Intelligent prediction model implementation method for deformation of adjacent existing sensitive structure

Abstract

The invention discloses an intelligent prediction model implementation method for deformation of an adjacent existing sensitive structure, and the method comprises the steps: S1, updating construction information and ground wall deflection information according to different prediction tasks; s2, the construction information and the ground wall deflection information are preprocessed; s3, constructing an intelligent prediction model based on a long short-term memory network and a convolutional neural network; and S4, inputting the data information into the intelligent prediction model so as to generate a prediction result of the ground wall deflection curve. According to the invention, continuous prediction of a plurality of future time steps is realized, and the prediction precision and stability are improved.

Description

Technical Field

[0001] The invention relates to a method for realizing an intelligent prediction model of deformation of an adjacent existing sensitive structure. Background Art

[0002] At present, my country is in a stage of rapid development of rail transit construction. In 2019, the number of new stations in the country reached 238, and rail transit projects of 59 lines in 12 regions were approved for construction, with a total project investment of about 970 billion yuan. As a result, a large number of station foundation pit projects have emerged, and the scale and depth of foundation pit excavation are also increasing. For example, the maximum excavation depth of a station on a subway line in a certain city has reached 29m. Foundation pit engineering is a spatial system consisting of soil and supporting structures, which is affected by many internal and external factors such as geological conditions, construction quality and surrounding environment. The on-site measured deformation data is a reflection of the comprehensive effects of various influences during the construction process. Analyzing and studying on-site deformation monitoring data has become an effective way for people to understand the deformation characteristics of foundation pits.

[0003] In the study of the prediction of the deformation law of foundation pit diaphragm wall under different soil properties and different support types, the prediction methods are mainly divided into three types: empirical method, numerical simulation method and machine learning method. The empirical method requires a large amount of monitoring data and is limited to using differential equations to establish discrete random models, which is not convenient for describing the essence and internal laws of the system change process; the numerical method has the accuracy of mathematical methods, but due to the complexity of the factors affecting the deformation of foundation pit diaphragm wall, the ambiguity of the physical mechanism, and the variability and uncertainty of the parameters, it is over-generalized when using this method, which reduces its practical value.

[0004] At present, most machine learning methods use the classic BP neural network. This back-propagation neural network structure is relatively simple, but it cannot accurately handle the temporal correlation of the input and has large errors when processing long sequence data.

[0005] Therefore, a method for implementing an intelligent prediction model for deformation of adjacent existing sensitive structures is provided. Summary of the invention

[0006] The purpose of the present invention is to overcome the existing defects and provide an implementation method of an intelligent prediction model for deformation of adjacent existing sensitive structures, thereby achieving continuous prediction of multiple future time steps and improving prediction accuracy and stability.

[0007] The technical solution to achieve the above purpose is:

[0008] A method for implementing an intelligent prediction model for deformation of adjacent existing sensitive structures, comprising:

[0009] Step S1, updating the construction information and ground-wall deflection information according to different prediction tasks;

[0010] Step S2, preprocessing the construction information and the ground-wall deflection information;

[0011] Step S3, constructing an intelligent prediction model based on long short-term memory network and convolutional neural network;

[0012] Step S4, inputting the data information into the intelligent prediction model to generate a prediction result of the ground-wall deflection curve.

[0013] Preferably, in step S2, the preprocessing operation includes but is not limited to normalization processing for eliminating the influence between different dimensions and orders of magnitude, wherein the normalization processing adopts Min-max normalization and Z-score normalization.

[0014] Preferably, in step S3, the optimizer in the intelligent prediction model is optimized by using the mean square error as the loss function, and the loss function MSE expression is as follows:

[0015]

[0016] In the formula, n represents the number of samples, and They are deformation prediction and measured results respectively.

[0017] Preferably, in step S4, the prediction process of the intelligent prediction model based on the long short-term memory network and the convolutional neural network includes:

[0018] Step S41, the pre-processed construction information and ground-wall deflection information are respectively extracted with time features through two independent encoders based on long short-term memory networks;

[0019] Step S42, the two extracted time features are spliced ​​in the channel dimension by feature fusion, and the construction information and the ground-wall deflection information are integrated into a unified feature;

[0020] Step S43, input the result after feature fusion into the 1D U-net model, extract the spatial structure information of the features and decode them step by step, so as to generate the prediction result of the ground-wall deflection curve.

[0021] Preferably, in step S41, the core of the long short-term memory network lies in its gating mechanism, including an input gate I t 、Forget Gate F t and output gate O t , the input gate is used to determine the importance of the current input information, the forget gate is used to control the past memories that need to be forgotten, and the output gate is used to determine the hidden state at the current moment, that is, the final output of the network; at time step t, its calculation formula is as follows:

[0022] I t=σ(W xi X t +W hi H t-1 +b i );

[0023] F t =σ(W xf X t +W hf H t-1 +b f );

[0024] O t =σ(W xo X t +W ho H t-1 +b o );

[0025]

[0026] H t =O t ⊙tanh(C t );

[0027] Where W is the weight parameter and b is the bias term. t and C t are the hidden variables and memory elements at time step t, X t is the input at time step t, H t-1 and C t-1 is the hidden state and memory element at time step t-1, I t 、F t , O t and They are the input gate, forget gate, output gate and candidate memory element at time t, σ is the sigmoid activation function, and the symbol ⊙ is the Hadamard product.

[0028] Preferably, the step S43 includes:

[0029] Step S431, inputting the result of feature fusion into the encoder of the 1D U-net model to extract the spatial structure information of the features;

[0030] Step S432, after feature extraction is completed, the fused features are sent to the upsampling decoder, which upsamples the low-resolution feature map to a high-resolution output, thereby generating a prediction result of the ground-wall deflection curve.

[0031] Preferably, in step S431, for the ground-connected wall deformation prediction task, the ground wall deflection obtained at each time point is one-dimensional sequence data, and the convolution layer extracts local features from the deformation sequence by introducing one-dimensional convolution to identify the deformation data in space and time, wherein the calculation formula of the one-dimensional convolution is as follows:

[0032]

[0033] Where y[i] represents the i-th element of the output feature map, w[k] is the k-th element of the convolution kernel, x[i+k] is the i+k-th element of the input data, and K is the size of the convolution kernel;

[0034] The pooling layer is usually set after the convolution layer to reduce the dimension of the feature map generated by the convolution. For a one-dimensional pooling operation, the pooling calculation formula is as follows:

[0035] y[i]=pool(x[i·s],x[i·s+1],...,x[i·s+p-1]);

[0036] Where y[i] represents the i-th element of the output feature map, pool represents the pooling operation, x is the input feature map, s is the step size, and p is the size of the pooling window;

[0037] The fully connected layer is located at the back of the convolutional neural network and is used for the final decision or prediction layer. It maps the features extracted by the previous convolutional layer and pooling layer to the output space to achieve classification or regression tasks. Assume that the input feature vector is z = [z1, z2, ..., z n ], the output is a, then the operation of the fully connected layer can be expressed as:

[0038] a=f(W·z+b);

[0039] Where W is the weight matrix, b is the bias vector, and f is the activation function.

[0040] Preferably, in step S432, a transposed convolution layer is introduced, that is, an upsampling layer is used as a substitute for the fully connected layer to upsample the low-resolution coding features to a high-resolution output while maintaining the original spatial structure, wherein the calculation formula of the one-dimensional transposed convolution is as follows:

[0041]

[0042] Where y[i] represents the i-th element of the output feature map, x[ik·s] is the ik·s-th element of the input feature map after inserting zeros, w[k] is the k-th element of the convolution kernel, and s is the step size.

[0043] Preferably, in step S43, the prediction result can be returned to step S1 for the next prediction iteration to perform prediction of future time steps.

[0044] The beneficial effects of the present invention are as follows: the present invention introduces construction information into the prediction model, increases the data dimension and information volume of the model input, enables the model to more comprehensively capture the time series characteristics, and after integrating the key information in the construction process, the model exhibits higher prediction accuracy in both long-term and short-term prediction tasks, thereby improving the prediction accuracy and stability of the model, and laying the foundation for achieving more accurate engineering safety management. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 It is a flow chart of a method for realizing an intelligent prediction model of deformation of an adjacent existing sensitive structure according to the present invention;

[0046] Figure 2 It is a specific flow chart of the prediction process of the intelligent prediction model based on long short-term memory network and convolutional neural network in the present invention;

[0047] Figure 3 It is a specific flow chart of inputting the result of feature fusion into the 1D U-net model, extracting the spatial structure information of the feature and decoding it step by step in the present invention;

[0048] Figure 4 It is a curve chart of the training loss change of the intelligent prediction model on 50 samples in the present invention;

[0049] Figure 5 It is a measured curve diagram of settlement of buildings adjacent to monitoring points JCJ02-02 and JCJ03-01 in an embodiment of the present invention;

[0050] Figure 6 It is a schematic diagram of the results of 20 consecutive steps ahead prediction by the intelligent prediction model and the LSTM model in an embodiment of the present invention. DETAILED DESCRIPTION

[0051] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings. In the description of the present invention, it should be noted that the terms "center", "up", "down", "left", "right", "vertical", "horizontal", "inside", "outside" and the like indicate directions or positional relationships based on the directions or positional relationships shown in the accompanying drawings, which are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific direction, be constructed and operated in a specific direction, and therefore cannot be understood as a limitation on the present invention. In addition, the terms "first", "second", and "third" are used for descriptive purposes only and cannot be understood as indicating or implying relative importance.

[0052] The present invention will be further described below in conjunction with the accompanying drawings.

[0053] like Figure 1 As shown, a method for realizing an intelligent prediction model of deformation of an adjacent existing sensitive structure includes:

[0054] Step S1, updating the construction information and ground-wall deflection information according to different prediction tasks.

[0055] Step S2, preprocessing the construction information and the ground and wall deflection information.

[0056] In the embodiment, the preprocessing operation includes but is not limited to normalization processing, which is used to eliminate the influence between different dimensions and orders of magnitude, wherein the normalization processing adopts Min-max normalization and Z-score normalization.

[0057] Normalization is a key step in data preprocessing, and its main purpose is to scale data with different features to a similar range, thereby avoiding training instability or degradation of model performance due to differences in feature dimensions. In the present invention, two common normalization methods are used for input data: Min-max normalization and Z-score normalization. These two normalization methods play a vital role in the data preprocessing process. Min-max normalization linearly converts the data to a specified range (usually [0,1]), ensuring that all input features have the same scale. This method can maintain the relative size relationship of the data while avoiding computational instability caused by eigenvalues ​​that are too large or too small. The formula is as follows:

[0058]

[0059] In the formula, x is the original data, x' is the normalized data, and x max and x min They are the minimum and maximum values ​​in the data set respectively; the advantage of Min-max normalization is that it is simple to calculate and the converted data has a clear range, which is suitable for scenarios where the relative proportional relationship of the data needs to be maintained. However, it also has a potential disadvantage, that is, when there are extreme values ​​in the data, the distribution of the normalized data may be affected. Z-score normalization (also called standardization) converts the data into a standard normal distribution with a mean of 0 and a standard deviation of 1, so that the distribution of the data is centralized and has a uniform scale. The formula is as follows:

[0060]

[0061] Where x is the original data, μ is the mean of the data, ε is the standard deviation of the data, and x' is the normalized data. The advantage of Z-score normalization is that it is insensitive to the distribution of the data, and is particularly suitable for processing data sets with large differences in the distribution of eigenvalues. By standardizing the data distribution of different features, Z-score normalization can effectively improve the model's ability to learn input features, especially in features with different dimensions and different data distributions. Normalization not only helps to improve the stability of model training, but also accelerates the optimization process. Without normalization, the magnitude difference of different eigenvalues ​​may cause inconsistent step sizes during gradient descent, which in turn affects the convergence speed and effect of the model. Through normalization, the data of all features are adjusted to a similar scale, and the optimizer can adjust the model parameters more smoothly, making the training process more efficient and stable.

[0062] Step S3, constructing an intelligent prediction model based on long short-term memory network and convolutional neural network.

[0063] In the embodiment, the optimizer in the intelligent prediction model is optimized by using the mean square error as the loss function, and the loss function MSE expression is as follows:

[0064]

[0065] In the formula, n represents the number of samples, and They are deformation prediction and measured results respectively.

[0066] like Figure 4 As shown in the figure, the training loss curve of the model on 50 samples is shown. All curves show that the training loss gradually decreases with the increase of training rounds. Compared with other optimizers, the Adam+Min-maxnormalization combination performs best in terms of training loss, with faster convergence speed and lower final loss value. Although the Adam+Z-score normalization combination has a lower final loss value, it fluctuates greatly during training. Therefore, the Adam+Min-max normalization combination is finally selected to train the model.

[0067] Step S4, inputting the data information into the intelligent prediction model to generate a prediction result of the ground-wall deflection curve.

[0068] like Figure 2 As shown in the figure, the prediction process of the intelligent prediction model based on long short-term memory network and convolutional neural network includes:

[0069] Step S41, the pre-processed construction information and ground-wall deflection information are respectively extracted with time features through two independent encoders based on long short-term memory networks.

[0070] In the embodiment, the core of the long short-term memory network lies in its gating mechanism, including an input gate I t 、Forget Gate F t and output gate O t , the input gate is used to determine the importance of the current input information, the forget gate is used to control the past memories that need to be forgotten, and the output gate is used to determine the hidden state at the current moment, that is, the final output of the network; at time step t, its calculation formula is as follows:

[0071] I t =σ(W xi X t +W hi H t-1 +b i ·;

[0072] F t =σ(W xf X t +W hf H t-1 +b f 0;

[0073] O t =σ(W xo X t +W ho H t-1 +b o );

[0074]

[0075] H t =O t ⊙tanh(C t );

[0076] Where W is the weight parameter and b is the bias term. t and C t are the hidden variables and memory elements at time step t, X t is the input at time step t, H t-1 and C t-1 is the hidden state and memory element at time step t-1, I t 、F t , O t and They are the input gate, forget gate, output gate and candidate memory element at time t, σ is the sigmoid activation function, and the symbol ⊙ is the Hadamard product.

[0077] In step S42, the two extracted time features are spliced ​​in the channel dimension through feature fusion, and the construction information and the ground-wall deflection information are integrated into a unified feature.

[0078] Step S43, input the result after feature fusion into the 1D U-net model, extract the spatial structure information of the features and decode them step by step, so as to generate the prediction result of the ground-wall deflection curve.

[0079] like Figure 3 As shown, step S43 includes:

[0080] Step S431, input the result of feature fusion into the encoder of the 1D U-net model to extract the spatial structure information of the features.

[0081] In the embodiment, for the ground-connected wall deformation prediction task, the ground-wall deflection obtained at each time point is one-dimensional sequence data. The convolution layer extracts local features from the deformation sequence by introducing one-dimensional convolution, and identifies the deformation data in space and time. The calculation formula of the one-dimensional convolution is as follows:

[0082]

[0083] Where y[i] represents the i-th element of the output feature map, w[k] is the k-th element of the convolution kernel, x[i+k] is the i+k-th element of the input data, and K is the size of the convolution kernel;

[0084] The pooling layer is usually set after the convolution layer to reduce the dimension of the feature map generated by the convolution. For a one-dimensional pooling operation, the pooling calculation formula is as follows:

[0085] y[i]=pool(x[i·s],x[i·s+1],...,x[i·s+p-1]);

[0086] Where y[i] represents the i-th element of the output feature map, pool represents the pooling operation, x is the input feature map, s is the step size, and p is the size of the pooling window;

[0087] The fully connected layer is located at the back of the convolutional neural network and is used for the final decision or prediction layer. It maps the features extracted by the previous convolutional layer and pooling layer to the output space to achieve classification or regression tasks. Assume that the input feature vector is z = [z1, z2, ..., z n ], the output is a, then the operation of the fully connected layer can be expressed as:

[0088] a=f(W·z+b);

[0089] Where W is the weight matrix, b is the bias vector, and f is the activation function.

[0090] Step S432, after feature extraction is completed, the fused features are sent to the upsampling decoder, which upsamples the low-resolution feature map to a high-resolution output, thereby generating a prediction result of the ground-wall deflection curve.

[0091] In the embodiment, a transposed convolution layer, that is, an upsampling layer is introduced as a substitute for the fully connected layer to upsample the low-resolution encoding features to a high-resolution output while maintaining the original spatial structure, wherein the calculation formula of the one-dimensional transposed convolution is as follows:

[0092]

[0093] Where y[i] represents the i-th element of the output feature map, x[ik·s] is the ik·s-th element of the input feature map after inserting zeros, w[k] is the k-th element of the convolution kernel, and s is the step size.

[0094] In the embodiment, the prediction result may be returned to step S1 for the next prediction iteration to perform prediction of future time steps.

[0095] Example Application

[0096] Deep foundation pit excavation is often accompanied by the uplift of the foundation pit bottom, deformation of the retaining structure, and movement of the surrounding strata, which in turn induces deformation of the surrounding existing buildings that affects their safety. Focusing on the settlement of existing stations around the foundation pit, based on the building settlement monitoring project, an intelligent prediction method for the settlement of surrounding existing buildings induced by foundation pit excavation is constructed. The monitoring points JCJ02-02 and JCJ03-01 are selected as the prediction objects. The measured settlement curves of the two measuring points are as follows: Figure 5 As shown in Figure 2, the data from the first 70 days were collected for model training, and the data from the 71st to the 90th day were used as the test data set.

[0097] Figure 6The results of 20-step forward prediction of the fusion model (i.e., the intelligent prediction model based on long short-term memory network and convolutional neural network) and the LSTM model; at the JCJ02-02 measurement point, the fusion model showed strong stability and accuracy in the 20-step forward prediction task. The mean absolute error (MAE) of the fusion model was 0.63 mm, and the mean absolute percentage error (MAPE) was 12.12%. In contrast, the MAE of the LSTM model was 1.23 mm, and the MAPE was 22.53%, which was significantly higher than that of the fusion model. The results show that the LSTM model has a significant error accumulation phenomenon in the long-step prediction task. Specifically, in the prediction of each step, the absolute error of the fusion model is basically controlled within 1.5 mm, and the maximum absolute error is 1.49 mm, while the absolute error of the LSTM model has reached 2.83 mm at the 20th step, showing a trend of gradual expansion of the error. In addition, the prediction curve of the LSTM model has a large deviation from the measured value in the overall trend, while the fusion model maintains a trend consistent with the measured curve. At the JCJ03-01 measuring point, the fusion model also showed excellent prediction performance. The MAE of the fusion model was 0.24mm and the MAPE was 6.17%. In comparison, the MAE of the LSTM model was 1.20mm and the MAPE was 29.62%. As the number of prediction steps gradually increased, the error of the LSTM model increased significantly, especially at the 6th and 20th steps, when the absolute errors reached 1.02mm and 2.05mm, respectively, and the MAPE exceeded 54%. In contrast, the fusion model had lower absolute and percentage errors throughout the entire prediction process, showing better prediction consistency and trend matching. The prediction curve of the LSTM model deviated significantly from the measured curve in the later stage of the advance prediction, showing a strong error accumulation effect, while the prediction curve of the fusion model maintained a trend closer to the measured data. The prediction results prove that the intelligent prediction method based on multi-source information fusion also has strong adaptability and prediction accuracy when dealing with multi-step advance prediction tasks of building settlement.

[0098] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that the technical solutions described in the above embodiments may still be modified, or some or all of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for implementing an intelligent prediction model for deformation of adjacent existing sensitive structures, characterized in that: include: Step S1, updating the construction information and ground-wall deflection information according to different prediction tasks; Step S2, preprocessing the construction information and the ground-wall deflection information; Step S3, constructing an intelligent prediction model based on long short-term memory network and convolutional neural network; Step S4, inputting the data information into the intelligent prediction model to generate a prediction result of the ground-wall deflection curve.

2. The method for realizing the intelligent prediction model of deformation of adjacent existing sensitive structures according to claim 1, characterized in that: In step S2, the preprocessing operation includes but is not limited to normalization processing, which is used to eliminate the influence between different dimensions and orders of magnitude, wherein the normalization processing adopts Min-max normalization and Z-score normalization.

3. The method for realizing the intelligent prediction model of deformation of adjacent existing sensitive structures according to claim 1, characterized in that: In step S3, the optimizer in the intelligent prediction model is optimized by using the mean square error as the loss function, and the loss function MSE expression is as follows: In the formula, n represents the number of samples, and They are deformation prediction and measured results respectively.

4. The method for realizing the intelligent prediction model of deformation of adjacent existing sensitive structures according to claim 1, characterized in that: In step S4, the prediction process of the intelligent prediction model based on the long short-term memory network and the convolutional neural network includes: Step S41, the pre-processed construction information and ground-wall deflection information are respectively extracted with time features through two independent encoders based on long short-term memory networks; Step S42, the two extracted time features are spliced ​​in the channel dimension by feature fusion, and the construction information and the ground-wall deflection information are integrated into a unified feature; Step S43, input the result after feature fusion into the 1D U-net model, extract the spatial structure information of the features and decode them step by step, so as to generate the prediction result of the ground-wall deflection curve.

5. The method for realizing the intelligent prediction model of deformation of adjacent existing sensitive structures according to claim 4 is characterized in that: In step S41, the core of the long short-term memory network lies in its gating mechanism, including input gate I t 、Forget Gate F t and output gate O t , the input gate is used to determine the importance of the current input information, the forget gate is used to control the past memories that need to be forgotten, and the output gate is used to determine the hidden state at the current moment, that is, the final output of the network; at time step t, its calculation formula is as follows: I t =σ(W xi X t +W hi H t-1 +b i ); F t =σ(W xf X t +W hf H t-1 +b f ); O t =σ(W xo X t +W ho H t-1 +b o ); H t =O t ⊙tanh(C t ); Where W is the weight parameter and b is the bias term. t and C t are the hidden variables and memory elements at time step t, X t is the input at time step t, H t-1 and C t-1 is the hidden state and memory element at time step t-1, I t 、F t , O t and They are the input gate, forget gate, output gate and candidate memory element at time t, σ is the sigmoid activation function, and the symbol ⊙ is the Hadamard product.

6. The method for realizing the intelligent prediction model of deformation of adjacent existing sensitive structures according to claim 4, characterized in that: The step S43 comprises: Step S431, inputting the result of feature fusion into the encoder of the 1D U-net model to extract the spatial structure information of the features; Step S432, after feature extraction is completed, the fused features are sent to the upsampling decoder, which upsamples the low-resolution feature map to a high-resolution output, thereby generating a prediction result of the ground-wall deflection curve.

7. The method for realizing the intelligent prediction model of deformation of adjacent existing sensitive structures according to claim 6, characterized in that: In step S431, for the ground-connected wall deformation prediction task, the ground-wall deflection obtained at each time point is a one-dimensional sequence data. The convolution layer extracts local features from the deformation sequence by introducing a one-dimensional convolution, and identifies the deformation data in space and time. The calculation formula of the one-dimensional convolution is as follows: Where y[i] represents the i-th element of the output feature map, w[k] is the k-th element of the convolution kernel, x[i+k] is the i+k-th element of the input data, and K is the size of the convolution kernel; The pooling layer is usually set after the convolution layer to reduce the dimension of the feature map generated by the convolution. For a one-dimensional pooling operation, the pooling calculation formula is as follows: y[i]=pool(x[i·s],x[i·s+1],...,x[i·s+p-1]); Where y[i] represents the i-th element of the output feature map, pool represents the pooling operation, x is the input feature map, s is the step size, and p is the size of the pooling window; The fully connected layer is located at the back of the convolutional neural network and is used for the final decision or prediction layer. It maps the features extracted by the previous convolutional layer and pooling layer to the output space to achieve classification or regression tasks. Assume that the input feature vector is z = [z1, z2, ..., z n ], the output is a, then the operation of the fully connected layer can be expressed as: a=f(W·z+b); Where W is the weight matrix, b is the bias vector, and f is the activation function.

8. The method for realizing the intelligent prediction model of deformation of adjacent existing sensitive structures according to claim 7, characterized in that: In step S432, a transposed convolution layer is introduced, that is, an upsampling layer is used as a substitute for the fully connected layer to upsample the low-resolution coding features to a high-resolution output while maintaining the original spatial structure, wherein the calculation formula of the one-dimensional transposed convolution is as follows: Where y[i] represents the i-th element of the output feature map, x[ik·s] is the ik·s-th element of the input feature map after inserting zeros, w[k] is the k-th element of the convolution kernel, and s is the step size.

9. The method for realizing the intelligent prediction model of deformation of adjacent existing sensitive structures according to claim 4, characterized in that: In step S43, the prediction result can be returned to step S1 for the next prediction iteration to perform prediction of future time steps.

Images