An underground engineering surrounding rock deformation rapid prediction and intelligent early warning method and system

Abstract

The present application relates to the technical field of underground engineering safety monitoring, surrounding rock deformation prediction, construction risk early warning and data-physical fusion modeling, and specifically discloses a kind of underground engineering surrounding rock deformation rapid prediction and intelligent early warning method and system, comprising: obtaining underground engineering basic data and construction monitoring data and pretreatment;In the design stage or construction initial stage, carry out multi-working condition numerical simulation calculation, construct numerical simulation parameter-response database;Combining database, pretreated construction monitoring data and surrounding rock physical constraint, train physical constraint neural information network;In the construction process, rely on real-time monitoring data to correct the model, input current construction parameters and real-time monitoring data, obtain the surrounding rock response prediction result of current and subsequent construction stage;After intelligent risk early warning model processing output risk state, generate construction dynamic control scheme;The present application can realize surrounding rock response prediction and risk state output, provide reference for construction scheme dynamic adjustment.

Description

Technical Field

[0001] This invention relates to the fields of underground engineering safety monitoring, surrounding rock deformation prediction, construction risk early warning and data-physical fusion modeling technology, and in particular to a method and system for rapid prediction and intelligent early warning of surrounding rock deformation in underground engineering. Background Technology

[0002] During underground engineering construction, the surrounding rock is susceptible to phenomena such as crown settlement, perimeter convergence, floor heave, increased deep displacement, abnormal stress on the support structure, and expansion of the plastic zone due to the effects of excavation unloading, stress redistribution, groundwater action, and support aging. If these deformations continue, they may induce engineering risks such as surrounding rock instability, support failure, collapse, roof fall, spalling, and water / mud inrush. Therefore, it is necessary to rapidly predict the deformation of the surrounding rock in underground engineering projects and conduct risk warnings and dynamic adjustments to construction plans accordingly.

[0003] Existing methods for predicting and warning of surrounding rock deformation mainly include monitoring threshold discrimination, numerical simulation analysis, and data-driven prediction. Monitoring threshold discrimination methods rely on indicators such as cumulative deformation, deformation rate, or deformation acceleration. While simple, these methods are difficult to adapt to different surrounding rock grades, construction stages, and support conditions. Numerical simulation methods can obtain surrounding rock displacement, stress, plastic zone, and support response, but complex model building and computational costs are high, making it difficult to directly meet the rapid prediction needs during construction. Data-driven prediction methods offer fast computation speeds, but when construction monitoring data is limited or abnormal samples are insufficient, model training stability and physical consistency are inadequate.

[0004] The existing technology has the following shortcomings: First, the computational cost of complex numerical simulation is high, and relying solely on construction monitoring data to train the prediction model is easily limited by the number of monitoring points and the number of abnormal samples, making it difficult to balance prediction efficiency and model generalization ability; Second, the existing methods have not yet formed a technical process for using numerical simulation parameter-response database, a small amount of construction monitoring data and physical constraints together for training and correction of physical constraint neural information network, and the early warning results are also difficult to output the main risk type, high-risk location and construction dynamic control scheme at the same time. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method for rapid prediction and intelligent early warning of surrounding rock deformation in underground engineering based on data-physical fusion. This method solves the problems of high computational costs in complex numerical simulations, insufficient stability of pure data-driven model training under limited construction monitoring data, lack of physical constraints in existing prediction models, and difficulty in supporting dynamic adjustments to construction plans with traditional early warning results.

[0006] On the one hand, a method for rapid prediction and intelligent early warning of surrounding rock deformation in underground engineering is provided, including: During the design or early construction phase, multi-condition numerical simulations are conducted to construct a numerical simulation parameter-response database; the numerical simulation parameter-response database, preprocessed construction monitoring data, and surrounding rock physical constraints are used to train a physical constraint neural network. During construction, real-time construction monitoring data is used to correct the trained physical constraint neural network, resulting in a corrected physical constraint neural network. The parameters of the current construction stage and real-time construction monitoring data are input into the corrected physical constraint neural network to obtain the surrounding rock response prediction results for the current and subsequent construction stages. Based on the surrounding rock response prediction results and the pre-processed construction monitoring data, a risk feature vector is constructed, input into the intelligent risk early warning model, and the risk status output result is obtained. A dynamic construction control plan is generated based on the risk status output.

[0007] Preferably, before constructing the numerical simulation parameter-response database, the method further includes: acquiring basic data and construction monitoring data of the underground engineering project, and preprocessing the basic data and construction monitoring data to form a standardized dataset.

[0008] Preferably, during the design phase or early construction phase, multi-condition numerical simulations are conducted to construct a numerical simulation parameter-response database. Specifically, based on the engineering geological information, support design information, boundary condition information, and construction condition information in the standardized dataset, the value ranges and combination schemes of the surrounding rock parameters, support parameters, boundary conditions, and construction conditions are determined. Based on the combination schemes, multi-condition numerical simulation calculations are performed to obtain the surrounding rock response results corresponding to different parameter combinations, and the parameter combinations and their corresponding surrounding rock response results are stored in the numerical simulation parameter-response database.

[0009] Preferably, the step of training the physical constraint neural network using the numerical simulation parameter-response database, construction monitoring data, and surrounding rock physical constraints specifically includes: during the training process, inputting samples from the numerical simulation parameter-response database and preprocessed construction monitoring data into the physical constraint neural network, establishing a mapping relationship between engineering parameters, construction parameters, support parameters, and surrounding rock response, and using the residuals corresponding to the surrounding rock physical constraints as a loss function term to participate in network parameter updates.

[0010] Preferably, a construction dynamic control scheme is generated based on the risk status output results; specifically, a construction dynamic control scheme is matched from the construction control strategy library based on the risk status output results. The construction control strategy library includes construction adjustment measures corresponding to different risk levels, main risk types, and high-risk locations.

[0011] Preferably, the parameters for the current construction stage include: engineering geological parameters, construction sequence parameters, and support structure parameters.

[0012] Preferably, it also includes: determining whether supplementary numerical simulation and model update are needed during construction: when the monitoring-prediction residual, physical constraint residual or prediction uncertainty exceeds a set threshold, or the current construction condition exceeds the coverage of the numerical simulation parameter-response database, the numerical simulation model is invoked for supplementary calculation, and the newly added parameter-response samples obtained from the supplementary calculation are added to the numerical simulation parameter-response database to update the physical constraint neural information network and intelligent risk warning model.

[0013] On the other hand, a rapid prediction and intelligent early warning system for surrounding rock deformation in underground engineering is provided, including: The numerical simulation parameter-response database module is used to construct the numerical simulation parameter-response database by conducting multi-condition numerical simulations during the design phase or the early stage of construction. The physical constraint neural network training module uses the numerical simulation parameter-response database, preprocessed construction monitoring data, and surrounding rock physical constraints to train the physical constraint neural network. The online correction module uses real-time construction monitoring data to correct the trained physical constraint neural network during construction, resulting in a corrected physical constraint neural network. The rapid prediction module for surrounding rock response inputs the parameters of the current construction stage and real-time construction monitoring data into the corrected physical constraint neural network to obtain the prediction results of the surrounding rock response for the current and subsequent construction stages. The intelligent risk early warning module constructs a risk feature vector based on the surrounding rock response prediction results and pre-processed construction monitoring data, inputs it into the intelligent risk early warning model, and obtains the risk status output result. The construction control module generates dynamic construction control plans based on the risk status output.

[0014] Furthermore, an electronic device is also provided, including: Memory, used for non-transitory storage of computer-readable instructions; and Processor, for executing the computer-readable instructions, When the computer-readable instructions are executed by the processor, they perform the method described in the first aspect above.

[0015] In another aspect, a storage medium is also provided for non-transitory storage of computer-readable instructions, wherein when the non-transitory computer-readable instructions are executed by a computer, the method described in the first aspect is performed.

[0016] The above technical solution has the following advantages or beneficial effects: This invention constructs a numerical simulation parameter-response database during the design or early construction phase, and trains a physical constraint neural network based on this database, enabling complex numerical simulation analysis to be completed primarily in the early stages. During construction, the trained model can be used to quickly obtain predicted results for surrounding rock displacement, stress, plastic zone, and support structure response, reducing the need for ad-hoc complex numerical simulation forward analysis and improving the efficiency of surrounding rock response prediction. Furthermore, a small amount of construction monitoring data is used to calibrate the physical constraint neural network, allowing the model to be modified to incorporate the current surrounding rock conditions, support status, and construction stage of the project, alleviating the problems of insufficient sample size and poor adaptability when training the model solely based on on-site monitoring data.

[0017] Furthermore, this invention introduces mechanical equilibrium, constitutive relations, boundary conditions, support structure constraints, and construction phase continuity constraints into the physical constraint neural information network, so that the prediction results are constrained by the laws of surrounding rock mechanics, reducing the risk of inconsistency between pure data-driven prediction results and the laws of engineering mechanics. Attached Figure Description

[0018] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0019] Figure 1 This is a flowchart of a method for rapid prediction and intelligent early warning of surrounding rock deformation in underground engineering, according to an embodiment of the present invention. Figure 2 This is an overall flowchart of a method for rapid prediction and intelligent early warning of surrounding rock deformation in underground engineering, according to an embodiment of the present invention. Figure 3 This is a flowchart illustrating the construction of a numerical simulation parameter-response database for a rapid prediction and intelligent early warning method for surrounding rock deformation in underground engineering, according to an embodiment of the present invention. Figure 4 This is a physical constraint neural network training framework diagram of a method for rapid prediction and intelligent early warning of surrounding rock deformation in underground engineering according to an embodiment of the present invention. Figure 5 This is a schematic diagram of the physical constraint neural network prediction and surrounding rock response identification of a rapid prediction and intelligent early warning method for surrounding rock deformation in underground engineering according to an embodiment of the present invention. Figure 6 This is a flowchart illustrating the risk feature construction and intelligent early warning process of a rapid prediction and intelligent early warning method for surrounding rock deformation in underground engineering, according to an embodiment of the present invention. Figure 7 This is a flowchart of database supplementation and model update for a method for rapid prediction and intelligent early warning of surrounding rock deformation in underground engineering according to an embodiment of the present invention. Figure 8This is a module composition diagram of a rapid prediction and intelligent early warning system for surrounding rock deformation in underground engineering, according to an embodiment of the present invention. Figure 9 The figures shown are experimental results of a method for rapid prediction and intelligent early warning of surrounding rock deformation in underground engineering according to an embodiment of the present invention. (a) represents the prediction error of displacement under retained working conditions; (b) represents the dimensionless mechanical equilibrium residual; and (c) represents the early synthetic observation correction and subsequent stage prediction. Detailed Implementation

[0020] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0021] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments of the invention. The terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion, for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0022] In this embodiment of the invention, "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 represent: A existing alone, A and B existing simultaneously, and B existing alone. Furthermore, in the description of this invention, "multiple" refers to two or more.

[0023] Furthermore, to facilitate a clear description of the technical solutions of the embodiments of the present invention, the terms "first" and "second" are used in the embodiments of the present invention to distinguish identical or similar items with essentially the same function and effect. Those skilled in the art will understand that the terms "first" and "second" do not limit the quantity or execution order, and the terms "first" and "second" are not necessarily different.

[0024] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0025] Example 1 Regarding methods for predicting and warning of surrounding rock deformation in underground engineering, some existing technologies disclose a mine disaster prediction and early warning method that combines on-site monitoring and numerical simulation. This method emphasizes the joint judgment of on-site monitoring and numerical simulation results, but does not disclose a predictive model that can be quickly invoked based on a numerical simulation parameter-response database. Other existing technologies disclose a method for inverting rock mechanics parameters and predicting subsequent excavation responses in underground powerhouses based on deep learning. This method emphasizes the inversion of rock mechanics parameters and the prediction of subsequent excavation responses, but does not disclose a method for rapidly predicting surrounding rock responses and outputting dynamic construction control schemes based on a physically constrained neural network. Some existing technologies disclose a method and device for joint intelligent inversion of surrounding rock parameters in soft rock tunnels. This method emphasizes iterative inversion of surrounding rock parameters, but does not disclose rapid predictive model training based on a multi-condition numerical simulation parameter-response database. Still other existing technologies disclose a method for predicting tunnel surrounding rock displacement using neural networks. This method mainly trains the neural network based on monitoring data, but does not disclose the use of numerical simulation samples, a small amount of monitoring data, and physical constraints for model learning.

[0026] To address the technical problems existing in current methods for predicting and warning of surrounding rock deformation in underground engineering, the following measures are proposed: First, reduce the dependence of underground engineering surrounding rock response prediction on repeated and complex numerical simulation calculations; second, improve the stability and adaptability of surrounding rock deformation prediction models under conditions of limited construction monitoring data; third, ensure that the prediction results meet the physical laws of surrounding rock mechanical equilibrium, constitutive relations, boundary conditions, and support constraints; and fourth, identify the main control risk types, high-risk locations, and construction control measures based on the prediction results.

[0027] Without altering the technical concept of this invention, those skilled in the art can adjust the data type, physical constraint form, rapid prediction model structure, database sample range, and early warning threshold settings according to the type of underground engineering, construction method, monitoring items, and support form.

[0028] Figure 1-2 This is a schematic diagram of a method for rapid prediction and intelligent early warning of surrounding rock deformation in underground engineering, as described in an embodiment of the present invention. Figure 1 and Figure 2 The method for rapid prediction and intelligent early warning of surrounding rock deformation in underground engineering, as shown in this embodiment, includes: Step 1: In the design phase or early construction phase, conduct multi-condition numerical simulations to build a numerical simulation parameter-response database.

[0029] Before constructing the numerical simulation parameter-response database, the process also includes: acquiring basic data and construction monitoring data of underground engineering projects, and preprocessing the basic data and construction monitoring data to form a standardized dataset.

[0030] Based on the engineering geological information, support design information, boundary condition information, and construction condition information in the standardized dataset, the value ranges and combination schemes of the surrounding rock parameters, support parameters, boundary conditions, and construction conditions are determined. Multi-condition numerical simulation calculations are then performed based on these combination schemes to obtain the surrounding rock response results corresponding to different parameter combinations. The parameter combinations and their corresponding surrounding rock response results are then stored in a numerical simulation parameter-response database. Figure 3 As shown: First, acquire basic data and construction monitoring data for the underground engineering project. Basic data includes one or more of the following: engineering geological parameters, surrounding rock classification information, groundwater conditions, initial in-situ stress parameters, excavation cross-section parameters, support structure parameters, and construction method parameters.

[0031] Construction monitoring data includes one or more of the following: surrounding rock deformation monitoring data, support structure response data, and construction sequence parameters.

[0032] Surrounding rock deformation monitoring data includes one or more of the following: crown settlement, perimeter convergence, floor heave, deep displacement, surface settlement, and internal displacement of the surrounding rock. Support structure response data include one or more of the following: anchor bolt axial force, anchor cable tension, steel arch stress, shotcrete strain, secondary lining stress, surrounding rock pressure, and contact pressure. Construction sequence data includes one or more of the following: excavation method, construction stage, cycle advance, excavation time, face distance, blasting parameters, mechanically assisted excavation parameters, advanced support parameters, grouting parameters, support construction time, and support closure time.

[0033] Next, the basic data and construction monitoring data are preprocessed to form a standardized dataset. Preprocessing operations include: time synchronization, outlier removal, missing value completion, spatial location matching, and dimensionless processing.

[0034] Time synchronization is used to unify data from different sampling frequencies to the same construction stage or the same time interval; spatial location matching is used to match the location of monitoring points, excavation outline, distance from the working face, and location of support structures; dimensionless processing is used to eliminate dimensional differences between different types of data; outlier removal is used to remove abnormal monitoring values ​​caused by sensor failure, abnormal acquisition, construction interruption, or communication errors, and can be performed using any of the following methods: engineering threshold method, three-standard deviation method, box plot method, or isolated forest method; missing value completion is used to complete missing data caused by sensor offline, data transmission interruption, or missing manual records, and can be performed using any of the following methods: linear interpolation, spline interpolation, Kalman filtering, or data estimation methods for adjacent construction stages.

[0035] The preprocessing of the present invention is preferably performed in the order of time synchronization, spatial location matching, outlier removal, missing value completion, and dimensionless processing; wherein, time synchronization and spatial location matching are used to ensure that data from different sources are in the same construction stage and have the same spatial correspondence, outlier removal and missing value completion are used to improve data integrity and reliability, and dimensionless processing is used to eliminate dimensional differences between different data types.

[0036] The initial stage of construction refers to the stage in which a small amount of monitoring data has been generated during the construction of underground engineering projects, but a large number of abnormal or disaster samples have not yet been accumulated.

[0037] The parameter combination includes one or more of the following: surrounding rock density, elastic modulus, Poisson's ratio, cohesion, internal friction angle, lateral pressure coefficient, groundwater pressure parameter, excavation disturbance parameter, support stiffness parameter, and boundary stress parameter.

[0038] The response results include one or more of the following: surrounding rock displacement field, surrounding rock stress field, plastic zone range, internal forces of support structure, deformation of support structure, ground settlement, deformation of existing structure, and response results of unmonitored areas.

[0039] The numerical simulation can employ any of the following methods: finite difference method, finite element method, discrete element method, finite element-discrete element coupling method, or other numerical methods capable of simulating the excavation and support process of underground engineering.

[0040] Preferably, the numerical simulation of the present invention is performed using the FLAC3D model.

[0041] The numerical simulation parameter-response database constructed in this invention is used to provide training samples for subsequent physical constraint neural information networks, so that the surrounding rock response prediction results can be obtained under normal working conditions without having to carry out a complete and complex numerical simulation calculation again.

[0042] Step 2: Train the physical constraint neural network using the numerical simulation parameter-response database, preprocessed construction monitoring data, and surrounding rock physical constraints. For example... Figure 4 As shown: First, a physical constraint neural information network is constructed. This network is based on a neural network and incorporates numerical simulation parameters-response samples from the numerical simulation parameter-response database, construction monitoring samples, and surrounding rock mechanical constraints into the training model to train the physical constraint neural information network.

[0043] The physically constrained neural network comprises an input layer, a feature mapping layer, a physical constraint layer, a fast prediction layer, and an output layer. The input layer receives standardized construction monitoring samples, construction sequence parameters, engineering geological parameters, support structure parameters, and numerical simulation parameter-response samples. The feature mapping layer extracts the correlation features between surrounding rock conditions, construction stages, support status, and monitoring responses. The physical constraint layer calculates mechanical equilibrium residuals, constitutive relation residuals, boundary condition residuals, support structure constraint residuals, and construction stage continuity residuals. The fast prediction layer predicts the surrounding rock response for the current and subsequent construction stages. The output layer outputs the fast prediction results of the surrounding rock response, including the surrounding rock displacement field, surrounding rock stress field, plastic zone expansion range, support structure response, location of abnormal responses in unmonitored areas, and prediction uncertainty.

[0044] The input layer, feature mapping layer, physical constraint layer, fast prediction layer, and output layer are functionally divided, and separate fully connected layers are not required. The physical constraint layer calculates the corresponding mechanical residuals based on the network's predicted response and participates in network parameter updates through the loss function; feature mapping and fast prediction are implemented by the backbone network and its output mapping. Numerical responses and monitoring observations are used as supervision values ​​corresponding to parameters, locations, and construction stages during training, and the response to be predicted is not used as known input during prediction.

[0045] Physically constrained neural networks can use multilayer perceptrons, residual neural networks, recurrent neural networks, temporal networks, or combinations thereof as the backbone network, and introduce surrounding rock mechanical constraints into the network training process through physical constraint layers.

[0046] The physical constraint neural network uses parameter and response samples from the numerical simulation parameter-response database as the main training samples, a small amount of construction monitoring data as correction samples, and introduces surrounding rock mechanical constraints for learning.

[0047] Among them, the surrounding rock mechanical constraints include one or more of the following: mechanical equilibrium constraints, constitutive relation constraints, boundary condition constraints, support structure constraints, and construction stage continuity constraints.

[0048] By introducing mechanical equilibrium, constitutive relations, boundary conditions, support structure constraints, and construction phase continuity constraints into the physical constraint neural information network, the prediction results are constrained by the laws of surrounding rock mechanics, reducing the risk of inconsistency between pure data-driven prediction results and the laws of engineering mechanics.

[0049] The inputs to the physical constraint neural information network include one or more of the following: engineering geological parameters, construction sequence parameters, support structure parameters, monitoring data features, and numerical simulation sample features. The monitoring data features are construction monitoring data after preprocessing and feature extraction, including one or more of the following: surrounding rock deformation features, support structure response features, deformation rate features, deformation acceleration features, and monitoring-prediction residual features.

[0050] The physical constraint neural information network outputs surrounding rock response prediction results, including one or more of the following: surrounding rock displacement field, surrounding rock stress field, plastic zone expansion range, support structure response, location of abnormal response in unmonitored areas, and prediction uncertainty in the current and subsequent construction stages.

[0051] Furthermore, the step of training the physical constraint neural network using the database, preprocessed construction monitoring data, and surrounding rock mechanical constraints specifically includes: during the training process, inputting samples from the numerical simulation parameter-response database and preprocessed construction monitoring data into the physical constraint neural network, establishing the mapping relationship between engineering parameters, construction parameters, support parameters, and surrounding rock response, and using the residuals corresponding to the surrounding rock physical constraints as loss function terms to participate in network parameter updates.

[0052] During the training process, the displacement field, stress field, strain field and support response output by the physical constraint neural information network are substituted into the mechanical equilibrium equation, constitutive relation, boundary conditions, support structure interaction relationship and construction stage continuity conditions to calculate the physical constraint residuals in the surrounding rock mechanical constraints, including mechanical equilibrium residuals, constitutive relation residuals, boundary condition residuals, support structure constraint residuals and construction stage continuity residuals.

[0053] The aforementioned physical constraint residuals are incorporated into the network loss function as physical constraint residual losses, and participate in network parameter updates together with the numerical simulation sample loss, construction monitoring data loss, response field consistency loss, and prediction uncertainty constraint loss. This ensures that the trained physical constraint neural network not only fits the numerical simulation samples and construction monitoring data but also satisfies the constraints of surrounding rock mechanics.

[0054] Furthermore, the loss function of the physical constraint neural information network includes one or more of the following: numerical simulation sample loss, construction monitoring data loss, physical constraint residual loss, response field consistency loss, and prediction uncertainty constraint loss.

[0055] The loss function of a physically constrained neural network satisfies:

[0056] in, Due to the loss of construction monitoring data, For numerical simulation parameters - response sample loss, The physical constraint residual loss includes mechanical equilibrium constraint residual, constitutive relation constraint residual, boundary condition constraint residual, support structure constraint residual, and construction stage continuity constraint residual; In response to field consistency loss, To predict the loss due to uncertainty constraints, These are the weighting coefficients.

[0057] Physical constraint residual loss includes one or more of the following: mechanical equilibrium residual, constitutive relation residual, boundary condition residual, support structure constraint residual, and construction stage continuity residual.

[0058] The construction monitoring data loss during the training phase and the online correction during construction are organized separately. The former, together with the numerical sample loss and the selected physical constraint residuals, is used to train the network parameters; the latter utilizes the deviation between newly added construction monitoring and prediction to correct the model parameters, input features, or auxiliary state parameters of the trained network. When using auxiliary state parameter correction, this correction does not replace the monitoring data supervision during the training phase. The specific selection and weight of each loss term are specified in the corresponding implementation method.

[0059] This invention constructs a numerical simulation parameter-response database during the design or early construction phase, and trains a physical constraint neural network based on the database, enabling complex numerical simulation analysis to be completed primarily in the early stages. During construction, the trained model can be used to quickly obtain predicted results for surrounding rock displacement, stress, plastic zone, and support structure response, reducing the need for ad-hoc complex numerical simulation forward analysis and improving the efficiency of surrounding rock response prediction.

[0060] Step 3: During construction, real-time construction monitoring data is used to correct the trained physical constraint neural network, resulting in the corrected physical constraint neural network. For example... Figure 5 As shown: Specifically, during construction, new real-time construction monitoring data will be input into the trained physical constraint neural network, and the model parameters, input features, response residuals, prediction uncertainties, or auxiliary state parameters will be corrected based on the deviation between the monitoring data and the prediction results.

[0061] The auxiliary state parameters include one or more of the following: equivalent elastic modulus of surrounding rock, equivalent Poisson's ratio, equivalent cohesion, equivalent internal friction angle, excavation disturbance correction coefficient, support stiffness correction coefficient, and boundary stress correction coefficient.

[0062] The auxiliary state parameters are used to interpret model prediction results, correct model inputs, or update the numerical simulation parameter-response database, and are not the sole prediction output of this invention.

[0063] Step 4: Input the parameters of the current construction stage and the real-time construction monitoring data into the corrected physical constraint neural network to obtain the surrounding rock response prediction results for the current and subsequent construction stages.

[0064] The parameters for the current construction phase include: engineering geological parameters, construction sequence parameters, and support structure parameters.

[0065] The results of surrounding rock response prediction include one or more of the following: surrounding rock displacement field, surrounding rock stress field, plastic zone expansion range, support structure response, location of abnormal response in unmonitored areas, and prediction uncertainty.

[0066] The location of abnormal response in unmonitored areas is determined based on one or more of the following: the location of predicted displacement exceeding the limit, the location of sudden change in displacement gradient, the location of stress concentration, the location of the plastic zone expansion boundary approaching the support structure, the location of the support structure internal force utilization rate exceeding the set ratio, and the location of the large monitoring-prediction residual.

[0067] This method allows for the prediction and identification of the surrounding rock response in unmonitored areas, even with limited monitoring point deployment.

[0068] This invention can predict the surrounding rock response field during the current and subsequent construction stages and identify the locations of abnormal responses in areas where no monitoring points have been set up. Compared to methods that rely solely on local monitoring points or a small number of monitoring sections, this invention improves the ability to identify potential risks in unmonitored areas.

[0069] Step 5: Based on the surrounding rock response prediction results and the pre-processed construction monitoring data, construct a risk feature vector, input it into the intelligent risk early warning model, and obtain the risk status output results; generate a construction dynamic control scheme based on the risk status output results.

[0070] like Figure 6 As shown, the risk status output results include one or more of the following: risk level, risk level probability, main risk type, high-risk location, early warning lead time, and prediction confidence level.

[0071] A risk feature vector is constructed based on the surrounding rock response prediction results and the pre-processed construction monitoring data. Specifically, the abnormal deformation characteristics are determined based on the relationship between the predicted displacement and the allowable displacement control value. Determine the deformation rate characteristics based on the predicted displacement changes between adjacent construction stages; The deformation acceleration characteristics are determined based on the changes in deformation rate between adjacent construction stages; specifically, the deformation acceleration characteristics are determined by the changes in deformation rate between adjacent construction stages or adjacent monitoring times. When the... Stage deformation value , No. Stage deformation value The corresponding time interval is When calculating the deformation rate, the deformation acceleration characteristics can be determined first based on the ratio of the difference in deformation rate between adjacent stages to the time interval.

[0072] The characteristics of plastic zone expansion are determined based on the extent of the plastic zone, the rate of plastic zone expansion, and the positional relationship between the plastic zone and the support structure. The stress characteristics of the support structure are determined based on the relationship between the internal forces or strains of the support structure and the load control values. Based on the deviation between the field monitoring data and the model prediction results, the monitoring-prediction residual characteristics are determined; Determine the characteristics of prediction uncertainty based on model ensemble, Bayesian neural network, Monte Carlo Dropout, quantile regression, or repeated training statistical methods.

[0073] Physical constraint residual characteristics are calculated using the physical constraint layer, including one or more of the following: mechanical equilibrium residuals, constitutive relation residuals, boundary condition residuals, support structure constraint residuals, and construction stage continuity residuals. Specifically, the displacement field, stress field, strain field, and support response predicted by the network are substituted into the corresponding physical constraint equations, and the deviation between the prediction results and the physical constraint conditions is calculated as the physical constraint residual characteristics.

[0074] Construction stage characteristics are determined based on construction sequence data, including one or more of the following: construction stage number, excavation sequence, cycle advance, face distance, support construction time, support closure time, excavation method, grouting parameters, and the status of completed construction sections. These construction stage characteristics characterize the phased impact of changes in the surrounding rock response as construction progresses.

[0075] The risk feature vector includes one or more of the following: deformation anomaly features, deformation rate features, deformation acceleration features, plastic zone expansion features, support structure stress features, monitoring-prediction residual features, physical constraint residual features, prediction uncertainty features, and construction stage features.

[0076] Furthermore, the risk feature vector is input into the intelligent risk early warning model to obtain the risk status output. The risk status output includes one or more of the following: risk level, risk level probability, main control risk type, high-risk location, early warning lead time, and prediction confidence level.

[0077] The intelligent risk warning model is used to output the risk status based on the risk feature vector, and can employ classification models, regression models, or ensemble learning models. Specifically, the intelligent risk warning model can be a random forest, gradient boosting tree, support vector machine, Bayesian classifier, multilayer perceptron, long short-term memory network, Transformer model, or a combination of the above models.

[0078] In one implementation, the intelligent risk warning model takes a risk feature vector as input and outputs risk level, risk level probability, main risk type, high-risk location, warning lead time, and prediction confidence level.

[0079] The risk levels are divided into four levels: normal, attention, alert, and danger.

[0080] The main types of risks include one or more of the following: abnormal settlement of the arch crown, abnormal convergence of the perimeter, abnormal heave of the base plate, abnormal expansion of the plastic zone, abnormal stress of the support structure, abnormal monitoring-prediction residuals, and abnormal physical constraint residuals.

[0081] High-risk locations include one or more of the following: arch crown, arch shoulder, arch waist, sidewall, base slab, local locations of the support structure, and locations of abnormal responses in unmonitored areas.

[0082] The probability of risk level can be determined by the classification output, probability normalization output, or voting results of the integrated model of the intelligent risk early warning model; the early warning lead time can be determined by the construction sequence or time interval when the prediction results of the subsequent construction stages reach the corresponding risk level; the prediction confidence can be determined by the prediction uncertainty, the degree of model integration dispersion, or the monitoring-prediction residual.

[0083] This invention determines risks based on surrounding rock response prediction results, monitoring-prediction residuals, physical constraint residuals, plastic zone expansion characteristics, support structure stress characteristics, and prediction uncertainties. It can output risk levels, main risk types, high-risk locations, and construction control suggestions, transforming early warning results from single threshold alarms into risk identification results oriented towards construction control.

[0084] Furthermore, based on the risk status output, a dynamic construction control scheme is matched from the construction control strategy library. The generation process is as follows: Based on the risk level, risk level probability, main control risk type, high-risk location, early warning lead time, and prediction confidence level, a dynamic construction control scheme is matched from the construction control strategy library.

[0085] The construction control strategy library is a pre-established rule base, including construction control measures corresponding to different risk levels, main risk types, and high-risk locations. The construction control strategy library can be established based on engineering design documents, construction organization plans, monitoring and control standards, and expert experience, and can also be updated based on monitoring feedback and the effectiveness of responses during the construction process.

[0086] The construction dynamic control plan includes one or more of the following: increasing monitoring frequency, shortening excavation advance, adjusting blasting parameters, adjusting mechanical excavation parameters, implementing support in advance, increasing support strength, local grouting reinforcement, drainage and pressure reduction, suspending construction or restricting excavation.

[0087] Finally, as Figure 7As shown, it is necessary to further determine whether additional numerical simulations and model updates are needed: When the monitoring-prediction residual, physical constraint residual, or prediction uncertainty exceeds the set threshold, or when the current construction condition exceeds the coverage of the numerical simulation parameter-response database, the numerical simulation model is invoked for supplementary calculation, and the newly added parameter-response samples obtained from the supplementary calculation are added to the numerical simulation parameter-response database to update the physical constraint neural information network and the intelligent risk early warning model.

[0088] When the monitoring-prediction residual, physical constraint residual, and prediction uncertainty do not exceed the set thresholds, and the current construction condition is within the coverage of the numerical simulation parameter-response database, no supplementary numerical simulation calculations are performed. Instead, the corrected physical constraint neural information network is used directly to quickly predict the surrounding rock response and provide risk warnings.

[0089] This invention can match construction control schemes based on risk assessment results. The construction control schemes include one or more of the following: increasing monitoring frequency, shortening excavation advance, adjusting construction parameters, strengthening support, local reinforcement, drainage and pressure reduction, suspending construction or restricting excavation, so that the prediction and early warning results can serve the dynamic adjustment of the construction scheme.

[0090] When the monitoring-prediction residual, physical constraint residual, or prediction uncertainty exceeds a set threshold, or when the current operating condition exceeds the coverage of the numerical simulation parameter-response database, this invention invokes the numerical simulation model for supplementary calculations and updates the database. This allows for maintaining model adaptability while avoiding unnecessary and complex numerical simulation analyses under normal operating conditions.

[0091] Furthermore, to train the physically constrained neural network, a parameter-response database was constructed using batch numerical simulation calculations. The numerical simulation parameter-response database includes parameter fields and response fields. Parameter fields characterize surrounding rock conditions, support conditions, construction conditions, and boundary conditions, while response fields characterize the response results of the surrounding rock and support structure under different parameter combinations. Tables 1 and 2 provide an example of one field setting for the numerical simulation parameter-response database.

[0092] Table 1. Examples of parameter fields in the numerical simulation parameter-response database

[0093] Table 2. Examples of response fields in the numerical simulation parameters-response database.

[0094] The table above is only an example of field settings for the numerical simulation parameter-response database. Those skilled in the art can adjust the parameter fields, response fields, and sample size according to the type of underground engineering, surrounding rock conditions, support type, construction method, and monitoring items.

[0095] In another embodiment, when the underground project is a shield tunnel, the construction monitoring data and construction sequence data can be replaced according to the characteristics of shield construction. The surrounding rock deformation monitoring data includes one or more of the following: surface settlement, deep horizontal displacement, building settlement, underground pipeline settlement, tunnel axis offset, segment convergence, and segment vertical displacement. The support structure response data includes one or more of the following: segment internal force, segment joint opening, segment misalignment, grouting pressure response behind the segment, and segment inter-ring deformation. The construction sequence data includes one or more of the following: shield thrust, shield torque, cutterhead rotation speed, propulsion speed, soil chamber pressure, screw conveyor rotation speed, shield tail clearance, synchronous grouting pressure, and synchronous grouting volume. The risk warning and dynamic construction control process in shield tunnel underground projects can be performed with reference to the above steps.

[0096] Furthermore, the experimental process of this invention is as follows: joint training and numerical verification of the elastic circular cavity, such as... Figure 9 As shown, (a) represents the displacement prediction error under retained working conditions; (b) represents the dimensionless mechanical equilibrium residual; and (c) represents the early synthetic observation correction and subsequent stage prediction.

[0097] This embodiment verifies the above training, correction, and risk output process under homogeneous, isotropic, plane strain linear elastic conditions using a circular cavity with an inner radius of 3 m and an outer radius of 18 m. The net unloading stress increment q of the cavity wall is positive with tensile stress, and the radial displacement increment at the outer boundary is zero; the convergent displacement v is positive with an inward orientation.

[0098] The inputs are [E,ν,q,s], where E ranges from 50 to 200 MPa, ν ranges from 0.20 to 0.35, q ranges from 0.05 to 0.50 MPa, and s = r / a ranges from 1 to 6. Each input is mapped to [E,ν,q,s] according to fixed upper and lower bounds. 1,1]. Let s=r / a, Eref=100MPa, qref=0.2 MPa, displacement scale u0=a qref / Eref=6 mm, network output w=v / u0. During the construction phase, the net unloading amount q=α(p0) is given. p i The predictions do not include time stickiness or path dependence.

[0099] The Physically Constrained Neural Network (PCN) is a 4-layer fully connected feedforward neural network, consisting of an input layer, two hidden layers, and an output layer (4-32-32-1 fully connected structure). It includes two 32-node Tanh hidden layers and a linear output node, with a total of 1249 trainable parameters. The PCN calculates the following dimensionless equilibrium and boundary residuals based on the predicted displacement, without independent fully connected weights.

[0100]

[0101]

[0102]

[0103] The internal derivatives are obtained using a second-order central difference with h=0.02, and the derivatives of the cavity walls are obtained using a second-order one-sided difference with the same step length. The numerical monitoring values ​​are generated by solving a second-order finite difference system with 401 radial nodes and a step size of 0.0125, with the same boundary conditions as the residual conditions described above.

[0104] The parameters were divided into 40 fitting conditions, 12 validation conditions, and 80 test conditions. For each of the 40 fitting conditions, displacements at s=1, 1.4, 2, 3.5, and 6 were recorded, generating 200 numerical monitoring records. For the same batch of fitting conditions, additional analytical displacements were superimposed with Gaussian noise of 0.03 mm standard deviation at s=1 and 1.6, generating 80 synthetic monitoring records. The synthetic monitoring seed was 20260911. The dimensionless mean square error for 80 records. The dimensionless mean square error of 200 records is averaged and then included in the joint loss.

[0105]

[0106]

[0107] The specific losses mentioned above correspond to λ1=1, λ2=1, λ3=0, and λ4=0 in the general loss expression. This invention selects three types of losses: monitoring, numerical samples, and physical residuals, without separately training the response field consistency and prediction uncertainty losses. Three internal physical points are set for each fitting condition, and resampling is performed every 50 rounds in the range [1.03, 5.97], with evaluation of the inner and outer boundary residuals; the gradients of the monitoring loss and the other losses are used to update the network parameters.

[0108] The physically constrained neural network employs Xavier uniform initialization, zero bias, and the Adam optimizer, with β1=0.9, β2=0.999, and ε=10. -8The learning rate was 0.001 for rounds 1-3000, 0.0003 for rounds 3001-4500, and 0.0001 for rounds 4501-5000, with a total of 5000 rounds run in full batch. The lowest verification error checkpoint was selected based on the analytical reference of 31 locations across 12 verification scenarios. 51 spatial locations were evaluated for each of the 80 test scenarios. The test was run once each with seeds 17, 29, and 43.

[0109] Under the same structure, numerical supervision, synthetic monitoring supervision, and training configuration, pure data BPNN only uses The physically constrained neural information network proposed in this invention employs... The displacement RMSEs for the three runs were 0.3772 ± 0.1264 mm and 0.1110 ± 0.0191 mm, respectively; the equilibrium residual RMSs were 0.44857 ± 0.20437 and 0.03802 ± 0.00410, respectively. The mean and sample standard deviation are reported for all results. These results only support the improvement in displacement estimation by physical constraints under this parameter domain and the supervised protocol.

[0110] Online calibration uses a six-stage synthetic sequence: analytical reference E=85 MPa, ν=0.27, maximum unloading 0.35 MPa, and α values ​​of 0.20, 0.35, 0.50, 0.65, 0.80, and 1.00 respectively; Gaussian noise with a standard deviation of 0.03 mm is added at two positions in the first three stages, with a seed of 20260909, resulting in 6 new observations. At the end of the third stage, the trained network weights and other inputs are fixed, and the mean square error of the observation-prediction in the first three stages is minimized within the range of E∈[50,200] MPa before predicting stages 4-6. The true value of 85 MPa is not used as the calibration input.

[0111] With an initial E=130 MPa, the average RMSE predicted by the three networks at two locations in stages 4-6 was 3.3793 mm, which was corrected to 0.2364 mm. The corrected curves for the first three stages are back-calculated and are not included in the evaluation of subsequent stages; this online correction does not replace the aforementioned monitoring loss training.

[0112] The intelligent risk early warning model employs a random forest with 100 trees, a maximum depth of 6, and a minimum of 2 leaf node samples. The input is the ratio of the predicted displacement of the tunnel wall in the current and next stages to the 12 mm demonstration control value. Four levels of labels—normal, attention, alert, and danger—are established using the maximum values ​​of the two ratios (0.6, 0.8, and 1.0) as boundaries. The model is trained using 800 synthetic features, and an additional 400 features are used to check consistency with the rules. The aforementioned control values, ratios, and labels are all pre-set and are not intended as engineering specification limits or historical disaster labels.

[0113] Based on risk level, tunnel wall location, and preset strategy library output, example measures include increasing monitoring frequency, reviewing subsequent construction and support arrangements, and submitting for manual review. The random forest output is the average probability of each tree category, not equivalent to the calibrated probability of disaster occurrence. In this example, the maximum convergence location is the tunnel wall; this is not used to demonstrate the location of anomalies in complex, unmonitored areas, seepage and plastic evolution, structural support response, automatic supplementary numerical calculations, or on-site early warning effects.

[0114] Example 2 like Figure 8 As shown, this embodiment provides a rapid prediction and intelligent early warning system for surrounding rock deformation in underground engineering, including: The numerical simulation parameter-response database module is used to construct the numerical simulation parameter-response database by conducting multi-condition numerical simulations during the design phase or the early stage of construction. The physical constraint neural network training module uses the numerical simulation parameter-response database, preprocessed construction monitoring data, and surrounding rock physical constraints to train the physical constraint neural network. The online correction module uses real-time construction monitoring data to correct the trained physical constraint neural network during construction, resulting in a corrected physical constraint neural network. The rapid prediction module for surrounding rock response inputs the parameters of the current construction stage and real-time construction monitoring data into the corrected physical constraint neural network to obtain the prediction results of the surrounding rock response for the current and subsequent construction stages. The intelligent risk early warning module constructs a risk feature vector based on the surrounding rock response prediction results and pre-processed construction monitoring data, inputs it into the intelligent risk early warning model, and obtains the risk status output result. The construction control module generates dynamic construction control plans based on the risk status output.

[0115] Example 3 This embodiment also provides an electronic device, including: one or more processors, one or more memories, and one or more computer programs; wherein, the processor is connected to the memory, and the one or more computer programs are stored in the memory. When the electronic device is running, the processor executes the one or more computer programs stored in the memory to cause the electronic device to perform the method described in Embodiment 1.

[0116] Example 4 This embodiment also provides a computer-readable storage medium for storing computer instructions, which, when executed by a processor, complete the method described in Embodiment 1.

[0117] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for rapid prediction and intelligent early warning of surrounding rock deformation in underground engineering, characterized in that, include: During the design or early construction phase, conduct multi-condition numerical simulations to build a numerical simulation parameter-response database. The physical constraint neural network was trained using the numerical simulation parameter-response database, preprocessed construction monitoring data, and surrounding rock physical constraints. During construction, real-time construction monitoring data is used to correct the trained physical constraint neural network, resulting in a corrected physical constraint neural network. The parameters of the current construction stage and real-time construction monitoring data are input into the corrected physical constraint neural network to obtain the surrounding rock response prediction results for the current and subsequent construction stages. Based on the surrounding rock response prediction results and the pre-processed construction monitoring data, a risk feature vector is constructed, input into the intelligent risk early warning model, and the risk status output result is obtained. A dynamic construction control plan is generated based on the risk status output.

2. The method for rapid prediction and intelligent early warning of surrounding rock deformation in underground engineering according to claim 1, characterized in that, Before constructing the numerical simulation parameter-response database, the process also includes: acquiring basic data and construction monitoring data of underground engineering projects, and preprocessing the basic data and construction monitoring data to form a standardized dataset.

3. The method for rapid prediction and intelligent early warning of surrounding rock deformation in underground engineering according to claim 1, characterized in that, During the design or early construction phase, conduct multi-condition numerical simulations to build a numerical simulation parameter-response database. Specifically, based on the engineering geological information, support design information, boundary condition information, and construction condition information in the standardized dataset, the value range and combination scheme of the surrounding rock parameters, support parameters, boundary conditions, and construction conditions are determined; based on the combination scheme, multi-condition numerical simulation calculations are carried out to obtain the surrounding rock response results corresponding to different parameter combinations, and the parameter combinations and their corresponding surrounding rock response results are stored in the numerical simulation parameter-response database.

4. The method for rapid prediction and intelligent early warning of surrounding rock deformation in underground engineering according to claim 1, characterized in that, The method involves training a physical constraint neural network using the numerical simulation parameter-response database, construction monitoring data, and surrounding rock physical constraints. Specifically, during the training process, samples from the numerical simulation parameter-response database and preprocessed construction monitoring data are input into the physical constraint neural network to establish a mapping relationship between engineering parameters, construction parameters, support parameters, and surrounding rock response. The residuals corresponding to the surrounding rock physical constraints are used as loss function terms to participate in network parameter updates.

5. The method for rapid prediction and intelligent early warning of surrounding rock deformation in underground engineering according to claim 1, characterized in that, A dynamic construction control plan is generated based on the risk status output; specifically, a dynamic construction control plan is matched from the construction control strategy library based on the risk status output. The construction control strategy library includes construction adjustment measures corresponding to different risk levels, main risk types, and high-risk locations.

6. The method for rapid prediction and intelligent early warning of surrounding rock deformation in underground engineering according to claim 1, characterized in that, The parameters for the current construction stage include: engineering geological parameters, construction sequence parameters, and support structure parameters.

7. The method for rapid prediction and intelligent early warning of surrounding rock deformation in underground engineering according to claim 1, characterized in that, Also includes: During construction, it is determined whether supplementary numerical simulation and model updates are needed: when the monitoring-prediction residual, physical constraint residual, or prediction uncertainty exceeds the set threshold, or when the current construction conditions exceed the coverage of the numerical simulation parameter-response database, the numerical simulation model is invoked for supplementary calculation, and the newly added parameter-response samples obtained from the supplementary calculation are added to the numerical simulation parameter-response database to update the physical constraint neural information network and intelligent risk early warning model.

8. A rapid prediction and intelligent early warning system for surrounding rock deformation in underground engineering, characterized in that, include: The numerical simulation parameter-response database module is used to construct the numerical simulation parameter-response database by conducting multi-condition numerical simulations during the design phase or the early stage of construction. The physical constraint neural network training module uses the numerical simulation parameter-response database, preprocessed construction monitoring data, and surrounding rock physical constraints to train the physical constraint neural network. The online correction module uses real-time construction monitoring data to correct the trained physical constraint neural network during construction, resulting in a corrected physical constraint neural network. The rapid prediction module for surrounding rock response inputs the parameters of the current construction stage and real-time construction monitoring data into the corrected physical constraint neural network to obtain the prediction results of the surrounding rock response for the current and subsequent construction stages. The intelligent risk early warning module constructs a risk feature vector based on the surrounding rock response prediction results and pre-processed construction monitoring data, inputs it into the intelligent risk early warning model, and obtains the risk status output result. The construction control module generates dynamic construction control plans based on the risk status output.

9. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the method described in any one of claims 1-7.

10. A computer device, characterized in that, The method includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the method described in any one of claims 1-7.

Images