A method for structural deformation prediction based on convolutional neural network

Abstract

The present invention discloses a method for predicting structural deformation based on a convolutional neural network, comprising the following steps: creating a three-dimensional voxel model of the image range around a deep foundation pit at different times; cutting a cross-section from the three-dimensional voxel model at the location of a deformation monitoring point to establish a structural change detection data set containing the cross-section and the deformation value of the deformation monitoring point at the current time; dividing the structural change detection data set into a training set and a test set, and constructing a change prediction convolutional neural network to predict structural deformation values from the cross-sections at different time states. The advantages of the present invention are that the structural deformation prediction method fully considers the correlation in three-dimensional spatial structure, utilizes the spatial relationship feature extraction capability of the convolutional neural network, and learns the mapping relationship between underground space material changes and structural deformation. Compared with other prediction algorithms, the input dimension considered is higher, the information is more comprehensive, and more in line with reality, thereby improving the accuracy and reliability of the prediction results.

Description

Technical Field

[0001] The present invention belongs to the field of geotechnical engineering structure safety, and specifically relates to a method for predicting structural deformation based on a convolutional neural network. Background Art

[0002] The safety status of structures surrounding foundation pits includes structural deformation and damage. Deformation prediction analysis involves a comprehensive three-dimensional analysis of the interactions between the foundation pit, soil, and pipelines. Structural deformation prediction generally includes statistical analysis, mechanical model analysis, hybrid models, and machine learning models.

[0003] Statistical analysis processes data from the perspective of time series analysis, but it often leads to overfitting, that is, the model has poor generalization ability.

[0004] Mechanical analysis and hybrid model methods use finite element analysis and other methods to simulate the structural stress conditions and compare them with the actual conditions. However, finite element analysis and other methods involve more simplifications and assumptions, which increases uncertainty and affects the reliability of structural deformation prediction.

[0005] The principle of machine learning models is to collect current structure-related data and use the established machine learning models to predict structural deformation. However, the current mainstream machine learning models only use the current influencing factor data and time series data of the structure, ignoring the correlation in the three-dimensional space of the structure.

[0006] Structural defects are generally predicted using machine learning models, such as Bayesian classification and other data mining techniques. Similar to structural deformation prediction, these only consider the current influencing factors and temporal correlation of the structure, without considering the spatial correlation characteristics of the structure.

[0007] Chinese patent CN111143934A proposes a structural deformation prediction method based on a temporal convolutional network, which ignores the correlation in three-dimensional space. Chinese patent CN113362389A adds image information of the measurement point location, as well as the cross-section, structural form, and lighting conditions of the shield subway tunnel on the basis of the original time series prediction. However, the structural deformation prediction of surrounding construction activities still ignores the correlation of the three-dimensional spatial structure, which affects the accuracy and reliability of the results. Summary of the Invention

[0008] The purpose of the present invention is to provide a method for predicting structural deformation based on a convolutional neural network in response to the above-mentioned shortcomings of the prior art. The method creates a three-dimensional voxel model of the image range around the deep foundation pit at different times; cuts a cross-section from the three-dimensional voxel model at the position of the deformation monitoring point to establish a structural change detection dataset of the cross-section and the deformation value of the deformation monitoring point at the current time; divides the structural change detection dataset into a training set and a test set, constructs a change prediction convolutional neural network, and predicts the structural deformation value from the cross-section at different time states.

[0009] The purpose of the present invention is achieved by the following technical solutions:

[0010] A method for predicting structural deformation based on a convolutional neural network, characterized in that the method comprises the following steps:

[0011] S1: Construct a three-dimensional voxel model (N, E, D, C) of the image range around the deep foundation pit at the construction state at time t; wherein N, E, and D represent the north coordinate, east coordinate, and depth in the three-dimensional coordinate system, respectively; C represents the material properties at the (N, E, D) coordinate position at the current construction state at time t, and the material properties include R, G, and B, where R represents the compression modulus or elastic modulus, G represents the cohesion or shear modulus, and B represents the internal friction angle;

[0012] S2: Based on the construction process of the deep foundation pit and the three-dimensional voxel model (N, E, D, C) in step S1, a time-series-based three-dimensional voxel model (N, E, D, C) is established; for each deformation monitoring point, a cross-sectional view (H, W, C) of the three-dimensional voxel model (N, E, D, C) corresponding to the deformation monitoring point at different observation times is extracted, where H and W are the height and width of the obtained cross-sectional view, respectively; and a structural change detection dataset of the deformation value y and the cross-sectional view (H, W, C) of the deformation monitoring point at the current time is established;

[0013] S3: Dividing the structural change detection dataset into a training set and a test set according to a certain ratio;

[0014] S4: The structural deformation prediction model is trained on the training set, features are extracted through convolution and pooling operations, and the final fully connected output is obtained. During the entire training process, the weighted cross entropy loss is used as the target optimization, and the stochastic gradient algorithm with adaptive learning rate is used for parameter optimization.

[0015] S5: testing the trained structural deformation prediction model on the test set, and determining the structural deformation prediction model to be finally applied;

[0016] S6: After determining the structural deformation prediction model to be finally applied, extract the current section view X2, the previous section view X1, and the previous measured cumulative deformation value y1 from the three-dimensional voxel model (N, E, D, C), input the current section view X2 and the previous section view X1 into the structural deformation prediction model to predict the deformation value △y, and thus calculate the cumulative deformation value y2=y1+△y under the current construction state.

[0017] The method for obtaining the cross-sectional view (H, W, C) is to extract the cross-sectional view (H, W, C) along the cross-sectional view where the line connecting the position of the deformation monitoring point and the nearest point of the foundation pit boundary is located.

[0018] The deformation value y in the structural change detection data set in step S2 is discretized. The discretization method is: mapping the deformation value y to the interval [0, 1] and discretizing it at intervals of 0.01. The discretization calculation formula is:

[0019]

[0020] Where: λ =-0.52, δ =0.73, ξ =0.10, γ =1.67.

[0021] The structural deformation prediction model includes an input layer, a first encoding layer, a second encoding layer, a first decoding layer, a second decoding layer, a third decoding layer and an output layer; the input layer divides the previous cross-sectional image X1 and the current cross-sectional image X2 into two paths and inputs them into the first encoding layer and the second encoding layer for encoding processing; the first encoding layer and the second encoding layer adopt DenseNet121 as the basic structure of the basic encoding layer; the first decoding layer and the second decoding layer are both composed of a maximum pooling layer and a fully connected layer, and receive the encoded previous cross-sectional image X1 and the current cross-sectional image X2 from the first encoding layer and the second encoding layer respectively, and output the cumulative deformation value y1 in the previous cross-sectional image X1 and the cumulative deformation value y2 in the current cross-sectional image X2 respectively; the third decoding layer takes the feature map splicing between the first encoding layer and the second encoding layer as input, and outputs △y after respective convolution, pooling and fully connected layer operations, and the third decoding layer outputs △y= y2- y1.

[0022] In step S3, 80% of the structural change detection dataset is divided into the training set and 20% is divided into the test set.

[0023] The advantages of the present invention are: the structural deformation prediction method fully considers the correlation in three-dimensional spatial structure, utilizes the spatial relationship feature extraction capability of convolutional neural network, and learns the mapping relationship between underground space material changes and structural deformation. Compared with other prediction algorithms, the input dimension considered is higher, the information is more comprehensive and more in line with reality, so the accuracy and reliability of the prediction results are better. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 Schematic diagram of the three-dimensional voxel model of the deep foundation pit under different construction processes in the present invention;

[0025] Figure 2 It is a cross-sectional view of the same deformation monitoring point at different times in the present invention;

[0026] Figure 3 Schematic diagram of the flow of the structural deformation prediction model in the present invention;

[0027] Figure 4 It is a schematic diagram of the basic structure of Dense Block in the present invention. DETAILED DESCRIPTION

[0028] The features of the present invention and other related features are further described in detail below through embodiments in conjunction with the accompanying drawings to facilitate understanding by those skilled in the art:

[0029] Example: Figure 1 、 2 As shown in , 3, and 4, this embodiment specifically relates to a method for predicting structural deformation based on a convolutional neural network, the method comprising the following steps:

[0030] (S1) Considering the characteristics of soil elastic modulus, foundation pit construction method, foundation pit retaining structure and reinforcement material stiffness around the foundation pit, a 3D voxel model (N, E, D, C) of the image range around the deep foundation pit under a certain construction state (i.e., time t) is constructed according to a certain resolution (usually 0.2m). The 3D voxel model is as follows: Figure 1 As shown, where:

[0031] N, E, and D represent the north coordinate, east coordinate, and depth in the three-dimensional coordinate system, respectively;

[0032] C represents the material properties of the (N, E, D) coordinate position at the current construction state at time t. Three values (R, G, and B) are used to represent the properties of different materials, such as the material properties of different soils and underground continuous walls.

[0033] The first value R is the compression modulus or elastic modulus. The compression modulus of soil is obtained through the compression modulus curve. The compression modulus of reinforcement is 150-200. The elastic modulus value is used for concrete and steel bars.

[0034] The second value G is cohesion or shear modulus. For soil, cohesion is used. The shear modulus of reinforced solid is 1 / 5 of the compressive strength, and the shear modulus of concrete is 1 / 20 of the compressive strength.

[0035] The third value B is the internal friction angle. The soil uses the measured solid block index, the reinforced body uses a value of 20-30 degrees, and concrete and steel use a uniform value of 45 degrees.

[0036] (S2) Based on the construction process of the deep foundation pit and the three-dimensional voxel model (N, E, D, C) in step S1, a time-series-based three-dimensional voxel model (N, E, D, C) is established, that is, each moment has its corresponding three-dimensional voxel model (N, E, D, C).

[0037] For each deformation monitoring point, the cross-section (H, W, C) of the corresponding three-dimensional voxel model (N, E, D, C) of the deformation monitoring point at any observation time is extracted. It can be understood that the cross-section (H, W, C) of the deformation monitoring point based on the time series is extracted, such as Figure 2 As shown in FIG, the method for obtaining the cross-sectional view (H, W, C) is to extract the cross-sectional view (H, W, C) along the cross-sectional view where the line connecting the position of the deformation monitoring point and the nearest point of the foundation pit boundary is located; wherein H and W are the height and width of the obtained cross-sectional view, respectively; C is the three values that characterize the material properties of different materials, namely the aforementioned R, G, and B values.

[0038] A structural change detection data set is established, which includes the cumulative deformation value of each deformation monitoring point at the current moment and the corresponding cross-sectional view (H, W, C).

[0039] In order to better learn the subsequent structural deformation prediction model, the deformation value y needs to be discretized. The specific method is to map the deformation value y to the [0,1] interval and discretize it at intervals of 0.01. The deformation value is discretized according to the following formula:

[0040]

[0041] Where: λ =-0.52, δ =0.73, ξ =0.10, γ =1.67.

[0042] (S3) The structural change detection dataset is divided into a training set and a test set according to the ratio, with the training set ratio being 80% and the test set ratio being 20%.

[0043] (S4) The structural deformation prediction model is trained on the training set, features are extracted through convolution and pooling operations, and finally the fully connected output is obtained. During the entire training process, the weighted cross entropy loss is used as the target optimization, and the stochastic gradient algorithm with adaptive learning rate is used for parameter optimization.

[0044] like Figure 3 As shown, the structural deformation prediction model includes an input layer, a first encoding layer, a second encoding layer, a first decoding layer, a second decoding layer, a third decoding layer, and an output layer:

[0045] Input layer: Obtain the profiles (H, W, C) of the same deformation monitoring point at two observation times, namely the previous profile X1 and the current profile X2. The previous profile X1 and the current profile X2 are divided into two paths and input into the first coding layer and the second coding layer for coding processing respectively;

[0046] The first coding layer and the second coding layer correspond to two inputs X1 and X2 respectively. The structure adopts DenseNet121 as the basic structure of the basic coding layer. The DenseNet structures of different depths are as follows Figure 4 As shown, the two layers share parameters;

[0047] The first decoding layer and the second decoding layer share parameters and are both composed of a maximum pooling layer and a fully connected layer. This layer takes as input the feature maps output from the first encoding layer and the second encoding layer, and outputs the cumulative deformation value y1 of the previous profile image X1 and the cumulative deformation value y2 of the current profile image X2, respectively.

[0048] The third decoding layer consists of three 3*3 convolution and 2*2 average pooling layers to form a feature pyramid structure. The input of each layer of the feature pyramid is the concatenation of the corresponding denseblock input feature maps in the first encoding layer and the second encoding layer. Finally, it inputs a pooling layer and a fully connected layer and outputs △y. The corresponding output of the third decoding layer is △y = y2-y1.

[0049] S5: Test the trained structural deformation prediction model on the test set. The test results use precision, recall, and F1 value as accuracy indicators for the test data. The model with the best F1 value is selected as the final structural deformation prediction model. The F1 value is the harmonic mean of precision and recall, which is the arithmetic mean divided by the geometric mean. The larger the F1 value, the better.

[0050] S6: After determining the final structural deformation prediction model, extract the current section X2, the previous section X1, and the previous measured cumulative deformation value y1 from any deformation monitoring point on the three-dimensional voxel model (N, E, D, C), input the current section X2 and the previous section X1 into the structural deformation prediction model to predict the deformation value △y, and thus calculate the cumulative deformation value y2=y1+△y under the current construction state.

[0051] The beneficial effects of this embodiment are: the structural deformation prediction method fully considers the correlation in three-dimensional spatial structure, utilizes the spatial relationship feature extraction capability of convolutional neural network, and learns the mapping relationship between material changes in underground space and structural deformation. Compared with other prediction algorithms, the input dimension considered is higher, the information is more comprehensive, and more in line with reality, so the accuracy and reliability of the prediction results are better.

Claims

1. A method for predicting structural deformation based on convolutional neural network, characterized in that The method comprises the following steps: S1: Construct a three-dimensional voxel model (N, E, D, C) of the image range around the deep foundation pit at the construction state at time t; wherein N, E, and D represent the north coordinate, east coordinate, and depth in the three-dimensional coordinate system, respectively; C represents the material properties at the (N, E, D) coordinate position at the current construction state at time t, and the material properties include R, G, and B, where R represents the compression modulus or elastic modulus, G represents the cohesion or shear modulus, and B represents the internal friction angle; S2: Based on the construction process of the deep foundation pit and the three-dimensional voxel model (N, E, D, C) in step S1, a time-series-based three-dimensional voxel model (N, E, D, C) is established; for each deformation monitoring point, a cross-sectional view (H, W, C) of the three-dimensional voxel model (N, E, D, C) corresponding to the deformation monitoring point at different observation times is extracted, where H and W are the height and width of the obtained cross-sectional view, respectively; and a structural change detection dataset of the deformation value y and the cross-sectional view (H, W, C) of the deformation monitoring point at the current time is established; S3: Dividing the structural change detection dataset into a training set and a test set according to a certain ratio; S4: The structural deformation prediction model is trained on the training set, features are extracted through convolution and pooling operations, and the final fully connected output is obtained. During the entire training process, the weighted cross entropy loss is used as the target optimization, and the stochastic gradient algorithm with adaptive learning rate is used for parameter optimization. S5: testing the trained structural deformation prediction model on the test set, and determining the structural deformation prediction model to be finally applied; S6: After determining the structural deformation prediction model to be finally applied, extract the current section view X2, the previous section view X1, and the previous measured cumulative deformation value y1 from the three-dimensional voxel model (N, E, D, C), input the current section view X2 and the previous section view X1 into the structural deformation prediction model to predict the deformation value △y, and thus calculate the cumulative deformation value y2=y1+△y under the current construction state.

2. The method for predicting structural deformation based on a convolutional neural network according to claim 1, characterized in that The method for obtaining the cross-sectional view (H, W, C) is to extract the cross-sectional view (H, W, C) along the cross-sectional view where the line connecting the position of the deformation monitoring point and the nearest point of the foundation pit boundary is located.

3. The method for predicting structural deformation based on convolutional neural network according to claim 1, characterized in that The deformation value y in the structural change detection data set in step S2 is discretized. The discretization method is: mapping the deformation value y to the interval [0, 1] and discretizing it at intervals of 0.

01. The discretization calculation formula is: Where: λ =-0.52, δ =0.73, ξ =0.10, γ =1.

67.

4. The method for predicting structural deformation based on convolutional neural network according to claim 1, characterized in that The structural deformation prediction model includes an input layer, a first encoding layer, a second encoding layer, a first decoding layer, a second decoding layer, a third decoding layer and an output layer; the input layer divides the previous cross-sectional image X1 and the current cross-sectional image X2 into two paths and inputs them into the first encoding layer and the second encoding layer for encoding processing; the first encoding layer and the second encoding layer adopt DenseNet121 as the basic structure of the basic encoding layer; the first decoding layer and the second decoding layer are both composed of a maximum pooling layer and a fully connected layer, and receive the encoded previous cross-sectional image X1 and the current cross-sectional image X2 from the first encoding layer and the second encoding layer respectively, and output the cumulative deformation value y1 in the previous cross-sectional image X1 and the cumulative deformation value y2 in the current cross-sectional image X2 respectively; the third decoding layer takes the feature map splicing between the first encoding layer and the second encoding layer as input, and outputs △y after respective convolution, pooling and fully connected layer operations, and the third decoding layer outputs △y= y2- y1.

5. The method for predicting structural deformation based on convolutional neural network according to claim 1, characterized in that In step S3, 80% of the structural change detection dataset is divided into the training set and 20% is divided into the test set.

Images