A method for predicting horizontal displacement of foundation pit based on TSNE-BP neural network

Abstract

A method for predicting horizontal displacement of a foundation pit based on a TSNE-BP neural network comprises: step 1: establishing a finite element design model for an actual project, determining inversion parameters and designing an orthogonal test to obtain training sample data and actual monitoring data; step 2: reading the training sample data and the actual monitoring data, constructing a TSNE-BP neural network in combination with a TSNE algorithm, determining relevant hyperparameters, reducing the dimension of high-dimensional input data and training the network; step 3: using the trained optimization model to perform soil parameter inversion on a test sample, taking a group of data with the lowest inversion standard error as a standard under the condition of a loss function Loss≤0.001 in the result to obtain inversion parameters, inputting the inversion parameters into Midas to calculate and obtain a predicted value of horizontal displacement of the foundation pit, and calculating the predicted standard error and the predicted displacement of subsequent working conditions; the invention has strong engineering significance in avoiding the occurrence of major engineering accidents and effectively shortening the construction period.

Description

Technical Field

[0001] The invention relates to a method for inverting foundation pit soil parameters and predicting horizontal displacement of deep foundation pit soil, and in particular to a method for inverting foundation pit soil parameters and predicting horizontal displacement of foundation pit using a TSNE algorithm which is firstly applied in the field of geotechnical engineering. Technical Background

[0002] With the continuous development of social economy and the continuous development of urban underground space, the design and construction of foundation pit engineering has become more and more complicated. In the design of foundation pit, many factors such as construction conditions, economic indicators, and surrounding environment need to be comprehensively considered. In addition, the excavation of foundation pit will also lead to problems such as uplift of the pit bottom, deformation of supporting and retaining structures, and water and soil seepage. In recent years, foundation pit engineering accidents have occurred frequently. The dense distribution of buildings, roads, and underground pipelines in urban centers with high demand for space has made the environmental conditions around deep foundation pits more severe, bringing unprecedented challenges to the balance between safety and economy for geotechnical engineering technicians. Therefore, how to scientifically design the construction process of the entire project, use on-site monitoring data, predict the deformation of the foundation pit engineering support structure, ensure the safety of the project during the entire construction process, and even achieve the purpose of shortening the construction period as much as possible under the premise of safety and saving costs has become a hot topic in current research.

[0003] The research on the prediction of horizontal displacement of foundation pits is mainly divided into: empirical method, numerical simulation method and machine learning method. The empirical method is based on a large amount of monitoring data to establish a discrete random model, which cannot systematically describe and explain the law of the change process; the numerical simulation method is relatively accurate, but due to the many influencing factors of foundation pit engineering, such as inaccurate soil parameters, the calculation results will have large errors. The machine learning method generally uses the traditional BP neural network, which has a relatively simple structure, has poor generalization ability and is prone to overfitting when processing high-dimensional data, resulting in unstable results and large errors. Summary of the invention

[0004] According to the deficiencies of the existing methods and technologies, the present invention proposes a method for predicting the horizontal displacement of a foundation pit based on a TSNE-BP neural network.

[0005] The TSNE-BP algorithm combines the numerical simulation method, which not only has the accuracy of numerical simulation in mathematical methods, but also improves the single structure of the BP neural network, the low generalization ability and easy overfitting problem in high-dimensional data processing. At the same time, the soil parameters and standard errors are inverted, and the inversion parameters are obtained based on the loss function and standard error as the value standard, which has higher prediction accuracy and stability.

[0006] The technical solution to implement the present invention is to provide a method for predicting horizontal displacement of a foundation pit based on a TSNE-BP neural network, comprising the following steps:

[0007] Step 1: Establish a finite element design model for the actual project, determine the inversion parameters and design an orthogonal experiment to obtain training sample data and actual monitoring data.

[0008] 1.1 According to the geological survey report and related information provided by the foundation pit project, Midas GTS NX was used to establish the relevant finite element model.

[0009] 1.2 Select monitoring points, determine the soil inversion parameters, conduct orthogonal test design, and put all combinations into Midas calculation to obtain the displacement of the corresponding monitoring points under each working condition and depth.

[0010] 1.3 Select the working conditions that require inversion, select N depths of calculated displacement data from shallow to deep as evenly as possible according to the foundation pit model, compare them with the actual monitoring data automatically collected by the sliding inclinometer, calculate the standard error, and organize them into a table.

[0011] Step 2: Read the training sample data and actual monitoring data, build a TSNE-BP neural network based on the TSNE algorithm, determine the relevant hyperparameters, reduce the dimensionality of the high-dimensional input data, and train the network.

[0012] 2.1 Use the xlrd module to read the monitoring table completed in step (1), and use the NumPy and PyTorch libraries to store the data as a tensor structure to obtain the required data set.

[0013] 2.2 Combined with the TSNE algorithm, the TSNE-BP neural network is established, and the network structure is determined as input layer-hidden layer 1-hidden layer 2-output layer. The input layer of the neural network is The standard error λ, where i represents the i-th working condition, i.e., N+1-dimensional input; a linear layer is set, and an activation layer is added to the hidden layer, with the modified linear unit ReLU as the activation function; a BN layer, i.e., a batch normalization layer, is added to speed up the convergence speed; the L1 norm loss function is used to calculate the error between the predicted value and the true value; the output layer outputs the secant stiffness E of each layer of soil 50 and the standard error of the inversion;

[0014] 2.3 Randomly set 3 groups of samples as test samples in the neural network, and the rest as training samples. Use the TSNE algorithm to reduce the dimensionality of high-dimensional input data and map it in three-dimensional space. The high-dimensional training data is converted into mapping points in multiple three-dimensional spaces. The measured data is similar to the measured data, and its clustering quality is observed. The measured mapping points in three-dimensional space and the 5 training sample mapping points closest to them are converted into the final measured mapping points with weights w, 1-w. The proportion of the 5 mapping points is distributed according to their distance from the measured mapping points.

[0015] The training samples are learned through the deep learning optimization algorithm, and the relevant hyperparameters are set. The hyperparameters include: the number of training rounds, the number of hidden layer nodes, the learning rate, and the weight w. By adjusting the hyperparameters, the optimized model is obtained through training.

[0016] For the adjustment of hyperparameters, the optimization algorithm Adam is used to update the gradient and hyperparameters. Its iterative formula is:

[0017] μ=β1μ+(1-β1)dθ (1)

[0018] s=β2s+(1-β2)dθ 2 (2)

[0019]

[0020] θ — hyperparameter to be trained;

[0021] η — learning rate;

[0022] dθ — gradient;

[0023] β1——first-order moment attenuation coefficient;

[0024] β2——second-order moment attenuation coefficient;

[0025] μ——exponentially weighted average of original gradients;

[0026] s - exponentially weighted average of the squared gradient;

[0027] ——Gradient normalization formula;

[0028] Step 3: Use the trained optimization model to invert the soil parameters of the test samples. In the result, take the set of data with the lowest inversion standard error as the standard when the loss function Loss≤0.001 to obtain the inversion parameters, and input the inversion parameters into Midas for calculation to obtain the predicted value of the horizontal displacement of the foundation pit, and calculate the predicted standard error and the predicted displacement of the subsequent working conditions.

[0029] The step 1.3 of the present invention mentions that the horizontal displacement monitoring data of the deep soil of the foundation pit is automatically collected by a sliding inclinometer, which is associated with the mobile phone terminal and the PC terminal to save the data in real time; the monitoring frequency is once / day, and twice / day when the deformation is abnormal.

[0030] The forward propagation formulas of the TSNE-BP neural network algorithm mentioned in step 2.2 of the present invention are:

[0031] Set the linear layer function: Linear(x) = Ax + B (4)

[0032] Batch normalization layer function: BN(x) = γ + β (5)

[0033] Activation layer function: ReLU(x) = max(0, x) (6)

[0034] The input is passed from the linear layer to the BN layer and then to the activation layer, that is, x = self.act1(self.bn1(self.hidden1(x))). (7)

[0035] Then it is passed to the second linear layer, batch normalization layer and activation layer, that is, x = self.act2(self.bn2(self.hidden2(x))). (8)

[0036] Where: A——weight matrix;

[0037] B——bias vector;

[0038] γ – learning stretching parameter;

[0039] β——offset parameter;

[0040] ——The mean of the input data x;

[0041] σ——variance of input data x;

[0042] Finally, the output layer outputs the data, that is, out = self.out(x). (9)

[0043] The method for inversion of foundation pit soil parameters and prediction of horizontal displacement based on TSNE-BP neural network mentioned in step 2.3 of the present invention has a preferred parameter scheme as follows:

[0044] β1=0.9, β2=0.999, ∈=1e -8 ,w=0.8.

[0045] The present invention uses the L1 norm loss function and the standard error as the value standard to obtain the inverted soil parameters, and combines the above-obtained measured value y of the monitoring point deformation with the horizontal displacement prediction value y obtained in step 3 P , calculate its prediction standard error as an evaluation index, and compare the calculated standard error with the inverted standard error to evaluate the prediction accuracy of the neural network. The standard error calculation formula mentioned in step 3 of the present invention is:

[0046]

[0047] The present invention determines the design model and inversion parameters according to the geological survey report and drawings through Midas GTS NX, and further obtains the training set samples by orthogonal test design for the inversion parameters, and obtains the test set samples through the intelligent inclinometer. TSNE-BP is used to learn the training samples, and the hyperparameters are adjusted to obtain the optimized model to perform parameter inversion on the test set samples. The inverted parameters are predicted to deform through numerical simulation to obtain the predicted value of the horizontal displacement of the foundation pit. The soil parameter inversion range is output, and the deformation prediction value is compared with the measured data to calculate the standard error. The results show that the TSNE-BP neural network algorithm proposed in the present invention is more accurate and stable than the traditional BP neural network output result, with higher prediction accuracy and better generalization ability. TSNE, as an algorithm used for the first time in the field of geotechnical engineering, has a better dimensionality reduction effect on the data set than the traditional SNE algorithm, and realizes the effect of visual dimensionality reduction of high-dimensional data, which is suitable for the inversion of foundation pit soil parameters and the prediction of horizontal displacement. It has strong engineering significance in avoiding the occurrence of major engineering accidents and effectively shortening the construction period.

[0048] Compared with the prior art, the present invention has the following advantages:

[0049] ① Aiming at the nonlinear characteristics of foundation pit engineering, the present invention adopts the TSNE-BP neural network algorithm combined with the Midas method for numerical simulation, which overcomes the problem that the soil parameters of foundation pit engineering are inaccurate and change with construction, resulting in low prediction accuracy of traditional prediction methods.

[0050] ② The TSNE-BP neural network algorithm proposed in the present invention can reduce the dimensionality of high-dimensional input data, thus overcoming the problem of overfitting in traditional BP neural network processing of high-dimensional data.

[0051] ③ The TSNE-BP neural network algorithm proposed in the present invention utilizes nonlinear dimensionality reduction, has the advantage of realizing high-dimensional data visualization, and is convenient for observing the clustering quality of high-dimensional input data in three-dimensional space.

[0052] ④ The TSNE-BP neural network algorithm proposed in the present invention makes the traditional SNE dimensionality reduction algorithm conform to the t distribution during dimensionality reduction, overcoming the dimensionality curse, that is, the problem that high-dimensional data appears very crowded in low-dimensional space, and making the distribution of data in low-dimensional space more consistent with high-dimensional space.

[0053] ⑤ The TSNE-BP neural network algorithm proposed in the present invention has better generalization ability.

[0054] ⑥ The TSNE-BP neural network algorithm proposed in the present invention shows better stability and higher prediction accuracy than the traditional BP neural network and numerical simulation method. It is an efficient optimization back-analysis method for foundation pit soil parameter inversion and horizontal displacement prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 This is a plan layout diagram of foundation pit monitoring points used for foundation pit soil parameter inversion in Example 1 of the present invention;

[0056] Figure 2 This is a cross-sectional view of a foundation pit standard section support design for foundation pit soil parameter inversion according to Example 1 of the present invention;

[0057] Figure 3 This is a flow chart of the technical route for inversion of foundation pit soil parameters provided in Example 1 of the present invention. DETAILED DESCRIPTION

[0058] The present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0059] Example 1

[0060] The present invention is applied in an engineering example of a foundation pit project in Hangzhou.

[0061] See attached Figure 1 , is the plan layout of the foundation pit monitoring points used for the foundation pit soil parameter inversion and horizontal displacement prediction research in this embodiment; the terrain of the project site is relatively small, and the absolute elevation of the entire site is between 5.13 and 5.79 meters. The perimeter of the foundation pit is about 210 meters, the foundation pit area is about 3,600 square meters, the designed foundation pit excavation depth is 10.2 meters, the surrounding structure adopts bored cast-in-place piles, the burial depth is 30.2 meters, the support system adopts two reinforced concrete supports, and the passive area around the foundation pit is reinforced by Three-axis cement mixing piles. The monitoring points CX02 and CX05 with large displacement on the east and south sides of the foundation pit were selected for analysis, and the secant stiffness E 50 As the inversion parameter, a four-parameter five-level orthogonal test design was designed. The parameter combination after the orthogonal test design was input into the finite element model, and the horizontal displacement of each working condition of CX02 and CX05 was calculated. Three working conditions were selected for analysis: the first layer of support construction was completed, the second layer of support construction was completed, and the excavation was carried out to the bottom of the foundation pit. The displacements of 12 depths were extracted from top to bottom, and the standard errors of each combination were calculated based on the corresponding displacements of the measured data. The displacements and standard errors of 12 depths under 25 combinations of three working conditions were used as the original sample data of the input layer for training.

[0062] See attached Figure 2, which is a cross-sectional view of the support design of a standard section of a foundation pit used for the inversion of foundation pit soil parameters and the prediction of horizontal displacement according to an embodiment of the present invention; the soil layers involved in the site from top to bottom are ① miscellaneous fill, ② silty clay intercalated with clayey silt, ③ silty silty clay, and ④ silty clay. The geomorphic unit of the site belongs to the Hangzhou-Jiaxing-Huzhou alluvial plain, with a soft foundation and a groundwater level buried at a depth of 1.21-3.00m. Among them, the silty silty clay has high compressibility, low shear strength, and rheological properties, and is the main soft soil layer that affects engineering construction.

[0063] See attached Figure 3 , which is a technical flow chart of foundation pit soil parameter inversion and horizontal displacement prediction provided by the embodiment of the present invention; Midas GTS NX is used to establish a foundation pit design model based on the geological survey report and related design drawings. Determine the inversion parameter E 50 , and an orthogonal test design is designed to obtain 25 soil layer parameter combinations and substitute them into Midas calculation to obtain training samples. The TSNE-BP neural network is used to learn the training samples, and hyperparameters such as the number of training rounds (EPOCH), learning rate (LR), number of hidden layer nodes (HIDDEN SIZE), and weight w of three-dimensional measured mapping points are set. The trained optimization model is used to invert the measured data samples to obtain the inversion parameters and standard errors of each soil layer under the measured data. The inversion parameters are substituted into Midas calculation to obtain the predicted displacement under the inversion parameters and calculate the predicted standard error under the predicted displacement, which is compared and analyzed with the design standard error and the inversion standard error, and finally the subsequent working conditions are predicted.

[0064] The specific implementation steps are as follows:

[0065] Step 1: Establish a finite element design model for the actual project, determine the inversion parameters and design an orthogonal experiment to obtain training sample data and actual monitoring data.

[0066] 1.1 According to the geological survey report and related drawings provided by the foundation pit project, Midas GTS NX was used to establish the relevant finite element model.

[0067] 1.2 Select monitoring points, determine the soil inversion parameters, conduct orthogonal test design, and put all combinations into Midas calculation to obtain the displacement of the corresponding monitoring points under each working condition and depth.

[0068] 1.3 Select the working conditions of the required inversion parameters: one-layer support construction, two-layer support construction and excavation to the bottom of the foundation pit. According to the foundation pit model, the calculated displacement data and actual monitoring data of 12 depths are selected from shallow to deep as evenly as possible, and the standard error is calculated and sorted into a table. The actual monitoring data is automatically collected by the sliding inclinometer, and the data is linked to the mobile phone and PC to save the data in real time. The monitoring frequency is once a day, and twice a day when the deformation is abnormal, which constitutes the test set sample of the network.

[0069] Step 2: Read the training sample data and actual monitoring data, build a TSNE-BP neural network based on the TSNE algorithm, determine the relevant hyperparameters, reduce the dimensionality of the high-dimensional input data, and train the network.

[0070] 2.1 Use the xlrd module to read the monitoring table completed in step 1, and use the NumPy and PyTorch libraries to store the data as a tensor structure to obtain the required data set.

[0071] 2.2 The neural network input layer is Standard error λ, where i represents the i-th working condition, that is, N+1-dimensional input, in this embodiment, N=12, i=3; the output layer outputs the secant stiffness E of each layer of soil 50 And the standard error of the inversion, this embodiment is a 4-layer soil.

[0072] Use the following formula to normalize the data set to a distribution with a mean of 0 and a variance of 1.

[0073]

[0074] X——normalized value;

[0075] x——parameter to be normalized;

[0076] μ – the mean value of the data set;

[0077] S – standard deviation of the data set;

[0078] 2.3 The TSNE algorithm is used to reduce the dimensionality of the training set samples, and the mapping points reduced to 3 dimensions are displayed in a 3D graph to observe the clustering quality of the measured mapping points and the training sample mapping points.

[0079] The TSNE-BP neural network is used to learn the training set samples, and the hyperparameters such as the number of training rounds, learning rate, number of hidden layer nodes, and measured mapping point weight w are set. The forward propagation formula is:

[0080] Set the linear layer function: Linear(x) = Ax + B (4)

[0081] Batch normalization layer function: BN(x) = γ + β (5)

[0082] Activation layer function: ReLU(x) = max(0, x) (6)

[0083] The input is passed from the linear layer to the BN layer and then to the activation layer, that is, x = self.act1(self.bn1(self.hidden1(x))). (7)

[0084] Then it is passed to the second linear layer, batch normalization layer and activation layer, that is, x = self.act2(self.bn2(self.hidden2(x))). (8)

[0085] Where: A——weight matrix;

[0086] B——bias vector;

[0087] γ – learning stretching parameter;

[0088] β——offset parameter;

[0089] ——The mean of the input data x;

[0090] σ——variance of input data x;

[0091] Finally, the output layer outputs the data, that is, out = self.out(x). (9)

[0092] The TSNE-BP network structure is deep, and parameter updating is somewhat difficult, so this embodiment uses the adaptive moment estimation algorithm Adma for optimization. The Adam algorithm can dynamically modify the learning rate of each parameter, and introduces the momentum method to make the parameters have a greater probability of jumping out of the local optimal solution while updating. Its iterative formula is as follows:

[0093] μ=β1μ+(1-β1)dθ (1)

[0094] s=β2s+(1-β2)dθ 2 (2)

[0095]

[0096] θ — hyperparameter to be trained;

[0097] η — learning rate;

[0098] dθ — gradient;

[0099] β1——first-order moment attenuation coefficient;

[0100] β2——second-order moment attenuation coefficient;

[0101] μ——exponentially weighted average of original gradients;

[0102] s - exponentially weighted average of the squared gradient;

[0103] ——Gradient normalization formula;

[0104] In this embodiment, β1 = 0.9, β2 = 0.999, ∈ = 1e-8 , w=0.8.

[0105] Three groups of data are randomly set as test samples for the data set to test whether the output of the neural network is reasonable.

[0106] Step 3: Use the trained optimization model to perform soil parameter inversion on the test samples to obtain the inversion parameters of each layer of soil.

[0107] Since there is a certain error between the output inversion value and the true value, the neural network evaluates the error through the L1 norm loss function, and uses the inversion value of the standard error to take the set of data with the lowest inversion standard error as the standard when the loss function Loss≤0.001 to obtain the inversion parameters. The inversion parameters of the corresponding working conditions are input into Midas for calculation to obtain the predicted value of the horizontal displacement of the foundation pit under the corresponding working conditions. The predicted standard error is calculated to evaluate the prediction effect of TSNE-BP, and the rationality of the inversion parameter value standard is explained.

[0108] The deformation prediction values ​​y of 12 depths P Compare it with the measured deformation value y and calculate its standard error. The formula is as follows:

[0109]

[0110] In order to further verify the optimization effect of the TSNE-BP neural network, it is compared with the traditional BP neural network. The soil parameters of each working condition are inverted using the traditional neural network, and the above steps are repeated to obtain its prediction standard error. The inversion and prediction effects of the two methods are compared.

[0111] Taking CX02 as an example, the soil parameter inversion results of TSNE-BP neural network and BP neural network are shown in Table 1.

[0112] Table 1. Inversion parameter ranges

[0113] Table 1.Table of inversion parameter range

[0114] Working conditions <![CDATA[E1(MPa)]]> <![CDATA[E2(MPa)]]> <![CDATA[E3(MPa)]]> <![CDATA[E4(MPa)]]> Working condition 1-T 2.801-3.447 4.656-5.330 1.764-2.331 6.907-7.561 Condition 1-BP 2.369-4.037 4.325-6.767 1.597-2.408 6.651-8.992 Working condition 2-T 2.756-3.609 5.031-5.886 1.753-2.339 6.570-7.229 Working condition 2-BP 2.227-3.889 4.456-6.986 1.437-2.310 6.291-8.865 Working condition three - T 2.787-3.531 4.339-5.026 1.690-2.259 6.205-7.142 Condition 3-BP 2.627-4.005 3.969-6.213 1.508-2.360 5.968-9.491

[0115] The results in Table 1 show that the integration of the TSNE algorithm effectively reduces the value range of the inversion parameters of each soil layer, especially for the silty clay layer, whose secant stiffness is relatively large. Simply using the BP neural network for parameter inversion is prone to fall into the local optimal situation and ignore the accuracy of the inversion parameters of the first few soil layers. Similarly, data processing of CX05 also confirms the above optimization effect.

[0116] The inversion parameters are obtained through TSNE-BP neural network and BP neural network, and the standard error results of horizontal displacement are shown in Table 2.

[0117] Table 2 Error analysis table

[0118] Table 2.Table of error analysis

[0119]

[0120] The results in Table 2 show that the TSNE-BP inversion model has a better fitting effect with the measured results, and the overall trend is close to the actual results. The standard error is reduced by 60% to 86% compared with the design model, and the actual inversion error is reduced by 18% to 50% compared with the single BP neural network. Especially when the deformation value is small, the actual inversion error of TSNE-BP is less than 0.5mm. As the construction progresses, the actual inversion error increases. Although the value is different from the predicted inversion error, it generally conforms to the law of predicted inversion error. It is determined that it can be used as a feature of the data set and play a certain auxiliary role in the selection of inversion data.

[0121] Substituting reasonable inversion parameters into the deformation of the next working condition, and comparing it with the actual deformation, the inversion parameter prediction error results of the TSNE-BP algorithm are shown in Table 3.

[0122] Table 3 Error analysis table

[0123] Table3.Table of error analysis

[0124]

[0125] The results in Table 3 show that the prediction error of CX05 is smaller than that of CX02 and the prediction errors of the two measuring points are slightly larger than the actual inversion errors, which are numerically smaller overall. In general, the prediction accuracy of the TSNE-BP algorithm is higher in the problem of predicting horizontal displacement within the entire depth range of the measuring point, which is a novel and effective method for the inversion of foundation pit soil parameters and the prediction of horizontal displacement.

Claims

1. A method for predicting horizontal displacement of foundation pit based on TSNE-BP neural network, characterized in that The following steps are involved: Step 1: Establish a finite element design model for the actual project, determine the inversion parameters and design an orthogonal experiment to obtain training sample data and actual monitoring data; 1.1 Use Midas GTS NX to establish relevant finite element models based on the geological survey report and related information provided by the foundation pit project; 1.2 Select monitoring points, determine soil inversion parameters, conduct orthogonal test design, and put all combinations into Midas calculation to obtain the displacement of the corresponding monitoring points under various working conditions and depths; 1.3 Select the working conditions required for inversion, select N calculated displacement data from shallow to deep as evenly as possible according to the foundation pit model, compare them with the actual monitoring data automatically collected by the sliding inclinometer, calculate the standard error, and organize them into a table; Step 2: Read the training sample data and actual monitoring data, build a TSNE-BP neural network based on the TSNE algorithm, determine the relevant hyperparameters, reduce the dimensionality of the high-dimensional input data and train the network; 2.1 Use the xlrd module to read the monitoring table completed in step (1), and use the NumPy and PyTorch libraries to store the data as a tensor structure to obtain the required data set; 2.2 Combined with the TSNE algorithm, a TSNE-BP neural network is established, and the network structure is determined as input layer-hidden layer 1-hidden layer 2-output layer; the input layer of the neural network is Standard error λ, where i represents the i-th condition, i.e., N+1-dimensional input; A linear layer is set, and an activation layer is added to the hidden layer, with the modified linear unit ReLU as the activation function; a BN layer, i.e., a batch normalization layer, is added to speed up the convergence speed; the L1 norm loss function is used to calculate the error between the predicted value and the true value; the output layer outputs the secant stiffness E of each layer of soil 50 And inversion standard error; the forward propagation formulas of the TSNE-BP neural network are: Set the linear layer function: Linear(x) = Ax + B (4) Batch Normalization layer function: Activation layer function: ReLU(x) = max(0, x) (6) The input is passed from the linear layer to the BN layer and then to the activation layer, that is, x = self.act1(self.bn1(self.hidden1(x))); (7) Then it is passed to the second linear layer, batch normalization layer and activation layer, that is, x = self.act2(self.bn2(self.hidden2(x))); (8) Where: A——weight matrix; B——bias vector; γ – learning stretching parameter; β——offset parameter; ——The mean of the input data x; σ——variance of input data x; Finally, the output layer outputs the data, i.e., out = self.out(x); (9) 2.3 Randomly set 3 groups of samples as test samples in the neural network, and the rest as training samples; use the TSNE algorithm to reduce the dimensionality of the high-dimensional input data and map it in the three-dimensional space. The high-dimensional training data is converted into mapping points in multiple three-dimensional spaces. The measured data is the same, and its clustering quality is observed; the measured mapping points in the three-dimensional space and the 5 training sample mapping points closest to them are converted into the final measured mapping points with weights w, 1-w, and the proportion of the 5 mapping points is distributed according to their distance from the measured mapping points; The training samples are learned through the deep learning optimization algorithm, and the relevant hyperparameters are set. The hyperparameters include: the number of training rounds, the number of hidden layer nodes, the learning rate, and the weight w. By adjusting the hyperparameters, the TSNE-BP neural network model is trained. For the adjustment of hyperparameters, the optimization algorithm Adam is used to update the gradient and hyperparameters. Its iterative formula is: μ=β1μ+(1-β1)dθ (1) s=β2s+(1-β2)dθ 2 (2) θ — hyperparameter to be trained; η — learning rate; dθ — gradient; β1——first-order moment attenuation coefficient; β2——second-order moment attenuation coefficient; μ——exponentially weighted average of original gradients; s – exponentially weighted average of the squared gradient; ——Gradient normalization formula; Step 3: Use the trained optimization model to invert the soil parameters of the test samples. In the result, take the set of data with the lowest inversion standard error as the standard when the loss function Loss≤0.001 to obtain the inversion parameters, and input the inversion parameters into Midas for calculation to obtain the predicted value of the horizontal displacement of the foundation pit, and calculate the predicted standard error and the predicted displacement of the subsequent working conditions.

2. The method for predicting horizontal displacement of foundation pit based on TSNE-BP neural network according to claim 1 is characterized in that: In step 1.3, the horizontal displacement monitoring data of the deep soil in the foundation pit is automatically collected by the sliding inclinometer, and the data is linked to the mobile phone and PC to save the data in real time; the monitoring frequency is once a day, and twice a day when the deformation is abnormal.

3. The method for predicting horizontal displacement of foundation pit based on TSNE-BP neural network according to claim 1 is characterized in that: In step 2.3 β1=0.9, β2=0.999, ∈=1e -8 ,w=0.

8.

4. The method for predicting horizontal displacement of foundation pit based on TSNE-BP neural network according to claim 1 is characterized in that: In step 3, the inverted soil parameters are obtained by taking the L1 norm loss function and the standard error as the value selection criteria, and the measured deformation value y of the monitoring point obtained above is combined with the predicted horizontal displacement value y obtained in step 3 P , calculate its standard error as an evaluation index, and compare the calculated standard error with the inverse standard error to evaluate the prediction accuracy of the neural network. The standard error calculation formula is: Here n is the number of input dimensions excluding the standard error.

Images