Tunnel cross-fault deformation prediction method based on machine learning

Abstract

The invention discloses a tunnel cross-fault deformation prediction method based on machine learning, and aims to solve the problems that cross-fault tunnel relative dislocation prediction is missing, monitoring data isomerism is difficult to unify and physical constraints are insufficient. According to the method, a fault plane coordinate system and an observation unified operator are constructed, a double-domain physical information neural operator or a Fourier neural operator is adopted to carry out direction weighted modeling, complementary constraints of contact and friction are applied to a fault interface, and implicit correction is implemented through micro-projection; and the multi-quantile, the over-limit probability and the time interval of the relative dislocation amount are output by combining the assimilation updating of uncertainty weighting and the quantile regression module, so that the technical effects of physical consistency and high-precision prediction of the cross-fault relative dislocation and completion of uncertainty quantization are realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of construction prediction, and in particular to a tunnel cross-fault deformation prediction method based on machine learning. BACKGROUND

[0002] In the tunnel cross-fault engineering, in order to control the deformation and risk, the existing technology mainly includes: the pre-judgment method based on engineering investigation and experience index, the analysis of surrounding rock and structure response by numerical simulation such as finite element, discrete element and finite difference, and the comprehensive application of convergence deformation, hoop strain, pore water pressure, ground deformation (synthetic aperture radar interferometry, etc.), navigation positioning displacement and microseismic event.

[0003] In recent years, data assimilation and digital twin have been gradually applied to underground engineering, using Kalman filter or variational assimilation fusion model and monitoring; at the same time, machine learning methods such as physical information neural network and Fourier neural operator are used for rapid proxy modeling and prediction of geomechanical field.

[0004] However, the relative displacement prediction of tunnel cross-fault has not formed a unified and reliable technical system.

[0005] The deficiencies of the prior art are:

[0006] (1) Lack of fault plane-oriented indicators and coordinate unification: most methods focus on overall convergence or displacement field, and it is difficult to extract and predict tangential relative displacement in fault plane coordinate system, and monitoring data is heterogeneous in spatial coordinates, time resolution and sampling frequency, lacking unified observation operator and uncertainty modeling;

[0007] (2) Insufficient coupling of physical constraints: machine learning methods often do not explicitly impose complementary constraints of contact and friction, and interface non-penetration and Coulomb friction cannot be guaranteed, which may lead to displacement penetration or distortion of shear response, and the difference between left and right surrounding rocks is significant, and domain decomposition and interface coupling are insufficient in rapid prediction;

[0008] (3) Weakness in quantifying uncertainty and risk output: existing prediction is mostly point estimation, lacking quantile level results and over-limit probability, time interval, and difficult to meet the needs of construction control and risk decision-making.

[0009] Therefore, a tunnel cross-fault deformation prediction method that can solve the above deficiencies of the prior art is needed to solve the problem for those skilled in the art. SUMMARY

[0010] One purpose of the present application is to provide a tunnel cross-fault deformation prediction method based on machine learning, aiming at the problems of lack of relative displacement prediction, difficulty in unifying monitoring data and insufficient coupling of interface physical constraints in the prior art, a technical scheme is proposed, which unifies multi-source monitoring in fault plane coordinate system and observation unification operator and depicts uncertainty, uses left and right surrounding rock dual-domain neural operator based on domain decomposition for directional weighting modeling, applies complementary constraints of non-penetration and Coulomb friction on the fault interface and performs implicit correction with differentiable projection, implements assimilation update once based on uncertainty weighted update gain, extracts tangential relative displacement sequence along the tunnel axis and outputs multi-quantile prediction and over-limit probability and time interval by quantile regression, and the present application has the effects of unified observation, consistent physics, high precision and quantifiable uncertainty in relative displacement prediction.

[0011] According to the tunnel cross-fault deformation prediction method based on machine learning, the method comprises the following steps:

[0012] S1, collect multi-source monitoring data and engineering background data of the cross-fault tunnel section, establish a fault plane coordinate system, divide the monitoring points and data into left and right surrounding rock domains according to the fault plane, construct an observation unification operator to project the multi-source monitoring data into displacement observation and its uncertainty in the fault plane coordinate system, and obtain the fault plane coordinate system, left surrounding rock domain monitoring data, right surrounding rock domain monitoring data and observation unification operator;

[0013] S2, input the fault plane coordinate system, left surrounding rock domain monitoring data, right surrounding rock domain monitoring data, engineering background data and observation unification operator, construct and train a dual-domain neural operator model containing left and right sub-models, and output uncorrected left and right surrounding rock domain displacement fields in a prediction time window;

[0014] S3, input the uncorrected left and right surrounding rock domain displacement fields and the fault plane coordinate system, apply complementary constraints of contact and friction on the fault interface, set the friction coefficient and stiffness boundary according to the survey and design parameters, perform implicit correction by differentiable projection, and output corrected left and right surrounding rock domain displacement fields;

[0015] S4, input the corrected left and right surrounding rock domain displacement fields, observation unification operator and multi-source monitoring data, project the displacement fields into prediction observations, compare the prediction observations with the multi-source monitoring data to calculate the update gain, and perform differentiable assimilation update on the displacement fields once, and output the assimilated left and right surrounding rock domain displacement fields;

[0016] S5, input the assimilated left and right surrounding rock domain displacement fields and the fault plane coordinate system, calculate the relative displacement time sequence in the tangential direction of the fault plane at discrete positions of the tunnel axis, and input the quantile prediction module to output multiple quantile prediction results of the relative displacement and its time sequence.

[0017] Optionally, step S1 is specifically:

[0018] The multi-source monitoring data includes convergence deformation, hoop strain, pore water pressure, surface deformation, navigation positioning displacement and microseismic event, and the engineering background data includes geological parameters and construction parameters;

[0019] The fault plane coordinate system defines the tangential direction of the fault plane, the normal direction of the fault plane and the tunnel axial direction;

[0020] The fault strike and the fault dip are obtained according to the geological survey line and the design data, the tangential direction of the fault plane is determined according to the fault strike, the normal direction of the fault plane is determined according to the fault dip, and the tunnel axial direction is determined according to the design alignment to establish the fault plane coordinate system;

[0021] The normal distance of each monitoring point relative to the fault plane is calculated based on the fault plane, and the monitoring points are divided into left and right surrounding rock areas according to the normal distance;

[0022] The multi-source monitoring data is executed with timestamp correction and unified time axis alignment, and is sampled and reconstructed on the unified time axis to form a continuous time sequence;

[0023] The observation unification operator is constructed to perform coordinate transformation on the surface deformation and the navigation positioning displacement and decompose them into displacement observations along the tangential direction of the fault plane, the normal direction of the fault plane and the tunnel axial direction, the convergence deformation and the hoop strain are converted into displacement observations according to the fault plane geometric parameters, the pore water pressure and the microseismic event are processed into time sequence input quantities on the unified time axis, and the noise covariance is set for each type of displacement observation to form the displacement observation uncertainty, thereby generating the fault plane coordinate system, the left surrounding rock area monitoring data, the right surrounding rock area monitoring data and the observation unification operator.

[0024] Optionally, step S2 is specifically:

[0025] The fault plane coordinate system, the left surrounding rock area monitoring data, the right surrounding rock area monitoring data and the observation unification operator are taken as inputs, the displacement observations processed by the observation unification operator are decomposed into components along the tangential direction of the fault plane, the normal direction of the fault plane and the tunnel axial direction in the fault plane coordinate system, and the geological parameters and the construction parameters are combined with the component-decomposed displacement observations;

[0026] constructing a dual-domain neural operator model, the dual-domain neural operator model comprising a sub-model for a left-side surrounding rock domain and a sub-model for a right-side surrounding rock domain, the sub-models adopting a spectral mapping structure of a physical information neural operator or a Fourier neural operator, training on input geological parameters, construction parameters and displacement observations processed by an observation unification operator, and setting directional weights in the spectral mapping structure for a fault plane tangential direction to enhance expression of relative displacement characteristics and for a fault plane normal direction to suppress physical inconsistency components;

[0027] training the dual-domain neural operator model with the displacement observations processed by the observation unification operator as a supervision signal, and generating uncorrected left-side surrounding rock domain displacement fields and uncorrected right-side surrounding rock domain displacement fields in a prediction time window.

[0028] Optionally, step S3 is specifically:

[0029] taking the uncorrected left-side surrounding rock domain displacement fields and the uncorrected right-side surrounding rock domain displacement fields as inputs, selecting a discrete interface point set on the fault plane according to a fault plane coordinate system, defining a displacement jump along the fault plane tangential direction and a normal gap along the fault plane normal direction at each discrete interface point, and setting complementary constraints of contact and friction, the complementary constraints comprising that the normal gap is not less than zero, the normal contact force is not less than zero, the product of the normal gap and the normal contact force is zero, and the tangential friction stress does not exceed the friction coefficient multiplied by the normal contact force and takes the upper limit when sliding occurs and the direction is consistent with the tangential direction of the displacement jump;

[0030] constructing a differentiable projection update feasible region with a given friction coefficient and stiffness boundary of survey and design parameters, performing differentiable projection update on the uncorrected left-side surrounding rock domain displacement fields and the uncorrected right-side surrounding rock domain displacement fields at the discrete interface points by implicit iteration until the complementary constraints and the stiffness boundary are satisfied, and outputting corrected left-side surrounding rock domain displacement fields and corrected right-side surrounding rock domain displacement fields.

[0031] Optionally, step S4 is specifically:

[0032] taking the corrected left-side surrounding rock domain displacement fields and the corrected right-side surrounding rock domain displacement fields as inputs, and projecting the corrected left-side surrounding rock domain displacement fields and the corrected right-side surrounding rock domain displacement fields into prediction observations by an observation unification operator;

[0033] taking multi-source monitoring data as actual observations, and calculating observation residuals between the prediction observations and the actual observations at each time on a unified time axis;

[0034] weighting the observation residuals according to displacement observation uncertainties, and solving an update gain in a least square sense;

[0035] The updated gain is used for once differentiable assimilation updating of the corrected left side wall rock area displacement field and the corrected right side wall rock area displacement field to obtain an assimilated left side wall rock area displacement field and an assimilated right side wall rock area displacement field.

[0036] Optionally, the step S5 is specifically:

[0037] The assimilated left side wall rock area displacement field and the assimilated right side wall rock area displacement field are taken as inputs, left side wall rock area displacement components and right side wall rock area displacement components are extracted according to the clear fault plane tangential direction of the fault plane coordinate system at discrete positions of the tunnel axis, and relative displacement amount time sequences are calculated;

[0038] The relative displacement amount time sequences are input into a quantile prediction module, relative displacement amount prediction values corresponding to multiple quantile levels and time sequences thereof are output, and monotonicity constraints are applied to each quantile level to ensure that a lower quantile is not greater than a higher quantile;

[0039] The quantile prediction module is a quantile regression network including a time sequence encoder and a multi-quantile prediction head, the time sequence encoder is any one of a one-dimensional convolutional network, a recurrent network or a Transformer or a combination thereof, is used for feature extraction on the relative displacement amount time sequences and optional external features, the multi-quantile prediction head is a parallel fully connected branch or a single network with quantile level as a conditional input, is used for outputting relative displacement amount prediction values of a preset quantile set, and the monotonicity constraint is a quantile non-decreasing constraint realized through reparameterization or monotonic projection, to ensure that a lower quantile is not greater than a higher quantile;

[0040] Based on a preset relative displacement amount threshold, threshold discrimination is performed on the relative displacement amount prediction values corresponding to the multiple quantile levels to obtain an overrun probability and an overrun time interval, and a relative displacement amount quantile prediction result is output.

[0041] The present application has the following advantages:

[0042] (1) Unified observation and improved accuracy: by establishing a fault plane coordinate system and an observation unified operator, convergent deformation, hoop strain, ground deformation, navigation displacement and other heterogeneous monitoring are unified into tangential, normal and axial displacements and uncertainties, and a double-domain neural operator direction weighted modeling is performed, so that the prediction accuracy of the cross-fault displacement field and the relative displacement amount is significantly improved;

[0043] (2) Physical consistency and stable convergence: complementary constraints of contact and friction are applied to the fault interface, and a differentiable projection implicit correction is adopted, left and right wall rock area decomposition and interface coupling are combined, displacement penetration and unreasonable shear are effectively inhibited, the adhesion and sliding state switching is accurately described, and the uncertainty weighted assimilation updating is introduced to improve the robustness and real-time adaptability of the model under the monitoring driving;

[0044] (3) Uncertainty quantification and risk warning: Calculate the relative displacement time series along the tunnel axis and output the multi-quantile results by quantile regression, impose monotonicity constraints to ensure quantile consistency, and directly give the overrun probability and time interval to realize risk quantification and early warning for construction control and design adjustment. BRIEF DESCRIPTION OF DRAWINGS

[0045] The accompanying drawings are included to provide a further understanding of the application and are incorporated in and constitute a part of this specification, illustrate embodiments of the application and are included to further explain the application and are not intended to limit the application. In the drawings:

[0046] Figure 1 A flowchart of a tunnel cross-fault deformation prediction method based on machine learning is proposed. DETAILED DESCRIPTION

[0047] The application will now be described in further detail with reference to the drawings. These drawings are simplified schematic diagrams that only schematically show the basic structure of the application and thus only show the components relevant to the application.

[0048] REFERENCE Figure 1 A tunnel cross-fault deformation prediction method based on machine learning, characterized by comprising the following steps:

[0049] S1, collect multi-source monitoring data and engineering background data of the cross-fault tunnel section, establish a fault plane coordinate system, divide the monitoring points and data into left and right surrounding rock areas according to the fault plane, construct an observation unification operator to project the multi-source monitoring data into displacement observation and its uncertainty in the fault plane coordinate system, and obtain the fault plane coordinate system, left surrounding rock area monitoring data, right surrounding rock area monitoring data and observation unification operator;

[0050] S2, input the fault plane coordinate system, left surrounding rock area monitoring data, right surrounding rock area monitoring data, engineering background data and observation unification operator, construct and train a dual-domain neural operator model containing left and right sub-models, and output uncorrected left and right surrounding rock area displacement fields in the prediction time window;

[0051] S3, input the uncorrected left and right surrounding rock area displacement fields and the fault plane coordinate system, impose complementary constraints of contact and friction on the fault interface, set the friction coefficient and stiffness boundary according to the survey and design parameters, perform implicit correction using a differentiable projection, and output the corrected left and right surrounding rock area displacement fields;

[0052] S4, inputting the corrected left and right side wall rock area displacement field, the observation unified operator and the multi-source monitoring data, projecting the displacement field into the predicted observation, comparing the multi-source monitoring data to calculate the update gain, and performing a differentiable assimilation update on the displacement field, outputting the assimilated left and right side wall rock area displacement field;

[0053] S5, inputting the assimilated left and right side wall rock area displacement field and the fault plane coordinate system, calculating the relative displacement time series in the tangential direction of the fault plane at the discrete positions of the tunnel axis, and inputting the quantile prediction module, outputting the multiple quantile prediction results and time series of the relative displacement.

[0054] In the specific embodiment, S1 is specifically:

[0055] First, collect multi-source monitoring data of the tunnel section across the fault and perform timestamp correction, unified time axis alignment and resampling, and denote the unified time axis as and form a joint observation vector at each time At the same time, collect engineering background data as features and constraints;

[0056] Then, establish a fault plane coordinate system, determine the fault plane tangential and normal directions according to the fault strike and dip, and determine the tunnel axial direction according to the tunnel design alignment, construct the coordinate transformation matrix and orthogonalize, wherein is the fault plane tangential unit vector, is the fault plane normal unit vector, is the tunnel axial unit vector, is the transformation matrix from the global coordinate to the fault plane coordinate;

[0057] On this basis, decompose any global displacement, using:

[0058] ;

[0059] wherein is the displacement component column vector in the fault plane coordinate system, is the transpose of the transformation matrix, is the global displacement vector, is the tangential component, is the normal component, is the axial component;

[0060] Divide the left and right wall rock areas according to the fault plane, select a reference point on the fault plane and calculate the signed normal distance from the monitoring point to the fault plane, using:

[0061] ;

[0062] wherein signed normal distance of the th monitoring point, global coordinate of the th monitoring point, reference point coordinate on the fault plane, according to which is classified into the left surrounding rock area, is classified into the right surrounding rock area;

[0063] For the coordinate unification of heterogeneous monitoring quantities, the tangential, normal and axial components of surface deformation and navigation positioning displacement are obtained through the above transformation, and the radial displacement is converted into the normal component according to the geometric parameters of the section, and the convergence deformation and hoop strain are approximately mapped into the normal component, and the following is used:

[0064] ;

[0065] wherein is the normal component obtained by mapping the radial displacement, is the hoop strain, is the equivalent radius of the tunnel section;

[0066] The pore water pressure and microseismic events are kept as time series input on the unified time axis to embody the external driving, and the noise covariance model of displacement observation is established, and the following is used:

[0067] ;

[0068] wherein is the observation noise at the unified time, is a zero-mean Gaussian distribution, is the noise covariance matrix, is an operator for constructing a diagonal matrix according to components, is the tangential observation noise variance, is the normal observation noise variance, is the axial observation noise variance;

[0069] Accordingly, the fault plane observation unification operator is defined to fuse and output the displacement observation in the fault plane coordinate system, and the following is used:

[0070] ;

[0071] wherein is the tangential-normal-axial displacement observation column vector, is the fault plane observation unification operator, is the joint observation vector at the unified time, is the set of section geometry and sensor installation parameters, and finally the fault plane coordinate system , the left surrounding rock area monitoring data set and the right surrounding rock area monitoring data set and the observation unification operator are obtained .

[0072] In the specific embodiment, S2 is specifically:

[0073] Based on the fault plane coordinate system and the observation unification operator and the geological parameters and the construction parameters, first, the displacement observation processed by the unification operator is decomposed into components in the fault plane coordinate system as the input and supervision of learning, for each monitoring point and time step, using:

[0074] ;

[0075] Wherein is the local displacement component column vector of the monitoring point and the time step contains the tangential component , the normal component and the axial component , Q is the transformation matrix from the global coordinate to the fault plane coordinate composed of the fault plane tangential unit vector and the normal unit vector and the tunnel axial unit vector, is the global displacement observation column vector output by the unification operator, is the global coordinate of the th monitoring point, is the th time step on the unified time axis;

[0076] Then the component-decomposed displacement observation and the geological parameters and the construction parameters are combined and divided into domain fields according to the left and right surrounding rock domains to form the domain feature field Wherein denotes the left surrounding rock domain and the right surrounding rock domain, is the domain feature field at the spatial position and the time contains , , and the domain geological parameter vector and the construction parameter vector ;

[0077] The double-domain neural operator model is constructed, and the spectral mapping structure of the Fourier neural operator is used to predict the field response of the domain feature field to obtain the final correction displacement field which highlights the fault plane tangential relative displacement and suppresses the normal physical inconsistency component. In the spectral mapping, an anisotropic direction weighting matrix and a direction selection weight are introduced, which are specifically:

[0078] ;

[0079] Wherein is the domain at the position and the time local displacement field column vector, and are Fourier transform and inverse transform operators for feature field, is the domain The spectral weight tensor on the frequency vector is used to learn spatial and temporal frequency response, is the frequency vector containing spatial and temporal frequency components, is the domain The directionally weighted diagonal matrix of the domain is the tangential direction weighting coefficient, is the normal direction weighting coefficient, is the axial direction weighting coefficient, is the operator for constructing the diagonal matrix by components.

[0080] The displacement observation processed by the observed unified operator is used as a supervised signal to train the dual-domain neural operator, and a loss function is defined in the sense of uncertainty-weighted least squares to realize end-to-end joint learning of spectral parameters and direction weights. Specifically, the following is used:

[0081]

[0082] where is the weighted least squares loss of the domain , is the predicted local displacement component extracted by at the monitoring point and the time step , is the local displacement component obtained by the observed unified operator, is the displacement observation noise covariance matrix, is its inverse matrix, is the vector transpose operator, is the summation operator for monitoring point index and time index. After training, the uncorrected left surrounding rock domain displacement field and the uncorrected right surrounding rock domain displacement field

[0083] are output within the prediction time window.

[0084] where and are the local displacement field column vectors of the left and right surrounding rock domains at position x and time .

[0085] In the specific embodiment, the S3 is specifically:

[0086] ​​​With the uncorrected displacement field as input, a set of discrete points are selected on the fault interface and the complementary constraints of contact and friction are imposed in the fault plane coordinate system and the implicit correction is performed by the differentiable projection. Firstly, the left and right uncorrected interface displacements are extracted at the discrete points on the interface and the normal gap and tangential displacement jump are calculated, using:

[0087] ;

[0088] wherein is the th interface discrete point, is the uniform time axis time, and are the left and right uncorrected interface displacement column vectors respectively, is the fault plane normal unit vector, is the fault plane tangential unit vector, is the vector transpose operator, is the uncorrected normal gap, is the uncorrected tangential displacement jump;

[0089] Then the interface equivalent stiffness and friction parameters are introduced and the correction and contact mechanics quantities are coupled by the compliance relation, using:

[0090] ;

[0091] wherein is the corrected normal gap, is the corrected tangential displacement jump, is the normal contact force, is the tangential friction stress, is the compliance matrix, is the operator for constructing the diagonal matrix by components, and are the normal and tangential equivalent stiffness respectively and satisfy the engineering given boundary, is the Coulomb friction coefficient;

[0092] The complementary and friction constraints are imposed at each interface point to ensure the non-penetration and friction cone conditions, using:

[0093] ;

[0094] wherein is the absolute value operator, and correspond to the non-penetration and non-stick, is the complementary condition, is the Coulomb friction cone constraint, for ensuring the consistency of the friction direction and the tangential displacement jump; ​

[0095] To obtain a differentiable and consistent solution, the implicit iteration with weighted projection is adopted to correct the solution, and the iteration format is:

[0096] ;

[0097] wherein is the interface state column vector, is the iteration step, is the weighted projection operator to the feasible region, is defined by the above complementary and friction constraints and the stiffness boundary, is the projection weight matrix, is the Euclidean norm, is the convergence threshold; is the corresponding component weight, is the Euclidean norm, is the convergence threshold. When all the interface points satisfy the feasible region, the contact force and the friction stress are written back to the displacement field of the left and right surrounding rock regions to obtain the correction result, and the following is adopted:

[0098]

[0099] ;

[0100] wherein denotes the left and right surrounding rock regions, and are the column vectors of the displacement field in the region before correction and after correction respectively, is the spatial position in the region, is a differentiable correction operator for distributing the interface contact to the nodes or elements in the region, is the set of all interface discrete points, and is the total number of discrete points.

[0101] In the specific embodiment, the S4 is specifically:

[0102] With the corrected displacement field as the prior, it is projected into the predicted observation on the unified time axis, the observation residual is calculated and the measured observation is obtained, and the update gain is obtained in the sense of least squares with uncertainty weighting, and then a differentiable assimilation update is performed on the displacement field to obtain the assimilated left and right surrounding rock region displacement field, which is specifically that the left and right region correction displacement fields are spliced according to the spatial grid and time index to be and the predicted observation is generated at the monitoring point and time level by using the observation unification operator, and the following is adopted:

[0103] ;

[0104] wherein is the monitoring point and the time level​​ Tangential-normal-axial displacement prediction observation column vector To map the displacement field to the observation space of the fault plane coordinate system, a unified observation operator is needed. The spliced ​​column vector of the displacement field after left and right domain correction. For the first Global coordinates of each monitoring point To unify the first on the timeline At any moment The transformation matrix from global coordinates to fault plane coordinates. This is a set of cross-sectional geometry and sensor installation parameters;

[0105] Multi-source monitoring data are used as actual observations, and residuals are calculated hourly on a unified time axis together with predicted observations. The following method is employed:

[0106] ;

[0107] in For time stamp The observation residual column vector To unify the column vector of measured displacement observations after operator processing;

[0108] The residuals are weighted based on the displacement observation uncertainty, and the update gain is obtained in the sense of weighted least squares.

[0109] ;

[0110] in For time stamp The update gain operator is a linear mapping from the observation space to the displacement field space. The prior covariance or smoothing regularization matrix of the displacement field increment is used to constrain the spatial distribution and amplitude. The linearized Jacobian matrix of the observation unified operator is denoted as... Used to characterize the sensitivity of observations to displacement fields, The displacement observation noise covariance matrix can be taken from the value set in step S1. It is a matrix inverse operator and can be coupled with a stable inverse in numerical implementation;

[0111] Perform a differentiable assimilation update on the corrected displacement field using the update gain to obtain the assimilated displacement fields in the left and right domains, using:

[0112] ;

[0113] in In spatial location With time The assimilated displacement field column vector, To correct the displacement field column vector, The displacement field increments driven by residuals are distributed on the spatial grid through linear mapping, ultimately resulting in... The output is divided into the left and right domains and output assimilated displacement fields of the left and right surrounding rock domains for use in subsequent steps.

[0114] In this specific embodiment, S5 specifically includes:

[0115] Using the assimilated displacement field as input, a set of discrete locations is selected along the tunnel axis. and on a unified timeline The above processing involves extracting the displacement components of the left and right surrounding rock domains along the tangential direction in the fault plane coordinate system and calculating the time series of relative displacement. The following method is used:

[0116] ;

[0117] in Position of axis With time Relative displacement, The unit vector of the tangential plane of the fault. and The right and left surrounding rock areas are respectively in and The assimilated displacement column vector at the location, The first on the tunnel axis discrete positions, The total number of discrete locations, To unify the first on the timeline A moment in time;

[0118] Will Input the quantile regression module and output the preset quantile set. Predicted value ;

[0119] in For the first The quantile level and satisfy , The number of quantile levels, For position With time Corresponding quantiles The predicted value of relative slippage;

[0120] To ensure that the quantiles are not in descending order, a reparameterized monotonicity constraint is used:

[0121] ;

[0122] in Unconstrained output of the network, to generate soft positive function with non-negative increment to ensure ;

[0123] Set a relative error threshold when quantifying risk and construct piecewise linear approximation cumulative distribution from quantile to obtain the exceedance probability When there is make ; ;

[0124] Wherein is the relative error threshold, is the approximate cumulative distribution function obtained by quantile interpolation, is the exceedance probability of position and time point , and is the adjacent quantile level;

[0125] Based on the risk threshold define the exceedance time interval set and take the continuous segment in as the exceedance time interval of position , and give the time series of as the multi-quantile prediction result of relative error.

[0126] The above is only the preferred specific implementation of the present application, but the protection scope of the present application is not limited to this, any skilled person in the art, according to the technical solution and the inventive concept of the present application, within the technical range disclosed by the present application, makes equivalent replacement or change, should be covered in the protection scope of the present application.

Claims

1. A method for predicting tunnel cross-fault deformation based on machine learning, characterized in that, The method comprises the following steps: S1, collecting multi-source monitoring data and engineering background data of a cross-fault tunnel section, establishing a fault plane coordinate system, dividing monitoring points and data into left and right surrounding rock areas according to the fault plane, constructing an observation unified operator to project the multi-source monitoring data into displacement observations and their uncertainties in the fault plane coordinate system, and obtaining the fault plane coordinate system, left surrounding rock area monitoring data, right surrounding rock area monitoring data and observation unified operator; S2, taking the fault plane coordinate system, left surrounding rock area monitoring data, right surrounding rock area monitoring data, engineering background data and observation unified operator as inputs, constructing and training a dual-domain neural operator model containing left and right sub-models, and respectively outputting uncorrected left and right surrounding rock area displacement fields in a prediction time window; S3, taking the uncorrected left and right surrounding rock area displacement fields and the fault plane coordinate system as inputs, applying complementary constraints of contact and friction on the fault interface, setting the friction coefficient and stiffness boundary according to the survey and design parameters, performing implicit correction using a differentiable projection, and outputting corrected left and right surrounding rock area displacement fields; S4, taking the corrected left and right surrounding rock area displacement fields, observation unified operator and multi-source monitoring data as inputs, projecting the displacement fields into predicted observations, comparing them with the multi-source monitoring data to calculate the update gain, and performing a differentiable assimilation update on the displacement fields to output assimilated left and right surrounding rock area displacement fields; S5, taking the assimilated left and right surrounding rock area displacement fields and the fault plane coordinate system as inputs, calculating the relative displacement time series in the tangential direction of the fault plane at discrete positions on the tunnel axis, and inputting them into a quantile prediction module to output multiple quantile prediction results and their time series of the relative displacement. 2.The tunnel cross-fault deformation prediction method based on machine learning according to claim 1, wherein, Step S1 is specifically: The multi-source monitoring data includes convergence deformation, hoop strain, pore water pressure, ground deformation, navigation positioning displacement and microseismic events, and the engineering background data includes geological parameters and construction parameters; The fault plane coordinate system defines the tangential direction of the fault plane, the normal direction of the fault plane and the tunnel axis; The fault strike and fault dip are obtained according to the geological survey line and design data, the tangential direction of the fault plane is determined according to the fault strike, the normal direction of the fault plane is determined according to the fault dip, and the tunnel axis is determined according to the design alignment to establish the fault plane coordinate system; The normal distance of each monitoring point relative to the fault plane is calculated based on the fault plane, and the monitoring points are divided into left and right surrounding rock areas accordingly; Time stamp correction and unified time axis alignment are performed on the multi-source monitoring data, and sampling reconstruction is performed on the unified time axis to form continuous time series; The observation unified operator is constructed to perform coordinate transformation and decomposition of ground deformation and navigation positioning displacement into displacement observations along the tangential direction of the fault plane, the normal direction of the fault plane and the tunnel axis, and the convergence deformation and hoop strain are converted into displacement observations according to the fault geometry parameters, the pore water pressure and microseismic events are processed into time series input on the unified time axis, and noise covariance is set for each type of displacement observation to form displacement observation uncertainty, thereby generating the fault plane coordinate system, left surrounding rock area monitoring data, right surrounding rock area monitoring data and observation unified operator. 3.The tunnel cross-fault deformation prediction method based on machine learning according to claim 1, wherein, Step S2 is specifically: The fault plane coordinate system, the left side surrounding rock area monitoring data, the right side surrounding rock area monitoring data, and the observation unification operator are taken as inputs, and the displacement observation processed by the observation unification operator is decomposed into components in the tangential direction of the fault plane, the normal direction of the fault plane, and the tunnel axial direction under the fault plane coordinate system, and the geological parameters and the construction parameters are combined with the displacement observation decomposed into the components. A double-domain neural operator model is constructed, the double-domain neural operator model includes a sub-model for the left side surrounding rock area and a sub-model for the right side surrounding rock area, the sub-models adopt a spectral mapping structure of a physical information neural operator or a Fourier neural operator, train the input geological parameters, construction parameters, and displacement observation processed by the observation unification operator, and set direction weights in the spectral mapping structure for the tangential direction of the fault plane to enhance the expression of relative displacement characteristics and for the normal direction of the fault plane to suppress the physical inconsistent component. The displacement observation processed by the observation unification operator is taken as a supervision signal to train the double-domain neural operator model, and uncorrected left side surrounding rock area displacement fields and uncorrected right side surrounding rock area displacement fields are generated in a prediction time window. 4.The method of claim 1, wherein, Step S3 is specifically: The uncorrected left side surrounding rock area displacement field and the uncorrected right side surrounding rock area displacement field are taken as inputs, a discrete interface point set is selected on the fault plane according to the fault plane coordinate system, a displacement jump in the tangential direction of the fault plane and a normal gap in the normal direction of the fault plane are defined at each discrete interface point, and complementary constraints of contact and friction are set, the complementary constraints include that the normal gap is not less than zero, the normal contact force is not less than zero, the product of the normal gap and the normal contact force is zero, and the tangential friction stress does not exceed the friction coefficient multiplied by the normal contact force and takes the upper limit when sliding occurs and the direction is consistent with the tangential direction of the displacement jump; A feasible region for differentiable projection update is formed by the friction coefficient and the stiffness boundary given by the survey and design parameters, and the uncorrected left side surrounding rock area displacement field and the uncorrected right side surrounding rock area displacement field are differentially projected and updated at the discrete interface points by implicit iteration until the complementary constraints and the stiffness boundary are satisfied, and the corrected left side surrounding rock area displacement field and the corrected right side surrounding rock area displacement field are output.

5. The machine learning based tunnel cross-fault deformation prediction method of claim 1, wherein, Step S4 is specifically: The corrected left side surrounding rock area displacement field and the corrected right side surrounding rock area displacement field are taken as inputs, and the observation unification operator is used to project the corrected left side surrounding rock area displacement field and the corrected right side surrounding rock area displacement field into prediction observations; The multi-source monitoring data are taken as actual observations, and observation residuals are calculated from the prediction observations at a time on a unified time axis; The observation residuals are weighted according to the displacement observation uncertainty, and an update gain is obtained in the least squares sense; The corrected left side surrounding rock area displacement field and the corrected right side surrounding rock area displacement field are differentially assimilated and updated once by the update gain, and the assimilated left side surrounding rock area displacement field and the assimilated right side surrounding rock area displacement field are obtained.

6. The machine learning based tunnel cross-fault deformation prediction method of claim 1, wherein, Step S5 is specifically: The displacement field of the assimilated left wall rock area and the displacement field of the assimilated right wall rock area are taken as inputs, and the displacement component of the left wall rock area and the displacement component of the right wall rock area are extracted according to the clear fault plane tangential direction of the fault plane coordinate system at discrete positions of the tunnel axis, and the relative displacement time sequence is calculated; The relative displacement time sequence is input into a quantile prediction module, and the relative displacement prediction values corresponding to multiple quantile levels and their time sequences are output, and monotonicity constraints are applied to each quantile level to ensure that lower quantiles are not greater than higher quantiles; The quantile prediction module is a quantile regression network including a time sequence encoder and a multi-quantile prediction head, the time sequence encoder is any one of a one-dimensional convolutional network, a recurrent network or a Transformer or a combination thereof, is used for feature extraction of the relative displacement time sequence and optional external features, the multi-quantile prediction head is a parallel fully connected branch or a single network with quantile level as conditional input, is used for outputting relative displacement prediction values of a preset quantile set, the monotonicity constraint is a quantile non-decreasing constraint realized by reparameterization or monotonic projection, to ensure that lower quantiles are not greater than higher quantiles; Based on a preset relative displacement threshold, the relative displacement prediction values of the multiple quantile levels are subjected to threshold discrimination to obtain an overrun probability and an overrun time interval, and a relative displacement quantile prediction result is output.

Images