CN111680343A - Deep foundation pit support structure deformation prediction method based on deep learning - Google Patents

Deep foundation pit support structure deformation prediction method based on deep learning Download PDF

Info

Publication number
CN111680343A
CN111680343A CN202010345711.8A CN202010345711A CN111680343A CN 111680343 A CN111680343 A CN 111680343A CN 202010345711 A CN202010345711 A CN 202010345711A CN 111680343 A CN111680343 A CN 111680343A
Authority
CN
China
Prior art keywords
data
foundation pit
deformation
training
deep learning
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN202010345711.8A
Other languages
Chinese (zh)
Other versions
CN111680343B (en
Inventor
徐杨青
张二勇
江威
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Wuhan Design and Research Institute of China Coal Technology and Engineering Group
Original Assignee
Wuhan Design and Research Institute of China Coal Technology and Engineering Group
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Wuhan Design and Research Institute of China Coal Technology and Engineering Group filed Critical Wuhan Design and Research Institute of China Coal Technology and Engineering Group
Priority to CN202010345711.8A priority Critical patent/CN111680343B/en
Publication of CN111680343A publication Critical patent/CN111680343A/en
Application granted granted Critical
Publication of CN111680343B publication Critical patent/CN111680343B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00—Computer-aided design [CAD]
    • G06F30/10—Geometric CAD
    • G06F30/13—Architectural design, e.g. computer-aided architectural design [CAAD] related to design of buildings, bridges, landscapes, production plants or roads
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00—Computer-aided design [CAD]
    • G06F30/20—Design optimisation, verification or simulation
    • G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/049—Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Evolutionary Computation (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Geometry (AREA)
  • Artificial Intelligence (AREA)
  • Software Systems (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Computer Hardware Design (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Mathematical Physics (AREA)
  • Data Mining & Analysis (AREA)
  • Computational Linguistics (AREA)
  • Health & Medical Sciences (AREA)
  • General Health & Medical Sciences (AREA)
  • Civil Engineering (AREA)
  • Architecture (AREA)
  • Structural Engineering (AREA)
  • Computational Mathematics (AREA)
  • Mathematical Analysis (AREA)
  • Mathematical Optimization (AREA)
  • Pure & Applied Mathematics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Medical Informatics (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The invention provides a deep foundation pit support structure deformation prediction method based on deep learning, which is characterized by analyzing factors influencing the deformation of a deep foundation pit support structure, determining a plurality of most important influencing factors, and carrying out initialization and normalization; constructing a tensor generator, and generating tensors and deformation quantities of training data, verification data and test data of historical data of the influence factors according to a preset backtracking period and a prediction period; the tensors of the training data and the verification data are input into the constructed deep learning network model for training, and after preset conditions are met, the trained deep learning network model is obtained; and inputting a plurality of most important influence factor tensors obtained actually into a final deep learning network model, predicting the deformation of the deep foundation pit support structure, and obtaining the deformation quantity of the deep foundation pit support structure so as to facilitate the timely repair of workers. The invention has the beneficial effects that: the prediction precision is improved, the construction unit is helped to predict the deformation of the enclosure structure, and the maintenance is carried out in time.

Description

Deep foundation pit support structure deformation prediction method based on deep learning
Technical Field
The invention relates to the field of deep foundation pit support, in particular to the technical field of inversion of deformation monitoring parameters, and particularly relates to a deep foundation pit support structure deformation prediction method based on deep learning.
Background
The deep foundation pit refers to a project which has an excavation depth of more than 5 meters (including 5 meters), or has a depth of less than 5 meters, but has particularly complicated geological conditions, surrounding environments and underground pipelines. With the rapid development of domestic underground railways and high-rise buildings, the application of deep foundation pit engineering is more and more extensive, and the prediction research on the enclosure deformation of the deep foundation pit is concerned day by day.
In the construction of underground deep foundation pit engineering, the deformation of the foundation pit has become a random, fuzzy and complex engineering problem under the influence of construction conditions, loading conditions, hydrological conditions, engineering geological conditions and other surrounding environmental factors, if the supporting structure is damaged in the process of excavation and foundation construction of the foundation pit, it will have serious consequences, therefore, monitoring work is very important in the construction process of foundation pit engineering, in the excavation process of the foundation pit, if the deformation of the foundation pit enclosure is predicted in advance by using a certain engineering monitoring method, therefore, the design method taking the deformation of the enclosure structure as a control measure is more and more emphasized by people. At present, for predicting the deformation of a deep foundation pit support structure, the existing main methods include four methods: time series analysis, regression analysis, artificial neural network, and gray system analysis;
the basic idea of the time series analysis method is to research and describe a certain time series as a random process, namely, a model is established on the basis of original data to describe the random process and carry out parameter estimation, model inspection and modification, and finally, the future time series value is reasonably predicted on the basis of the past value and the current value of the known time series. Regression analysis is the prediction or examination of the changes in unknowns based on the mathematical expressions that establish the relationships between the variables. The artificial neural network method is not based on any mathematical model, and only continuously trains and learns the method for distinguishing the effective information from the ineffective information through a large amount of existing data experience, so that the analysis and the processing of related similar data are achieved, and the future data can be predicted. Grey system analysis to solve the problem of uncertainty in the minority of data.
The deformation prediction of the foundation pit support structure is a complex problem, the number of influence factors is large, for high-dimensional characteristics, the weight of different characteristics in the prediction is difficult to determine by the existing method, and the problem is simulated by using a proper model, so that the precision is not high. Deep learning is a machine learning technology which is at the leading edge of research in recent years, and completes feature learning by simulating the cognition of human brain neurons on real objects. The method is widely applied to problems of pattern recognition, classification, regression analysis and the like. Compared with the traditional method, deep learning is good at extracting a large number of features through mass data and establishing the relationship between the features and the problem pieces. Therefore, deep learning is introduced into deformation monitoring of the deep foundation pit support structure, deformation influence factors of the deep foundation pit support structure are analyzed and researched, and a proper deep learning network framework is constructed, so that the method has important significance for deformation prediction of the deep foundation pit support structure.
Disclosure of Invention
In order to solve the problems, the invention provides a deep foundation pit support structure deformation prediction method based on deep learning, which mainly comprises the following steps:
s101: analyzing factors influencing deformation of the deep foundation pit support structure, determining a plurality of most important influence factors, dividing the plurality of most important influence factors into dynamic influence factors and static influence factors according to sampling frequency of the influence factors, and initializing and normalizing historical data of the plurality of most important influence factors;
s102: constructing a tensor generator, and generating tensors and deformation quantities of training data, verification data and test data respectively from the selected historical data of the influence factors according to a preset backtracking period and a preset prediction period;
s103: constructing a deep learning network model of coupling of a convolutional neural network and a long-short term memory network, wherein tensors of static influence factors are input into branches of the convolutional neural network, and tensors of dynamic influence factors are input into branches of the long-short term memory network;
s104: inputting tensors of training data and verification data into the deep learning network model for training, and obtaining and storing the trained deep learning network model after preset conditions are met;
s105: inputting the tensor of the test data into a trained deep learning network model to verify the prediction precision; and inputting a plurality of most important influence factor tensors obtained actually into a final deep learning network model, predicting the deformation of the deep foundation pit enclosure structure, and outputting the deformation quantity of the deep foundation pit enclosure structure so as to facilitate the timely repair of workers.
Further, analyzing factors influencing the deep foundation pit support structure from seven major factors of hydrogeology, space geometry, a support structure system, construction conditions, construction load, foundation pit exposure conditions and surrounding environment; the method comprises the following steps of screening out the most important 14 influence factors from seven major factors, wherein the 14 influence factors are geological conditions, hydrological conditions, the length of the side where an inclination measuring hole is located, the ratio of the side length of the inclination measuring hole to the width of a foundation pit, the excavation depth of the foundation pit, the type of a support structure, support stress of the support, the soil penetration depth of the support structure, precipitation, load conditions, the excavation exposure time of the foundation pit, peripheral dynamic load, peripheral static load and peripheral settlement;
according to the deep foundation pit monitoring standard, different influence factors have different observation frequencies, and the 14 influence factors are divided into two categories according to the observation frequencies: a dynamic impact factor and a static impact factor; wherein the static impact factors include: geological conditions, hydrological conditions, the length of the side where the inclination measuring hole is located, the ratio of the side length of the inclination measuring hole to the width of the foundation pit, the type of the enclosure structure and the soil penetration depth of the enclosure structure; the dynamic influence factors include: the method comprises the following steps of foundation pit excavation depth, enclosure support stress, precipitation, load conditions, foundation pit excavation exposure time, peripheral dynamic load, peripheral static load and peripheral settlement.
Further, the normalization process is as follows: assigning values to all the influence factors, assigning values according to the actual observed values when the actual observed values exist, assigning values according to empirical values when no observed values exist, keeping static influence factors unchanged in the observation process, and averagely setting the observation frequency of the dynamic influence factors twice a day, namely, at intervals of 12 hours; and normalizing the historical data of each influence factor to the range of [ -1,1] according to the value range of each observation value, and connecting each influence factor according to the time sequence to construct a dynamic influence factor array and a static influence factor array.
Further, step S102 specifically includes:
s201: designing a generator G for generating data tuples T (samples, displacements), wherein the displacements represent the horizontal displacement of monitoring points of batches, the samples represent a batch set of influence factors corresponding to the displacements of the monitoring points, and the batch number is batch size; the generator G contains the following parameters: a backtracking period BP, a prediction period FP, a Start time Start date and an End time End date; the backtracking period is used for limiting the acting time of the research influence factors on the deformation points, the prediction period is used for limiting the future deformation prediction time, and the starting time and the ending time are used for determining the time range of the data tuples; the specific generation mode is as follows: in the time interval [ Start date + BP, End date]And randomly selecting blocksize quantity of deformation displacement displacements from β each time, and acquiring all monitoring data samples in backtracking period BP of all influence factors influencing the displacements aiming at each deformation displacement in the batch displacements, wherein β is an array formed by sorting building enclosure horizontal displacement monitoring point data in a construction area according to time, and the array is β [2 × t × N ]]Wherein N is the total number of the monitoring points of the horizontal displacement of the building area enclosure structure,
Figure BDA0002470111080000031
in the monitoring period of t, m is the number of the inclination measuring holes in the building area enclosure structure, h is the buried depth of the inclination measuring holes, and h is1Observing intervals for the inclination measuring holes;
s202: dividing the whole monitoring period t into three time intervals of Training period, Validation period and Test period according to a certain proportion of time length,aiming at three time intervals, utilizing a generator G to respectively construct a training data generator GtrainingVerification data generator GvalidationAnd a test data generator GtestRespectively, for generating training data, verification data and test data.
Further, step S103 specifically includes:
s301: the deep learning network model is provided with two branches, wherein the first branch corresponds to a dynamic influence factor tensor, and the second branch corresponds to a static influence factor tensor; the first branch comprises a long-time memory network and a short-time memory network and is used for analyzing the influence of the time sequence influence factors on the deformation of the enclosure structure, and the second branch comprises a convolutional neural network and is used for mining the relation between the characteristics of the static influence factors and the deformation;
s302: the shape input _ shape of an input branch tensor of the long-time memory network is (time steps, data), the time steps are the observation times of each dynamic influence factor in a backtracking period, the data dim is a matrix dimension formed by all the dynamic influence factors observed each time, the unit of each layer of hidden units of the long-time memory network is 128, an activation function is hyperbaric value (tanh (x)), a cyclic activation function is sigmoid, and a first-layer parameter return _ sequences parameter and a second-layer parameter return _ sequences parameter are set as True; the output of the long-time and short-time memory network is LSTM _ OUT;
s303: the convolutional neural network consists of a two-dimensional convolutional neural network Conv2D, and a batch normalization layer, a flattening layer and a full connection layer are immediately followed by Conv 2D; the kernel size of the two-dimensional convolutional network is (1,2), the number of the convolutional kernels is 128, the activation function is a Rectified Linear Unit, ReLU, and the batch normalization layer can prevent the gradient of the convolutional neural network from being reduced in the training process and avoid the problem of gradient disappearance, so that the training time is reduced; the flat layer is used to transform the multidimensional input into a one-dimensional vector; inputting the vector after the flattening layer into a full connection layer for realizing the nonlinear combination of the characteristics, wherein the output of the convolutional neural network is CNN _ OUT;
s304: connecting the output LSTM _ OUT and CNN _ OUT by using a connection function, inputting the connected vector into a dropout layer, wherein the dropout layer is used for avoiding overfitting, and the drop rate is set to be 0.5; finally, outputting a prediction result through a full-connection layer with the unit number of 1; the activation function of the fully connected layer is sigmoid.
Further, step S104 specifically includes: training a deep learning network model using a fit _ generator function, which receives GtrainingTraining the generated data tuples and receiving GvalidationTraining the generated data tuples, setting Adam as an optimizer for training, setting training times as epoch, setting a loss function as Mean Square Error (MSE), controlling the training period number by monitoring the loss function in order to prevent overfitting in the training process, finishing training when the loss function is not reduced any more in continuous n periods, obtaining a trained deep learning network model at the moment, and storing the trained deep learning network model.
Further, in step S105, the deep learning network model trained in step S104 is used to predict the test data, and the test data passes GtestGenerating the number of generated test samples as size _ test, and comparing the predicted data with GtestComparing and analyzing the generated reference data; RMSE (root Mean Squared error), MAE (Mean absolute error) and R are selected2As the accuracy evaluation indexes, three evaluation index calculation formulas are as follows:
Figure BDA0002470111080000051
Figure BDA0002470111080000052
Figure BDA0002470111080000053
in the above formula, xiThe actual value is represented by the value of,
Figure BDA0002470111080000054
the predicted value is represented by a value of the prediction,
Figure BDA0002470111080000055
represents the average of all predicted values; i represents a test deformation data serial number, and i is 1,2, …, m; m is GtestAnd generating the total number of the test deformation data.
The technical scheme provided by the invention has the beneficial effects that: the technical scheme provided by the invention can greatly improve the prediction precision, and is beneficial to helping a construction unit to accurately predict the deformation of the enclosure structure, thereby providing protective measures in a targeted manner.
Drawings
The invention will be further described with reference to the accompanying drawings and examples, in which:
fig. 1 is a flowchart of a method for predicting deformation of a deep foundation pit support structure based on deep learning in an embodiment of the present invention;
FIG. 2 is a schematic diagram of monitoring data of two adjacent stages of a certain inclination measuring hole in a building envelope in the embodiment of the invention;
FIG. 3 is a block diagram of a deep learning network model in an embodiment of the present invention;
FIG. 4 is a graph of accuracy of deep learning network model training and validation in an embodiment of the present invention;
FIG. 5 is a diagram of the comparison of predicted value to true value accuracy using the method of the present invention in an embodiment of the present invention;
FIG. 6 is a comparison graph of prediction accuracy for several reference methods in an embodiment of the present invention.
Detailed Description
For a more clear understanding of the technical features, objects and effects of the present invention, embodiments of the present invention will now be described in detail with reference to the accompanying drawings.
The embodiment of the invention provides a deep foundation pit support structure deformation prediction method based on deep learning.
Referring to fig. 1 to 6, fig. 1 is a flowchart of a deep foundation pit support structure deformation prediction method based on deep learning in an embodiment of the present invention, fig. 2 is a schematic diagram of monitoring data of two adjacent stages of a certain inclination measuring hole in a support structure in an embodiment of the present invention, fig. 3 is a structural diagram of a deep learning network model in an embodiment of the present invention, fig. 4 is a graph of deep learning network model training and verification accuracy in an embodiment of the present invention, fig. 5 is a graph of predicted value and actual value accuracy comparison using the method of the present invention in an embodiment of the present invention, and fig. 6 is a graph of prediction accuracy comparison of several reference methods in an embodiment of the present invention; taking a deep foundation pit of a subway station in Wuhan City as an example, the length of the foundation pit main body is 227.8m, and the width is 23.1 m. And (7) excavating a foundation pit by 16m (H), wherein the depth of the enclosure structure into the soil is 14m, the foundation pit is constructed by an open excavation method, and the part of the foundation pit is constructed by a cover excavation method. The building enclosure selects an underground continuous wall with the thickness of 1000mm and six internal supports, the monitoring period is from 10 months in 2014 to 10 months in 2016, two years are total, the observation frequency is averaged once for 12 hours, 15 monitoring sections are arranged in the foundation pit, each monitoring section is provided with 2 inclination measuring holes for monitoring the deformation of the building enclosure, 30 inclination measuring holes are totally arranged, the inclination measuring holes are 30m deep, a building enclosure deformation observation point is arranged every 0.5m from top to bottom, the total number of the observation points is N which is 30/0.5 which is 1800, and the observation value array in the monitoring period is beta [1800 multiplied by 2 by 365] whichis beta [1314000 ]. As shown in fig. 2, two adjacent deformation records of a certain inclination measuring hole are observed, and "+" indicates that the hole is displaced into the foundation pit; "-" indicates a displacement outward of the pit.
S101: analyzing factors influencing deformation of the deep foundation pit support structure, determining a plurality of most important influence factors, dividing the plurality of most important influence factors into dynamic influence factors and static influence factors according to sampling frequency of the influence factors, and initializing and normalizing historical data of the plurality of most important influence factors;
seven major factors of hydrogeology (F1), space geometry (F2), a support structure system (F3), construction conditions (F4), construction load (F5), foundation pit exposure condition (F6) and surrounding environment (F7) are mainly considered in the analysis of the influence factors of the deep foundation pit support structure. The method comprises the steps of screening 14 influence factors from seven major factors, wherein the influence factors are geological conditions (F1-1), hydrological conditions (F1-2), the length of the side where an inclination measuring hole is located (F2-1), the ratio of the side length of the inclination measuring hole to the width of a foundation pit (F2-2), the excavation depth of the foundation pit (F2-3), the type of a support structure (F3-1), support stress of the support structure (F3-2), the soil penetration depth of the support structure (F3-3), precipitation (F4-1), load conditions (F5), exposure time of excavation of the foundation pit (F6), dynamic load of the periphery (F7-1), static load of the periphery (F7-2) and deformation of the periphery (F7-3).
According to the deep foundation pit monitoring specification, different influence factors have different observation frequencies, and the 14 influence factors can be divided into two categories, namely dynamic influence Factors (FD) and static influence Factors (FS) according to the observation frequencies. Wherein the static factors include: f1-1, F1-2, F2-1, F2-2, F3-1 and F3-3; the dynamic factors include: f2-3, F3-2, F4-1, F5, F6, F7-1, F7-2 and F7-3. The array initialization is performed for each factor according to the example, specifically referring to the following table:
Figure BDA0002470111080000071
Figure BDA0002470111080000081
the array dimension formed by the combined FS influence factors is as follows: [6,1 ]; the array dimensions formed by the FD influence factors in combination are as follows: [365*2*2,21]
S102: constructing a tensor generator, and generating tensors and deformation quantities of training data, verification data and test data from the historical data of the most important influence factors according to a preset backtracking period and a preset prediction period;
in step S102, the method specifically includes:
s201: designing a tensor generator G for generating data tuples T (displacements) to prepare for data input of each training period of the deep learning network, wherein the displacements represent the horizontal displacement of monitoring points of batches, the samples represent a batch set of influence factors corresponding to the displacement of the monitoring points, and the batch number is batch size. The generator G contains important parameters: the Backtracking Period (BP) is used for limiting the acting time of research influence factors on deformation points, in this example, BP is 720 hours, that is, all influence factor data in 720 hours of Backtracking are obtained for deformation data of a certain building envelope, and the Forecast Period (FP) is used for limiting the forecast time of future deformation, in this example, FP is 12 hours, that is, deformation after 12 hours is forecasted. The Start time Start date and the End time End date are used to determine the time range of the data tuple. The specific generation mode is that in a time interval [ Start date + BP, End date ], deformation quantities displacements with the batch size of 1024 are randomly selected from beta each time, and all monitoring data samples in a backtracking period BP of all influence factors influencing the displacements are acquired aiming at each deformation quantity displacement in the batch displacements.
S202: dividing the whole monitoring period t into three time intervals of a Training period, a verification period and a Test period according to a time length proportion of 5:2:3, and respectively constructing a Training data generator G by using a generator G aiming at the three time intervalstrainingVerification data generator GvalidationAnd a test data generator Gtest。GtrainingHas a time interval of [2014.11-2015.10 ]]，GvalidationHas a time interval of [2015.10-2016.2 ]]，GtrainingHas a time interval of [2016.2-2016.10 ]]。
S103: constructing a deep learning network model of coupling of a convolutional neural network and a long-short term memory network, wherein tensors of static influence factors are input into branches of the convolutional neural network, and tensors of dynamic influence factors are input into branches of the long-short term memory network;
in step S103, a deep learning network model is established based on the Keras library. The deep learning network model has two branches, the first branch corresponding to the FD tensor and the second branch corresponding to the FS tensor. The first branch mainly comprises a stack Short Term Memory (LSTM) Network for analyzing the influence of the time sequence influence factor on the deformation of the enclosure structure, and the LSTM is one of the Recurrent Neural Networks (RNNs), but can solve a very prominent problem in RNN training: gradient disappearance/explosion. The second branch mainly contains a convolutional neural network CNN for mining the relationship between the characteristics of the static impact factors and the deformation.
As shown in fig. 3, a branched stacked LSTM network is mainly constructed by three layers of LSTM networks. The shape input _ shape of the input branch tensor is (time steps, data dim), the time steps is the number of times of observation of each FD in the trace-back period, the trace-back period of the example is 720 hours, and the observation frequency is once 12 hours, so the time steps is 60, the data dim is the matrix dimension formed by all the FDs observed each time, in the example, the data dim is [1,21], each LSTM network layer hidden unit is 128, the activation function is hyperbaric generator (tank), the cyclic activation function is sigmoid, and the first and second layers of LSTM parameters, return _ sequences parameter is set as True. The output of network branch one is LSTM _ OUT.
In fig. 3, the branched bi-CNN network mainly consists of a two-dimensional convolutional neural network Conv2D, after which a batch normalization layer (batch normalization layer), a flattening layer (flattening layer) and a fully connected layer (dense layer) follow the Conv 2D; the kernel size of the two-dimensional convolutional network is (1,2), the number of the convolutional kernels is 128, the activation function is a Rectified Linear Unit, ReLU, and the batch normalization layer can prevent the gradient of the convolutional neural network from being reduced in the training process and avoid the problem of gradient disappearance, so that the training time is reduced; the flat layer is used to "flatten" the input, i.e., to convert the multidimensional input into a one-dimensional vector; the vector after the flattening layer is input into a full connection layer for realizing the nonlinear combination of the features, and the output of the branch two is CNN _ OUT.
ReLU(x)＝max(0,x)
FIG. 3 is a diagram in which the outputs LSTM _ OUT and CNN _ OUT of the two branches are connected by a connection function concatenate, and the connected vectors are input into a dropout layer, which is used to avoid overfitting, and the drop rate is set to 0.5; and finally, outputting a prediction result through a full-connection layer with the unit number of 1. The activation function of the fully connected layer is sigmoid.
S104: inputting tensors of training data and verification data into the deep learning network model for training, and obtaining and storing the trained deep learning network model after preset conditions are met;
in step S104, the method specifically includes: the model is trained in batches using the fit _ generator function, which receives GtrainTraining the generated data tuples and receiving GvalidationVerifying the generated data tuple, setting an optimizer for training as Adam, setting the training times epoch as 100, and setting the loss function val _ loss as Mean Square Error (MSE)or). And controlling the number of training cycles by monitoring the loss function by adopting early stopping in order to prevent overfitting in the training process, finishing training when the loss function does not decrease any more in 15 continuous cycles, and storing the model when the loss function is the lowest. In the example, the training process is shown in fig. 4, two curves represent the loss function of the training data and the loss function of the verification data, respectively, and the optimal model (Val _ loss ═ 0.0819) is reached at the 15 th epoch in the whole training period of 30 epochs, and the model is saved for subsequent prediction and testing.
S105: inputting the tensor of the test data into a trained deep learning network model to verify the prediction precision; inputting a plurality of most important influence factor tensors obtained actually into a finally trained deep learning network model, predicting the deformation of the deep foundation pit enclosure structure, and outputting the deformation quantity of the deep foundation pit enclosure structure so as to facilitate the timely repair of workers;
when the model trained in step S104 is used to predict the test data, the test data passes through GtestGenerating, setting the number of randomly selected samples to 1000, and comparing the predicted data with GtestAnd comparing the generated reference data, and analyzing the precision. RMSE (root Mean Squared error), MAE (Mean Absolute error) and R are selected2As an index for precision evaluation. The three evaluation index calculation formulas are as follows:
Figure BDA0002470111080000101
Figure BDA0002470111080000102
Figure BDA0002470111080000103
in the above formula, xiThe actual value is represented by the value of,
Figure BDA0002470111080000104
the predicted value is represented by a value of the prediction,
Figure BDA0002470111080000105
represents the average of all predicted values; i represents a test deformation data serial number, and i is 1,2, …, m; and m is the total number of the test deformation data.
FIG. 5 is a comparison graph of the true value and the predicted value, and the prediction accuracy of the present invention is: r2＝0.83，RMSE＝0.29，MAE＝0.25。
In order to verify the prediction accuracy of the invention, several common methods are selected to be combined with the indexes in the S501 for comparison and verification. These common methods include: the prediction accuracy of the comparison method is shown in fig. 6, which is an Autoregressive moving averaging model (ARMA), Multiple regression (Multiple regression), Gray system (Gray system), and BP neural network (BP neural network) for time series analysis. Compared with the reference method, the prediction precision of the method is superior to that of all reference methods, and the superiority of the method is reflected, and the method is specifically as follows:
R2 RMSE MAE
ARMA 0.62 0.49 0.19
Multiple regression 0.35 0.27 0.76
Gray system 0.50 0.39 0.51
BP neural network 0.39 0.31 0.68
the method of the invention 0.83 0.29 0.25
The invention has the beneficial effects that: the technical scheme provided by the invention fills the blank of prediction research on deformation of the deep foundation pit maintenance structure by deep learning, can greatly improve the prediction precision, and is beneficial to helping a construction unit to accurately predict the deformation of the enclosure structure, thereby providing protective measures in a targeted manner.
The above description is only for the purpose of illustrating the preferred embodiments of the present invention and is not to be construed as limiting the invention, and any modifications, equivalents, improvements and the like that fall within the spirit and principle of the present invention are intended to be included therein.

Claims (7)

1. A deep foundation pit support structure deformation prediction method based on deep learning is characterized by comprising the following steps: the method comprises the following steps:
s101: analyzing factors influencing the deformation of the deep foundation pit support structure, determining a plurality of most important influence factors, dividing the plurality of most important influence factors into dynamic influence factors and static influence factors according to the sampling frequency of the influence factors, and initializing and normalizing historical data of the plurality of most important influence factors;
s102: constructing a tensor generator, and generating tensors and deformation quantities of training data, verification data and test data from the historical data of the most important influence factors according to a preset backtracking period and a preset prediction period;
s103: constructing a deep learning network model of coupling of a convolutional neural network and a long-short term memory network, wherein tensors of static influence factors are input into branches of the convolutional neural network, and tensors of dynamic influence factors are input into branches of the long-short term memory network;
s104: inputting tensors of training data and verification data into the deep learning network model for training, and obtaining and storing the trained deep learning network model after preset conditions are met;
s105: inputting the tensor of the test data into a trained deep learning network model to verify the prediction precision; and inputting a plurality of most important influence factor tensors obtained actually into a final deep learning network model, predicting the deformation of the deep foundation pit enclosure structure, and outputting the deformation quantity of the deep foundation pit enclosure structure so as to facilitate the timely repair of workers.
2. The deep foundation pit support structure deformation prediction method based on deep learning of claim 1, wherein the method comprises the following steps: in the step S101, analyzing factors influencing the deep foundation pit support structure from seven major factors of hydrogeology, space geometry, a support structure system, construction conditions, construction load, foundation pit exposure conditions and surrounding environment; the method comprises the following steps of screening out the most important 14 influence factors from seven major factors, wherein the 14 influence factors are geological conditions, hydrological conditions, the length of the side where an inclination measuring hole is located, the ratio of the side length of the inclination measuring hole to the width of a foundation pit, the excavation depth of the foundation pit, the type of a support structure, support stress of the support, the soil penetration depth of the support structure, precipitation, load conditions, the excavation exposure time of the foundation pit, peripheral dynamic load, peripheral static load and peripheral settlement;
according to the deep foundation pit monitoring standard, different influence factors have different observation frequencies, and the 14 influence factors are divided into two categories according to the observation frequencies: a dynamic impact factor and a static impact factor; wherein the static impact factors include: geological conditions, hydrological conditions, the length of the side where the inclination measuring hole is located, the ratio of the side length of the inclination measuring hole to the width of the foundation pit, the type of the enclosure structure and the soil penetration depth of the enclosure structure; the dynamic influence factors include: the method comprises the following steps of foundation pit excavation depth, enclosure support stress, precipitation, load conditions, foundation pit excavation exposure time, peripheral dynamic load, peripheral static load and peripheral settlement.
3. The deep foundation pit support structure deformation prediction method based on deep learning of claim 1, wherein the method comprises the following steps: in step S101, the normalization process is: assigning values to all the influence factors, assigning values according to the actual observed values when the actual observed values exist, assigning values according to empirical values when no observed values exist, keeping static influence factors unchanged in the observation process, and averagely setting the observation frequency of the dynamic influence factors twice a day, namely, at intervals of 12 hours; and normalizing the historical data of each influence factor to the range of [ -1,1] according to the value range of each observation value, and connecting each influence factor according to the time sequence to construct a dynamic influence factor array and a static influence factor array.
4. The deep foundation pit support structure deformation prediction method based on deep learning of claim 1, wherein the method comprises the following steps: in step S102, the method specifically includes:
s201: designing a generator G for generating data tuples T (samples, displacements), wherein the displacements represent the horizontal displacement of monitoring points of batches, the samples represent a batch set of influence factors corresponding to the displacements of the monitoring points, and the batch number is batch size; the generator G contains the following parameters: a backtracking period BP, a prediction period FP, a Start time Start date and an End time End date; the backtracking period is used for limiting the acting time of the research influence factors on the deformation points, the prediction period is used for limiting the future deformation prediction time, and the starting time and the ending time are used for determining the time range of the data tuples; the specific generation mode is as follows: in the time interval [ Start date + BP, End date]In that, the deformation displacements of batchsize quantity are randomly selected from β each timeFor each deformation quantity displacement in the batch displacements, acquiring all monitoring data samples in backtracking period BP of all influence factors influencing the displacement, wherein β is an array formed by sequencing horizontal displacement monitoring point data of the enclosure structure in a construction area according to time, and the array is β [2 × t × N ]]Wherein N is the total number of the monitoring points of the horizontal displacement of the building area enclosure structure,
Figure FDA0002470111070000021
in the monitoring period of t, m is the number of the inclination measuring holes in the building area enclosure structure, h is the buried depth of the inclination measuring holes, and h is1Observing intervals for the inclination measuring holes;
s202: dividing the whole monitoring period t into three time intervals of a Training period, a Validation period and a Test period according to a certain proportion of time duration, and respectively constructing a Training data generator G by using a generator G aiming at the three time intervalstrainingVerification data generator GvalidationAnd a test data generator GtestRespectively, for generating training data, verification data and test data.
5. The deep foundation pit support structure deformation prediction method based on deep learning of claim 1, wherein the method comprises the following steps: in step S103, the method specifically includes:
s301: the deep learning network model is provided with two branches, wherein the first branch corresponds to a dynamic influence factor tensor, and the second branch corresponds to a static influence factor tensor; the first branch comprises a long-time memory network and a short-time memory network and is used for analyzing the influence of the time sequence influence factors on the deformation of the enclosure structure, and the second branch comprises a convolutional neural network and is used for mining the relation between the characteristics of the static influence factors and the deformation;
s302: the shape input _ shape of an input branch tensor of the long-time memory network is (time steps, data), the time steps are the observation times of each dynamic influence factor in a backtracking period, the data dim is a matrix dimension formed by all the dynamic influence factors observed each time, the unit of each layer of hidden units of the long-time memory network is 128, an activation function is hyperbaric value (tanh (x)), a cyclic activation function is sigmoid, and a first-layer parameter return _ sequences parameter and a second-layer parameter return _ sequences parameter are set as True; the output of the long-time and short-time memory network is LSTM _ OUT;
s303: the convolutional neural network consists of a two-dimensional convolutional neural network Conv2D, and a batch normalization layer, a flattening layer and a full connection layer are immediately followed by Conv 2D; the kernel size of the two-dimensional convolutional network is (1,2), the number of the convolutional kernels is 128, the activation function is a Rectified Linear Unit, ReLU, and the batch normalization layer can prevent the gradient of the convolutional neural network from being reduced in the training process and avoid the problem of gradient disappearance, so that the training time is reduced; the flat layer is used to transform the multidimensional input into a one-dimensional vector; inputting the vector after the flattening layer into a full connection layer for realizing the nonlinear combination of the characteristics, wherein the output of the convolutional neural network is CNN _ OUT;
s304: connecting the output LSTM _ OUT and CNN _ OUT by using a connection function, inputting the connected vector into a dropout layer, wherein the dropout layer is used for avoiding overfitting, and the drop rate is set to be 0.5; finally, outputting a prediction result through a full-connection layer with the unit number of 1; the activation function of the fully connected layer is sigmoid.
6. The deep foundation pit support structure deformation prediction method based on deep learning of claim 4, wherein the method comprises the following steps: in step S104, the method specifically includes: training a deep learning network model using a fit _ generator function, which receives GtrainingTraining the generated data tuples and receiving GvalidationTraining the generated data tuples, setting Adam as an optimizer for training, setting training times as epoch, setting a loss function as Mean Square Error (MSE), controlling the training period number by monitoring the loss function in order to prevent overfitting in the training process, finishing training when the loss function is not reduced any more in continuous n periods, obtaining a trained deep learning network model at the moment, and storing the trained deep learning network model.
7. The deep foundation pit periphery based on deep learning of claim 4The method for predicting the deformation of the protective structure is characterized by comprising the following steps: in step S105, the deep learning network model trained in step S104 is used for predicting test data, and the test data passes through GtestGenerating the number of generated test samples as size _ test, and comparing the predicted data with GtestComparing and analyzing the generated reference data; selecting RMSE, MAE and R2As the accuracy evaluation indexes, three evaluation index calculation formulas are as follows:
Figure FDA0002470111070000041
Figure FDA0002470111070000042
Figure FDA0002470111070000043
in the above formula, xiThe actual value is represented by the value of,
Figure FDA0002470111070000044
the predicted value is represented by a value of the prediction,
Figure FDA0002470111070000045
represents the average of all predicted values; i represents a test deformation data serial number, and i is 1,2, …, m; m is GtestAnd generating the total number of the test deformation data.
CN202010345711.8A 2020-04-27 2020-04-27 Deep foundation pit support structure deformation prediction method based on deep learning Active CN111680343B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010345711.8A CN111680343B (en) 2020-04-27 2020-04-27 Deep foundation pit support structure deformation prediction method based on deep learning

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010345711.8A CN111680343B (en) 2020-04-27 2020-04-27 Deep foundation pit support structure deformation prediction method based on deep learning

Publications (2)

Publication Number Publication Date
CN111680343A true CN111680343A (en) 2020-09-18
CN111680343B CN111680343B (en) 2022-12-13

Family

ID=72433846

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010345711.8A Active CN111680343B (en) 2020-04-27 2020-04-27 Deep foundation pit support structure deformation prediction method based on deep learning

Country Status (1)

Country Link
CN (1) CN111680343B (en)

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112529315A (en) * 2020-12-17 2021-03-19 杭州鲁尔物联科技有限公司 Landslide prediction method, device, equipment and storage medium
CN114492211A (en) * 2022-04-15 2022-05-13 中国石油大学(华东) Residual oil distribution prediction method based on autoregressive network model
CN114592525A (en) * 2022-03-16 2022-06-07 深圳宏业基岩土科技股份有限公司 Intelligent monitoring prevention and control system for inner support of deep and large foundation pit
CN117236518A (en) * 2023-11-10 2023-12-15 中国矿业大学 Prediction method for geological deformation of subway along line
CN117249803A (en) * 2023-08-09 2023-12-19 中交第四航务工程局有限公司 Deep foundation pit supporting upright post settlement monitoring and early warning method

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
RU2220251C2 (en) * 2001-08-06 2003-12-27 Калустян Эдуард Семенович Method establishing deformation characteristics of bedrock by data of observations on location
CN103116804A (en) * 2013-01-31 2013-05-22 苏州科技学院 Fuzzy neural network model and intelligent prediction method for deep excavation deformation
CN107832568A (en) * 2017-12-12 2018-03-23 青岛理工大学 The assay method of deep pit digging soil layer dynamic level resistance coefficient

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
RU2220251C2 (en) * 2001-08-06 2003-12-27 Калустян Эдуард Семенович Method establishing deformation characteristics of bedrock by data of observations on location
CN103116804A (en) * 2013-01-31 2013-05-22 苏州科技学院 Fuzzy neural network model and intelligent prediction method for deep excavation deformation
CN107832568A (en) * 2017-12-12 2018-03-23 青岛理工大学 The assay method of deep pit digging soil layer dynamic level resistance coefficient

Non-Patent Citations (5)

* Cited by examiner, † Cited by third party
Title
熊春宝等: "深基坑变形影响因子分析及预测方法研究", 《石家庄铁道大学学报(自然科学版)》 *
程楠等: "基于改进神经网络的深基坑工程变形预报", 《河南科技大学学报(自然科学版)》 *
葛长峰等: "人工神经网络在预测深基坑周边地表沉降变形中的应用研究", 《防灾减灾工程学报》 *
袁志明等: "基于LSTM-GM神经网络模型的深基坑沉降变形预测", 《江西理工大学学报》 *
陈炳志: "基坑变形灰色人工神经网络预测模型及其应用", 《山东科技大学学报(自然科学版)》 *

Cited By (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112529315A (en) * 2020-12-17 2021-03-19 杭州鲁尔物联科技有限公司 Landslide prediction method, device, equipment and storage medium
CN112529315B (en) * 2020-12-17 2024-05-28 杭州鲁尔物联科技有限公司 Landslide prediction method, landslide prediction device, landslide prediction equipment and storage medium
CN114592525A (en) * 2022-03-16 2022-06-07 深圳宏业基岩土科技股份有限公司 Intelligent monitoring prevention and control system for inner support of deep and large foundation pit
CN114492211A (en) * 2022-04-15 2022-05-13 中国石油大学(华东) Residual oil distribution prediction method based on autoregressive network model
CN114492211B (en) * 2022-04-15 2022-07-12 中国石油大学(华东) Residual oil distribution prediction method based on autoregressive network model
CN117249803A (en) * 2023-08-09 2023-12-19 中交第四航务工程局有限公司 Deep foundation pit supporting upright post settlement monitoring and early warning method
CN117249803B (en) * 2023-08-09 2024-05-28 中交第四航务工程局有限公司 Deep foundation pit supporting upright post settlement monitoring and early warning method
CN117236518A (en) * 2023-11-10 2023-12-15 中国矿业大学 Prediction method for geological deformation of subway along line
CN117236518B (en) * 2023-11-10 2024-05-24 中国矿业大学 Prediction method for geological deformation of subway along line

Also Published As

Publication number Publication date
CN111680343B (en) 2022-12-13

Similar Documents

Publication Publication Date Title
CN111680343B (en) Deep foundation pit support structure deformation prediction method based on deep learning
CN105973594B (en) A kind of rolling bearing fault Forecasting Methodology based on continuous depth confidence network
Meletti et al. The new Italian seismic hazard model (MPS19)
CN108197432B (en) Gene regulation and control network reconstruction method based on gene expression data
CN110046379B (en) Structure overall damage identification method based on space-frequency information
CN109583621A (en) A kind of PSO-LSSVM short-term load forecasting method based on improvement variation mode decomposition
CN104951836A (en) Posting predication system based on nerual network technique
CN109613611B (en) Method and system for determining input seismic waves for structural seismic time-course analysis
CN112001110B (en) Structural damage identification monitoring method based on vibration signal space real-time recurrent graph convolutional neural network
Ding et al. Stationary subspaces-vector autoregressive with exogenous terms methodology for degradation trend estimation of rolling and slewing bearings
CN112528365B (en) Method for predicting healthy evolution trend of underground infrastructure structure
CN109635245A (en) A kind of robust width learning system
CN114548509A (en) Multi-type load joint prediction method and system for multi-energy system
CN112131794A (en) Hydraulic structure multi-effect optimization prediction and visualization method based on LSTM network
CN112818601A (en) Hydroelectric generating set health assessment method based on GA-BP neural network and error statistical analysis
CN112836789A (en) Ground connection wall deformation dynamic prediction method based on composite neural network algorithm
Roy et al. Demand forecasting in smart grid using long short-term memory
CN111798055A (en) Variable weight combined photovoltaic output prediction method based on grey correlation degree
CN116703644A (en) Attention-RNN-based short-term power load prediction method
CN115359197A (en) Geological curved surface reconstruction method based on spatial autocorrelation neural network
CN115577854A (en) Quantile regression wind speed interval prediction method based on EEMD-RBF combination
Lee et al. Seismic Acceleration Estimation Method at Arbitrary Position Using Observations and Machine Learning
Yan et al. AQI prediction based on CEEMD-WOA-Elman neural network
CN112685900A (en) Power load simulation method for representing impact load power characteristics
Torky DEEP LEARNING TECHNIQUES FOR STRUCTURAL RESPONSE PREDICTION DURING STRONG GROUND MOTIONS

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant