure regulation coefficient. This represents the current maximum stress value of the rock mass;

[0117] During the coordination process, the actual pressure inside the well is collected in real time through pressure sensors. ,when Preset well pressure difference threshold At that time, the opening degree of the two-stage pressure regulating valve group is automatically adjusted.

[0118] Furthermore, the specific method of incremental training is as follows: preset the excavation cycle start threshold. Each completed After each excavation cycle, the rock mass dataset from this cycle, along with the corresponding support scheme parameters and stability level determination results, are used to form new training samples. Incremental learning algorithms are then used to update the parameters of the random forest classification model and the LightGBM gradient boosting regression model.

[0119] In the random forest model, incremental updates combine adding new decision trees with retaining the optimal decision tree, and preset an importance threshold for each decision tree. Preserve the importance of features in historical training before scoring The number of new decision trees is equal to the number of initial decision trees. ;

[0120] The incremental update of the LightGBM model adopts a hot start mode, which initializes the new model based on the parameters of the historical model and performs a small number of iterations of training on the newly added samples, reducing training time while ensuring that the model is adapted to the latest geological conditions.

[0121] This embodiment details a rock mass stability control method for surge tank wells with ongoing support during excavation. A rock mass dataset is constructed by pre-setting distributed fiber optic sensors around the excavation face to collect rock mass stress, displacement, and fracture development rates. A historical dataset is then built based on historical engineering characteristic parameters and stability levels. Based on this historical dataset, a stability judgment model is trained at the central control terminal using a parameter grid optimization method combined with a random forest algorithm, and a support scheme determination model is trained using a LightGBM gradient boosting regression algorithm. The rock mass dataset is input into the stability judgment model, which outputs stable, critical, and instability warning levels. A three-level dynamic control mechanism is activated at the critical or instability warning level. Combining the stability level and rock mass data, the support scheme determination model generates support parameters and adjusts the fluid pressure within the well. After the excavation cycle is completed, the data is incorporated into a sample library for incremental training of the two models, enabling adaptive learning and ensuring continuous adaptation to changes in geological conditions. This maintains high judgment accuracy throughout the process, providing strong support for the safety, efficiency, and economy of surge tank well excavation.

[0122] Example 2 describes in detail the experimental process of this method in a surge tank project of a pumped storage power station, as follows:

[0123] like Figure 5 As shown, the experimental surge tank was buried at a depth of 280m, surrounded by Class IV tuff with a fracture density of 9 fractures / m and a groundwater pressure of 0.5MPa. The excavation system adopted a step-type excavation process, with an excavation cycle set at 3m / cycle. It was equipped with a small hydraulic tunneling machine with an excavation rate of 0.8m / h.

[0124] Eight distributed FBG-600 fiber optic sensors were deployed in a 360° ring around the excavation face. The sensors were buried at a depth of 2m and spaced 1.5m apart. They collected data on rock stress, displacement and fracture development rate in real time, and the data was transmitted to the main control system through a fiber optic demodulator.

[0125] The main control system has a built-in random forest stability judgment model and a LightGBM support scheme judgment model trained based on 200 sets of historical engineering data. It is equipped with a quick-setting shotcrete equipment, a hollow grouting anchor drilling rig and an in-well water pressure regulating pump. It can adjust the support parameters and water pressure in real time according to the model output.

[0126] Stress, displacement, and crack rate data were collected every 10 seconds. Support cost data were statistically analyzed for each cycle. One excavation cycle was conducted per day for a three-month experimental test. At the same time, existing methods, such as three-dimensional laser scanning and LightGBM static data method, were designed as control groups in other areas of the project and conducted as control experiments for a three-month period.

[0127]

[0128] As shown in the table above, the average displacement and the average displacement over three months in the area controlled by this method are both lower than those of the other two existing methods. Furthermore, the support cost of this method is 107,500 yuan lower than that of the LightGBM static data method and 172,500 yuan lower than that of the three-dimensional laser scanning method. The total delay time is only 12.5 hours, and the average support strength matching degree is also higher than that of the other two methods. The average accuracy rate of judgment reaches 94%.

[0129] according to Figure 6 It can be seen that the rock mass displacement controlled by this method is generally smaller than that of the other two methods, which proves the reliable control capability of the scheme;

[0130] according to Figure 7 It can be seen that, as geological conditions deteriorate, the cumulative displacement of this method is still controlled at 10.2 mm, and the average change in rock mass displacement is only 1.3 mm, which is much lower than the other two methods, proving the outstanding stability of this scheme in long-term control.

[0131] according to Figure 8 As can be seen, the accuracy of the proposed method remained at 88% in the third month, while the accuracy of the other two methods dropped to 71% and 65% respectively. This proves that the incremental training mechanism of this application effectively resists the accuracy decay caused by geological deterioration and ensures the accuracy of control throughout the process.

[0132] This embodiment details the experimental process of this method in the surge tank engineering of a pumped storage power station. Through one excavation cycle per day for three months, the experiment fully verified the significant advantages of this method in rock mass control, support cost, response time, support strength matching degree and judgment accuracy. It also proved the cost and efficiency advantages, accuracy and versatility advantages, and long-term stability advantages of this method.

[0133] The above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. For those skilled in the art, the present invention can have various modifications and variations. Any changes, modifications, substitutions, integrations, and parameter changes made to these embodiments within the spirit and principles of the present invention, without departing from the principles and spirit of the present invention, through conventional substitutions or to achieve the same function, fall within the scope of protection of the present invention.

Claims

1. A method for controlling the stability of rock mass in a surge tank well with ongoing support during excavation, characterized in that, include: Distributed fiber optic sensors are pre-installed around the excavation face of the surge tank well to synchronously collect stress values, displacement, and fracture development rates of the rock mass in the excavation area, forming a rock mass dataset; historical datasets are constructed using characteristic parameters and stability levels from historical projects. Based on historical datasets, a rock mass stability judgment model was trained at the central control terminal by combining the parametric grid optimization method with the random forest algorithm, and a support scheme judgment model was trained by the LightGBM gradient boosting regression algorithm. Based on the rock mass dataset, the current rock mass stability level is output through the rock mass stability judgment model, including the stable level, critical level and instability warning level. When the rock mass is judged to be at the critical level or instability warning level, the dynamic control mechanism is activated. Based on the current rock mass stability level and rock mass data, the optimal support parameters are generated through the support scheme determination model. The initial support thickness of the quick-setting shotcrete, the spacing of the hollow grouting anchors, and the grouting pressure of the anchor channel are determined. The opening degree of the preset two-stage pressure regulating valve group at the bottom of the pressure regulating well is controlled to regulate the fluid pressure in the well. After completing the excavation cycle, the central control terminal will include the data from this cycle into the model training sample library to incrementally train the random forest classification model and gradient boosting regression model until all the surge tank excavation operations are completed.

2. The method according to claim 1, characterized in that, The preset distributed fiber optic sensor specifically refers to: a preset sensor deployment depth multiple. Multiple rings of distributed fiber optic sensors are deployed at equal angular intervals along the excavation outline of the surge tank. These sensors simultaneously collect stress values, displacement, and fracture development rates of the rock mass in the excavation area. The depth of each sensor ring is equal to the excavation radius. The distributed fiber optic sensor acquires rock mass strain data using BOTDR technology and calculates stress values ​​using the following formula: ； in, This represents the rock mass stress value. The elastic modulus of the rock mass. The rock mass strain values ​​are collected by distributed fiber optic sensors; the displacement is calculated by integrating the strain distribution curve of the fiber optic sensor; and the fracture development rate is determined by the ratio of the stress change gradient to the time difference in adjacent acquisition cycles.

3. The method according to claim 1, characterized in that, The process of training a rock mass stability assessment model by fusing a parametric grid optimization method with a random forest algorithm is as follows: The feature parameters and rock mass stress anisotropy coefficients in the historical dataset are standardized to construct the training feature set, as shown in the formula: ; in, Let j be the standardized feature parameter of the i-th sample. Let j be the feature parameter of the original i-th sample. Let be the mean of the j-th feature parameter. Let $\frac{j}{j}$ be the standard deviation of the $j$-th feature parameter. The parameter grid optimization method is used to optimize the key parameters of the random forest algorithm, including the number of decision trees, the maximum depth of decision trees, and the minimum number of samples required for node splits. The accuracy of 5-fold cross-validation is used as the evaluation index to select the optimal parameter combination. A random forest classification model is constructed based on the optimal parameter combination. The model output is the current rock mass stability level. During model training, core influencing features are selected based on feature importance, using the following formula: ; in, Score the importance of the j-th feature. The reduction in the Gini coefficient when splitting the j-th feature in the k-th decision tree. For the first The total number of features.

4. The method according to claim 2, characterized in that, The rock mass stress anisotropy coefficient is specifically defined as the coefficient obtained through the maximum principal stress of the rock mass. Minimum principal stress of rock mass and the mean principal stress of the rock mass Calculate the rock mass stress anisotropy coefficient The formula is: ; During model construction, a bagging method should be used for sample sampling, with the sampling ratio being a preset proportion of the total sample size. The number of features selected for each decision tree is set to the total number of features. ,in The total number of features is controlled by the purity threshold of the leaf nodes.

5. The method according to claim 1, characterized in that, The LightGBM gradient boosting regression algorithm is used to train the support scheme determination model. Specifically, it introduces the coupling coefficient between rock mass stress and displacement as a key input feature, as shown in the formula: ; in, The coupling coefficient is... This represents the change in rock mass stress between adjacent data collection periods. This represents the change in rock mass displacement between adjacent data collection periods. Density of the rock mass; During the training process of the support scheme determination model, the gradient descent method is used to minimize the loss function, and the formula is: ; in, For loss function, The number of training samples. Let i be the actual support parameters for the i-th sample. These are the support parameters predicted by the model; and feature binning is used during model training to divide continuous features into multiple preset intervals, thereby improving the model's ability to fit nonlinear relationships.

6. The method according to claim 1, characterized in that, The dynamic control mechanism includes a three-level response strategy, specifically: based on the operating conditions of the pressure regulating well, the opening degree of the first-level valve of the two-stage pressure regulating valve group is preset. , , Rock mass stress growth rate threshold ; When the condition is determined to be critical, the first-level response is initiated, and the opening of the first-level valve of the two-stage pressure regulating valve group is adjusted to... The secondary valve remains closed, and the initial support thickness of the quick-setting shotcrete is increased by a certain value based on the model prediction. When it is determined to be at the instability warning level and the rock mass stress growth rate is less than At that time, the secondary response is initiated, and the opening of the primary valve is adjusted to... The opening of the secondary valve is adjusted to... Furthermore, the spacing between hollow grouting anchor rods was reduced, and the grouting pressure was increased; When it is determined to be at the instability warning level and the rock mass stress growth rate is not less than At that time, a three-level response is initiated: the opening of the first-level valve is adjusted to... The opening of the secondary valve is adjusted to... Excavation work was suspended and advanced support was implemented, which adopted pipe roof support.

7. The method according to claim 1, characterized in that, The optimal support parameters are generated through the support scheme determination model, specifically by determining the initial support thickness of the rapid-setting shotcrete based on the rock mass fracture development rate, using the following formula: ; in, The corrected initial support thickness, To determine the initial support thickness predicted by the support scheme model, For correction factor, The rate of rock mass fracture development; Based on the rock mass pull-out resistance requirements, the spacing of the hollow grouting anchors and the grouting pressure of the anchor channels are determined using the following formula: ; in, For the pull-out resistance of the anchor bolt, The diameter of the anchor rod. This refers to the anchorage length of the anchor bolt. The bond strength between the rock mass and the mortar; The grouting volume is calculated based on the rock mass porosity, using the following formula: ; in, This refers to the grouting volume for a single anchor bolt. This represents the rock mass volume corresponding to the anchorage section of the anchor bolt. The porosity of the rock mass This represents the grouting filling coefficient.

8. The method according to claim 1, characterized in that, The opening adjustment of the two-stage pressure regulating valve group needs to be coordinated with the fluid pressure in the well. The pressure control formula is as follows: ; in, The target fluid pressure inside the well. This represents the initial fluid pressure inside the well. This is the pressure regulation coefficient. This represents the current maximum stress value of the rock mass; During the coordination process, the actual pressure inside the well is collected in real time through pressure sensors. ,when Preset well pressure difference threshold At that time, the opening degree of the two-stage pressure regulating valve group is automatically adjusted.

9. The method according to claim 1, characterized in that, The specific method of incremental training is as follows: preset the excavation cycle start threshold. Each completed After each excavation cycle, the rock mass dataset from this cycle, along with the corresponding support scheme parameters and stability level determination results, are used to form new training samples. Incremental learning algorithms are then used to update the parameters of the random forest classification model and the LightGBM gradient boosting regression model. In the random forest model, incremental updates combine adding new decision trees with retaining the optimal decision tree, and preset an importance threshold for each decision tree. Preserve the importance of features in historical training before scoring The number of new decision trees is equal to the number of initial decision trees. ; The incremental update of the LightGBM model adopts a hot start mode, which initializes the new model based on the parameters of the historical model and only performs a small number of rounds of iterative training on the newly added samples.

Images