Adjacent shield tunnel segment convergence deformation machine learning prediction method and device

Abstract

The invention discloses a machine learning prediction method and device for convergence deformation of adjacent shield tunnel segments, and the method achieves the accurate prediction of the convergence deformation of the adjacent tunnel segments through a long and short term memory neural network algorithm in machine learning in combination with shield tunneling parameters and convergence deformation monitoring data. Shield tunneling parameters and existing adjacent tunnel segment convergence deformation monitoring data in the tunneling period of a new tunnel are obtained, the data are preprocessed, and then a segment deformation prediction database is constructed; secondly, constructing a long-short-term memory neural network prediction model, establishing a time sequence correlation relationship between tunneling parameters of a new tunnel and convergence deformation of an existing tunnel segment, and performing model training and verification; thirdly, searching an optimal hyper-parameter by adopting a grid search optimization algorithm, and determining an optimal model for predicting convergence deformation of the existing tunnel segment; and 4, on the basis of the training model, carrying out prediction and result evaluation on convergence deformation of other ring segments.

Description

Technical Field

[0001] The present invention relates to machine learning and tunnel deformation prediction, and in particular to a machine learning prediction method and device for adjacent shield tunnel segment convergence deformation. Background Art

[0002] At present, my country's urban rail transit is still in a stage of rapid development. At the same time, with the continuous increase in the total mileage of subway operations, the operation and maintenance of the opened lines are also receiving increasing attention.

[0003] The excavation of a new tunnel will inevitably lead to the unloading of the surrounding strata, thereby destroying the balance between the existing tunnel segments and the surrounding soil. The additional stress generated during the shield tunneling process is transmitted through the soil and applied to the existing tunnel segments, inducing further deformation. The cross-section of the shield tunnel is a circular structure made of high-strength concrete and bolts, and its overall stiffness is relatively high. Therefore, any significant deformation of the shield tunnel structure can easily lead to cracks and damage, induce concrete spalling and segment water leakage, and thus affect the safe operation of subway trains during the operation period and the service life of the shield tunnel. The Shanghai Metro Line 4 connecting channel accident is a typical case of secondary disasters caused by adjacent construction activities.

[0004] In addition, when a new tunnel is excavated, the deformation of the existing tunnel segments will change with the relative position of the new shield during the excavation process. Traditional methods are difficult to accurately describe this time-varying effect, while deep learning algorithms have the advantage of being able to accurately handle nonlinear correlations between time series data. Summary of the invention

[0005] In order to overcome the above-mentioned shortcomings of the prior art, the present invention provides a machine learning prediction method and device for convergence deformation of adjacent shield tunnel segments.

[0006] The present invention establishes the time series correlation between the excavation parameters of a new shield tunnel and the convergence deformation of the existing tunnel segments. The prediction results are compared with the measured data of the convergence deformation of the existing shield tunnel segments obtained during the construction of a subway shield tunnel, verifying the prediction performance of the method.

[0007] The present invention uses an LSTM (Long Short-Term Memory Neural Network) prediction algorithm, which solves the problem by controlling the gating mechanism (forget gate, input gate and output gate) of the information accumulation rate and introducing a new internal state to transfer linear cyclic information to the external state of the hidden layer. It significantly improves the long-term dependency problem often encountered in traditional prediction models and realizes accurate prediction of the convergence deformation of adjacent shield tunnel segments.

[0008] The machine learning prediction method for convergence deformation of adjacent shield tunnel segments proposed in the present invention uses seven groups of tunneling parameters generated by shield tunneling, including cutter head speed (SH), cutter head torque (TH), advance speed (AT), penetration (PT), total thrust average (GT), relative position (RP) and face earth pressure average (EP), as input variables, and takes the convergence deformation of adjacent shield tunnel segments as output variables; during training, a certain ring of adjacent tunnel segment convergence deformation data is used as a training set, and another ring of segment convergence deformation data is used as a test set to verify the prediction accuracy of the model.

[0009] A first aspect of the present invention relates to a machine learning prediction method for convergence deformation of adjacent shield tunnel segments, comprising the following steps:

[0010] A. Construct a segment convergence and deformation prediction database; obtain 7 sets of shield tunneling parameters during the excavation of the new tunnel and the convergence and deformation monitoring data of the existing adjacent tunnel segments, and perform data preprocessing before the prediction model calculation to enhance the prediction ability of the model, that is, process the missing and noisy data in the excavation parameters to generate a data sequence suitable for model prediction, and then construct a segment convergence and deformation prediction database;

[0011] B. Construct a long short-term memory neural network prediction model, establish the time series correlation between 7 sets of shield tunneling parameters and the convergence deformation of existing tunnel segments, conduct model training, and use the LSTM model to train and verify the data set;

[0012] C. Grid search algorithm to optimize the neural network model; A grid search optimization algorithm is used to identify the optimal ratio of hyperparameters, thereby determining the best model for predicting the convergence deformation of existing tunnel segments;

[0013] D. Predict the convergence deformation of adjacent segments; based on the training model, predict the convergence deformation of the remaining segments and evaluate the results.

[0014] Among them, the 7 groups of shield tunneling parameters described in step A include: cutter head rotation speed, cutter head torque, propulsion speed, penetration, total thrust, average value and relative position of face soil pressure.

[0015] The specific implementation steps of the present invention are as follows.

[0016] The step A of constructing a segment convergence deformation prediction database includes:

[0017] A1. Install segment convergence and deformation monitoring equipment represented by laser displacement meters in nearby existing tunnels;

[0018] A2. During the excavation of a new shield tunnel, 7 sets of shield tunneling parameter data were collected;

[0019] A3. During tunnel excavation, the convergence and deformation of existing tunnel segments are continuously monitored to obtain segment convergence and deformation data;

[0020] A4. Process the missing and noisy data in the monitoring data to generate a data sequence suitable for model prediction, and then build a segment convergence deformation prediction database.

[0021] The step B of constructing a long short-term memory neural network prediction model includes:

[0022] B1. Establish a long short-term memory neural network prediction model, with the number of input layer variables being 7 sets of shield tunneling parameters; the number of output layers being 1, corresponding to the convergence deformation of adjacent tunnel segments monitored in A3;

[0023] B2. Establish the time-series correlation between the excavation parameters of the new tunnel and the convergence deformation of the existing adjacent tunnel segments;

[0024] B3. Train the neural network model, select the shield tunneling parameters in the database and the deformation monitoring data of a certain ring of adjacent tunnels as the training set, and after inputting 7 sets of shield tunneling parameters, output the deformation prediction data and optimize it in combination with the actual monitoring data.

[0025] The grid search algorithm described in step C optimizes the neural network model, including:

[0026] C1. Set the initial grid search algorithm parameters to form a hyperparameter combination grid;

[0027] C2. For the LSTM neural network training model in B3, perform model training and evaluation on different hyperparameter combinations in the hyperparameter combination grid to find the optimal hyperparameters;

[0028] C3. Assign the optimal hyperparameters to the neural network algorithm model to obtain the best model after grid search algorithm optimization.

[0029] The step D of predicting the convergence deformation of adjacent segments includes:

[0030] D1. Select the segments in the adjacent tunnel that have not participated in the model training, use their deformation data as output, and use the corresponding new tunnel excavation parameters as input;

[0031] D2. Input the excavation parameters in D1 into the optimized long short-term memory neural network prediction model to predict the convergence deformation of the adjacent tunnel segments;

[0032] D3. Use MSE, RMSE, R 2 The model prediction results are evaluated by other indicators.

[0033] The second aspect of the present invention relates to a machine learning prediction device for convergence and deformation of adjacent shield tunnel segments, comprising a memory and one or more processors, wherein the memory stores executable code, and when the one or more processors execute the executable code, they are used to implement a machine learning prediction method for convergence and deformation of adjacent shield tunnel segments of the present invention.

[0034] A third aspect of the present invention relates to a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements a machine learning prediction method for convergence deformation of adjacent shield tunnel segments of the present invention.

[0035] The advantages of the present invention are:

[0036] 1. The LSTM algorithm can accurately predict the convergence deformation of adjacent tunnels caused by the excavation of new tunnels;

[0037] 2. The LSTM model can predict long-term data through the gating mechanism (forget gate, input gate, and output gate) that controls the information accumulation rate, improving the long-term dependency problem often encountered by traditional models;

[0038] 3. The algorithm prediction does not require human intervention and can output the convergence deformation prediction results of adjacent tunnel segments in real time during shield tunneling;

[0039] 4. The prediction method can be applied to shield tunnel projects of different types and regions and has a certain degree of versatility. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 It is a schematic diagram of the LSTM model algorithm principle in the present invention.

[0041] Figure 2 It is a schematic diagram of the adjacent tunnel segment convergence deformation prediction model architecture of the present invention.

[0042] Figure 3 It is a general implementation flow chart of the present invention.

[0043] Figure 4a and Figure 4b is a prediction result diagram of the present invention, wherein Figure 4a This is a time series comparison chart of model prediction and measured results. Figure 4b It is a scatter plot comparison of model prediction and actual measured results.

[0044] Figure 5 It is a device diagram of the present invention. DETAILED DESCRIPTION

[0045] The specific implementation of the present invention is further described below in conjunction with the accompanying drawings.

[0046] Example 1

[0047] This embodiment relates to a machine learning prediction method for convergence deformation of adjacent shield tunnel segments, and the specific steps are as follows:

[0048] A. Construct a segment convergence deformation prediction database;

[0049] A1. Install segment convergence and deformation monitoring equipment represented by laser displacement meters in nearby existing tunnels;

[0050] A2. During the excavation of a new shield tunnel, seven sets of shield tunneling parameter data (cutterhead speed, cutterhead torque, propulsion speed, penetration, total thrust, average value and relative position of face earth pressure) were collected;

[0051] A3. During tunnel excavation, the convergence and deformation of existing tunnel segments are continuously monitored to obtain segment convergence and deformation data;

[0052] A4. Process the missing and noisy data in the monitoring data to generate a data sequence suitable for model prediction, and then build a segment convergence deformation prediction database, including Figure 2 Input data on the left and output verification data on the right;

[0053] B. Construct a neural network prediction model;

[0054] B1. Establish a long short-term memory neural network prediction model. The model architecture is as follows Figure 1 As shown in the figure, the number of input layer variables is 7 sets of shield tunneling parameters, such as Figure 2 As shown on the left, they correspond to the parameters in A2; the number of output layers is 1, such as Figure 2 As shown on the right, it corresponds to the convergence deformation of the adjacent tunnel segments monitored in A3;

[0055] B2. Establish the time series correlation between 7 sets of shield tunneling parameters and the convergence deformation of existing adjacent tunnel segments, namely Figure 2 The corresponding relationship between the data on both sides in the time series;

[0056] B3. Train the neural network model, select the shield tunneling parameters in the database and the deformation monitoring data of a certain ring of adjacent tunnels as the training set, and after inputting 7 sets of shield tunneling parameters, output the deformation prediction data, and optimize it in combination with the actual monitoring data;

[0057] C. Grid search algorithm optimizes the neural network model;

[0058] C1. Set the initial grid search algorithm parameters to form a hyperparameter combination grid;

[0059] C2. The architecture of the long short-term memory neural network training model in B3 is as follows Figure 1As shown, the model is trained and evaluated for different hyperparameter combinations in the hyperparameter combination grid to find the optimal hyperparameters;

[0060] C3. Assign the optimal hyperparameters to the neural network algorithm model to obtain the best model after grid search algorithm optimization. The algorithm model architecture is as follows: Figure 1 As shown in the figure, the overall framework has not changed, but the hyperparameter values ​​have changed. The setting of hyperparameter values ​​has a great impact on the model training results.

[0061] D. Predict the convergence deformation of adjacent segments;

[0062] D1. Select the segments in the adjacent tunnels that have not participated in the model training, use their deformation data as output, and use the corresponding new tunnel excavation parameters as input. The prediction process is as follows: Figure 2 ,The difference is that the LSTM model has been trained during the prediction, and only the tunneling parameters need to be input into the prediction;

[0063] D2. Input the excavation parameters in D1 into the optimized long short-term memory neural network prediction model to predict the convergence deformation of the adjacent tunnel segments;

[0064] D3. Use MSE, RMSE, R 2 The model prediction results are evaluated by other indicators.

[0065] A specific embodiment is as follows:

[0066] (1) Seven sets of new tunnel excavation parameters (cutterhead speed, cutterhead torque, propulsion speed, penetration, total thrust, average and relative position of face earth pressure) were collected and stored in Excel. At each time point, each set of excavation parameters had a corresponding value. During the excavation of the new tunnel, deformation monitoring was performed on two rings of segments in the adjacent tunnel to obtain the convergence deformation monitoring values ​​of the two rings of segments. MATLAB was used to check each data point and process the missing and abnormal data in the monitoring data. For example, when processing abnormal data, the average value of the ten data points before and after each data point was used as the standard for judging the data abnormal value. If the value of the data point exceeded the average value by more than ten times, it was considered as a sensor failure and deleted.

[0067] (2) A long short-term memory (LSTM) neural network model was established in Python. The LSTM model consisted of one LSTM layer and one fully connected layer. The mean absolute error (MAE) was used as the loss function. The Adam optimizer was selected as the optimizer, and the time step was set to 1, indicating a set of data points every 10 minutes. The tunneling parameters and the deformation monitoring data of one of the rings were imported as input and output data, respectively, and the time series correspondence of the data was established for training. The hyperparameters in the model training were initially set to: Learning_rate = 0.001, Momentum = 0.9, Bitch_size = 150;

[0068] (3) In terms of network structure, the LSTM model sets a forget gate to delete information features that are not needed for training in long time series, uses an input gate to adjust the amount of new information added to the long-term state, and uses an output gate to manage access to long-term state information. Through the above gating mechanism, it can overcome the gradient explosion and gradient vanishing problems in traditional models, and capture the temporal correlation between current and future information, significantly improving the accuracy of processing long-term time series data;

[0069] (4) Establish a grid search optimization algorithm to train and evaluate models for different hyperparameter combinations in the hyperparameter combination grid to find the optimal hyperparameters and improve model accuracy;

[0070] (5) After the optimized neural network model is established, the convergence deformation data of the other segment that is not involved in the model training is used as the output, and the tunnel excavation parameter data is input according to the data time series correspondence. The prediction results are compared and verified with the actual values ​​of deformation monitoring;

[0071] (6) Figure 4a and Figure 4b The results of the prediction model are compared with the actual results. The LSTM model produces an MSE of 0.010, RMSE of 0.101, MAE of 0.079, and R 2 It is 0.966, and the predicted value is very close to the measured value. It can be seen that this model has good prediction performance and can accurately predict the convergence deformation of adjacent tunnel segments.

[0072] In summary, the long short-term memory neural network prediction model established according to this embodiment is perfect and can achieve accurate prediction of the convergence deformation of adjacent shield tunnel segments.

[0073] Example 2

[0074] Reference Figure 5This embodiment relates to a machine learning prediction device for convergence and deformation of adjacent shield tunnel segments, including a memory and one or more processors, wherein the memory stores executable code, and when the one or more processors execute the executable code, they are used to implement a machine learning prediction method for convergence and deformation of adjacent shield tunnel segments of Example 1.

[0075] Example 3

[0076] This embodiment relates to a computer-readable storage medium having a program stored thereon. When the program is executed by a processor, a machine learning prediction method for convergence deformation of adjacent shield tunnel segments of Embodiment 1 is implemented.

[0077] The contents described in this specification are merely an enumeration of the implementation forms of the inventive concept. The protection scope of the present invention is not limited to the specific forms described in the implementation cases, but also includes equivalent technical means that can be associated with the technical personnel in this field based on the inventive concept.

Claims

1. A machine learning prediction method for convergence deformation of adjacent shield tunnel segments, comprising the following steps: A. Construct a segment convergence and deformation prediction database; obtain 7 sets of shield tunneling parameters during the excavation of the new tunnel and the convergence and deformation monitoring data of the existing adjacent tunnel segments, and perform data preprocessing before the prediction model calculation to enhance the prediction ability of the model, that is, process the missing and noisy data in the excavation parameters to generate a data sequence suitable for model prediction, and then construct a segment convergence and deformation prediction database; B. Construct a long short-term memory neural network prediction model; establish the time series correlation between 7 sets of shield tunneling parameters and the convergence deformation of existing tunnel segments, conduct model training, and use the LSTM model to train and verify the data set; C. Grid search algorithm to optimize the neural network model; A grid search optimization algorithm is used to identify the optimal ratio of hyperparameters, thereby determining the best model for predicting the convergence deformation of existing tunnel segments; D. Predict the convergence deformation of adjacent segments; based on the training model, predict the convergence deformation of the remaining segments and evaluate the results.

2. A machine learning prediction method for convergence deformation of adjacent shield tunnel segments as claimed in claim 1, characterized in that: The seven groups of shield tunneling parameters described in step A include: cutter head rotation speed, cutter head torque, propulsion speed, penetration, total thrust, average value and relative position of face soil pressure.

3. A machine learning prediction method for convergence deformation of adjacent shield tunnel segments as claimed in claim 2, characterized in that: The step A of constructing a segment convergence deformation prediction database includes: A1. Install segment convergence and deformation monitoring equipment represented by laser displacement meters in nearby existing tunnels; A2. During the excavation of a new shield tunnel, data of the seven sets of shield excavation parameters are collected; A3. During tunnel excavation, the convergence and deformation of existing tunnel segments are continuously monitored to obtain segment convergence and deformation data; A4. Process the missing and noisy data in the monitoring data to generate a data sequence suitable for model prediction, and then build a segment convergence deformation prediction database.

4. A machine learning prediction method for convergence deformation of adjacent shield tunnel segments as claimed in claim 3, characterized in that: The construction of the long short-term memory neural network prediction model in step B includes: B1. establishing a long short-term memory neural network prediction model, the number of input layer variables is the 7 groups of shield tunneling parameters; the number of output layers is 1, corresponding to the convergence deformation of the adjacent tunnel segments monitored in step A3; B2. Establish the time-series correlation between the excavation parameters of the new tunnel and the convergence deformation of the existing adjacent tunnel segments; B3. Train the neural network model, select the shield tunneling parameters in the database and the deformation monitoring data of a certain ring of adjacent tunnels as the training set, and after inputting the 7 sets of shield tunneling parameters, output the deformation prediction data and optimize it in combination with the actual monitoring data.

5. A machine learning prediction method for convergence deformation of adjacent shield tunnel segments as claimed in claim 4, characterized in that: The grid search algorithm described in step C optimizes the neural network model, including: C1. Set the initial grid search algorithm parameters to form a hyperparameter combination grid; C2. For the long short-term memory neural network training model in step B3, perform model training and evaluation on different hyperparameter combinations in the hyperparameter combination grid to find the optimal hyperparameters; C3. Assign the optimal hyperparameters to the neural network algorithm model to obtain the best model after grid search algorithm optimization.

6. A machine learning prediction method for convergence deformation of adjacent shield tunnel segments as claimed in claim 5, characterized in that: The prediction of the convergence deformation of adjacent segments in step D includes: D1. Select the segments in the adjacent tunnel that have not participated in the model training, use their deformation data as output, and use the corresponding new tunnel excavation parameters as input; D2. Input the excavation parameters in step D1 into the optimized long short-term memory neural network prediction model to predict the convergence deformation of adjacent tunnel segments; D3. Use MSE, RMSE, R 2 The model prediction results are evaluated by other indicators.

7. A machine learning prediction device for convergence deformation of adjacent shield tunnel segments, characterized in that: It includes a memory and one or more processors, wherein the memory stores executable code, and when the one or more processors execute the executable code, it is used to implement a machine learning prediction method for convergence deformation of adjacent shield tunnel segments according to any one of claims 1 to 6.

8. A computer-readable storage medium, characterized in that: A program is stored thereon, and when the program is executed by a processor, a machine learning prediction method for convergence deformation of adjacent shield tunnel segments as described in any one of claims 1-6 is implemented.

Images