BPNN parameter inversion and LSTM tunnel deformation prediction methods

Abstract

This invention relates to the field of tunnel deformation prediction technology, specifically to a BPNN parameter inversion and LSTM tunnel deformation prediction method. Based on the principle of orthogonal experimental design, multiple parameter combinations are designed for the M-C constitutive model and the HSS constitutive model, respectively, and numerical calculations are performed to obtain vertical displacement data of tunnel measuring points. A BPNN model is constructed, using the vertical displacement data as input and the corresponding constitutive model parameters as output, and the BPNN model is trained. Field-measured tunnel settlement data is input into the trained BPNN model to invert and obtain the optimal constitutive parameters of the M-C and HSS constitutive models. The optimal constitutive parameters obtained through inversion are used to perform forward numerical simulation of the entire pipe jacking construction process to obtain the evolution law of tunnel vertical displacement. Based on a long short-term memory neural network, the vertical displacement monitoring values ​​of the downstream tunnel measuring points are input to obtain the predicted results of future tunnel settlement. This invention improves the accuracy of the predicted values.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of tunnel deformation prediction, and particularly relates to a BPNN parameter inversion and LSTM tunnel deformation prediction method. BACKGROUND

[0002] In the construction process of a tunnel and underground engineering, the original stress balance of the rock-soil body is broken due to excavation, which can cause deformation of the surrounding rock. If the deformation is too large or out of control, it can cause catastrophic accidents such as collapse and ground subsidence. Therefore, real-time and accurate prediction of tunnel deformation is the key to ensuring construction safety, optimizing support parameters, and controlling environmental impact.

[0003] A traditional method for predicting tunnel deformation is to input comprehensive tunnel data into a trained auto-encoding network model to obtain tunnel features, and then input the tunnel features into a trained long short-term memory (LSTM) network model to obtain a tunnel deformation prediction result. However, when predicting tunnel deformation using only the LSTM network model, the prediction is completely based on data correlation and lacks the support of physical and mechanical principles. When the prediction environment changes, such as changes in construction technology or the construction passing through different layers, similar situations have never occurred in the historical training data, resulting in a large deviation between the predicted result and the measured value. The generalization ability of the LSTM network model for predicting tunnel deformation is weak. SUMMARY

[0004] To solve the technical problem that the deviation between the predicted result and the measured value is large when only the LSTM network model is used for prediction in situations that have never occurred in the training data, the present application provides a BPNN parameter inversion and LSTM tunnel deformation prediction method, and the technical solution is as follows:

[0005] Based on the principle of orthogonal experiment design, multiple groups of parameter combinations are designed for the M-C and HSS constitutive models, and numerical calculations are performed to obtain tunnel measurement point vertical displacement data;

[0006] A BPNN model is constructed, with the tunnel measurement point vertical displacement data as input and the corresponding constitutive model parameters as output, and the BPNN model is trained. The field measured tunnel settlement data is input into the trained BPNN model to obtain the optimal constitutive parameters of the M-C and HSS constitutive models through inversion;

[0007] The optimal constitutive parameters obtained through inversion are used for forward numerical simulation of the whole process of pipe jacking construction to obtain the evolution law of tunnel vertical displacement;

[0008] Based on long short-term memory neural network, historical monitoring data is taken as a training set, and the vertical displacement monitoring value of the downlink tunnel measuring point is input to obtain the prediction result of the future settlement of the tunnel.

[0009] Further, a three-dimensional calculation model including the pipe, the subway tunnel and the stratum structure is established, and the structured and unstructured mixed grid is divided.

[0010] Further, the establishment of the three-dimensional calculation model comprises:

[0011] Rhino software is used for geometric modeling, the grid is divided through the Griddle plug-in, and is exported to FLAC3D for numerical calculation; the model boundary condition is set as: the left and right and front and back plane normal displacement constraint, the bottom surface three-way displacement constraint, and the top surface is free.

[0012] Further, the long short-term memory neural network comprises: normalizing the monitoring data, and adopting a sliding window to construct an input-output sequence.

[0013] Further, the long short-term memory neural network comprises: utilizing the Adam optimization method to optimize the model hyperparameters.

[0014] Further, the long short-term memory neural network based on the historical monitoring data as the training set comprises:

[0015] The historical monitoring data in the neural network training data set is divided into measured period data and prediction period data, and the measured period data is standardized.

[0016] The embodiment of the application has at least the following beneficial effects:

[0017] The long short-term memory LSTM deep learning algorithm is adopted, the downlink tunnel settlement monitoring value is taken as the neural network training data set, the monitoring data is divided into a training set and a validation set according to a proportion, the LSTM network model is trained and verified to predict the settlement deformation of the measuring point. The results show that the change rule of the predicted value obtained by the application is very close to the measured value, the error range of the model in predicting the deformation of the subway tunnel in the pipe construction process can be controlled in the millimeter level, is within the allowable error range, and the implementation process is convenient and fast, without too much manual intervention, can replace the time-consuming and complex numerical calculation under certain conditions, and the grid and the constitutive parameter are sensitive. BRIEF DESCRIPTION OF DRAWINGS

[0018] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, and the advantages thereof, below will briefly introduce the drawings needed in the embodiments or prior art description. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 The method flow chart of the BPNN parameter inversion and the LSTM tunnel deformation prediction method provided by an embodiment of the present application;

[0020] Figure 2 The schematic diagram of the pipe jacking pipe passing through the stratum of the subway No. 6 line provided by an embodiment of the present application;

[0021] Figure 3 The schematic diagram of the main modeling process of the Rhino and Griddle mixed grid provided by an embodiment of the present application;

[0022] Figure 4 The schematic diagram of the corresponding calculation model of FLAC3D provided by an embodiment of the present application;

[0023] Figure 5 The flow chart of the calculation and parameter inversion corresponding to the calculation model provided by an embodiment of the present application;

[0024] Figure 6 The schematic diagram of the 3-layer BPNN topological relationship adopted by the HSS constitutive parameter inversion provided by an embodiment of the present application;

[0025] Figure 7 The schematic diagram of the 3-layer BPNN topological relationship adopted by the M-C constitutive parameter inversion provided by an embodiment of the present application;

[0026] Figure 8 The schematic diagram of the work flow of the soil material parameter inversion by using the BPNN neural network provided by an embodiment of the present application;

[0027] Figure 9 The schematic diagram of the neural network parameter fitting result of the HSS model provided by an embodiment of the present application;

[0028] Figure 10 The schematic diagram of the neural network parameter fitting result of the M-C model provided by an embodiment of the present application;

[0029] Figure 11 The flow chart of the pipe jacking construction simulation calculation provided by an embodiment of the present application;

[0030] Figure 12A schematic diagram of an LSTM network construction and prediction process provided by an embodiment of the present application. DETAILED DESCRIPTION

[0031] In order to further illustrate the technical means and effects taken by the present application to achieve the predetermined purposes, the BPNN parameter inversion and LSTM tunnel deformation prediction method according to the present application, the specific implementation, structure, features and effects thereof, are described in detail as follows in combination with the drawings and preferred embodiments. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. In addition, the specific features, structures or characteristics in one or more embodiments can be combined in any suitable form.

[0032] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the present application belongs.

[0033] The specific scheme of the BPNN parameter inversion and LSTM tunnel deformation prediction method provided by the present application is specifically described below in combination with the drawings.

[0034] Please refer to Figure 1 , which shows the step flowchart of the BPNN parameter inversion and LSTM tunnel deformation prediction method provided by an embodiment of the present application. The method comprises the following steps:

[0035] In step S100, based on the orthogonal test design principle, a plurality of parameter combinations are designed for M-C constitutive model and HSS constitutive model respectively, numerical calculation is performed, and tunnel measuring point vertical displacement data is obtained.

[0036] Please refer to Figure 2 , Figure 2 A schematic diagram of the top pipe passing through the stratum of the subway line 6.

[0037] The embodiment of the present application is based on a certain water intake upper moving project. According to the schematic diagram of the top pipe passing through the stratum of the subway line 6, the top pipe construction 50m range line is taken as the calculation model x direction length, the double launch top pipe net distance horizontal line center is taken as the coordinate origin, and the x positive and negative directions are respectively taken as 25m. According to the top pipe line and the position relationship diagram of the subway line 6, the top pipe jacking process is divided into 50m test section, 50m subway protection area before and after passing through, and 20.2m core area passing through. The model y direction length is set to 170m. According to the geological profile in the schematic diagram of the top pipe passing through the stratum of the subway line 6, the model z direction line is divided into seven layers in the top pipe construction 50m range line, which are respectively miscellaneous fill, denoted as ZTT, simple fill, denoted as STT, silty clay, denoted as FZNT, silt clay with silt, denoted as YNZFZNTJFT, sandy silty clay, denoted as HSFZNT, round gravel, denoted as YL, and strongly weathered mudstone, denoted as QFHNY.

[0038] A three-dimensional calculation model containing the pipe jacking, the subway tunnel and the stratum structure is established, and a structured and unstructured mixed grid is divided. According to the positional relationship between the pipe jacking line and the subway tunnel, the model subway tunnel gradually increases in depth from west to east in space, the inner diameters of the two tunnels are both 5.5 m, the lining segment thickness is 0.35 m, the outer diameter is 6.2 m, the clear distance between the up and down lines tunnels is 14 m, the lining segment is C50 concrete material, and the segment width is 1.2 m; the left pipe jacking is DN2800, the right pipe jacking is DN3200, the diameters of the left and right pipe jacking heads are 2.87 m and 3.27 m respectively, the lengths are 4.55 m and 4.8 m respectively, the diameters of the left and right steel pipe jacking segments are 2.8 m and 3.2 m respectively, the wall thicknesses are 0.03 m and 0.034 m respectively, and the steel pipe is Q355B material; the drag-reducing mud system is an important link for controlling the jacking force of the long-distance pipe jacking, can effectively reduce the pipe jacking friction, thereby reducing the number of use of the relay ring in the pipe jacking construction process, and if the injected lubricating mud can form a relatively complete mud sleeve around the pipe, the friction reduction effect is better. In addition, when the mud sleeve around the pipe wall is formed, the purpose of controlling the settlement can be achieved by reducing the soil carrying phenomenon.

[0039] Since the pipe jacking cutter head diameter is greater than the steel pipe by 2 cm, the gap generated after passing through the soil needs to be filled by the mud sleeve, if a mud pressure equivalent to the soil pressure is maintained between this ring sleeve and the jacking pipe, the mud bears the entire soil pressure, so that the soil pressure is no longer directly loaded on the pipe wall, but indirectly loaded on the pipe wall through the mud. In the embodiment of the application, the mud pressure is controlled at 0.1-0.24 MPa, and the mud sleeve is modeled according to the equivalent layer, and is activated and grouted during the pipe jacking process.

[0040] The calculation model is established by using a professional three-dimensional modeling software Rhino 7.0, first, the tunnel and the pipe jacking geometry are established in Rhino 7.0, then the mesh division is performed by using the Mesh function in Rhino 7.0, finally, the structured and unstructured mixed grid is generated by using the mesh division plug-in Griddle 2.0 developed by ITASCA company in Rhino 7.0, the corresponding grid data file of FLAC3D is exported, and the calculation is performed by importing FLAC3D, and the main modeling process is shown in Figure 3 , Figure 3 The schematic diagram of the main modeling process of Rhino and Griddle mixed grid.

[0041] The specific modeling steps are as follows:

[0042] 1. Tunnel contour line;

[0043] Establish the lining layer, draw the lining cross-section profile through the tunnel lining coordinate, and generate the tunnel axis by using the single track sweep according to the actual tunnel direction.

[0044] 2. Tunnel surface generation;

[0045] Close the tunnel profile by using the "cap" command, generate the tunnel surface and hide the profile line. Create a new tunnel lining surface layer, and copy all the objects to the lining surface layer.

[0046] 3. Tunnel mesh generation;

[0047] Establish the tunnel lining mesh layer, input the block mesh command in the command window, set the maximum edge length and other parameters, change the mesh layer to the lining mesh layer, set the output mode to FLAC3D format and name it as lining.

[0048] 4. Generate inner and outer lining mesh;

[0049] Establish two layers, namely the outer lining mesh and the inner lining mesh, extract the surface by using the GExtract command, set the maximum break angle Max BreakAngle = 89, and form the inner and outer two layers of mesh.

[0050] 5. Generate the surrounding rock area;

[0051] Select the inner lining mesh and copy it to the inner lining mesh layer; select the outer lining mesh and copy it to the outer lining mesh layer, close the lining mesh layer, copy the outer lining mesh layer and the object, and close all layers. Establish the surrounding rock area layer, input the spatial coordinates of the surrounding rock area in the command window, and generate the surrounding rock area.

[0052] 6. Calculate the model mesh generation;

[0053] Use the unstructured volume mesh divider (GVol) and unstructured surface mesh divider tool (Gsurf) in Griddle 2.0 mesh division software, generate the surrounding rock area mesh through setting the hard boundary and layer conversion operation, mesh merging and segmentation.

[0054] 7. Pipe jacking mesh generation;

[0055] Use the single track sweep to establish the total length of the pipe jacking multi-surface, generate the surface through the cross-sectional curve, combine the surface along the length direction of the pipe jacking, use the surface segmentation to divide the total length object, delete the array surface, use the BR command to generate the mesh, and import it into the simulation calculation software Flac 3D for grouping.

[0056] Please refer to Figure 4 , Figure 4The schematic diagram of the corresponding calculation model of FLAC3D is shown in FIG. 1. After importing FLAC3D, the soil layer is grouped, and the model is divided into 516420 units in total. The mechanical parameters of the lining, steel pipe and equivalent layer material are shown in Table 1.

[0057] Table 1

[0058]

[0059] After the calculation model is imported into FLAC3D, the model boundary is divided into upper and lower, left and right, front and back planes by automatic surface division. The normal displacement of the nodes located in the left and right planes (vertical to the x-axis) and the front and back planes (vertical to the y-axis) is constrained, the displacement of the nodes in the bottom plane in the x, y and z directions is constrained, and the top plane is a free surface. In the calculation process, the elastic, elastoplastic and plastic hardening models are used. The elastic model is a linear elastic model, the elastoplastic model is a Mohr-Coulomb (M-C) constitutive model, and the plastic hardening model is a HSS model considering small strain modulus.

[0060] Please refer to Figure 5 as shown, Figure 5 The flow chart of the calculation and parameter inversion corresponding to the calculation model is shown in FIG. 2. Specifically, first, the measured data of the descending tunnel is obtained: the geometric model is imported, the boundary conditions are set, the constitutive parameters are determined, the ground stress is calculated, the constitutive model is modified, the orthogonal experiment is designed, the constitutive parameters are determined, and the settlement of the measurement point is calculated; then the parameter inversion is performed: the calculation data is obtained, the training sample is established, the neural network is constructed, the network model is trained, the measured data is substituted, the constitutive parameters are predicted, the constitutive parameters are modified, and the settlement of the measurement point is calculated.

[0061] Orthogonal design is an effective mathematical statistics method, which uses orthogonal array to arrange and analyze multi-factor experiment. In the embodiment of the present application, a set of small and representative soil mechanical parameter combinations are constructed by using orthogonal design, which are uniformly distributed in the parameter space and used as the input of training and test samples. Therefore, the number of forward analysis and the calculation cost involved in the inverse analysis are greatly reduced. In the orthogonal design method, each inversion parameter is divided into different values in its value range, and the number of segmentation values represents the level of each inversion parameter. When the number of inversion parameters is the same, the more the levels of inversion parameters, the more the number of runs in the orthogonal design method. For example, if there are 3 inversion parameters and 2 levels are set for each inversion parameter, the number of runs based on the orthogonal design method is 4. The orthogonal experiment must meet the following two requirements: ① Different numbers in each column show the same number of times, such as numbers 1 and 2 appear twice in each column. ② The number of times of occurrence of each combination is equal in any combination of 2 columns in the level component. Therefore, if there are 3 inversion parameters and 3 levels are set for each inversion parameter, the number of runs based on the orthogonal design method is 9. In addition, the number of runs cannot be less than the square of the level. Orthogonal design usually uses L α (b) c is expressed, which means that a represents the number of experiments, b represents the number of inversion parameters, and c represents the level of each inversion parameter. Therefore, L 25 (5) 5 The meaning is described as follows: 25 groups of experiments are designed by the orthogonal design method, the number of inversion parameters is 5, and the level of each parameter is 5.

[0062] According to the sensitivity of different constitutive model parameters, the HSS orthogonal experiment parameters for small strain stiffness G0, small strain stiffness G0 decay to 70% corresponding to shear strain γ 0.7 , failure ratio R f , cohesive force c, and internal friction angle φ are designed as 5-parameter 4-level orthogonal experiment, that is, L 16 (5) 4 , a total of 16 experiments are performed. The HSS constitutive model orthogonal experiment parameters are shown in Table 2.

[0063] Table 2

[0064]

[0065] The M-C constitutive model uses elastic modulus E, Poisson's ratio υ, cohesive force c, and internal friction angle φ to design a 4-parameter 3-level orthogonal experiment, that is, L9(4) 3, and a total of 9 experiments were conducted; and the HSS model has more parameters, therefore, according to the research of Gu Xiaojing et al.

[99] on the correlation between the small-strain hardening constitutive model parameters of Shanghai soft clay and the soil layer pore ratio e in the geological exploration data, the research results are used to calculate the HSS constitutive model parameters, and the orthogonal test parameter levels are determined according to the upper and lower limits of the statistical parameters. The orthogonal test table of the M-C constitutive parameters is shown in Table 3:

[0066] Table 3

[0067]

[0068] Step S200, constructing a BPNN model, taking the vertical displacement data of the tunnel measuring point as input and the corresponding constitutive model parameters as output, training the BPNN model; inputting the field measured tunnel settlement data into the trained BPNN model to obtain the optimal constitutive parameters of the M-C and HSS constitutive models.

[0069] The back propagation neural network (BPNN) has the ability to capture and map potential features hidden in a large amount of irregular data. Therefore, in the case of unknown analytical functions, it can be used as a prediction tool to determine the nonlinear function relationship between input and output data.

[0070] The numerical calculation measuring point displacement values required for training the BPNN neural network based on the HSS constitutive model are shown in Table 4:

[0071] Table 4

[0072]

[0073] The numerical calculation process of the measuring point displacement values sets a monitoring point directly above each lining segment of the downlink tunnel to record the tunnel settlement deformation during the pipe jacking process. According to the 16 parameter combinations designed by the orthogonal test, the vertical displacement of the measuring point is calculated when the DN3200 is jacked for 50m and the DN2800 machine head is completed into the hole. Similarly, according to the results of the orthogonal test design, 9 groups of numerical calculations are performed using the M-C constitutive model to obtain the corresponding vertical displacement of the measuring point as the input data of the trained neural network, and the output data is the different constitutive model parameters, as shown in Table 3.

[0074] Please refer to Figure 6 and Figure 7 , Figure 6 The schematic diagram of the 3-layer BPNN topology relationship adopted for the HSS constitutive parameter inversion in the embodiments of the present application is shown in Figure 7 The schematic diagram of the 3-layer BPNN topology relationship adopted for the M-C constitutive parameter inversion is shown.

[0075] The HSS model has 45 neuron nodes in the input layer, 8 neuron nodes in the hidden layer, and 5 neuron nodes in the output layer; the M-C model has 45 neuron nodes in the input layer, 7 neuron nodes in the hidden layer, and 4 neuron nodes in the output layer. The number of neurons in the input layer is equal to the number of numerical characteristic parameters, and correspondingly, the number of neurons in the output layer is the same as the number of inversion parameters.

[0076] Although the mapping relationship between the input and output data is difficult to describe with an analytical function, it can be easily expressed using a BPNN. The BPNN needs to be trained using training samples before running. For the composition of the training samples, first, the input data is designed by the orthogonal design method. Using the orthogonal design method can greatly reduce the number of samples and make the individuals in the samples evenly distributed in the entire search interval, with uniform comparability. According to the orthogonal design table, the M-C and HSS models only need to run 25 groups of finite difference calculations. Then, the corresponding output data can be obtained by numerical modeling. The input data required for training the BP neural network of the M-C model is shown in Table 5:

[0077] Table 5

[0078]

[0079] During the running of the BPNN model, the input and output data are usually normalized in the range of -1 to 1. The number of hidden layers and the number of neurons in each layer depend on the complexity of the problem to be solved. Generally, only one hidden layer is needed. There is no general rule for accurately setting the number of neurons in the hidden layer, and the method of repeated experiments according to the empirical formula is mainly used. The empirical formula generally uses the square root of the sum of the number of neurons in the input layer and the number of neurons in the output layer, as shown in the following formula:

[0080] ; wherein, is the number of neurons in the input layer; is the number of neurons in the output layer; and N is the number of neurons in the hidden layer.

[0081] Please refer to Figure 8 , Figure 8 is a schematic diagram of the workflow of using the BPNN neural network to invert the material parameters of the soil body.

[0082] Please refer to Figure 9 and Figure 10 , Figure 9 is a schematic diagram of the neural network parameter fitting results of the HSS model, Figure 10 is the neural network parameter fitting results of the M-C model. From Figure 9It can be seen that the correlation coefficients of the training set, the validation set, the test set and the overall set of the HSS model BPNN network are about 0.99, 0.99, 0.98 and 0.98 respectively, and the fitting result is better. From the fitting results of the training set, the validation set and the test set of the HSS model BPNN network, it can be seen that the fitting results are good. Figure 10 It can be seen that the correlation coefficients of the training set, the validation set, the test set and the overall set are all about 0.99, and the fitting result is better. After the model training is completed, the measured data can be used for parameter inversion.

[0083] Please refer to Tables 6 and 7. For the measured data of the vertical displacement of the downlink tunnel, the measured data set is used as the input data, and the trained model is used to predict the constitutive parameters of the HSS and M-C models as the optimal constitutive parameters.

[0084] Table 6

[0085]

[0086] Table 7

[0087]

[0088] The inversion results of the HSS are shown in Table 8. The inversion result of the small strain stiffness G0 in the HSS model is about 50.76, the shear strain γ is about 5.0, the failure ratio Rf is about 0.64, the cohesion c is about 0.32 kPa, and the friction angle φ is about 35.6°. The inversion parameters are within the value range of the orthogonal test parameters. 0.7

[0089] Table 8

[0090]

[0091] Similarly, the inversion parameters of the M-C constitutive model are shown in Table 9.

[0092] Table 9

[0093]

[0094] Step S300, the optimal constitutive parameters obtained by inversion are used for forward numerical simulation of the whole process of pipe jacking construction, and the evolution law of the vertical displacement of the tunnel is obtained.

[0095] Based on the HSS and M-C constitutive models, the BPNN inversion parameters are used for numerical simulation calculation of the pipe jacking process. Please refer to Figure 11 , Figure 11 ​The flow chart of the simulation calculation of pipe jacking construction; the simulation calculation flow of pipe jacking includes: establishing a geometric model, giving boundary conditions, calculating ground stress, emptying speed displacement, updating the constitutive model, servo excavation, machine head jacking parameter assignment, pipe joint jacking parameter assignment, equivalent layer parameter assignment, applying mud pressure, DN3200 jacking 50m, when DN3200 jacking 50m is not met, servo excavation, machine head jacking parameter assignment, etc. are performed again, until it is detected again whether DN3200 jacking 50m is met, when DN3200 jacking 50m is met, bidirectional jacking simulation, equivalent layer parameter assignment are performed, and finally the calculation is completed. In the construction process, DN3200 is first jacked at the starting point, DN2800 is started after jacking 50m, and then the two pipes are jacked simultaneously. In the field construction process, an average of 12m is jacked per day, the pipe joint length is 6m, and the machine head length is 5m. In the excavation process, servo control is used, and the units in the excavation range are gradually simulated by reducing the stiffness, stress and density to ensure that the virtual inertia effect in the static solution process will not cause unrealistic damage to the surrounding area of the excavation. When the reduction factor of the given area is zero, the mechanical constitutive model of the unit in the area is set to empty, and the initial and boundary conditions are deleted. The soil is given HSS or M-C constitutive model, the reinforcement zone, pipe joint, machine head, equivalent layer and subway tunnel lining are given elastic constitutive model, wherein the pipe joint and machine head are Q355B type steel material, the reinforcement zone and tunnel lining are concrete C50, and the equivalent layer is mud in the excavation process, the constitutive parameters are shown in Table 1, when the pipe jacking is completed, the equivalent layer constitutive parameters are updated to concrete C50. The resistance reducing mud grouting pressure is controlled at 0.1MPa during the pipe jacking process, and the normal grouting pressure is applied to the outside of the pipe joint while the pipe joint is excavated. Since the pipe jacking passes through the subway section with a shallow depth, the influence of underground water is not considered in the simulation process, and it is assumed that the jacking pressure and soil pressure balance on the excavation face during the jacking process, and no jacking pressure is applied to the excavation face.

[0096] The vertical displacement change law of the downlink and uplink subway tunnel measuring points during the DN3200 exit process is similar, both first subsidence, then rebound, and finally subsidence and tend to be stable. Due to the different depths of the subway tunnels, the vertical displacement values of different measuring points are different, the vertical displacement of the measuring point with a deeper depth on the downlink is larger, and with the increase of the depth, the vertical displacement of the measuring point gradually decreases, and after the final deformation stabilizes, the maximum vertical displacement is about 12.4mm, and the minimum is about 7.3mm; the uplink is close to the position of the pipe jacking, and the disturbance during the jacking process is large, and after the final deformation stabilizes, the maximum vertical displacement is about 16.9mm, and the minimum is about 12.3mm. Since the DN3200 pipe jacking is located on the right side of the model, corresponding to the larger depth of the subway tunnel, the vertical displacement at the larger depth is more significant, while the tunnel at the shallower depth is on the left side of the model, and the DN2800 is not excavated, so the disturbance is smaller.

[0097] According to the vertical displacement change law of the DN3200 top-in 50m subway tunnel measuring point, the vertical displacement change law of the DN2800 downlink and uplink subway tunnel measuring point during the hole-out process. It can be found that the vertical displacement value of the measuring point is not significantly affected during the DN3200 top-in 50m and the DN2800 hole-out process. The maximum vertical settlement of the downlink tunnel measuring point after deformation stabilization is about 12.4mm, and the minimum value is about 7.3mm. The maximum vertical displacement of the uplink tunnel measuring point after deformation stabilization is about 16.9mm, and the minimum value is about 12.3mm.

[0098] According to the vertical displacement change law of the DN3200 up-through uplink subway tunnel measuring point, it can be found that the vertical displacement value of the measuring point has a slight fluctuation. The maximum vertical settlement of the downlink tunnel measuring point after deformation stabilization is about 11.3mm, and the minimum value is about 6.83mm. The maximum vertical displacement of the uplink tunnel measuring point after deformation stabilization is about 17.4mm, and the minimum value is about 12.8mm.

[0099] According to the vertical displacement change law of the DN3200 up-through downlink subway tunnel measuring point, it can be found that the maximum vertical settlement of the downlink tunnel measuring point after deformation stabilization is about 11.3mm, and the minimum value is about 6.83mm. The up-through construction disturbance is small. The maximum vertical displacement of the uplink tunnel measuring point after deformation stabilization is about 17.5mm, and the minimum value is about 12.8mm.

[0100] According to the vertical displacement change law of the DN2800 up-through uplink subway tunnel measuring point, it can be found that the maximum vertical settlement of the downlink tunnel measuring point after deformation stabilization is about 11.3mm, and the minimum value is about 6.83mm. The up-through construction disturbance is small. The maximum vertical displacement of the uplink tunnel measuring point after deformation stabilization is about 17.4mm, and the minimum value is about 12.7mm.

[0101] According to the vertical displacement change law of the DN2800 up-through downlink subway tunnel measuring point, it can be found that the maximum vertical settlement of the downlink tunnel measuring point after deformation stabilization is about 11.3mm, and the minimum value is about 6.82mm. The up-through construction disturbance is small. The maximum vertical displacement of the uplink tunnel measuring point after deformation stabilization is about 17.4mm, and the minimum value is about 12.7mm. Overall, the influence of the DN2800 up-through on the existing subway tunnel is small, and the vertical displacement of the uplink and downlink tunnel measuring points is almost unchanged, which also shows that the top-in construction process adopted in the embodiment has good effect in controlling the deformation of the existing structure.

[0102] According to the vertical displacement variation law of the subway tunnel measuring points after the double launching pipe penetrates, it can be found that the maximum vertical settlement of the downlink tunnel measuring points is about 11.3 mm and the minimum is about 6.81 mm after the deformation stabilizes; the maximum vertical displacement of the uplink tunnel measuring points is about 17.4 mm and the minimum is about 12.7 mm after the deformation stabilizes. After the pipe jacking construction is completed, the vertical deformation of the existing tunnel measuring points is relatively stable, the uplink process has no obvious disturbance, and the jacking construction technology used has good effect in controlling deformation.

[0103] According to the vertical displacement variation law of the subway tunnel measuring points when DN3200 emerges from the hole, the vertical displacement variation law of the downlink and uplink subway tunnel measuring points during the DN3200 emergence process is similar, which is first settlement, then rebound and finally tends to be stable. Due to the different depths of the subway tunnels, the vertical displacement values of different measuring points are different. The vertical displacement of the measuring points with shallow depth in the downlink is larger, and with the increase of the depth, the vertical displacement of the measuring points gradually decreases. The maximum vertical displacement is about 17.3 mm and the minimum is about 15.7 mm after the final deformation stabilizes. The uplink line is close to the position of the pipe jacking, and the disturbance during the jacking process is larger. The maximum vertical displacement is about 28.0 mm and the minimum is about 26.2 mm after the final deformation stabilizes. Compared with the calculation results of the HSS constitutive model, the settlement of the uplink and downlink tunnel monitoring points calculated by the M-C constitutive model is obviously larger, and when DN3200 emerges from the hole, the vertical displacement difference between different measuring points in the downlink is obvious, and the vertical displacement difference between the measuring points in the uplink is not large.

[0104] According to the vertical displacement variation law of the uplink subway tunnel measuring points during the DN3200 uplink, it can be found that the vertical displacement variation law of the downlink and uplink subway tunnel measuring points during the DN2800 emergence process. It can be found that the vertical displacement values of the measuring points during the DN3200 jacking of 50 m and the DN2800 emergence process are significantly disturbed. The downlink tunnel measuring points have large deformation fluctuations, and the maximum vertical settlement of the measuring points is about 35.7 mm and the minimum is about 26.5 mm when the calculation is completed. The maximum vertical displacement of the uplink tunnel measuring points is about 37.5 mm and the minimum is about 35.2 mm after the deformation stabilizes. The uplink tunnel is close to the jacking position, and the disturbance is obviously larger.

[0105] According to the vertical displacement variation law of the uplink subway tunnel measuring points during the DN3200 uplink, it can be found that due to the influence of excavation unloading, the vertical displacement values of the measuring points rebound, the maximum vertical settlement of the downlink tunnel measuring points is about 31.6 mm and the minimum is about 22.9 mm after the deformation stabilizes; the maximum vertical displacement of the uplink tunnel measuring points is about 32.5 mm and the minimum is about 30.8 mm after the deformation stabilizes. It can be seen that the M-C model is more sensitive to excavation unloading response. Although the grouting pressure is applied outside the pipe joint, the tunnel uplink process still produces obvious disturbance.

[0106] According to the vertical displacement change law of the DN3200 upper through lower metro tunnel measuring point, it can be found that the maximum vertical settlement of the lower tunnel measuring point is about 31.7 mm after deformation stabilization, and the minimum value is about 21.6 mm; the maximum vertical displacement of the upper tunnel measuring point is about 32.7 mm after deformation stabilization, and the minimum value is about 30.7 mm. Compared with the upper through upper tunnel, the disturbance of the pipe jacking through the lower tunnel is significantly reduced, and the vertical displacement of the measuring point fluctuates less.

[0107] According to the vertical displacement change law of the DN2800 upper through upper metro tunnel measuring point, it can be found that the maximum vertical settlement of the lower tunnel measuring point is about 29.3 mm after deformation stabilization, and the minimum value is about 20.2 mm; the maximum vertical displacement of the upper tunnel measuring point is about 31.1 mm after deformation stabilization, and the minimum value is about 28.6 mm. When DN2800 passes through the upper metro tunnel, the vertical displacement of the measuring point rebounds due to excavation unloading. The difference of the vertical displacement of the lower tunnel measuring point along the tunnel axis direction is larger, and the maximum vertical displacement difference is about 9.1 mm. The vertical displacement difference of the upper tunnel measuring point along the axis direction is smaller, but the overall deformation is larger.

[0108] According to the vertical displacement change law of the DN2800 upper through lower metro tunnel measuring point, it can be found that the maximum vertical settlement of the lower tunnel measuring point is about 29.0 mm after deformation stabilization, and the minimum value is about 19.9 mm; the maximum vertical displacement of the upper tunnel measuring point is about 31.1 mm after deformation stabilization, and the minimum value is about 28.7 mm. Overall, the influence of DN2800 on the existing metro tunnel is small, and the fluctuation range of the vertical displacement of the upper and lower tunnel measuring points is not large, indicating that after DN2800 passes through the upper tunnel, the stress and displacement of the rock-soil body are readjusted to a stable state, and when it passes through the lower tunnel, the unloading effect is no longer significant.

[0109] According to the vertical displacement change law of the double-launch pipe jacking through the metro tunnel measuring point, it can be found that the maximum vertical settlement of the lower tunnel measuring point is about 28.9 mm after deformation stabilization, and the minimum value is about 19.6 mm; the maximum vertical displacement of the upper tunnel measuring point is about 30.8 mm after deformation stabilization, and the minimum value is about 28.5 mm. After the pipe jacking construction is completed, the vertical deformation of the existing tunnel measuring point is relatively stable, and there is no obvious disturbance after the pipe jacking is completed. However, the vertical displacement of the upper and lower tunnel measuring points calculated by the M-C constitutive model responds obviously to the pipe jacking construction process, and the deformation gradually stabilizes after the upper through upper tunnel is completed, but the settlement after deformation stabilization is significantly increased compared with the HSS model calculation result, of which the settlement of the lower tunnel measuring point is 3 times that of the HSS model, and the upper tunnel is nearly 2 times that of the HSS model.

[0110] Step S400, based on the long short-term memory neural network, taking the historical monitoring data as the training set, inputting the field measured tunnel settlement data, and obtaining the prediction result of the future settlement of the tunnel.

[0111] For the long short-term memory neural network (LSTM) model, the specific construction process of the model is as follows:

[0112] (1) Prepare the obtained downlink tunnel measurement point settlement monitoring value data, and use the downlink tunnel measurement point settlement monitoring value as the neural network training data set. It should be noted that part of the data of the numerical simulation forward calculation result is used as the LSTM training data, which is equivalent to expanding the training data in the training data set;

[0113] (2) The data in the neural network training data set is divided into measured period data and prediction period data, wherein the measured period data is the data in the training set, and the prediction period data is the data in the validation set. For the LSTM model, if the sequence distribution of the input data (measured period data) is not standard or the change amplitude is too large, the learning and convergence speed of the network will be slowed down, therefore, the measured period data needs to be standardized. Normalization processing is usually used, that is, the data is mapped to the range of 0~1, and the formula is as follows:

[0114]

[0115] Wherein, is the normalized measured period data; is the current measured period data; is the minimum measured period data in the measured period data sequence; is the maximum measured period data in the measured period data sequence.

[0116] Then the input data (measured period data) format of the LSTM is constructed through the sliding window;

[0117] (3) The Adam optimization method is used to optimize the model super parameter. The optimization algorithm is a kind of algorithm that can realize the minimization or maximization of the objective function. It should be noted that the objective function is also called the loss function. The objective function is often a mathematical combination of model parameters and data. For example, given the data and its corresponding label, the model constructed in the embodiment of the application is a linear model. With the model, the predicted output can be obtained according to the input, and the difference between the predicted value and the true value can be calculated, which is the loss function. The purpose of the embodiment of the application is to find the appropriate value of the above loss function to reach the minimum. The smaller the loss value is, the closer the model is to the true situation. The Adam method is proposed based on the above idea. For each parameter, it not only has its own learning rate, but also has its own Momentum quantity. In this way, the update of each parameter is more independent in the training process, which improves the model training speed and the stability of the training. Generally speaking, for some deep neural networks or very complex neural networks, the use of Adam or other adaptive (Adaptive) methods can also converge faster. The related formula of Adam is as follows:

[0118]

[0119]

[0120]

[0121]

[0122]

[0123] where generally is set to 0.9, is set to 0.999; is the first moment estimate for the iteration step t; is the second moment of the iteration step t; is the parameter of the model at the iteration step t; is the corrected first moment estimate, and ΔJ(θ) is the gradient of the loss function J with respect to the parameter θ; is the second moment estimate for the iteration step t; is the second moment estimate for the iteration step t-1; is the corrected second moment estimate, is an arbitrarily small positive number, is the square root of the corrected second moment estimate, used to adaptively scale the learning rate.

[0124] (4) The measured period data is input into the trained LSTM model for prediction, and the prediction result is compared with the prediction period data to determine the prediction accuracy. The LSTM model uses a sliding time window method for multi-step prediction. First, j values are predicted, then the j predicted values are used as the last j period input data for the second step prediction, and so on, until the multi-step prediction result is obtained. Finally, the prediction result is normalized and compared with the prediction period data to evaluate the prediction accuracy. Please refer to Figure 12 , Figure 12 The schematic diagram of LSTM network construction and prediction process.

[0125] The LSTM deep learning algorithm is used to predict the settlement curve of the monitored subway shield segment XJC2200, the LSTM deep learning algorithm is used to predict the settlement curve of the monitored subway shield segment XJC2190, the LSTM deep learning algorithm is used to predict the settlement curve of the monitored subway shield segment XJC2180, the LSTM deep learning algorithm is used to predict the settlement curve of the monitored subway shield segment XJC2170, the LSTM deep learning algorithm is used to predict the settlement curve of the monitored subway shield segment XJC2160, the LSTM deep learning algorithm is used to predict the settlement curve of the monitored subway shield segment XJC2150, the LSTM deep learning algorithm is used to predict the settlement curve of the monitored subway shield segment XJC2140, the LSTM deep learning algorithm is used to predict the settlement curve of the monitored subway shield segment XJC2130, the LSTM deep learning algorithm is used to predict the settlement curve of the monitored subway shield segment XJC2120, the LSTM deep learning algorithm is used to predict the settlement curve of the monitored subway shield segment XJC2110, and the LSTM deep learning algorithm is used to predict the settlement curve of the monitored subway shield segment XJC2100. The monitoring data is obtained by arranging a measuring point every 10 rings from the 2200th ring to the 2100th ring of the descending tunnel. The settlement monitoring of the measuring point in the descending tunnel started on August 14 and ended on October 12. A total of 50 monitoring data were obtained for each group. When constructing the neural network, the data set was divided into a training set and a validation set, of which 10% of the monitoring data was used to verify the prediction result, i.e. the last five settlement points of each group of data were predicted and verified by monitoring data. The overall trend of the prediction data of the eleven measuring points is close to the measured data, although there is a deviation in the numerical value, but the error is within the acceptable range.

[0126] The division of the training period of the measuring point is shown in Table 10:

[0127] Table 10

[0128]

[0129] The vertical displacement neural network prediction value of the measuring point from October 6 to October 12 is shown in Table 11:

[0130] Table 11

[0131]

[0132] The vertical displacement values of each measuring point fluctuate up and down to different degrees, the fluctuation trend of the prediction value is close to the measured value, it is explained that the network model can learn the general rule of the monitoring data, the error range of the prediction of the pipe jacking construction process of the existing subway tunnel can be controlled in the millimeter level by selecting a suitable deep learning model, within the error range of the deformation prediction, and the implementation process is convenient and fast, without too much manual intervention. Based on the suitable prediction model, after the displacement of each measuring point is predicted, the overall trend of the tunnel measuring point settlement can be obtained, under the premise of ensuring the prediction accuracy, the neural network can replace the numerical calculation method in some prediction problems, and the disadvantages of time-consuming and complex numerical calculation and the influence of the grid and the constitutive parameter sensitivity on the calculation result are overcome. However, the prediction accuracy of the neural network depends on the amount of measured data to a great extent, in the embodiment of the present application, the measured data is very limited, the monitoring investment can be further increased, a large amount of monitoring data can be obtained, and then the prediction accuracy of the network model can be improved.

[0133] In summary, the present application aims at the problem that the FLAC3D preprocessing cannot realize fine modeling of complex geological bodies, the rhino three-dimensional modeling software is used, the geometric model is meshed through Griddle2.0, the calculation model of the pipe jacking is established based on the mixed structured and unstructured grid according to the geological data of the pipe jacking site, the simulation of the pipe jacking process is realized through program control, the servo mode is adopted, the soil excavation unloading process is simulated by means of the attenuation of the soil stiffness, stress and density with the calculation time step, and the unrealistic damage caused by one-time instantaneous excavation is avoided.

[0134] Based on the orthogonal experiment design and principle, two constitutive models, namely a Mohr-Coulomb (M-C) constitutive model and a Hardening Soil-Small (HSS) constitutive model, are adopted, according to the sensitivity of different constitutive model parameters, an orthogonal test scheme is designed, 9 groups of calculation schemes of the M-C model and 16 groups of calculation schemes of the HSS model are designed, the constitutive parameters and the vertical displacement of the downlink tunnel measuring point of 25 groups of different simulation schemes are trained through the BPNN neural network, the constitutive parameters are inversed according to the measured data of the downlink tunnel, and finally the M-C and HSS constitutive parameters are obtained.

[0135] The evolution law of the vertical displacement of the uplink and downlink tunnel measuring points is obtained by inverting the constitutive parameters for forward numerical calculation. The calculation results show that the influence of different pipe jacking processes on the vertical displacement of the existing tunnel measuring points is different, the influence of excavation disturbance on the vertical displacement of the tunnel measuring points can be well controlled by using the HSS model, the vertical displacement of the measuring points fluctuates greatly only when the pipe jacking exits the tunnel, while the deformation is stable during the jacking process, the maximum vertical settlement of the downlink tunnel measuring points is about 11.3 mm, and the minimum value is about 6.81 mm; the maximum vertical displacement of the uplink tunnel measuring points is about 17.4 mm, and the minimum value is about 12.7 mm. The vertical displacement values of the uplink and downlink tunnel measuring points calculated by the M-C model are larger, and the deformation fluctuation is larger, after the DN3200 jacks 50 m, the deformation gradually stabilizes, the maximum vertical settlement of the downlink tunnel measuring points is about 35.7 mm, and the minimum value is about 26.5 mm; the maximum vertical settlement of the uplink tunnel measuring points is about 37.5 mm, and the minimum value is about 35.2 mm. The maximum and minimum values of the downlink tunnel measuring point settlement calculated by the M-C are about 3 times the results of the HSS calculation, and the results of the uplink tunnel calculated by the M-C are about 2 times the results of the HSS calculation.

[0136] The long short-term memory (LSTM) deep learning algorithm is used, the downlink tunnel settlement monitoring value is used as the neural network training data set, and the monitoring data is divided into a training set and a validation set in proportion, and the LSTM network model is trained and verified for the prediction of the settlement deformation of the measuring point. The results show that the change law of the predicted value is very close to the measured value, the error range of the model for predicting the deformation of the existing subway tunnel during pipe jacking construction can be controlled within millimeters, within the allowable error range, and the implementation process is convenient and fast, without too much manual intervention, which can replace time-consuming and complex numerical calculation under certain conditions, and is sensitive to grid and constitutive parameters. However, the prediction accuracy of the neural network depends largely on the amount of measured data, and the measured data in the embodiment of the application is very limited, and a large amount of monitoring data can be obtained by further increasing the monitoring investment, and the prediction accuracy of the network model can be improved.

[0137] It should be noted that the above-mentioned embodiment of the application is only for description, and does not represent the advantages and disadvantages of the embodiment. The process depicted in the drawing does not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0138] Each embodiment in the specification is described in a progressive manner, and the same or similar parts of each embodiment can be referred to each other, and each embodiment mainly describes the difference from other embodiments.

Claims

1. A method of BPNN parameter inversion and LSTM tunnel deformation prediction, characterized in that, The method comprises the following steps: Based on the orthogonal test design principle, a plurality of groups of parameter combinations are designed for the M-C constitutive model and the HSS constitutive model respectively, numerical calculation is performed, and tunnel measuring point vertical displacement data are obtained; A BPNN model is constructed, the tunnel measuring point vertical displacement data are taken as inputs, the corresponding constitutive model parameters are taken as outputs, the BPNN model is trained, and the BPNN model is input with the field measured tunnel settlement data to obtain the optimal constitutive parameters of the M-C and HSS constitutive models through inversion; The optimal constitutive parameters obtained through inversion are used for forward numerical simulation of the whole process of pipe jacking construction, and tunnel vertical displacement evolution rules are obtained; wherein, the data of the forward calculation results of the numerical simulation are taken as the data for training of the long short-term memory neural network; Based on the long short-term memory neural network, the historical monitoring data and the numerical simulation forward calculation results are taken as the training set, the field measured tunnel settlement data are input into the downlink tunnel measuring point vertical displacement monitoring value, and the prediction result of the future settlement of the tunnel is obtained.

2. The method of claim 1, wherein, A three-dimensional calculation model comprising a pipe jacking, a subway tunnel and a stratum structure is established, and a structured and unstructured mixed grid is divided.

3. The method of claim 2, wherein, The establishment of the three-dimensional calculation model comprises: Rhino software is used for geometric modeling, a Griddle plug-in is used for grid division, and numerical calculation is performed in FLAC3D; the model boundary conditions are set as: left and right and front and back plane normal displacement constraints, bottom three-direction displacement constraints, and top surface free.

4. The method of claim 1, wherein, The long short-term memory neural network comprises: normalizing the monitoring data, and using a sliding window to construct an input-output sequence.

5. The method of claim 1, wherein, The long short-term memory neural network comprises: using an Adam optimization method to optimize the model hyperparameters.

6. The method of claim 1, wherein, The long short-term memory neural network based on historical monitoring data as a training set comprises: The historical monitoring data in the neural network training data set are divided into measured period data and prediction period data, and the measured period data are standardized.

Images