阅读说明：本技术 一种基于cv-lstm组合模型的地连墙变形动态预测方法 (Ground continuous wall deformation dynamic prediction method based on CV-LSTM combined model ) 是由 刘维 赵华菁 管浩 王航远 于 2021-09-28 设计创作，主要内容包括：本发明公开了一种基于CV-LSTM组合模型的地连墙变形动态预测方法。选取监测点,采集基坑工程地连墙的变形历史监测数据,整理形成监测表；采用CV-LSTM组合模型对变形监测数据中的训练样本进行学习,将训练得到的最优模型对测试集样本进行变形预测,得到地连墙变形预测值。将按本发明技术方案得到变形预测值与变形实测值进行对比,计算评价指标,结果显示,本发明提供的CV-LSTM组合模型相比传统BP神经网络表现出了较高的预测精度,较单独的LSTM深度网络具有更好的泛化能力,适用于地连墙变形的动态预测问题,可为施工现场实现信息化管理提供参考。(The invention discloses a ground connection wall deformation dynamic prediction method based on a CV-LSTM combined model. Selecting a monitoring point, collecting deformation history monitoring data of the foundation pit engineering diaphragm wall, and arranging to form a monitoring table; and learning the training samples in the deformation monitoring data by adopting a CV-LSTM combined model, and performing deformation prediction on the test set samples by using the optimal model obtained by training to obtain a prediction value of the deformation of the diaphragm wall. Compared with the traditional BP neural network, the CV-LSTM combined model provided by the invention has higher prediction precision, better generalization capability than an independent LSTM deep network, is suitable for the dynamic prediction problem of diaphragm wall deformation, and can provide reference for realizing informatization management of a construction site.)

1. A ground connection wall deformation dynamic prediction method based on CV-LSTM neural network algorithm is characterized by comprising the following steps:

the method comprises the following steps:

selecting monitoring points, collecting deformation history monitoring data of the foundation pit engineering diaphragm wall, and recording deformation observation values collected by each monitoring point asRepresenting the deformation value of the measuring point i on the t day to form a time sequence of an observed value; arranging the monitoring data to form a monitoring table;

step two:

(1) reading the monitoring table formed in the step (1) by using an xlrd module in a PyTorch frame package, and storing data into a tensor structure by using a tensor function to obtain a data set;

(2) the prediction combination model is established by adopting a K-fold cross validation method and an LSTM neural network algorithm, and the input layer of the prediction combination model isThe output layer isWherein N is the input information length, and M is the prediction time span;

the prediction combination model divides a data set into K subsets by adopting a K-fold cross validation method, wherein K-1 subsets are used as training set samples for training in turn, and the rest 1 subsets are used as test set samples for testing;

the prediction combination model adopts an LSTM neural network algorithm to learn a training set sample, sets the super parameters including the number of training rounds EPOCH, the learning rate LR and the number of HIDDEN layer neurons HIDDEN _ SIZE, and obtains an optimal model through training by adjusting the super parameters;

the method for adjusting the hyper-parameters comprises the following steps: updating network parameters in formula iteratively by adopting Adam algorithmwAnd adjusting, wherein an iterative update formula is as follows:

；

wherein the content of the first and second substances,wnetwork parameters to be trained;is the learning rate;dwis a gradient;is the first moment attenuation coefficient;is the second moment attenuation coefficient; v is the exponentially weighted average of the original gradients; s is an exponentially weighted average of the squares of the gradients;normalization processing for gradient;

step three:

and carrying out deformation prediction on the test set sample by using the optimal model obtained by training to obtain a deformation prediction value of the diaphragm wall.

2. The ground wall deformation dynamic prediction method based on the CV-LSTM neural network algorithm as claimed in claim 1, wherein: the forward propagation formulas of the LSTM neural network algorithm are respectively as follows:

the first module is 'forget gate' and is used for calculating forget proportion of neuron state information at a last moment:

the second module is the "input gate" for new information written into the ratio of neuron states:

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

wherein the content of the first and second substances,C ^(t) andh ^(t) respectively representtA temporal neuron state and a hidden state;f ^(t) ，i ^(t) and o ^(t) Respectively representtA forgetting gate, an input gate and an output gate at a moment;W，brespectively are a weight matrix and an offset vector in each gating module;srepresenting a sigmoid activation function; tanh represents the hyperbolic tangent activation function.

3. The ground wall deformation dynamic prediction method based on the CV-LSTM neural network algorithm as claimed in claim 1, wherein:，， 。

4. the ground wall deformation dynamic prediction method based on the CV-LSTM neural network algorithm as claimed in claim 1, wherein: acquiring deformation history monitoring data of the foundation pit engineering diaphragm wall, wherein the deformation history monitoring data comprises horizontal displacement of the diaphragm wall monitored by a movable inclinometer buried in wall concrete; the monitoring frequency is 1 time per day, the excavation depth is increased to 2 times per day after exceeding 20 m, and the frequency is 2-3 times per day when the deformation is abnormal.

5. The ground wall deformation dynamic prediction method based on the CV-LSTM neural network algorithm as claimed in claim 1, wherein: measuring true deformation value of each monitoring point of diaphragm wallŷ _iObtaining the deformation prediction value of the diaphragm wall according to the third stepy _i And true valueŷ _iAnd respectively calculating precision evaluation indexes MSE, MAE and MAPE by taking MSE, MAE and MAPE as loss functions, and carrying out prediction precision evaluation on the prediction combination model, wherein the calculation formula is as follows:

，

，

。

Technical Field

The invention relates to a diaphragm wall deformation prediction method, in particular to a diaphragm wall deformation dynamic prediction method based on a CV-LSTM combined model.

Background

At present, China is in the rapid development stage of rail transit construction, a large number of station foundation pit projects appear, and the excavation scale and depth of foundation pits are continuously increased. For example, the maximum excavation depth of a station of a certain number of subway lines in a certain city reaches 29 m. The foundation pit engineering is a space system comprising a soil body and a supporting structure, and is influenced by a plurality of internal and external factors such as geological conditions, construction quality, surrounding environment and the like. The field measured deformation data is the embodiment of various comprehensive effects in the construction process. Therefore, analyzing and researching the on-site deformation monitoring data becomes an effective way for people to know the deformation characteristics of the foundation pit.

In the aspect of the prediction of the deformation rule of the underground diaphragm wall of the foundation pit under different soil properties and different support types, the prediction method is mainly divided into the following three types: empirical methods, numerical simulations, and machine learning. The empirical method needs a large amount of monitoring data, is limited to establishing a discrete random model by using a differential equation, and is inconvenient for describing the essence and the internal law of the system change process; the numerical method has accuracy on the mathematical method, but due to the complexity of the deformation influence factors of the underground diaphragm wall of the foundation pit, the fuzziness of a physical mechanism and the changeability and uncertainty of parameters, the method is excessively generalized when in use, and the practical value is reduced. At present, most of machine learning methods adopt classical BP neural networks, the structure of the neural network which propagates reversely is relatively simple, but the relevance of input time cannot be accurately processed, and the error is large when long sequence data is processed; the LSTM algorithm for processing sequence data has the advantages of convenience for sequence modeling and long-term memory capability, but has the defect of poor generalization capability.

Disclosure of Invention

Aiming at the defects in the prior art, the invention provides the diaphragm wall deformation prediction method based on the CV-LSTM combined model, and the combined model has higher stability, is suitable for long-term and dynamic prediction of diaphragm wall deformation, and has higher prediction precision and better generalization capability.

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

the method comprises the following steps:

selecting monitoring points, collecting deformation history monitoring data of the foundation pit engineering diaphragm wall, and recording deformation observation values collected by each monitoring point asRepresenting the deformation value of the measuring point i on the t day to form a time sequence of an observed value; arranging the monitoring data to form a monitoring table;

step two:

(1) reading the monitoring table formed in the step (1) by using an xlrd module in a PyTorch frame package, and storing data into a tensor structure by using a tensor function to obtain a data set;

(2) the prediction combination model is established by adopting a K-fold cross validation method and an LSTM neural network algorithm, and the input layer of the prediction combination model isThe output layer isWherein N is the input information length, and M is the prediction time span;

the prediction combination model divides a data set into K subsets by adopting a K-fold cross validation method, wherein K-1 subsets are used as training set samples for training in turn, and the rest 1 subsets are used as test set samples for testing;

the prediction combination model adopts an LSTM neural network algorithm to learn a training set sample, sets the super parameters including the number of training rounds EPOCH, the learning rate LR and the number of HIDDEN layer neurons HIDDEN _ SIZE, and obtains an optimal model through training by adjusting the super parameters;

the method for adjusting the hyper-parameters comprises the following steps: updating network parameters in formula iteratively by adopting Adam algorithmwAnd adjusting, wherein an iterative update formula is as follows:

；

wherein the content of the first and second substances,wnetwork parameters to be trained;is the learning rate;dwis a gradient;is the first moment attenuation coefficient;is the second moment attenuation coefficient; v is the exponentially weighted average of the original gradients; s is an exponentially weighted average of the squares of the gradients;normalization processing for gradient;

step three:

and carrying out deformation prediction on the test set sample by using the optimal model obtained by training to obtain a deformation prediction value of the diaphragm wall.

In the technical scheme of the invention, the forward propagation formulas of the LSTM neural network algorithm are respectively as follows:

the first module is 'forget gate' and is used for calculating forget proportion of neuron state information at a last moment:

the second module is the "input gate" for new information written into the ratio of neuron states:

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

wherein the content of the first and second substances,C ^(t) andh ^(t) respectively representtA temporal neuron state and a hidden state;f ^(t) ，i ^(t) and o ^(t) Respectively representtA forgetting gate, an input gate and an output gate at a moment;W，brespectively are a weight matrix and an offset vector in each gating module;srepresenting a sigmoid activation function; tanh represents the hyperbolic tangent activation function.

The invention relates to a ground connection wall deformation dynamic prediction method based on CV-LSTM neural network algorithm, which has a preferable scheme that:，， 。

the invention collects the deformation history monitoring data of the foundation pit engineering diaphragm wall, including the horizontal displacement of the diaphragm wall monitored by a movable inclinometer buried in the concrete of the wall body; the monitoring frequency is 1 time per day, the excavation depth is increased to 2 times per day after exceeding 20 m, and the frequency is 2-3 times per day when the deformation is abnormal.

The invention measures the deformation true value of each monitoring point of the diaphragm wallŷ _iObtaining the deformation prediction value of the diaphragm wall according to the third stepy _i And true valueŷ _iAnd respectively calculating precision evaluation indexes MSE, MAE and MAPE by taking MSE, MAE and MAPE as loss functions, and carrying out prediction precision evaluation on the prediction combination model, wherein the calculation formula is as follows:

，

，

。

compared with the prior art, the invention has the following advantages:

1. aiming at the characteristics of long period and nonlinearity of monitoring data, the invention adopts a CV-LSTM combined model prediction model method to overcome the defects of overfitting and gradient explosion of the traditional neural network.

2. The combined model provided by the invention has higher accuracy for long-term prediction of diaphragm wall deformation due to the advantage of processing long sequence data.

3. The combined model provided by the invention shows better stability and higher prediction precision, and is more suitable for dynamic prediction of diaphragm wall deformation.

4. The underground continuous wall deformation prediction method provided by the invention has better generalization capability.

Drawings

FIG. 1 is a schematic diagram of a foundation pit plane and an inclination measuring point arrangement for deformation prediction of a diaphragm wall according to an embodiment of the present invention;

FIG. 2 is a schematic cross-sectional design view of a standard section of a foundation pit for predicting deformation of an underground diaphragm wall according to an embodiment of the invention;

FIG. 3 is a flow chart of a deformation prediction combination model according to an embodiment of the present invention;

in the figure, the soil is filled with impurities, the silt clay is pulverized, the silt sand is mixed with the silt, the silt clay is mixed with the silt clay, the silt clay is mixed with the silt, and the silt sand is mixed with the silt.

Detailed Description

The invention is described in detail below with reference to the figures and specific embodiments.

Example 1

The method is applied to the engineering example of deep foundation pit engineering of a subway station in urban rail transit.

Referring to fig. 1, a schematic diagram of a foundation pit plane and an inclination measuring point for predicting deformation of a diaphragm wall in this embodiment is shown; the subway station foundation pit is in the east-west direction, and the nearest distance between the north side of the foundation pit and the existing building is only 1.7 m. The foundation pit enclosure structure adopts an underground continuous wall with the thickness of 1.0 m. The standard section of the foundation pit adopts 6 inner supports, the total length of the sub-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. 2329 groups of underground diaphragm wall inclination measurement monitoring data totaling from 8 months and 8 days in 2018 to 7 months and 9 days in 2019 are selected as monitoring points of the main structure CX 10-CX 19 of the foundation pit B to be used as original samples for prediction training, and the horizontal displacement of the underground diaphragm wall is monitored by a movable inclinometer buried in wall concrete.

Referring to fig. 2, a schematic cross-sectional design diagram of a standard section of a foundation pit for predicting deformation of an underground diaphragm wall in the embodiment is shown; the method comprises the following steps of sequentially distributing (from top to bottom) miscellaneous fill, silt clay, silt sand, silt clay and silt sand from top to bottom, wherein the burial depth of a stable water level is 1.20-1.90 m. The silty clay layers distributed in the field are mainly soft plastic, have the characteristics of high compressibility, high sensitivity, low shear strength and the like, have obvious rheological property and are main soft soil layers influencing engineering construction. And 6 inner supports are vertically arranged on the standard section of the foundation pit B.

Referring to fig. 3, a flow chart of a deformation prediction combination model provided in this embodiment is shown; and acquiring a data set, complementing missing data, and dividing the data set by adopting a Cross-validation method. The LSTM neural network is used for learning the training set sample, and the super parameters such as the number of training rounds (EPOCH), the Learning Rate (LR), the number of HIDDEN layer neurons (HIDDEN _ SIZE) and the like are set. And performing deformation prediction on the test set sample by using the trained optimal model to obtain a short-term deformation prediction value and a long-term deformation prediction value of the diaphragm wall, and calculating an accuracy evaluation index.

The specific implementation steps are as follows:

the method comprises the following steps:

selecting monitoring points, collecting deformation history monitoring data of a foundation pit engineering diaphragm wall, wherein the inclination measurement monitoring frequency of the diaphragm wall is 1 time per day in principle, the excavation depth is increased to 2 times per day after exceeding 20 m, the deformation is 2-3 times per day when abnormal, the monitoring data is arranged to form a monitoring daily newspaper, partial missing data is caused due to the fact that the monitoring points shield a gland and the like in the construction process, the data is completed by linear interpolation, and the deformation observation value collected by each monitoring point is recorded asAnd representing the deformation value of the measuring point i on the t day to form a time sequence of the deformation observation value, and sorting the monitoring data to form a monitoring table.

Step two:

and reading in data in the monitoring daily report by using an xlrd module in the PyTorch framework package, and storing the monitoring data into a tensor structure by using a tensor function. The input layer of the predictive combined model isThe output layer of the prediction combination model isWhere N inputs the information length and M represents the prediction time span.

A Cross-validation method is adopted to divide a data set, a 10-fold Cross-validation method is adopted for CX 10-CX 19 inclination measuring points in the original sample of the inclination measuring deformation of the diaphragm wall, each measuring point can be used as a test set to be predicted once, and therefore the generalization of a prediction model is relatively and objectively reflected.

Learning a training set sample by adopting an LSTM neural network, setting hyper-parameters such as the number of training rounds (EPOCH), the Learning Rate (LR), the number of HIDDEN layer neurons (HIDDEN _ SIZE) and the like, wherein forward propagation formulas are respectively as follows:

the first module is 'forget gate' and is used for calculating forget proportion of neuron state information at a last moment:

the second module is the "input gate" for new information written into the ratio of neuron states:

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

wherein the content of the first and second substances,C ^(t) andh ^(t) respectively representtTemporal neuron state sum concealmentA state;f ^(t) ，i ^(t) and o ^(t) Respectively representtA forgetting gate, an input gate and an output gate at a moment;W，brespectively are a weight matrix and an offset vector in each gating module;srepresenting a sigmoid activation function; tanh represents the hyperbolic tangent activation function.

The LSTM network has a deep structure and is difficult to update parameters, so the embodiment adopts an Adaptive Moment Estimation algorithm (Adam) for optimization. Adam dynamically modifies the learning rate of each parameter on one hand, and introduces a momentum method on the other hand, so that parameter updating has more chances to jump out of a local optimal solution. The iterative update formula is as follows:

；

wherein the content of the first and second substances,wnetwork parameters to be trained;is the learning rate;dwis a gradient;is the first moment attenuation coefficient;the second moment attenuation coefficient. v is the exponentially weighted average of the original gradients; s is an exponentially weighted average of the squares of the gradients;normalization processing for gradient; in this example，，，。

Step three:

performing deformation prediction on the test set sample by using the trained optimal combination model to obtain the deformation prediction value of the diaphragm wally _i 。

Due to output of predicted valuey _i And true valueŷ _iCertain errors exist among the combined models, the neural network evaluates the errors through a Loss Function (Loss Function), and the smaller the three values of the commonly used Loss functions MSE, MAE and MAPE, the better precision of the combined model is shown. And performing deformation prediction on the test set sample by using the trained optimal combination model so as to obtain a short-term deformation prediction value and a long-term deformation prediction value of the diaphragm wall. And calculating evaluation indexes MSE, MAE and MAPE values of the combined model under a prediction task, and evaluating the prediction precision of the CV-LSTM combined model.

Predicting the deformation predicted value and the deformation measured valuey _i And true valueŷ _i And comparing, and respectively calculating the precision evaluation indexes MSE, MAE and MAPE, wherein the calculation formula is as follows:

。

in order to further verify the effectiveness of the CV-LSTM combined model, a classical BP neural network is selected for comparison, and 4 prediction tasks are designed to reflect the prediction effect of the prediction model under different input information lengths and prediction step lengths.

Task 1: predicting the deformation amount after 1 day according to the deformation monitoring values of the previous 3 days (N =3, M = 1);

task 2: predicting the deformation amount after 7 days according to the deformation monitoring value of the previous 3 days (N =3, M = 7);

task 3: predicting the deformation amount after 1 day according to the deformation monitoring value of the previous 15 days (N =15, M = 1);

and task 4: the deformation amount after 7 days (N =15, M = 7) was predicted from the deformation monitoring value of the previous 15 days.

Prediction error MSE (unit/mm) value of 10-fold cross validation²) The specific results are shown in Table 1.

TABLE 1

。

The results in Table 1 show that the CV-LSTM combination model performed better than the BP combination model in all prediction tasks, and smaller error values were obtained on the test set.