CN113792372A - A Dynamic Prediction Method of Ground Connecting Wall Deformation Based on CV-LSTM Combined Model - Google Patents

Abstract

The invention discloses a dynamic prediction method for ground connection wall deformation based on a CV-LSTM combined model. Select the monitoring points, collect the deformation history monitoring data of the foundation pit project, and organize and form a monitoring table; use the CV-LSTM combined model to learn the training samples in the deformation monitoring data, and use the optimal model obtained by training on the test set samples. Perform deformation prediction to obtain the predicted value of the deformation of the ground connecting wall. The deformation prediction value obtained according to the technical solution of the present invention is compared with the deformation measured value, and the evaluation index is calculated. The LSTM deep network has better generalization ability and is suitable for the dynamic prediction problem of the deformation of the ground connection wall, which can provide a reference for the realization of information management on the construction site.

Description

A Dynamic Prediction Method of Ground Connecting Wall Deformation Based on CV-LSTM Combination Model

technical field

The invention relates to a method for predicting the deformation of a ground connection wall, in particular to a dynamic prediction method for the deformation of a ground connection wall based on a CV-LSTM combined model.

Background technique

At present, my country is in the stage of rapid development of rail transit construction, resulting in a large number of station foundation pit projects, and the excavation scale and depth of foundation pits are also increasing. For example, the maximum excavation depth of a station of a certain subway line in a certain city reaches 29 m. Foundation pit engineering is a space system including soil and supporting structure, which is affected by many internal and external factors such as geological conditions, construction quality and surrounding environment. The field measured deformation data is the embodiment of the comprehensive effect of various influences in the construction process. Therefore, analyzing and studying on-site deformation monitoring data has become an effective way for people to understand the deformation characteristics of foundation pits.

In the research on the prediction of the deformation law of the foundation pit wall under different soil properties and different support types, the prediction methods are mainly divided into the following three types: empirical method, numerical simulation method and machine learning. The empirical method requires a large amount of monitoring data, and is limited to the use of differential equations to establish discrete random models, which is inconvenient to describe the nature and internal laws of the system change process; the numerical method has the accuracy of the mathematical method, but because the foundation pit is connected to the wall The complexity of the influencing factors of deformation, the ambiguity of the physical mechanism, and the variability and uncertainty of the parameters make the method overgeneralized and reduce the practical value. At present, most of the machine learning methods use the classic BP neural network. The structure of this back-propagation neural network is relatively simple, but it cannot accurately handle the correlation of input time and the error is large when processing long sequence data; An LSTM algorithm for processing sequence data, the advantage is that it is convenient for sequence modeling and has the ability of long-term memory, but it also has the disadvantage of poor generalization ability.

SUMMARY OF THE INVENTION

Aiming at the deficiencies in the prior art, the present invention provides a method for predicting the deformation of a ground connecting wall based on a CV-LSTM combined model, the combined model has higher stability and is suitable for long-term and dynamic prediction of the ground connecting wall deformation, and It has higher prediction accuracy and better generalization ability.

The technical scheme for realizing the purpose of the present invention is to provide a dynamic prediction method for the deformation of the ground connecting wall based on the CV-LSTM neural network algorithm, which comprises the following steps:

step one:

，表示测点i在第t天的变形值，形成观测值的时间序列；将监测数据整理形成监测表；Select the monitoring points, collect the historical monitoring data of the deformation of the ground connecting wall of the foundation pit project, and record the deformation observations collected at each monitoring point as

, represents the deformation value of the measuring point i on the t day, forming the time series of the observation value; organize the monitoring data to form a monitoring table;

Step 2:

(1) Use the xlrd module in the PyTorch framework package to read the monitoring table formed in step (1), and use the tensor function to store the data as a tensor structure to obtain a data set;

，输出层为

，其中，N为输入信息长度，M为预测时间跨度；(2) The K-fold cross-validation method and the LSTM neural network algorithm are used to establish the prediction combination model. The input layer of the prediction combination model is

, the output layer is

, where N is the length of the input information and M is the prediction time span;

The prediction combination model adopts the K-fold cross-validation method to divide the data set into K subsets, in which K-1 subsets are taken as training set samples in turn for training, and the remaining 1 subset is used as a test set sample, for testing;

For the prediction combination model, the LSTM neural network algorithm is used to learn the training set samples, and hyperparameters are set, including the number of training rounds EPOCH, the learning rate LR, and the number of hidden layer neurons HIDDEN_SIZE. optimal model;

The method for adjusting hyperparameters is as follows: using the Adam algorithm to adjust the network parameter w in the iterative update formula, and the iterative update formula is:

；

;

为学习率；dw为梯度；

为一阶矩衰减系数；

为二阶矩衰减系数；v为原始梯度的指数加权平均值；s为梯度平方的指数加权平均值；

为梯度的归一化处理；Among them, w is the network parameter to be trained;

is the learning rate; dw is the gradient;

is the first-order moment attenuation coefficient;

is the second-order moment decay coefficient; v is the exponentially weighted average of the original gradient; s is the exponentially weighted average of the square of the gradient;

is the normalization of the gradient;

Step 3:

The optimal model obtained by training is used to predict the deformation of the test set samples, and the predicted value of the deformation of the ground connecting wall is obtained.

In the technical solution of the present invention, the forward propagation formulas of the LSTM neural network algorithm are respectively:

The first module is the "forget gate", which is used to calculate the forgetting proportion of the neuron state information at the previous moment:

The second module is the "input gate", which uses new information to write the proportion of neuron states:

The third module is the "output gate", which will determine the information to be output as a hidden state:

Among them, C ^(t) and h ^(t) represent the neuron state and hidden state at time t , respectively; f ^(t) , i ^(t) and o ^(t) represent the forgetting gate, input gate and output gate at time t , respectively ; W , b are the weight matrix and bias vector in each gating module respectively; s is the sigmoid activation function; tanh is the hyperbolic tangent activation function.

，

，

。A preferred solution of the method for dynamically predicting the deformation of the ground connection wall based on the CV-LSTM neural network algorithm of the present invention is:

,

,

.

The invention collects the deformation history monitoring data of the ground connecting wall of the foundation pit project, including the horizontal displacement of the connecting wall monitored by the movable inclinometer buried in the wall concrete; the monitoring frequency is once a day, and the excavation depth exceeds 20 m Then increase to 2 times/day, and 2 to 3 times/day when deformation is abnormal.

The present invention measures the real deformation value ŷ _i of each monitoring point of the ground connecting wall, obtains the predicted value y _i and the real value ŷ _i of the ground connecting wall deformation according to step 3, takes MSE, MAE and MAPE as loss functions, and calculates the accuracy evaluation index MSE respectively , MAE and MAPE to evaluate the prediction accuracy of the prediction combination model. The calculation formula is as follows:

，

,

，

,

。

.

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

1. Aiming at the characteristics of long period and nonlinearity of monitoring data, the present invention adopts the prediction model method of CV-LSTM combined model, which overcomes the shortcomings of traditional neural network over-fitting and gradient explosion.

2. The combined model provided by the present invention has a high precision for long-term prediction of the deformation of the ground connection wall due to its advantage of processing long-sequence data.

3. The combined model provided by the present invention exhibits better stability and higher prediction accuracy, and is more suitable for dynamic prediction of the deformation of the ground connecting wall.

4. The ground connection wall deformation prediction method provided by the present invention has better generalization ability.

Description of drawings

1 is a schematic diagram of the layout of the foundation pit plane and the inclination measuring points used for the prediction of the deformation of the ground connecting wall according to the embodiment of the present invention;

2 is a schematic cross-sectional design of a standard section of a foundation pit used for predicting the deformation of a ground connecting wall according to an embodiment of the present invention;

3 is a flowchart of a deformation prediction combination model provided by an embodiment of the present invention;

In the figure, ① miscellaneous fill, ③ silty clay, ④ silt with silt, ⑤ silt with silty clay, ⑥ silty clay, and ⑦ silt with silt.

Detailed ways

The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

Example 1

The present invention is applied in the engineering example of the deep foundation pit project of a subway station of a certain urban rail transit.

Referring to FIG. 1, it is a schematic diagram of the layout of the foundation pit plane and the inclination measuring points used for the prediction of the deformation of the ground connecting wall in this embodiment; the foundation pit of the subway station is in the east-west direction, and the north side of the foundation pit is only 1.7 meters away from the existing building. m. The enclosure structure of the foundation pit adopts a 1.0 m thick underground diaphragm wall. The standard section of the foundation pit adopts 6 internal supports. The total length of the sub-foundation pit B is about 103.0 m, the width of the standard section is 23.1 m, and the excavation depth is 24.16 m. From August 8, 2018 to July 9, 2019, a total of 2329 sets of ground-connecting wall inclination monitoring data from monitoring points CX10-CX19 of the main structure of foundation pit B were selected as the original samples for prediction training. Monitoring is carried out by an active inclinometer in the wall concrete.

Referring to FIG. 2, the schematic diagram of the cross-sectional design of the standard section of the foundation pit used for the prediction of the deformation of the ground connecting wall in the present embodiment; the site is sequentially distributed from top to bottom with ① miscellaneous fill, ③ silty clay, and ④ silt mixed with silt sand , ⑤ silt with silty clay, ⑥ silty clay and ⑦ silt with silt, and the buried depth of the stable water level is 1.20-1.90 m. The silty clay layer distributed in the site is mainly soft plastic, with high compressibility, high sensitivity, low shear strength, etc., and has significant rheology, which is the main soft soil layer affecting the construction of the project. The standard section of foundation pit B is provided with 6 inner supports vertically.

Referring to FIG. 3 , the flow chart of the deformation prediction combination model provided in this embodiment is obtained; a data set is acquired, missing data is filled, and a Cross-validation cross-validation method is used to divide the data set. The LSTM neural network is used to learn the training set samples, and hyperparameters such as the number of training rounds (EPOCH), the learning rate (LR), and the number of hidden layer neurons (HIDDEN_SIZE) are set. The trained optimal model is used to predict the deformation of the test set samples, and the short-term and long-term deformation prediction values of the ground connecting wall are obtained, and the accuracy evaluation index is calculated.

The specific implementation steps are as follows:

step one:

，表示测点i在第t天的变形值，形成变观测值的时间序列，将监测数据整理形成监测表。Select the monitoring points and collect the historical monitoring data of the deformation of the ground connecting wall of the foundation pit project. In principle, the monitoring frequency of the ground connecting wall inclination measurement is once a day. After the excavation depth exceeds 20 m, it will be increased to 2 times per day, and 2 times when the deformation is abnormal. ～3 times/day, the monitoring data is organized to form a monitoring daily report. Due to the occlusion of the monitoring points during the construction process, some missing data are caused by linear interpolation, and the deformation observations collected at each monitoring point are recorded as

, represents the deformation value of measuring point i on the t day, and forms a time series of variable observation values, and organizes the monitoring data into a monitoring table.

Step 2:

，预测组合模型的输出层为

，其中N输入信息长度，M表示预测时间跨度。Use the xlrd module in the PyTorch framework package to read the data in the monitoring daily report, and use the tensor function to store the monitoring data as a tensor structure. The input layer of the prediction combination model is

, the output layer of the prediction combination model is

, where N is the length of the input information and M is the prediction time span.

The cross-validation method is used to divide the data set, and the 10-fold cross-validation method is used for the CX10~CX19 inclination measurement points in the original sample of the inclination measurement deformation of the ground connecting wall. Each measurement point can be predicted once as a test set, so that It relatively objectively reflects the generalization of the prediction model.

The LSTM neural network is used to learn the training set samples, and hyperparameters such as the number of training rounds (EPOCH), learning rate (LR), and the number of hidden layer neurons (HIDDEN_SIZE) are set. The forward propagation formulas are:

The first module is the "forget gate", which is used to calculate the forgetting proportion of the neuron state information at the previous moment:

The second module is the "input gate", which uses new information to write the proportion of neuron states:

The third module is the "output gate", which will determine the information to be output as a hidden state:

Among them, C ^(t) and h ^(t) represent the neuron state and hidden state at time t , respectively; f ^(t) , i ^(t) and o ^(t) represent the forgetting gate, input gate and output gate at time t , respectively ; W , b are the weight matrix and bias vector in each gating module respectively; s is the sigmoid activation function; tanh is the hyperbolic tangent activation function.

The LSTM network has a deep structure and it is difficult to update parameters. Therefore, in this embodiment, an adaptive moment estimation algorithm (Adam, Adaptive Moment Estimation) is used for optimization. On the one hand, Adam dynamically modifies the learning rate of each parameter, and on the other hand, introduces the momentum method, so that the parameter update has more opportunities to jump out of the local optimal solution. Its iterative update formula is as follows:

；

;

为学习率；dw为梯度；

为一阶矩衰减系数；

为二阶矩衰减系数。v为原始梯度的指数加权平均值；s为梯度平方的指数加权平均值；

为梯度的归一化处理；本实施例中，

，

，

。Among them, w is the network parameter to be trained;

is the learning rate; dw is the gradient;

is the first-order moment attenuation coefficient;

is the second-order moment attenuation coefficient. v is the exponentially weighted average of the original gradients; s is the exponentially weighted average of the squared gradients;

is the normalization processing of the gradient; in this embodiment,

,

,

.

Step 3:

Use the trained optimal combination model to predict the deformation of the test set samples, and obtain the predicted value y _i of the deformation of the ground connecting wall.

Since there is a certain error between the output predicted value y _i and the real value ŷ _i , the neural network evaluates the error through the loss function. The smaller the three values of the commonly used loss functions MSE, MAE and MAPE, the combined model have better accuracy. Use the trained optimal combination model to predict the deformation of the test set samples, so as to obtain the short-term deformation prediction value and long-term deformation prediction value of the ground connecting wall. Calculate the evaluation indicators MSE, MAE and MAPE values of the combined model under the prediction task, and evaluate the prediction accuracy of the CV-LSTM combined model.

Compare the predicted deformation value and the measured deformation value y _i with the real value ŷ _i , and calculate the accuracy evaluation indexes MSE, MAE and MAPE respectively. The calculation formulas are as follows:

。

.

In order to further verify the effectiveness of the CV-LSTM combined model, the classical BP neural network was selected for comparison, and four prediction tasks were designed to reflect the prediction effect of the prediction model under different input information lengths and prediction step sizes.

Task 1: Predict the deformation amount after 1 day (N=3, M=1) according to the deformation monitoring value of the first 3 days;

Task 2: Predict the deformation amount after 7 days according to the deformation monitoring values of the first 3 days (N=3, M=7);

Task 3: Predict the deformation amount after 1 day according to the deformation monitoring value of the first 15 days (N=15, M=1);

Task 4: Predict the deformation amount after 7 days (N=15, M=7) based on the deformation monitoring values of the first 15 days.

The specific results of the prediction error MSE value (unit/mm ² ) of the 10-fold cross-validation are shown in Table 1.

Table 1

。

.

The results in Table 1 show that the CV-LSTM combined model outperforms the BP combined model in all prediction tasks, and obtains smaller error values on the test set.

Claims (5)

1. a kind of ground connection wall deformation dynamic prediction method based on CV-LSTM neural network algorithm is characterized in that comprising the following steps: step one: Select the monitoring points, collect the historical monitoring data of the deformation of the ground connecting wall of the foundation pit project, and record the deformation observations collected at each monitoring point as

, represents the deformation value of the measuring point i on the t day, forming the time series of the observation value; organize the monitoring data to form a monitoring table; Step 2: (1) Use the xlrd module in the PyTorch framework package to read the monitoring table formed in step (1), and use the tensor function to store the data as a tensor structure to obtain a data set; (2) The K-fold cross-validation method and the LSTM neural network algorithm are used to establish the prediction combination model. The input layer of the prediction combination model is

, the output layer is

, where N is the length of the input information and M is the prediction time span; The prediction combination model adopts the K-fold cross-validation method to divide the data set into K subsets, in which K-1 subsets are taken as training set samples in turn for training, and the remaining 1 subset is used as a test set sample, for testing; For the prediction combination model, the LSTM neural network algorithm is used to learn the training set samples, and hyperparameters are set, including the number of training rounds EPOCH, the learning rate LR, and the number of hidden layer neurons HIDDEN_SIZE. optimal model; The method for adjusting hyperparameters is as follows: using the Adam algorithm to adjust the network parameter w in the iterative update formula, and the iterative update formula is:

; Among them, w is the network parameter to be trained;

is the learning rate; dw is the gradient;

is the first-order moment attenuation coefficient;

is the second-order moment decay coefficient; v is the exponentially weighted average of the original gradient; s is the exponentially weighted average of the square of the gradient;

is the normalization of the gradient; Step 3: The optimal model obtained by training is used to predict the deformation of the test set samples, and the predicted value of the deformation of the ground connecting wall is obtained. 2. a kind of ground connection wall deformation dynamic prediction method based on CV-LSTM neural network algorithm according to claim 1, is characterized in that: the forward propagation formula of LSTM neural network algorithm is respectively: The first module is the "forget gate", which is used to calculate the forgetting proportion of the neuron state information at the previous moment:

The second module is the "input gate", which uses new information to write the proportion of neuron states:

The third module is the "output gate", which will determine the information to be output as a hidden state:

Among them, C ^(t) and h ^(t) represent the neuron state and hidden state at time t , respectively; f ^(t) , i ^(t) and o ^(t) represent the forgetting gate, input gate and output gate at time t , respectively ; W , b are the weight matrix and bias vector in each gating module respectively; s is the sigmoid activation function; tanh is the hyperbolic tangent activation function. 3. a kind of ground connection wall deformation dynamic prediction method based on CV-LSTM neural network algorithm according to claim 1, is characterized in that:

,

,

. 4. a kind of ground connecting wall deformation dynamic prediction method based on CV-LSTM neural network algorithm according to claim 1, is characterized in that: collect the deformation history monitoring data of foundation pit engineering ground connecting wall, including by being buried in the wall The horizontal displacement of the ground connecting wall monitored by the movable inclinometer in the concrete; the monitoring frequency is 1 time per day, and it is increased to 2 times per day when the excavation depth exceeds 20 m, and 2 to 3 times per day when the deformation is abnormal. 5. a kind of ground connecting wall deformation dynamic prediction method based on CV-LSTM neural network algorithm according to claim 1, is characterized in that: measure the deformation true value ŷ _i of each monitoring point of ground connecting wall, obtain ground according to step 3. For the predicted value y _i and the real value ŷ _i of the connecting wall deformation, MSE, MAE and MAPE are used as loss functions to calculate the accuracy evaluation indicators MSE, MAE and MAPE respectively, and the prediction accuracy of the combined prediction model is evaluated. The calculation formula is as follows:

,

,

.

Images