Foundation pit support structure lateral displacement prediction method based on dictionary learning fusion monitoring data

Abstract

The invention provides a foundation pit support structure lateral displacement prediction method based on dictionary learning fusion monitoring data. The foundation pit support structure lateral displacement prediction method comprises the steps that S1, a physical parameter database is established; s2, performing conversion to obtain weak prior distribution of a compressibility parameter ES; s3, sampling by using Latin hypercube to obtain N alternative parameter combinations of the soil body; s4, establishing a finite element model of the target foundation pit; s5, inputting a parameter combination to calculate corresponding side displacement data of the enclosure structure, and forming an over-complete dictionary D; s6, reading measured side displacement data of the enclosure structure to form a monitoring data matrix Y; S7, extracting corresponding data in the over-complete dictionary D, and forming a dimension transformation matrix A matched with the dimension of the monitoring data matrix Y; s8, obtaining an important atom set and a sparse solution vector x by using an orthogonal matching pursuit algorithm program; and S9, obtaining a side displacement prediction curve of the building envelope corrected by combining the monitoring data. The method has the advantages of high prediction precision, low calculation cost and high response speed.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of intelligent construction, and in particular to a method for predicting lateral displacement of a foundation pit retaining structure based on dictionary learning and fusion of monitoring data. BACKGROUND

[0002] In cities with high building density, the analysis of the surrounding influence and risk control of foundation pit excavation are core issues in engineering construction. Foundation pit support, as a temporary support project, often has sudden destruction. The lateral displacement of the foundation pit retaining structure is an important indicator for evaluating the risk of the foundation pit. Excessive lateral displacement of the retaining structure is usually accompanied by structural failure risk, irreversible damage to the surrounding ground subsidence and existing roads and buildings. If the lateral displacement caused by the excavation of the foundation pit can be predicted in advance, it will have high value for evaluating the safety of the foundation pit project and disposing the surrounding influence risk in advance. However, the mechanical properties of rock-soil mass have high uncertainty, and the spatial variability of its parameters is difficult to reveal from a limited number of test holes in the design stage. In addition, complex construction disturbance changes the parameters, making it very difficult to predict the lateral displacement of the retaining structure during construction.

[0003] The existing methods for predicting the lateral displacement of the retaining structure can be mainly divided into theoretical methods, empirical methods, numerical simulation methods, and machine learning methods represented by data-driven proxy models. (1) Theoretical methods are usually limited by simplified theoretical assumptions and are difficult to directly guide specific engineering practice. (2) Empirical methods are mostly based on the summary and induction of existing engineering data, and important factors are selected to derive empirical formulas. These methods are often broad and lack precision, and can only be applied in specific site conditions. (3) Numerical simulation methods represented by finite element methods have low computational efficiency and cannot be well combined with real-time monitoring data. They are difficult to handle the uncertainty of rock-soil parameters. (4) Pure data-driven machine learning proxy models ignore the underlying physical laws, resulting in insufficient model generalization ability and prediction results that may deviate significantly from engineering reality. In addition, training the model requires a large amount of high-quality data set, and the reliability of the prediction model is not stable when the data set is small in the early stage of the project.

[0004] Therefore, there is an urgent need for a method that can dynamically combine real-time engineering monitoring data to reduce the uncertainty of rock-soil parameters and accurately and quickly predict the lateral displacement of the foundation pit retaining structure. SUMMARY

[0005] To address the problems of low prediction accuracy, high computational cost, and slow response speed in existing displacement prediction methods for foundation pit retaining structures, this invention proposes a method for predicting lateral displacement of foundation pit retaining structures based on dictionary learning and fusion of monitoring data. Using a dictionary learning framework, it eliminates the need to train a surrogate model. A complete dictionary is rapidly constructed through parallel computation using a stochastic finite element method, allowing deployment on platforms with low computational power. Monitoring data is incorporated in real-time for model updates, providing rapid, high-precision predictions for foundation pit retaining structures. Since no forward computation model needs to be invoked during the update and prediction phases, this invention enables lightweight deployment in engineering projects.

[0006] The specific technical solution of the present invention is as follows:

[0007] A method for predicting the lateral displacement of foundation pit retaining structures based on dictionary learning and fusion of monitoring data includes the following steps:

[0008] S1: Based on in-situ geological survey and laboratory test data during the design phase, establish a database of physical parameters of the soil and rock masses involved in the foundation pit excavation;

[0009] S2: Use a transformation model to convert the physical parameters in the physical parameter database into the compressibility parameters E of the HSS constitutive model. S The weak prior distribution;

[0010] S3: The prior distribution of the compressibility parameters of the HSS constitutive model is obtained by Latin hypercube sampling to obtain N possible combinations of parameters for soil.

[0011] S4: Establish a geometric model based on the foundation pit support design drawings, establish a finite element model of the target foundation pit according to the construction plan, and control the finite element calculation with a command flow program;

[0012] S5: The finite element calculation program is called in parallel by the Python program, and the input parameters are combined to calculate the corresponding lateral displacement data of the enclosure structure, forming an overcomplete dictionary D;

[0013] S6: Set up monitoring points at the foundation pit site, and read the measured lateral displacement data of the retaining structure under the current working conditions according to the set monitoring frequency. Combine this data with the previously read measured lateral displacement data of the retaining structure to form a monitoring data matrix Y.

[0014] S7: Based on the dimension of the monitoring data matrix Y, use XOR operation to align the encoding, extract the corresponding data from the overcomplete dictionary D, and form a dimension transformation matrix A that matches the dimension of the monitoring data matrix Y;

[0015] S8: Use the orthogonal matching tracking algorithm program, input dimension transformation matrix A and monitoring data matrix Y, output the set of important matching atoms and sparse solution vector x;

[0016] S9: Linearly combine the sparse solution vector x and the overcomplete dictionary D to obtain the predicted lateral displacement curve of the enclosure structure after correction based on monitoring data, and provide interpretable results based on the importance atoms.

[0017] Preferably, the physical parameters include compression modulus, void ratio and porosity, shear wave velocity, standard penetration test blow count, soil layer thickness, soil unit weight, permeability coefficient, and groundwater stability level.

[0018] Preferably, in step S2, the physical parameters are related to the compressive modulus E. S The conversion model is as follows:

[0019] Clay soil: E s = -5.15ln e + 3.53MPa;

[0020] Sandy soil: E s = 111.0exp(-2.89e)MPa;

[0021] The compressibility parameter E of the HSS constitutive model 50 E oed E ur G0 with compressibility modulus E S As an intermediate variable, the transformation model formula is as follows:

[0022] For any soil mass: E 50 =E oed =Es;

[0023] For soft soil: E ur =6~8Es, G0=3~5E ur ;

[0024] For clay, silty clay, and silt: E ur =3~5Es, G0=3~5E ur ;

[0025] For sandy soil and residual granite soil: E ur =3~4Es, G0=3~5E ur .

[0026] Preferably, in step 2, the compression modulus E is set to upper and lower thresholds of 0.5 times and 1.5 times, respectively. S The corresponding weak prior distribution is a uniform distribution.

[0027] Preferably, the finite element model in step S4 adopts the HSS constitutive model, and the finite element model is established using Plaxis2D or Plaxis3D software. The compressive modulus E is obtained by combining the N parameters obtained from the Latin hypercube sampling results in step 3. S As input variable Es ={E s1 E s2 ,...,E sn}, where n is the number of soil layers in the finite element model, and the intermediate variable E is completed in the program. S The conversion of the compressibility parameters E50, Eoed, Eur, and G0 of the HSS constitutive model results in the final output value: Z = {Z1, Z2, ..., Z...}. T}, where T is the number of sub-cases in the finite element simulation, i.e., the number of construction days, and Z T Z is the lateral displacement data of M points with a vertical resolution of 0.5 meters along the depth obtained on day T. Therefore, Z is a T×M single lateral displacement data matrix.

[0028] Preferably, step S5 employs parallel computing technology, utilizing the Plaxis software API to write a Python program that directly calls the pre- and post-processing modules. The main program performs parallel computation, batch extracts the lateral displacement data z of the enclosure structure, and completes the overcomplete dictionary D = {d1, d2, ..., d...}. i ,...d N The assembly phase of}, where d i This represents the i-th dictionary atom, corresponding to the lateral displacement Z of the enclosure structure obtained from the i-th call to the calculation module. i Given a complete dictionary D that is an N×T×M matrix, an index is built based on a fixed resolution, with a spatial index length of M and a temporal index length of T.

[0029] Preferably, in step S7, the measured lateral displacement data of the retaining structure under the current working condition is read and combined with the previously read measured lateral displacement data of the retaining structure to form a monitoring data matrix Y = {y1, y2, ..., y t}, where t is the number of monitoring days with existing records, and y t ={y t1 ,y t2 ,...,y tNz} along the depth measurement point, N z The number of vertical measurement points is Nz. Since dictionary learning has the advantage of being able to predict even without monitoring points, Nz can be less than M, and the monitoring point data can be discontinuous.

[0030] Preferably, in step S8, the orthogonal matching tracking algorithm adopts the solution method of sparse representation model in the field of compressed sensing. For any monitoring data matrix Y, its reconstructed representation is as follows:

[0031] Y = Dx

[0032] x is the corresponding sparse solution vector. Since the dimension of the monitoring data matrix Y is much smaller than that of the overcomplete dictionary matrix D, the problem is underdetermined, and solving for x cannot yield a unique solution. That is, for the overcomplete dictionary matrix D, there are infinitely many solution vectors that satisfy this formula. Therefore, one method to solve this formula is to add sparsity constraints:

[0033]

[0034] To ensure that the elements of the sparse solution vector x are sufficiently sparse, i.e., with a sufficiently small number of non-zero elements, the coefficients of most atoms are close to or equal to 0. The few significantly non-zero atoms are called important atoms, representing those that are significantly important for reconstructing the monitoring data matrix Y in the dictionary and possess intrinsic structural consistency. Simultaneously, the reconstruction accuracy must be sufficiently high. Therefore, the objective function is transformed into...

[0035]

[0036] Where η is the reconstruction error and ε is the minimum threshold of the reconstruction error;

[0037] To solve this L0 norm model, the sequential basis selection method can be used. Among them, the greedy algorithm represented by the orthogonal matching pursuit (OMP) algorithm selects the optimal atom from the atom of the dictionary each time to sparsely approach the monitoring data matrix Y.

[0038] Since the Orthogonal Matching Pursuit (OMP) algorithm requires maintaining a similar dimensional structure between the input dictionary matrix before flattening and the original signal, the actual dimension is t×N. z The monitoring data matrix Y is used to remove unrelated indices using XOR, and the corresponding day and depth data are extracted from the overcomplete dictionary D to form a new dimension transformation matrix A. The dimension transformation matrix A has dimensions of N×t×N. z The order of the dictionary atoms is not changed; only spatial and temporal information is extracted.

[0039] Preferably, the OMP algorithm process can be described as follows:

[0040] Input: Dimension transformation matrix A, monitoring data matrix Y, sparsity level s;

[0041] Output: Support set S and sparse solution vector x;

[0042] Step 1: Initialize by setting i = 0, sparse solution vector x (0) =0, residual r (0) =Y, where the support set S represents the indices of all atoms selected in the current step, and is the initial support set of the atoms.

[0043] Step 2: i = i + 1, find the value of the residual r.(i-1) The most relevant atom, and this atom cannot be a selected atom:

[0044]

[0045] In the formula, a j r is a dictionary atom (i-1) This is the residual from the previous iteration;

[0046] The third step is to update the support set by incorporating the newly selected optimal atom into the support set.

[0047] S (i) =S (i-1) ∪{j};

[0048] Solve the sparse solution vector x in the current step using the least squares method. (i) :

[0049]

[0050] Representative in supporting set S (i) The sub-dictionary under the index is equivalent to the set of selected atoms, and the sparse solution vector x (i) Only in S (i) The corresponding index is not 0, and all others are 0;

[0051] Step 4: Update the residuals

[0052] r (i) =Y-Ax (i) ;

[0053] Step 5: Convergence Determination

[0054] If the current residual satisfies ||r (i) If ||2≤ε or |S|=s, then return x. (i) Otherwise, go back to step two.

[0055] Thus, the corresponding set of importance atoms and the sparse solution vector x are obtained based on the index of the support set S.

[0056] As a preferred option, the prediction results are incorporated after real-time updates of the monitoring data.

[0057]

[0058] Interpretability analysis: By tracing the important atom samples according to the index corresponding to the support set S, the physical parameters of the corresponding atoms can be found in the physical parameter database. This allows us to determine the relative change of the corrected physical parameters compared to the original physical parameters, providing parameter selection suggestions for subsequent engineering.

[0059] The beneficial effects of this invention are:

[0060] 1. Non-proxy model: It does not require a large number of samples to train the proxy model (such as ensemble machine learning model, deep neural network model, etc.) nor does it require a complicated proxy model hyperparameter tuning process. Therefore, it is suitable for rapid deployment in engineering environments with small sample sizes. Foundation pit support projects are often temporary support projects with short construction periods and insufficient data. This invention is more feasible in such engineering applications.

[0061] 2. The overall framework is more efficient and has a lower computational cost: A complete dictionary is built through parallel computing finite element program, and there is no need to call the finite element model again during the update stage. Therefore, it can be deployed on a lightweight platform and has a computational efficiency far higher than traditional back analysis methods, which can help with rapid risk assessment of foundation pit engineering.

[0062] 3. High accuracy and spatial information consideration: As can be verified in the examples, after the framework of this invention is updated, the prediction results are much higher than those of traditional finite element models. Furthermore, since the overcomplete dictionary incorporates the physical prior knowledge of finite elements, it can provide prediction results for the entire displacement curve containing spatial correlation.

[0063] 4. The results are interpretable: Each time the framework is updated, it can provide parameter interpretability analysis. By incorporating real-time monitoring data, it can identify information about parameter selection in important atoms. Therefore, it can provide real-time correction suggestions for soil and rock parameter values ​​and provide corrected parameter values ​​for engineering projects in nearby sites. Attached Figure Description

[0064] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0065] Figure 1 This is a schematic diagram of the entire process of the present invention;

[0066] Figure 2 This is a flowchart of the OMP algorithm of the present invention;

[0067] Figure 3 A schematic diagram of the plan for implementing the second type of foundation pit;

[0068] Figure 4 A cross-sectional schematic diagram for implementing the second type of foundation pit;

[0069] Figure 5 The update strategy diagram for implementing steps 1, 2, and 3 in step 2;

[0070] Figure 6 The corrected results of the lateral displacement curves for four representative working conditions in step-1 of implementation 2 are shown in the figure.

[0071] Figure 7 The corrected results of the lateral displacement curves under four representative working conditions in step-2 of implementation 2 are shown in the figure.

[0072] Figure 8 The diagram shows the corrected lateral displacement curves for four representative operating conditions in step-3 of section 2. Detailed Implementation

[0073] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0074] Example 1

[0075] Reference Figure 1 and Figure 2 A method for predicting the lateral displacement of foundation pit retaining structures based on dictionary learning and fusion of monitoring data includes the following steps:

[0076] S1: Based on in-situ geological survey and laboratory test data during the design phase, establish a database of physical parameters of the soil and rock masses involved in the foundation pit excavation;

[0077] S2: Use a transformation model to convert the physical parameters in the physical parameter database into the compressibility parameters E of the HSS constitutive model. S The weak prior distribution;

[0078] S3: The prior distribution of the compressibility parameters of the HSS constitutive model is obtained by Latin hypercube sampling to obtain N possible combinations of parameters for soil.

[0079] S4: Establish a geometric model based on the foundation pit support design drawings, establish a finite element model of the target foundation pit according to the construction plan, and control the finite element calculation with a command flow program;

[0080] S5: The finite element calculation program is called in parallel by the Python program, and the input parameters are combined to calculate the corresponding lateral displacement data of the enclosure structure, forming an overcomplete dictionary D;

[0081] S6: Set up monitoring points at the foundation pit site, and read the measured lateral displacement data of the retaining structure under the current working conditions according to the set monitoring frequency. Combine this data with the previously read measured lateral displacement data of the retaining structure to form a monitoring data matrix Y.

[0082] S7: Based on the dimension of the monitoring data matrix Y, use XOR operation to align the encoding, extract the corresponding data from the overcomplete dictionary D, and form a dimension transformation matrix A that matches the dimension of the monitoring data matrix Y;

[0083] S8: Use the orthogonal matching tracking algorithm program, input dimension transformation matrix A and monitoring data matrix Y, output the set of important matching atoms and sparse solution vector x;

[0084] S9: Linearly combine the sparse solution vector x and the overcomplete dictionary D to obtain the predicted lateral displacement curve of the enclosure structure after correction based on monitoring data, and provide interpretable results based on the importance atoms.

[0085] The physical parameters include compression modulus, void ratio and porosity, shear wave velocity, standard penetration test blow count, soil layer thickness, soil unit weight, permeability coefficient, and groundwater stability level.

[0086] In step S2, the physical parameters and the compressive modulus E S The conversion model is as follows:

[0087] Clay soil: E s = -5.15ln e + 3.53MPa;

[0088] Sandy soil: E s = 111.0exp(-2.89e)MPa;

[0089] The compressibility parameter E of the HSS constitutive model 50 E oed E ur G0 with compressibility modulus E S As an intermediate variable, the transformation model formula is as follows:

[0090] For any soil mass: E 50 =E oed =Es;

[0091] For soft soil: E ur =6~8Es, G0=3~5E ur ;

[0092] For clay, silty clay, and silt: E ur =3~5Es, G0=3~5E ur ;

[0093] For sandy soil and residual granite soil: E ur =3~4Es, G0=3~5E ur .

[0094] In a preferred embodiment, step 2 uses a compression modulus E with upper and lower thresholds of 0.5 times and 1.5 times, respectively. S The corresponding weak prior distribution is a uniform distribution.

[0095] The finite element model in step S4 adopts the HSS constitutive model and is established using Plaxis 2D or Plaxis 3D software. The compressibility modulus E is obtained by combining the N parameters from the Latin hypercube sampling results in step 3. S As input variable E s ={E s1 E s2 , ..., E sn}, where n is the number of soil layers in the finite element model, and the intermediate variable E is completed in the program. S The conversion of the compressibility parameters E50, Eoed, Eur, and G0 of the HSS constitutive model results in the final output value: Z = {Z1, Z2, ..., Z...}, where Z is the lateral displacement of the retaining structure of the foundation pit extracted every 0.5 meters along the depth. T}, where T is the number of sub-cases in the finite element simulation, i.e., the number of construction days, and Z T Z is the lateral displacement data of M points with a vertical resolution of 0.5 meters along the depth obtained on day T. Therefore, Z is a T×M single lateral displacement data matrix.

[0096] In step S5, parallel computing technology is employed. Using the Plaxis software API, a Python program is written to directly call the pre- and post-processing modules. The main program performs parallel computation, batch extracting the lateral displacement data z of the enclosure structure, and completing the overcomplete dictionary D = {d1, d2, ..., d...}. i ,...d N The assembly phase of}, where d i This represents the i-th dictionary atom, corresponding to the lateral displacement Z of the enclosure structure obtained from the i-th call to the calculation module. i Given a complete dictionary D that is an N×T×M matrix, an index is built based on a fixed resolution, with a spatial index length of M and a temporal index length of T.

[0097] In step S7, the measured lateral displacement data of the retaining structure under the current working condition is read and combined with the previously read measured lateral displacement data of the retaining structure to form a monitoring data matrix Y = {y1, y2, ..., y...}. t}, where t is the number of monitoring days with existing records, and y t ={y t1 y t2 ,...,y tNz} along the depth measurement point, N zThe number of vertical measurement points is Nz. Since dictionary learning has the advantage of being able to predict even without monitoring points, Nz can be less than M, and the monitoring point data can be discontinuous.

[0098] In step S8, the orthogonal matching tracking algorithm adopts the solution method of sparse representation model in the field of compressed sensing. For any monitoring data matrix Y, its reconstructed representation is as follows:

[0099] Y = Dx

[0100] x is the corresponding sparse solution vector. Since the dimension of the monitoring data matrix Y is much smaller than that of the overcomplete dictionary matrix D, the problem is underdetermined, and solving for x cannot yield a unique solution. That is, for the overcomplete dictionary matrix D, there are infinitely many solution vectors that satisfy this formula. Therefore, one method to solve this formula is to add sparsity constraints:

[0101]

[0102] To ensure that the elements of the sparse solution vector x are sufficiently sparse, i.e., with a sufficiently small number of non-zero elements, the coefficients of most atoms are close to or equal to 0. The few significantly non-zero atoms are called important atoms, representing those that are significantly important for reconstructing the monitoring data matrix Y in the dictionary and possess intrinsic structural consistency. Simultaneously, the reconstruction accuracy must be sufficiently high. Therefore, the objective function is transformed into...

[0103]

[0104] Where η is the reconstruction error and ε is the minimum threshold of the reconstruction error;

[0105] To solve this L0 norm model, the sequential basis selection method can be used. Among them, the greedy algorithm represented by the orthogonal matching pursuit (OMP) algorithm selects the optimal atom from the atom of the dictionary each time to sparsely approach the monitoring data matrix Y.

[0106] Since the Orthogonal Matching Pursuit (OMP) algorithm requires maintaining a similar dimensional structure between the input dictionary matrix before flattening and the original signal, the actual dimension is t×N. z The monitoring data matrix Y is used to remove unrelated indices using XOR, and the corresponding day and depth data are extracted from the overcomplete dictionary D to form a new dimension transformation matrix A, with dimensions N×t×N. z The order of the dictionary atoms is not changed; only spatial and temporal information is extracted.

[0107] The OMP algorithm can be described as follows:

[0108] Input: Dimension transformation matrix A, monitoring data matrix Y, sparsity level s;

[0109] Output: Support set S and sparse solution vector x;

[0110] Step 1: Initialize by setting i = 0, sparse solution vector x (0) =0, residual r (0) =Y, where the support set S represents the indices of all atoms selected in the current step, and is the initial support set of the atoms.

[0111] Step 2: i = i + 1, find the value of the residual r. (i-1) The most relevant atom, and this atom cannot be a selected atom:

[0112]

[0113] In the formula, a j r is a dictionary atom (i-1) This is the residual from the previous iteration;

[0114] The third step is to update the support set by incorporating the newly selected optimal atom into the support set.

[0115] S (i) =S (i-1 )∪{j};

[0116] Solve the sparse solution vector x in the current step using the least squares method. (i) :

[0117]

[0118] Representative in supporting set S (i) The sub-dictionary under the index is equivalent to the set of selected atoms, and the sparse solution vector x (i) Only in S (i) The corresponding index is not 0, and all others are 0;

[0119] Step 4: Update the residuals

[0120] r (i) =Y-Ax (i) ;

[0121] Step 5: Convergence Determination

[0122] If the current residual satisfies |r (i) If ||2≤ε or |S|=s, then return x. (i) Otherwise, go back to step two.

[0123] Thus, the corresponding set of importance atoms and the sparse solution vector x are obtained based on the index of the support set S.

[0124] As a preferred option, the prediction results are incorporated after real-time updates of the monitoring data.

[0125]

[0126] Interpretability analysis: By tracing the important atom samples according to the index corresponding to the support set S, the physical parameters of the corresponding atoms can be found in the physical parameter database. This allows us to determine the relative change of the corrected physical parameters compared to the original physical parameters, providing parameter selection suggestions for subsequent engineering.

[0127] Among them, the HSS constitutive model (Hardening Soil Small-strain model) is an elastoplastic constitutive model for geotechnical engineering. It is developed based on the hardened soil model (HS model) and describes the mechanical behavior of soil under complex stress paths by introducing small strain characteristics. It is especially suitable for deformation analysis of underground engineering.

[0128] Among them, the compression modulus (E) S The stress-strain ratio (SSR) is the ratio of vertical compressive stress to total vertical strain in soil when it is completely laterally indeformable. It is an important indicator of soil compressibility and a crucial parameter for calculating foundation settlement. It is generally used to calculate the final settlement of a foundation under approximately one-dimensional deformation conditions, and experimentally, it can be determined by the slope of the initial segment of the stress-strain curve.

[0129] Example 2

[0130] This embodiment is an engineering example applied to a foundation pit project in Fuzhou, as detailed below:

[0131] See Figure 3 The plan shows that the standard section of the foundation pit is 22.8 meters wide, while the enlarged section (6#) is 30 meters wide. Pit 6# has the shortest clearance to the existing Metro Line 1 and requires the most stringent deformation control. The QCX20 measuring point in pit 6# was selected for analysis.

[0132] See Figure 4 The cross-sectional view shows that Pit 6 was constructed using the open-cut method, with a maximum excavation depth of H. e =18.1m, pit width B=30m, retaining structure adopts Φ1400@1000 full casing full rotation interlocking pile, internal support system is 5 layers of internal support, the first and third layers of internal support from top to bottom are 0.9m×0.9m concrete support, the second layer is Φ800mm thick t=16mm steel support, the fourth and fifth layers are Φ800mm thick t=16mm servo steel support.

[0133] According to geological survey data, the soil and rock masses involved in Pit 6, in descending order of depth, are: miscellaneous fill, silt, silty clay, residual sandy clayey soil, and completely weathered granite. The physical parameters of the soil and rock masses include: compression modulus, void ratio and porosity, shear wave velocity, standard penetration test blow count, soil layer thickness, soil unit weight, permeability coefficient, and stable groundwater level.

[0134] The geological survey has provided the compression modulus of the corresponding soil layer; therefore, upper and lower threshold values ​​of 0.5 times and 1.5 times the compression modulus E are used, respectively. S The prior distribution is a uniform distribution.

[0135] The compression modulus E of N parameter combinations is obtained using Latin hypercube sampling (LHS) based on a prior distribution. S Use a transformation model to map it to the parameter space of the HSS constitutive model, where N can be 50 or 100.

[0136] Based on the foundation pit support design drawings, a geometric model is established. Based on the construction plan, a finite element model matching the actual construction conditions is created. The foundation pit finite element model is established using Plaxis 2D. The soil types all adopt the HSS constitutive model to describe the nonlinear behavior of the soil during excavation. Soil parameters are defined as random variables. All of the above processes can be calculated by writing Python program command flow control.

[0137] The variable input value is E s ={E s1 E s2 , ..., E sx}, where n is the number of soil layers in the finite element model, and the intermediate variable E is completed in the program. S To the compressibility parameter E of the HSS constitutive model 50 E oed E ur The conversion of G0 ultimately outputs the lateral displacement results of the retaining structure of the foundation pit, extracted every 0.5 meters along the depth: Z = {Z1, Z2, ..., Z...} T}, where T is the number of sub-cases in the finite element simulation (number of construction days), Z T Z is the lateral displacement data of M points with a vertical resolution of 0.5 meters along the depth obtained on day T. Therefore, Z is a T×M single lateral displacement data matrix.

[0138] This example uses dual-process parallel computation, therefore the computation time is N / 2 × the time of a single finite element calculation. An overcomplete dictionary D = {d1, d2, ..., d...} is completed. i ,...d N The assembly phase of}, where d i This represents the i-th dictionary atom, corresponding to the lateral displacement Z of the enclosure structure obtained from the i-th call to the calculation module. iThe complete dictionary D is an N×T×M matrix.

[0139] According to embodiments of the present invention, there is no upper limit to the number of concurrent applications; the specific limit depends on the computing device capabilities.

[0140] Based on the monitoring points and frequency deployed on-site, during the foundation pit construction period, the monitoring frequency is once per day as required by specifications. The vertical resolution of the inclinometer tubes is 0.5 meters, meaning lateral displacement data is read every 0.5 meters. The measured lateral displacement data of the retaining structure are used to form a monitoring data matrix Y = {y1, y2, ..., y...}. t}, where t is the number of monitoring days with existing records. Note that y t ={y t1 y t2 ,...,y tNz Along the depth measurement points, according to the spatial index {1, 2, 3, ..., M} in the overcomplete dictionary D, N is determined by the rule. z This represents the number of vertical measurement points within the index.

[0141] The monitoring data matrix Y and the overcomplete dictionary D are saved as CSV files. The NumPy library is used to read the data and store it as a tensor structure. The spatial and temporal indices of the overcomplete dictionary D and the monitoring data matrix Y are aligned using the 0xR operation. Data rows without monitoring data are filtered out, based on a dimension of t×N. z The monitoring data matrix Y is used to extract data corresponding to the number of days and depth from the overcomplete dictionary D to form a new dimension transformation matrix A. The dimension transformation matrix A has dimensions of N×t×N. z .

[0142] In this example, four discontinuous representative operating conditions of the QCX20 were used as monitoring conditions. Lateral displacement values ​​were extracted every 0.5 meters along a depth of 20 meters, i.e., t = 4, N. z =40. To recreate the actual process of monitoring data input during excavation in a real-world project, a real-time update strategy is constructed as follows: Figure 5 As shown, steps 1, 2, and 3 correspond to the excavation conditions of 7m, 8.5m, and 14m of the foundation pit, respectively.

[0143] The Orthogonal Matching Pursuit (OMP) algorithm is used, the specific principle of which is described above. The inputs are the dimensionality transformation matrix A and the monitoring data matrix Y, with a sparsity level of 5. Each update performs only one OMP calculation, resulting in low computational cost and a time consumption of approximately 0.02 seconds. The output support set S and sparse solution vector x are shown in Table 1.

[0144] Table 1

[0145]

[0146] The output sparse solution vector x is linearly combined with the original overcomplete dictionary D to obtain the predicted lateral shift value:

[0147]

[0148] like Figures 6-8 The figure shows the prediction results of the lateral displacement curves under four representative working conditions. As can be seen from the figure, the lateral displacement prediction results updated by dictionary learning are more accurate than the random finite element lateral displacement prediction results before the update. In the unexcavated working condition, it is closer to the measured results, which verifies the effectiveness of dictionary learning and fusion of monitoring data in correcting the lateral displacement prediction of the foundation pit retaining structure.

[0149] The interpretability results of dictionary learning can be obtained by backtracking the dictionary atoms in the support set S, as shown in Table 2. The important atoms are arranged in descending order of weight, corresponding to the normalized compressibility parameters of the first four soil layers (the baseline value of the normalized parameter is set to 1, corresponding to the initial geological exploration parameters: when the parameter is less than 1, it indicates that the original parameter needs to be reduced; when it is greater than 1, it indicates that the original parameter needs to be amplified). The interpretability analysis results show that the parameter values ​​in the third column are all less than 1, reflecting an overestimation of the compressibility index of the third layer of silty clay by the initial geological exploration parameters. Therefore, a qualitative suggestion can be given: in the deformation prediction considering the impact of construction disturbance within this site area, the compressibility index of the third layer of silty clay should be reduced.

[0150] Table 2

[0151]

[0152] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting the lateral displacement of foundation pit retaining structures based on dictionary learning and fusion of monitoring data, characterized in that: Includes the following steps: S1: Based on in-situ geological survey and laboratory test data during the design phase, establish a database of physical parameters of the soil and rock masses involved in the foundation pit excavation; S2: Use a transformation model to convert the physical parameters in the physical parameter database into the compressibility parameters E of the HSS constitutive model. S The weak prior distribution; S3: The prior distribution of the compressibility parameters of the HSS constitutive model is obtained by Latin hypercube sampling, resulting in N possible combinations of parameters for the soil. S4: Establish a geometric model based on the foundation pit support design drawings, establish a finite element model of the target foundation pit according to the construction plan, and control the finite element calculation with a command flow program; S5: The finite element calculation program is called in parallel by the Python program, and the input parameters are combined to calculate the corresponding lateral displacement data of the enclosure structure, forming an overcomplete dictionary D; S6: Set up monitoring points at the foundation pit site, and read the measured lateral displacement data of the retaining structure under the current working conditions according to the set monitoring frequency. Combine this data with the previously read measured lateral displacement data of the retaining structure to form a monitoring data matrix Y. S7: Based on the dimension of the monitoring data matrix Y, use XOR operation to align the encoding, extract the corresponding data from the overcomplete dictionary D, and form a dimension transformation matrix A that matches the dimension of the monitoring data matrix Y; S8: Use the orthogonal matching tracking algorithm program, input dimension transformation matrix A and monitoring data matrix Y, output the set of important matching atoms and sparse solution vector x; S9: Linearly combine the sparse solution vector x and the overcomplete dictionary D to obtain the predicted lateral displacement curve of the enclosure structure after correction based on monitoring data, and provide interpretable results based on the importance atoms.

2. The method for predicting the lateral displacement of foundation pit retaining structures based on dictionary learning and fusion of monitoring data as described in claim 1, characterized in that: The physical parameters include compression modulus, void ratio and porosity, shear wave velocity, standard penetration test blow count, soil layer thickness, soil unit weight, permeability coefficient, and groundwater stability level.

3. The method for predicting the lateral displacement of foundation pit retaining structures based on dictionary learning and fusion of monitoring data as described in claim 1, characterized in that: In step S2, the physical parameters and the compressive modulus E S The conversion model, The formula is as follows: Clay soil: E s = -5.15lne + 3.53MPa; Sandy soil: E s = 111.0 exp(-2.89e) MPa; The compressibility parameter E of the HSS constitutive model 50 E oed E ur G0 with compressibility modulus E S As an intermediate variable, the transformation model formula is as follows: For any soil mass: E 50 =E oed =Es; For soft soil: E ur =6~8Es, G0=3~5E ur ; For clay, silty clay, and silt: E ur =3~5Es, G0=3~5E ur ; For sandy soil and residual granite soil: E ur =3~4Es, G0=3~5E ur .

4. The method for predicting the lateral displacement of foundation pit retaining structures based on dictionary learning and fusion of monitoring data as described in claim 1, characterized in that: In step 2, the upper and lower threshold values ​​of the compression modulus E are 0.5 times and 1.5 times, respectively. S The corresponding weak prior distribution is a uniform distribution.

5. The method for predicting the lateral displacement of foundation pit retaining structures based on dictionary learning and fusion of monitoring data as described in claim 1, characterized in that: The finite element model in step S4 adopts the HSS constitutive model and is established using Plaxis2D or Plaxis3D software. The compressibility modulus E is obtained by combining the N parameters from the Latin hypercube sampling results in step 3. S As input variable E s ={E s1 E s2 ,...,E sn }, where n is the number of soil layers in the finite element model, and the intermediate variable E is completed in the program. S The conversion of the compressibility parameters E50, Eoed, Eur, and G0 of the HSS constitutive model results in the final output value: Z = {Z1, Z2, ..., Z...}, where Z is the lateral displacement of the retaining structure of the foundation pit extracted every 0.5 meters along the depth. T }, where T is the number of sub-cases in the finite element simulation, i.e., the number of construction days, and Z T Z is the lateral displacement data of M points with a vertical resolution of 0.5 meters along the depth obtained on day T. Therefore, Z is a T×M single lateral displacement data matrix.

6. The method for predicting the lateral displacement of foundation pit retaining structures based on dictionary learning and fusion of monitoring data as described in claim 1, characterized in that: In step S5, parallel computing technology is employed. Using the Plaxis software API, a Python program is written to directly call the pre- and post-processing modules. The main program implements parallel computation, batch extracting the lateral displacement data z of the enclosure structure, and completing the overcomplete dictionary D = {d1, d2, ..., d...}. i ,...d N The assembly phase of}, where d i This represents the i-th dictionary atom, corresponding to the lateral displacement Z of the enclosure structure obtained from the i-th call to the calculation module. i Given a complete dictionary D that is an N×T×M matrix, an index is built based on a fixed resolution, with a spatial index length of M and a temporal index length of T.

7. The method for predicting the lateral displacement of foundation pit retaining structures based on dictionary learning and fusion of monitoring data as described in claim 1, characterized in that: In step S7, the measured lateral displacement data of the retaining structure under the current working condition is read and combined with the previously read measured lateral displacement data of the retaining structure to form a monitoring data matrix Y = {y1, y2, ..., y...} t }, where t is the number of monitoring days with existing records, and y t ={y t1 y t2 , ..., y tNz } along the depth measurement point, N z Nz represents the number of vertical measurement points. Since dictionary learning has the advantage of being able to predict even without monitoring points, Nz can be less than M, and the monitoring point data can be discontinuous.

8. The method for predicting the lateral displacement of foundation pit retaining structures based on dictionary learning and fusion of monitoring data as described in claim 1, characterized in that: In step S8, the orthogonal matching tracking algorithm adopts the solution method of sparse representation model in the field of compressed sensing. For any monitoring data matrix Y, its reconstructed representation is as follows: Y = Dx x is the corresponding sparse solution vector. Since the dimension of the monitoring data matrix Y is much smaller than that of the overcomplete dictionary matrix D, the problem is underdetermined, and solving for x cannot yield a unique solution. That is, for the overcomplete dictionary matrix D, there are infinitely many solution vectors that satisfy this formula. Therefore, one method to solve this formula is to add sparsity constraints: To ensure that the elements of the sparse solution vector x are sufficiently sparse, i.e., with a sufficiently small number of non-zero elements, the coefficients of most atoms are close to or equal to 0. The few significantly non-zero atoms are called important atoms, representing those that are significantly important for reconstructing the monitoring data matrix Y in the dictionary and possess intrinsic structural consistency. Simultaneously, the reconstruction accuracy must be sufficiently high. Therefore, the objective function is transformed into... Where η is the reconstruction error and ε is the minimum threshold of the reconstruction error; To solve this L0 norm model, the sequential basis selection method can be used. Among them, the greedy algorithm represented by the orthogonal matching pursuit (OMP) algorithm selects the optimal atom from the atom of the dictionary each time to sparsely approach the monitoring data matrix Y. Since the Orthogonal Matching Pursuit (OMP) algorithm requires maintaining a similar dimensional structure between the input dictionary matrix before flattening and the original signal, the actual dimension is t×N. z The monitoring data matrix Y is used to remove unrelated indices using XOR, and the corresponding day and depth data are extracted from the overcomplete dictionary D to form a new dimension transformation matrix A. The dimension transformation matrix A has dimensions of N×t×N. z The order of the dictionary atoms is not changed; only spatial and temporal information is extracted.

9. The method for predicting the lateral displacement of foundation pit retaining structures based on dictionary learning and fusion of monitoring data as described in claim 1, characterized in that: The process of the OMP algorithm can be described as follows: Input: Dimension transformation matrix A, monitoring data matrix Y, sparsity level s; Output: Support set S and sparse solution vector x; Step 1: Initialize by setting i = 0, sparse solution vector x (0) =0, residual r (0) =Y, where the support set S represents the indices of all atoms selected in the current step, and is the initial support set of the atoms. Step 2: i = i + 1, find the value of the residual r. (i-1) The most relevant atom, and this atom cannot be a selected atom: In the formula, a j r is a dictionary atom (i-1) This is the residual from the previous iteration; The third step is to update the support set by incorporating the newly selected optimal atom into the support set. S (i) =S (i-1) ∪{j}; Solve the sparse solution vector x in the current step using the least squares method. (i) : Representative in supporting set S (i) The sub-dictionary under the index is equivalent to the set of selected atoms, and the sparse solution vector x (i) Only in S (i) The corresponding index is not 0, and all others are 0; Step 4: Update the residuals r (i) =Y-Ax (i) ; Step 5: Convergence Determination If the current residual satisfies ||r (i) If ||2≤ε or |S|=s, then return x. (i) Otherwise, go back to step two. Thus, the corresponding set of importance atoms and the sparse solution vector x are obtained based on the index of the support set S.

10. The method for predicting the lateral displacement of foundation pit retaining structures based on dictionary learning and fusion of monitoring data as described in claim 1, characterized in that: Prediction results after incorporating real-time updates of monitoring data Interpretability analysis: By tracing the important atom samples according to the index corresponding to the support set S, the physical parameters of the corresponding atoms can be found in the physical parameter database. This allows us to determine the relative change of the corrected physical parameters compared to the original physical parameters, providing parameter selection suggestions for subsequent engineering.

Images