A deformation prediction method for shield tunneling through buildings based on multiple factors

Abstract

The present invention discloses a deformation prediction method for shield construction passing through buildings based on multiple factors. A multi-factor matrix is ​​established by using settlement data, tilt data, the distance between the construction face and the building, shield construction parameter data, and geological information data of the construction area obtained at multiple times. The matrix is ​​used as the input of a prediction model to obtain the predicted value of the settlement data of the building, thereby realizing deformation prediction of shield construction passing through buildings based on multiple factors. The present invention combines CNN and LSTM models to not only correlate the characteristics of multiple factors, but also correlate the progress of shield construction with the deformation of the building through the factor of the distance between the shield construction and the building. These technologies can not only effectively reduce construction risks, but also improve overall construction efficiency and economic benefits, making settlement prediction more intelligent, dynamic, and refined. It can greatly improve the prediction accuracy and safety of the settlement of adjacent buildings during shield construction.

Description

Technical Field

[0001] The present invention relates to the technical field of building deformation prediction, and in particular to a deformation prediction method for shield construction passing through a building based on multiple factors. Background Art

[0002] Driven by urban modernization, subway tunnel projects continue to emerge in urban construction, marking the continuous improvement and modernization of urban transportation systems. Subway tunnels often need to pass through urban built-up areas, and shield tunneling is an important method for subway tunnel construction. The potential settlement of adjacent buildings and underground structures caused by tunnel construction is a very important consideration. During shield construction, as soil and rock formations are destroyed, significant settlement occurs, causing structural damage to buildings and raising safety issues. Therefore, studying the settlement and deformation of adjacent buildings during shield construction is of great significance, aiming to develop effective control and prediction measures to ensure the safety and sustainability of engineering construction.

[0003] However, the existing statistical model prediction method is unable to accurately predict the deformation of related adjacent buildings during shield construction. The current deformation prediction of adjacent buildings during shield construction has problems such as insufficient model accuracy, incomplete monitoring data, difficulty in coupling multiple factors, lack of real-time prediction capabilities and model adaptability. Summary of the Invention

[0004] The present invention provides a deformation prediction method for shield construction passing through buildings based on multiple factors to overcome the above technical problems.

[0005] In order to achieve the above object, the technical solution of the present invention is:

[0006] A deformation prediction method for shield construction passing through a building based on multiple factors includes the following steps:

[0007] S1: Set up an automated monitoring system during shield construction to obtain building settlement and tilt data during shield tunnel excavation; simultaneously obtain the shield construction location coordinates, shield construction parameter data, and geological information data of the construction area;

[0008] S2: Obtain the distance between the construction face and the building based on the location coordinates of the shield construction;

[0009] S3: Obtain a multi-factor matrix based on the settlement data, tilt data, distance between the construction face and the building, shield construction parameter data, and geological information data of the construction area obtained at multiple times;

[0010] S4: Construct a prediction model based on CNN network and LSTM unit for predicting the settlement data of the building, so as to obtain the predicted value of the settlement data according to the multi-factor matrix, and realize the deformation prediction of the shield construction passing through the building based on multiple factors.

[0011] Furthermore, the prediction model includes an optimization unit, a convolution unit, a sample division unit, an LSTM unit, a first judgment unit, a fully connected unit, a second judgment unit, a fitness value calculation unit and an output unit.

[0012] The convolution unit is used to obtain a one-dimensional spatiotemporal feature sequence of factors in the multi-factor matrix according to the multi-factor matrix;

[0013] The input end of the sample division unit is connected to the output end of the convolution unit, and is used to divide the one-dimensional spatiotemporal feature sequence into a training set and a test set;

[0014] The input end of the LSTM unit is connected to the output end of the sample division unit, and is used to train the LSTM unit according to the training set and the test set, and obtain a predicted spatiotemporal feature value sequence based on the trained LSTM unit;

[0015] The input end of the fitness value calculation unit is connected to the output end of the LSTM unit, and is used to obtain the root mean square error value between the predicted spatiotemporal feature value sequence of the current training step and the predicted spatiotemporal feature value sequence of the previous training step;

[0016] The input end of the first judgment unit is connected to the output end of the fitness value calculation unit, and is used to judge whether the root mean square error value is less than a set error threshold:

[0017] The output end of the first judgment unit is connected to the input end of the LSTM unit and the input end of the fully connected unit respectively, and is used to retrain the LSTM unit according to the training set and the test set when the root mean square error value between the predicted spatiotemporal feature value sequence of the current training step and the predicted spatiotemporal feature value sequence of the previous training step is not less than a set error threshold, and obtain a new predicted spatiotemporal feature value sequence based on the retrained LSTM unit; and when the root mean square error value between the predicted spatiotemporal feature value sequence of the current training step and the predicted spatiotemporal feature value sequence of the previous training step is less than the set error threshold, obtain a predicted time series matrix through the fully connected unit;

[0018] The input end of the fully connected unit is connected to the output end of the first judgment unit;

[0019] The input end of the second judgment unit is connected to the output end of the fully connected unit, and is used to obtain the predicted value of the building settlement data according to the predicted time series matrix.

[0020] The output end of the second judgment unit is connected to the input end and the output unit of the optimization unit respectively, and is used to update the size of the convolution kernel of the convolution unit and the number of neurons in the LSTM unit through the optimization unit when the difference between the predicted value of the building settlement data and the measured value of the building settlement data is not less than the set prediction error threshold; and output the predicted value of the building settlement data through the output unit when the difference between the predicted value of the building settlement data and the measured value of the building settlement data is less than the set prediction error threshold,

[0021] The output end of the optimization unit is connected to the input end of the convolution unit and the LSTM unit respectively; so as to transmit the updated convolution kernel size of the convolution unit and the number of neurons in the updated LSTM unit to the convolution unit and the LSTM unit respectively.

[0022] Furthermore, the geological information data includes groundwater level and rock formation level parameters; the rock formation level parameters include surrounding rock level and stratum thickness corresponding to the surrounding rock level.

[0023] Furthermore, the shield construction parameter data includes soil bin pressure, excavation speed and propulsion force.

[0024] Furthermore, the distance between the construction face and the building is obtained by the following formula:

[0025]

[0026] Where: S is the distance between the shield construction location and the building location; x represents the horizontal coordinate of the shield construction location; y represents the vertical coordinate of the shield construction location.

[0027] Beneficial effects: The present invention provides a deformation prediction method for shield construction passing through buildings based on multiple factors. A multi-factor matrix is ​​established through settlement data, tilt data, the distance between the construction face and the building, shield construction parameter data and geological information data of the construction area obtained at multiple times. The multi-factor matrix is ​​used as the input of the prediction model to obtain the predicted value of the settlement data of the building, thereby realizing deformation prediction of shield construction passing through buildings based on multiple factors. By real-time monitoring and feedback of relevant data of buildings during shield tunnel excavation, the prediction accuracy and safety of settlement of adjacent buildings during shield construction are greatly improved. Through the combination of CNN and LSTM models, not only can the multi-factor features be correlated, but the progress of shield construction can be correlated with the deformation of the building through the factor of the distance between the shield construction and the building. These technologies can not only effectively reduce construction risks, but also improve overall construction efficiency and economic benefits, making settlement prediction more intelligent, dynamic and refined. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following is a brief introduction to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0029] Figure 1 is a flow chart of the deformation prediction method of the present invention;

[0030] Figure 2 Schematic diagram of the flow chart of the prediction model training method in an embodiment of the present invention;

[0031] Figure 3 Schematic diagram of the prediction method flow of the prediction model in an embodiment of the present invention;

[0032] Figure 4 Schematic diagram of the prediction model structure in an embodiment of the present invention;

[0033] Figure 5 Schematic diagram of the distance between the shield construction face and the building in an embodiment of the present invention;

[0034] Figure 6 This is a schematic diagram of the structure of the LSTM neural network model and the CNN convolutional neural network after fusion in an embodiment of the present invention;

[0035] Figure 7 Schematic diagram of settlement data prediction results in an embodiment of the present invention;

[0036] Figure 8 Schematic diagram of prediction result error in an embodiment of the present invention;

[0037] Figure 9 4 is a flow chart of a deformation prediction method in an embodiment of the present invention. DETAILED DESCRIPTION

[0038] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0039] This embodiment provides a deformation prediction method for shield construction passing through a building based on multiple factors, including the following steps: Figure 1 and Figure 9 As shown:

[0040] S1: Set up an automated monitoring system for nearby buildings and the ground during shield construction to obtain settlement and tilt data of nearby buildings during shield tunnel excavation; and simultaneously obtain the shield construction location coordinates, shield construction parameter data, and geological information data of the construction area;

[0041] The shield construction parameter data includes soil bin pressure, excavation speed and propulsion force;

[0042] The geological information data includes groundwater level and rock formation level parameters; the rock formation level parameters include surrounding rock level and formation thickness corresponding to the surrounding rock level;

[0043] The geological information data includes the surrounding rock grade, the stratum thickness corresponding to the surrounding rock grade, and the groundwater level, wherein the stratum grade corresponds to the classification of the surrounding rock, including grades 1 to 5. The groundwater level and the surrounding rock grade together constitute the geological information data;

[0044] The rock layer level parameter in this embodiment is obtained based on the stratum distribution information within the selected construction area obtained from geological surveys, and the types of surrounding rock layers within the construction area are counted. The rock layers within the tunnel construction area and the overlying strata are classified according to the construction specifications, and the classification is as follows:

[0045] Grade I (Excellent): The rock is intact, compact in structure, has no obvious joints, has high shear strength and good stability.

[0046] Level II (good): The rock is basically intact, with a tight structure, a small number of small joints and good stability.

[0047] Level III (medium): The rock is relatively complete, with average structure, many small joints and average stability.

[0048] Level IV (poor): The rock is broken, the structure is loose, the joints are developed, and the stability is poor.

[0049] Grade V (poor): The rock is severely broken, the structure is loose, landslides are likely to occur, and the stability is extremely poor.

[0050] During tunnel construction, the thickness F of the rock formations within the tunnel face and the overlying strata is collected at set intervals. The surrounding rock formation type E and thickness F are recorded and arranged into 2×5 vector groups based on the stratum distribution information. If a certain grade of surrounding rock is absent, its thickness is set to 0. This ultimately forms structured geological information data (E, F).

[0051] The automated monitoring system includes: several automated monitoring and collection boxes, several box-type fixed inclinometers, several static levels, a transmitting box, a remote server, and a terminal; the box-type fixed inclinometers are fixedly mounted on the parapet of the building roof; the static levels are respectively fixedly mounted on the parapet of the building roof; the box-type fixed inclinometers and the static levels are both electrically connected to the automated monitoring and collection box; the automated monitoring and collection box is connected to the transmitting box; the transmitting box is in communication with the remote server; and the remote server is in communication with the terminal.

[0052] Specifically, an automated monitoring system is set up at different locations of the same building that the shield construction passes through. The monitoring plan includes monitoring the settlement, deformation and inclination of the building, obtaining the settlement data and inclination data of the building. At the same time, the underground shield machine collects geological information data during the excavation process and records shield construction parameter data, including soil bin pressure, excavation speed and propulsion force.

[0053] In this embodiment, the settlement and tilt data of the building, as well as the soil bin pressure, tunneling speed, and thrust, acquired by the automated deformation monitoring system, are collected at set time intervals. The distance between the construction face and the building is calculated based on the shield construction coordinates. The geological information data collected for the construction area at different time periods is structured into data. This data includes rock formation level parameters and groundwater level. The rock formation level parameters are the surrounding rock grades one to five and the assigned stratum thickness corresponding to the surrounding rock grade. The groundwater level and rock formation level parameters together constitute the geological information data.

[0054] S2: Obtain the distance between the construction face and the building based on the location coordinates of the shield construction;

[0055]

[0056] Where: S is the distance between the shield construction location and the building location; x represents the horizontal coordinate of the shield construction location; y represents the vertical coordinate of the shield construction location;

[0057] Specifically, in this embodiment, a two-dimensional coordinate system is established on the horizontal plane where the monitoring point is located, wherein the center point of the horizontal plane where the building monitoring point is located is selected as the origin, a wall of the building is selected as a reference, the X-axis is parallel to the wall, and the Y-axis is perpendicular to the wall, and the shield machine position information is collected based on the monitoring time node.

[0058] S3: obtaining a multi-factor matrix based on the settlement data, tilt data, shield construction parameter data, and geological information data of the construction area obtained at multiple times;

[0059] Specifically, we selected m types of data from the time the shield tunnel entered the building's impact zone until it completely passed through the building. These data included settlement data at various points in the building, tilt data, geological information, the distance between the construction face and the building, shield machine soil chamber pressure, tunneling speed, and thrust. Each data type contained n values ​​over time, forming a multi-factor time series. These multiple time series were then combined to form a multi-factor matrix A' of size m × n.

[0060] Specifically, in this embodiment, the shield construction face entering the area with a radius of 10m near the building is considered to have entered the affected area. The distance between the construction face and the building is calculated as follows: Figure 5 As shown, a rectangular coordinate system with the building center as the origin is established to calculate the coordinate position of the shield machine tunnel face and obtain the distance between the construction face and the building. By considering the shield machine's construction position, it helps improve the spatial accuracy of the prediction, dynamically respond to construction changes, capture local features, and better simulate the actual construction environment. This makes settlement prediction more timely and accurate, providing strong support and risk management capabilities for engineering construction.

[0061] S4: Constructing a prediction model for predicting the settlement data of a building, so as to obtain the predicted value of the settlement data according to the multi-factor matrix, and realize deformation prediction of shield construction passing through a building based on multiple factors.

[0062] Preferably, the prediction model includes an optimization unit, a convolution unit, a sample division unit, an LSTM unit, a first judgment unit, a fully connected unit, a second judgment unit, a fitness value calculation unit and an output unit.

[0063] The input end of the convolution unit is connected to the output end of the sorting unit and the optimization unit respectively, and is used to obtain a one-dimensional spatiotemporal feature sequence of factors in the multi-factor matrix according to the convolution kernel size output by the optimization unit and the multi-factor matrix;

[0064] Specifically, the convolution unit determines the convolution unit parameters by using the optimal convolution kernel size output by the optimization unit, and then performs a convolution operation on the multi-factor matrix to obtain a one-dimensional spatiotemporal feature sequence between the factors.

[0065] The input end of the sample division unit is connected to the output end of the convolution unit, and is used to divide the one-dimensional spatiotemporal feature sequence into a training set and a test set;

[0066] The input end of the LSTM unit is connected to the output end of the sample division unit, and is used to train the LSTM unit according to the training set and the test set, and obtain a predicted spatiotemporal feature value sequence based on the trained LSTM unit;

[0067] Specifically, the LSTM unit trains and updates the prediction model based on the divided samples of the input one-dimensional spatiotemporal feature sequence through the divided training set and test set, and predicts the feature value results to predict the spatiotemporal feature value sequence.

[0068] The input end of the fitness value calculation unit is connected to the output end of the LSTM unit, and is used to obtain the root mean square error value between the predicted spatiotemporal feature value sequence of the current training step and the predicted spatiotemporal feature value sequence of the previous training step;

[0069] The input end of the first judgment unit is connected to the output end of the fitness value calculation unit, and is used to judge whether the root mean square error value is less than a set error threshold:

[0070] The output end of the first judgment unit is connected to the input end of the LSTM unit and the input end of the fully connected unit respectively, and is used to retrain the LSTM unit according to the training set and the test set when the root mean square error value between the predicted spatiotemporal feature value sequence of the current training step and the predicted spatiotemporal feature value sequence of the previous training step is not less than the set error threshold, and obtain a new predicted spatiotemporal feature value sequence based on the retrained LSTM unit; and when the root mean square error value between the predicted spatiotemporal feature value sequence of the current training step and the predicted spatiotemporal feature value sequence of the previous training step is less than the set error threshold, obtain the predicted time series matrix B through the fully connected unit;

[0071] The input end of the fully connected unit is connected to the output end of the first judgment unit; it is used to expand the one-dimensional predicted spatiotemporal eigenvalue sequence to obtain the predicted time series matrix B, and restore the one-dimensional sequence to a multidimensional time series.

[0072] The input end of the second judgment unit is connected to the output end of the fully connected unit, and is used to obtain the predicted value of the building settlement data according to the predicted time series matrix B.

[0073] The output end of the second judgment unit is connected to the input end and the output unit of the optimization unit respectively, and is used to update the size of the convolution kernel of the convolution unit and the number of neurons in the LSTM unit through the optimization unit when the difference between the predicted value of the building settlement data and the measured value of the building settlement data is not less than the set prediction error threshold; and output the predicted value of the building settlement data through the output unit when the difference between the predicted value of the building settlement data and the measured value of the building settlement data is less than the set prediction error threshold,

[0074] The output end of the optimization unit is connected to the input end of the convolution unit and the LSTM unit respectively; so as to transmit the updated convolution kernel size of the convolution unit and the number of neurons in the updated LSTM unit to the convolution unit and the LSTM unit respectively.

[0075] Specifically, such as Figure 2-Figure 4 As shown, this embodiment uses the convolution kernel size and the number of hidden layer neurons in the prediction model as the population individuals of the optimization algorithm in the optimization unit, the multi-factor time series input matrix, and sets the initial population individuals. As shown, the multi-factor matrix A' is input into the prediction model, and the one-dimensional spatiotemporal feature sequence 1×a is obtained through convolution operation. The one-dimensional spatiotemporal feature sequence obtained by convolution is input into the hidden layer for training and time series prediction. Where a represents the spatiotemporal feature sequence extracted from multiple factors;

[0076] Specifically, in the hidden layer, the input one-dimensional eigenvalue time series 1×a is divided into training and test sample sets based on the sample partitioning unit in the previous step. By updating the weights controlled by the gate structure, a time series prediction model is obtained after training and learning, resulting in a one-dimensional predicted spatiotemporal eigenvalue sequence 1×b. Here, b represents the predicted spatiotemporal eigenvalue sequence. When the predicted spatiotemporal eigenvalue sequence meets the fitness function value setting conditions, that is, the root mean square error between the predicted spatiotemporal eigenvalue sequence of the current training step and the predicted spatiotemporal eigenvalue sequence of the previous training step is no less than the set error threshold, the LSTM is retrained. Otherwise, the system enters the fully connected layer and expands the obtained one-dimensional predicted spatiotemporal eigenvalue sequence 1×b through fully connected layer operations to obtain the predicted time series matrix B. The optimization unit then calculates the fitness value of each individual in the population, updates the position of the individuals in the population, and selects the individuals with the best fitness value as the next generation of the population.

[0077] Finally, the prediction model outputs k predicted time series data in the prediction time series matrix B, where the size of the output matrix B is m×k, and k is the number of time points for prediction. That is, if the settlement data and tilt data of the building at k time points need to be predicted, the size of the output prediction time series matrix B is m×k, where m is the number of factors in the multi-factor matrix.

[0078] Specifically, since the monitoring data will continuously expand with the construction progress, the types of factors that can be obtained will change as the construction process progresses, and the resulting time series will also increase. The corresponding change in the model is that m and n in the size m×n of the input matrix A will change. To ensure that the prediction model is not affected by the size of the multi-factor matrix A' input, the optimization unit automatically updates the number of neurons and the convolutional kernel size in the prediction algorithm, and finds the optimal values of the number of neurons and the convolutional kernel size to adapt to input matrices of different sizes. After the initially set number of neurons and convolutional kernel size are input into the optimization unit, the optimal values are found based on the size of the input matrix of the entire prediction model, and then the optimal values of the number of neurons and the convolutional kernel size are output.

[0079] Specifically, in the optimization process of the optimization unit in this embodiment, the positions of the population individuals are initialized. The individuals in the population are divided into three categories. One category is the explorers, who are responsible for finding food and guiding the population to areas rich in food. Another category is the followers, who will constantly observe the producers during the foraging process. Once they find that the producers have found better food, they will immediately abandon their current food and turn to the producers. In addition, there is a certain proportion of guardians in the population. When there is no danger, they move around randomly in the population, and once they detect danger, they immediately move to a safe area. During the iteration process, they update their position information according to certain rules. The positions of the explorers are randomly generated within the allowable range of the number of neurons and the convolutional kernel size. Calculate the fitness for the position of each individual (i.e., each combination of the number of neurons and the convolutional kernel size). During each iteration process, the explorers are responsible for finding new potential optimal solutions in the search space. The position update method of the explorers is as follows:

[0080]

[0081] In the formula: is the position of the i-th explorer in the j-th dimension parameter in the t-th iteration; iter max is the maximum number of iterations, α is a random number within [0,1]; R2 is the early warning value, where R2 ∈ [0,1]; ST is the safety value, where ST ∈ [0.5,1]; Q is a random number obeying the normal distribution; L is a 1×q matrix, where each element is 1, where q is a positive integer related to the parameter dimension of the explorer; i represents the index number of the explorer; j represents the parameter dimension of the number of neurons / convolutional kernel size; t represents the number of iteration times. When R2 < ST, this means there are no natural enemies around, and the explorers can conduct global searches. If R2 ≥ ST, it means that some individuals have discovered predators, and all individuals have to take relevant actions. During the foraging process, some followers will constantly monitor the explorers. Once the explorers find better food, they will immediately leave their current positions to compete for food. The position update method of the followers is as follows:

[0082]

[0083] Where: is the position of the optimal explorer in the tth iteration; is the current global worst position; represents the position of the j-th dimension follower at the t-th iteration; i2 represents the index number of the follower; n2 is the population size. A is a 1×g matrix, where g represents a positive integer equal to the dimension of the optimal explorer parameter, and each element has a random amplitude of 1 or -1. Here, A+ is defined as follows:

[0084] A + =A T (AA T ) -1

[0085] When i2>n / 2, it indicates that the i2th follower with a lower fitness value is in poor condition and needs to fly to other places to find food.

[0086] Specifically, assuming that some individuals in the population are aware of the danger, the initial positions of these individuals (alerts) are randomly generated in the population:

[0087]

[0088] Where: is the current global optimal position, β is the step size control parameter, and its value is a random number that obeys the normal distribution with a mean of 0 and a variance of 1; f i is the fitness value, f g 、f w are the current optimal and worst fitness values ​​respectively; ε is a constant to avoid the denominator being 0. i >f g When f i =f g When , this indicates that the individuals in the middle of the population are aware of the danger and need to move closer to other individuals to avoid being preyed upon. p is a random number in the range [-1, 1], which represents the direction of individual movement and is also the step size control parameter.

[0089] Specifically, since the long-term and short-term prediction unit can only be trained on one-dimensional time series, in order to ensure that long-term and short-term prediction can be performed on multi-dimensional time series, the convolution unit of this embodiment extracts features between multiple factors through the CNN convolution layer of the convolution unit, compresses the input multi-dimensional factor sequence into a one-dimensional sequence, and outputs the one-dimensional sequence. The formula for calculating the size of the extracted eigenvalue matrix is ​​as follows:

[0090] w`=l(w+2o-s)+1

[0091] Where w is the size of the input multi-factor matrix, which is an m*n matrix; s is the convolution kernel size; l is the stride length; and o is the number of zero-padding layers. The convolution kernel has the same width as w pixels in the horizontal direction, and the extracted eigenvectors are one-dimensional. w' represents the size of the extracted eigenvalue matrix.

[0092] Furthermore, the one-dimensional sequence output by the convolution unit is input to the LSTM unit. The LSTM unit includes three structures: memory gate, forget gate, and output gate. The number of neurons is calculated by the following formula:

[0093] R=k1*(m1+n1)

[0094] Where k1 is a constant that usually depends on the specific task and network structure; m1 is the dimension of the input features; n1 is the number of output categories or the dimension of the target value. R is the number of neurons;

[0095] Specifically, in this embodiment, in the LSTM unit, during the time series prediction process, the sample division unit constructs training samples from the feature value sequence and performs training based on the samples. The specific process is as follows:

[0096] Input one-dimensional spatiotemporal feature sequence 1×a=(x1,x2,x3…x a ), construct the input dataset D of the training sample and the output dataset E of the training sample through the time sliding window according to the input feature sequence 1×a, the input dataset D=[(x1,x2…x d ),(x2,x3…x d+1 )…(x a-d-1 ,x a-d …x a-1 )], output data set E = (x d+1 ,x d+2 …x a ); d is the time sliding window; x a represents the ath feature in the extracted feature sequence; x d Indicates the dth feature extracted. During the training process, the prediction data set C = (c d+1 ,c d+2 …c a ),c a Represents the ath feature of the prediction; calculate the error of the predicted data through the dataset E and the predicted data; perform reverse training based on the error to update the model parameters and obtain the optimal prediction model.

[0097] After learning and training, it is necessary to judge whether the trained model meets the requirements, and the mean square error function value is used for judgment.

[0098] The first judgment unit selects the mean square error function value. The formula of the mean square error function is: where x f Represents the predicted feature value, x t represents the actual eigenvalue, and n3 represents the n3th set of predicted and actual eigenvalues. The MSE satisfies the condition that it is less than a set threshold. Ultimately, the optimal parameters of multiple prediction models need to be selected, and the root mean square error function is used to determine whether the optimal parameters have been achieved.

[0099] The second judgment unit uses a root mean square error function. The formula of the root mean square error function is: where y f Indicates the predicted deformation value of the building, y t represents the monitored building deformation value; n4 represents the n4th set of predicted and actual feature values. The condition is that the RMSE is less than a set threshold.

[0100] Specifically, this embodiment also discloses a system for predicting deformation of a building during shield construction based on multiple factors, including a data acquisition unit, a data sorting unit, and a prediction unit.

[0101] The data acquisition unit's hardware components include sensors: a static level, an inclinometer, a wireless module, a data terminal, and a storage device. The unit collects building deformation data and shield construction parameters at intervals. Deformation data includes settlement and tilt deformation, while shield construction parameters include tunneling speed, soil bin pressure, and thrust.

[0102] The data collation unit uses the data samples obtained by the acquisition unit to transform the geological information collected at different time periods into structured data. It selects m types of data from the time the shield tunneling machine enters the building's impact zone until it completely passes through the building. These data include settlement data and tilt data at various points in the building, structured geological information data, shield machine soil chamber pressure, and thrust force. Each data type contains n data values. These multiple time series are combined to form a multi-factor matrix A' with a size of m × n.

[0103] The prediction unit is used to obtain the predicted value of the settlement data according to the multi-factor matrix A'.

[0104] A specific embodiment of the present invention is as follows:

[0105] In this example, a nearby building in the Xinleng section of the Nanchang Metro Section 02 East Extension Project is selected. The SSA-CNN-LSTM deformation prediction method for the nearby building, which takes into account the position of the shield machine and temporal and spatial factors, is used to predict the settlement value of the building monitoring point. Figure 6 shown.

[0106] (1) Collect deformation data of each monitoring point of the building and the position information of the shield machine during shield construction to obtain the original data. From October 20, 2023 to October 23, 2023, the monitoring data of the building settlement monitoring points CJ2-1, CJ2-2, CJ2-3, CJ2-5, and CJ2-6 (corresponding to settlement variables: J1, J2, J3, J4, and J5) and the tilt monitoring points CX2-1, CX2-2, CX2-3, CX2-5, and CX2-6 (corresponding to tilt variables: Q1, Q2, Q3, Q4, and Q5) are collected every hour, as well as the position relationship data between the center point of the shield construction face and the building, and the geological information structured data, a total of 12 groups, each with 150 data.

[0107] (2) The collected raw data were processed to remove noise, and the pre-processed data were integrated into a multivariate time series. The settlement value of the CJ2-1 monitoring point, namely the settlement variable J1, was selected as the prediction target and set as the dependent variable. The other variables included the settlement data of each monitoring point (four variables), the tilt data (five variables) and the positional relationship data of the center point of the shield construction face and the building (two variables). These 11 variables were used as independent variables.

[0108] (3) In order to demonstrate the superiority of the model, the first 90 data of the multivariate time series were selected as input, and the settlement values ​​of the last 60 CJ2-1 monitoring points were compared with the predicted values ​​of the model.

[0109] (4) Input 90 multivariate time series into the model for learning and training. The training process includes:

[0110] The time series is input into the convolution unit in the model. The convolution layer of the convolution unit extracts features from the time series of 11 independent variables and 1 dependent variable. The individual initial convolution kernel size is set to 12×1, and the extracted feature value sequence is 90×1. That is, for the data collected for 900 hours, the feature values ​​of the shield machine position and other spatial position deformations and the target position deformation are extracted every hour, integrating the construction progress, spatial correlation and time series dynamic information.

[0111] The extracted features are input into the LSTM unit in the prediction model. The LSTM unit performs rolling training on the extracted features and performs multi-step prediction on the feature values.

[0112] Eigenvalue sequence A=(x1,x2,x3…x 90 ), set the time sliding window to 3, and construct the input data set D of the training sample and the output data set E of the training sample through the time sliding window. The input data set D = [(x1, x2, x3), (x2, x3, x4)… (x 87 ,x 88 ,x89 )], output data set E=(x4,x5…x 90 );. Through the multiple training processes of the constructed sample set, the prediction data set C=(c4,c5…c 90 ), train the prediction model, perform reverse training based on the error between the test set and the prediction set to update the model parameters and obtain the optimal prediction model.

[0113] The optimization unit is used to update the weights and hyperparameters in the neural network to improve the performance and accuracy of the model. The root mean square error function is selected to judge whether the output prediction result meets the standard. If it meets the standard, the deformation value of the prediction target is output by connecting the fully connected layer. Otherwise, the parameters are optimized and retrained.

[0114] Initial parameter settings of the parameter optimization algorithm: the initial value of the population size is set to 100; the maximum number of iterations is set to 500 iterations; the dimension is 2, the same as the number of optimization parameters (number of neurons and convolution kernel size); the search range, the number of neurons range is [30,200], the convolution kernel size, the range may be [9x1,9x9]; the safety factor controls the escape behavior of individuals when discovering predators, set to 0.8; the proportion of discoverers determines the proportion of individuals in the population that serve as discoverers, set to 0.2; the ratio of joiners to discoverers determines the ratio of joiners (individuals following the discoverer) to discoverers, set to 0.5; the step size affects the movement step of individuals in the search space, the step size is 10.

[0115] Calculate the root mean square error (RMSE) of each individual's position. Select the individual with the smallest RMSSE as the leader, and the other individuals move closer to the leader, updating their positions. Repeat the above steps until convergence reaches a sufficiently small error threshold. The RMSSE threshold is set to 10.

[0116] After the training optimization is completed and meets the requirements, the settlement prediction value and optimal parameters of the monitoring point CJ2-1 for the 91-150 hour period are output.

[0117] The specific prediction results are as follows Figure 7 and Figure 8 As shown in Figure 2, the advantages of using a fusion neural network for settlement prediction lie in its powerful nonlinear processing capabilities, multi-source data fusion, automated feature extraction, and adaptive learning capabilities. These features give the neural network greater prediction accuracy and robustness under complex engineering conditions, enabling it to effectively address building settlement issues in complex environments.

[0118] The present embodiment provides a method for predicting deformation of shield tunneling through buildings based on multiple factors. A multi-factor matrix is ​​established by using settlement data, tilt data, the distance between the construction face and the building, shield construction parameter data, and geological information data of the construction area obtained at multiple times. This matrix is ​​used as the input of a refined prediction model to obtain the predicted value of the building's settlement data. By fusion analysis of multi-source data, deformation prediction of shield tunneling through buildings based on multiple factors is achieved. By real-time monitoring and feedback of relevant data of buildings during shield tunnel excavation, the prediction accuracy and safety of settlement of adjacent buildings during shield construction are greatly improved. By combining CNN and LSTM models, not only can the multi-factor features be correlated, but the progress of shield construction can be correlated with building deformation through the factor of the distance between shield construction and buildings. These technologies can not only effectively reduce construction risks, but also improve overall construction efficiency and economic benefits, making settlement prediction more intelligent, dynamic, and refined.

[0119] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A deformation prediction method for shield tunneling through buildings based on multiple factors, characterized by: The following steps are involved: S1: Set up an automated monitoring system during shield construction to obtain building settlement and tilt data during shield tunnel excavation; simultaneously obtain the shield construction location coordinates, shield construction parameter data, and geological information data of the construction area; S2: Obtain the distance between the construction face and the building based on the location coordinates of the shield construction; S3: Obtain a multi-factor matrix based on the settlement data, tilt data, distance between the construction face and the building, shield construction parameter data, and geological information data of the construction area obtained at multiple times; S4: Constructing a prediction model based on a CNN network and LSTM units for predicting the settlement data of a building, so as to obtain a predicted value of the settlement data according to the multi-factor matrix, and realize deformation prediction of a shield tunneling through a building based on multiple factors; The prediction model includes an optimization unit, a convolution unit, a sample division unit, an LSTM unit, a first judgment unit, a fully connected unit, a second judgment unit, a fitness value calculation unit and an output unit; The convolution unit is used to obtain a one-dimensional spatiotemporal feature sequence of factors in the multi-factor matrix according to the multi-factor matrix; The input end of the sample division unit is connected to the output end of the convolution unit, and is used to divide the one-dimensional spatiotemporal feature sequence into a training set and a test set; The input end of the LSTM unit is connected to the output end of the sample division unit, and is used to train the LSTM unit according to the training set and the test set, and obtain a predicted spatiotemporal feature value sequence based on the trained LSTM unit; The input end of the fitness value calculation unit is connected to the output end of the LSTM unit, and is used to obtain the root mean square error value between the predicted spatiotemporal feature value sequence of the current training step and the predicted spatiotemporal feature value sequence of the previous training step; The input end of the first judgment unit is connected to the output end of the fitness value calculation unit, and is used to judge whether the root mean square error value is less than a set error threshold: The output end of the first judgment unit is connected to the input end of the LSTM unit and the input end of the fully connected unit respectively, and is used to retrain the LSTM unit according to the training set and the test set when the root mean square error value between the predicted spatiotemporal feature value sequence of the current training step and the predicted spatiotemporal feature value sequence of the previous training step is not less than a set error threshold, and obtain a new predicted spatiotemporal feature value sequence based on the retrained LSTM unit; and when the root mean square error value between the predicted spatiotemporal feature value sequence of the current training step and the predicted spatiotemporal feature value sequence of the previous training step is less than the set error threshold, obtain a predicted time series matrix through the fully connected unit; The input end of the fully connected unit is connected to the output end of the first judgment unit; The input end of the second judgment unit is connected to the output end of the fully connected unit, and is used to obtain the predicted value of the building settlement data according to the predicted time series matrix. The output end of the second judgment unit is connected to the input end and the output unit of the optimization unit respectively, and is used to update the size of the convolution kernel of the convolution unit and the number of neurons in the LSTM unit through the optimization unit when the difference between the predicted value of the building settlement data and the measured value of the building settlement data is not less than a set prediction error threshold; When the difference between the predicted value of the building settlement data and the measured value of the building settlement data is less than a set prediction error threshold, the predicted value of the building settlement data is output through the output unit, The output end of the optimization unit is connected to the input end of the convolution unit and the input end of the LSTM unit respectively; The updated convolution kernel size of the convolution unit and the updated number of neurons in the LSTM unit are transmitted to the convolution unit and the LSTM unit respectively.

2. The method for predicting deformation of a building during shield construction based on multiple factors according to claim 1, characterized in that: The geological information data includes groundwater level and rock formation level parameters; the rock formation level parameters include surrounding rock level and formation thickness corresponding to the surrounding rock level.

3. The method for predicting deformation of a building during shield construction based on multiple factors according to claim 1, characterized in that: The shield construction parameter data includes soil bin pressure, excavation speed and propulsion force.

4. The method for predicting deformation of a building during shield construction based on multiple factors according to claim 1, characterized in that: The formula for obtaining the distance between the construction face and the building is as follows: Where: The distance between the shield construction location and the building location; x The horizontal coordinate indicating the shield construction position; y The vertical coordinate representing the shield construction position.

Images