Settlement time sequence prediction method and system for deep foundation pit excavation adjacent building

Abstract

The invention discloses a settlement amount time sequence prediction method and system for deep foundation pit excavation adjacent buildings. The method comprises the steps that original monitoring data of on-site building settlement are acquired; performing data preprocessing on the obtained original monitoring data; constructing a recurrent neural network; model input and output parameters are determined through principal component analysis; optimizing the recurrent neural network based on an optimizer; performing hyper-parameter optimization based on an optimization result; performing settlement time sequence prediction by using the optimized recurrent neural network; the system comprises a data acquisition module, a preprocessing module, a model construction module, an analysis module, an optimization module, a parameter optimization module and a prediction module. By constructing the settlement prediction model, dynamic modeling and accurate prediction of the settlement trend of the building in the excavation process of each stage of the foundation pit are realized; historical settlement monitoring data and multi-layer soil body excavation information are combined, and multi-source input parameters are introduced, so that the adaptability of the model to complex working conditions is enhanced.

Description

Technical Field

[0001] This invention relates to the field of settlement time series prediction technology, and in particular to a method and system for settlement time series prediction of adjacent buildings during deep foundation pit excavation. Background Technology

[0002] With the acceleration of urbanization, the demand for underground space development is increasing, and foundation pit engineering is gradually developing towards greater depths, more complex geological conditions, and denser surrounding environments. However, during the excavation of foundation pits, factors such as the release of soil stress can easily lead to uneven settlement of adjacent buildings, which may even cause structural cracking or collapse in severe cases. Therefore, how to accurately predict the dynamic settlement trend of buildings during construction has become a key technical challenge to ensure project safety and the stability of the surrounding environment.

[0003] Traditional settlement prediction methods mainly rely on field monitoring, empirical formulas, and numerical simulations. However, these methods have significant limitations and are insufficient to meet the high-precision and time-series requirements of engineering projects. First, while field monitoring can directly acquire settlement data during construction, it has a certain lag, making it difficult to provide timely warnings and effective decision support. Second, empirical formula methods derive empirical relationships based on historical engineering data and geological conditions; however, due to insufficient consideration of complex geological conditions and construction techniques, they fail to accurately reflect the nonlinear settlement characteristics under complex construction environments, resulting in low prediction accuracy. Third, while numerical simulation methods can simulate the stratum deformation process in relatively detailed terms, they depend on the precise setting of soil constitutive model parameters (such as elastic modulus and Poisson's ratio). In actual engineering projects, soil types exhibit high spatial heterogeneity and complexity, and errors in parameter values ​​can easily lead to distorted prediction results. In addition, the settlement process of adjacent buildings caused by foundation pit excavation is extremely complex. It is affected by multiple factors such as soil type, groundwater changes, soil excavation depth, support structure type and building self-weight. Traditional methods often rely on manual experience to adjust the model, which is difficult to adapt to the ever-changing construction environment and thus cannot accurately reflect the temporal characteristics of building settlement.

[0004] In recent years, with the development of artificial intelligence methods, given their excellent performance in handling complex data and dynamic environmental changes, they have been increasingly applied in settlement prediction, such as Support Vector Machine (SVM) and Random Forest algorithms. While these methods can train models using historical data, they cannot describe the impact of time on the output target, ignoring the time dependence of settlement evolution and typically representing settlement as a single numerical value. In fact, settlement data is usually a continuous sequence that changes over time; that is, the settlement at one time moment has a significant impact on the settlement at the next time moment.

[0005] In response, Recurrent Neural Networks (RNNs), a deep learning model used in artificial intelligence, possesses greater potential due to its advantages in processing time series data. It can effectively capture the dynamic features and potential patterns within time series data. By introducing state information from hidden layers, RNNs allow the settlement data from the previous time step to influence the settlement prediction at the current time step, thereby more accurately capturing the trend of settlement changes over time and improving prediction accuracy.

[0006] However, current settlement prediction methods are limited to the settlement of the excavation area itself, with limited predictions of settlement of adjacent buildings caused by the excavation. Therefore, leveraging the powerful potential of RNNs in processing time-series data, developing a time-series prediction method for settlement of adjacent existing buildings during deep foundation pit excavation is of great significance for ensuring the safety of existing buildings, improving the accuracy of settlement prediction, and optimizing construction plans.

[0007] However, the key challenges in predicting the settlement of adjacent buildings over time are as follows: (1) The settlement data of buildings has obvious time-series characteristics. The settlement at the current time step is often affected by the settlement changes at the previous time step. During the settlement process, due to the existence of various nonlinear factors, the model needs to have good time-series modeling capabilities and be able to effectively capture the settlement trend from historical settlement data, so as to achieve accurate prediction of future settlement.

[0008] (2) Different soil layers have significant differences in physical properties and mechanical response, which leads to complex changes in the settlement behavior of buildings at different excavation stages. Therefore, the constructed model should not only be able to learn from the settlement data accumulated during the excavation of existing soil layers, but also have a certain transfer learning and dynamic adaptation capabilities in order to realize reasonable deduction and prediction of the impact on subsequent soil layer excavation.

[0009] (3) In actual engineering projects, the deployment of monitoring points adjacent to buildings is limited, and the collected time-series settlement data are often accompanied by high-frequency noise and missing values, which can easily lead to overfitting or underfitting of the model during training. Therefore, the RNN model needs to be combined with appropriate network structure design, hyperparameter configuration and optimization strategies to ensure that the model can maintain stable and reliable prediction performance in each excavation stage. Summary of the Invention

[0010] The purpose of this invention is to overcome the shortcomings of the existing technology and to propose a method and system for predicting the settlement of adjacent buildings during deep foundation pit excavation.

[0011] To achieve the above objectives, the present invention adopts the following technical solution: A method for time-series prediction of settlement of adjacent buildings during deep foundation pit excavation includes the following specific steps: S1: Obtain raw monitoring data on the settlement of buildings on site; S2: Perform data preprocessing on the acquired raw monitoring data; S3: Determine input and output parameters based on preprocessed monitoring data; S4: Construct a recurrent neural network (RNN) and divide it into training and test sets; S5: Select an appropriate optimization algorithm to optimize the recurrent neural network (RNN); S6: Perform hyperparameter optimization based on the optimization results; S7: Comparison and analysis of model prediction results with measured data and verification of evaluation indicators; S8: Use the optimized recurrent neural network to predict the time series of sedimentation.

[0012] As a further technical solution of the present invention, in S2, data preprocessing includes performing integrity checks on the original monitoring data, identifying and removing missing data and abnormal monitoring point data, converting the monitoring data into a normalized cumulative settlement-time curve, using the mean error CMSE and a preset threshold for consistency judgment, gradually filtering out data with learning value, and then using the three-times root mean square method to smooth the local outliers in the curve, effectively removing outliers.

[0013] As a further technical solution of the present invention, in S3, the main engineering and geological factors affecting building settlement are systematically analyzed, and soil parameters, groundwater level and excavation parameters are selected as input parameters of the model; the output parameter of the model is the cumulative settlement of the building, that is, the total settlement of the building since excavation is predicted at each time step; the soil parameters include but are not limited to soil weight, internal friction angle and cohesion; the excavation parameters include but are not limited to excavation location and excavation volume.

[0014] As a further technical solution of the present invention, in S4, based on the construction of a recurrent neural network (RNN), a long short-term memory network (LSTM) unit is introduced to enhance the ability to remember long-term dependencies, and a locally connected subnetwork is added to capture spatial local features, forming a hybrid neural network model that integrates RNN, LSTM and locally connected networks; the first 70% of the data is used for model training, so that the model can fully learn the historical subsidence change pattern, and the last 30% of the data is used as a test set to test the model's predictive ability on unseen data.

[0015] S41: Input layer: Receives preprocessed building settlement-related data, including soil parameters (soil weight). internal friction angle Cohesion ), groundwater level And excavation parameters (excavation location coordinates) Data such as excavation volume (V) are arranged and combined according to time series to form an input vector sequence. , where t represents the time step; S42: RNN layer: Consists of several RNN units, each of which receives the hidden state from the previous time step. and the input at the current moment Calculate the hidden state at the current time step. : ,in: For example, the sigmoid function; Let be the weight matrix input to the hidden layer, with dimension . Input size Hidden layer size The weight matrix from hidden layer to hidden layer has a dimension of . Hidden layer size Hidden layer size Here is the hidden layer bias vector, with dimension . Hidden layer size ; S43: LSTM Layer: An LSTM layer is added on top of the RNN layer. An LSTM unit contains a forget gate, an input gate, and an output gate. The forget gate determines how much past information needs to be retained, the input gate controls the degree to which current input information is written, and the output gate determines the amount of output information. The specific calculation is as follows: S431: The Gate of Oblivion : ,in: It is the weight matrix of the forget gate; It is the bias vector of the forget gate; This is the sigmoid function, with an output range between 0 and 1, representing the degree of forgetting; S432: Input Gate : It is the weight matrix of the input gate; It is the bias vector of the input gate, used to control the degree of update of the input information; S433: Candidate memory cell status : ,in: It is the weight matrix of the candidate memory cell states. It is a bias vector. The function compresses the input to a range between -1 and 1; S434: Memory cell status update: ,in: It is the state of memory cells from the previous moment; It is the updated state of the memory cells at the current moment; Indicates element-level multiplication; S435: Output Gate : It is the weight matrix of the output gate. It is the bias vector of the output gate, which controls the amount of output information; S436: Output in Hidden State : The hidden state integrates the memory cell state and the control information of the output gate; S44: Local Connectivity Subnetwork: The outputs of the RNN and LSTM layers are concatenated and then input into the local connectivity subnetwork. The local connectivity subnetwork consists of several fully connected layers. The neurons in each fully connected layer are only connected to a subset of neurons in the previous layer, used to extract local features; the first fully connected layer... The calculation formula is: ,in: This represents the output of the RNN layer. This represents the output of the LSTM layer. It is the weight matrix of the first fully connected layer. It is the bias vector; the calculation of subsequent fully connected layers is similar, performing feature extraction and nonlinear transformation layer by layer; S45: Multi-feature fusion module: fuses the local features extracted by the local connectivity sub-network with the temporal features extracted by the RNN and LSTM layers using a weighted summation method. ,in: This represents the final model output; This represents the feature output of a locally connected subnetwork; This represents the temporal feature output of RNN and LSTM layers; and These are weighting coefficients used to balance the importance of local features and temporal features, and are learned through training data. S46: Attention Mechanism: An attention mechanism is added after the multi-feature fusion module to weight the fused features and highlight important feature information; attention weights. The calculation formula is: ,in: It is the attention weight matrix. It is a bias term, which uses the softmax function to normalize the weights to between 0 and 1; the weighted features Represented as: This allows the model to focus more on feature information related to settlement prediction; S47: Output Layer: Receives the processed feature information and outputs the predicted cumulative settlement of the building. The calculation formula is: ,in: It is the weight matrix of the output layer. It is a bias term.

[0016] As a further technical solution of the present invention, in S5, a hybrid optimization algorithm based on AdamW and simulated annealing mechanism is used to optimize network parameters.

[0017] S51: First, AdamW is used for basic optimization. AdamW is an improved adaptive gradient optimization algorithm that can more reasonably reduce weights while maintaining the ability to estimate first and second moments. In each training iteration, AdamW is used to adaptively adjust the learning rate and weight reduction to achieve efficient convergence. The specific AdamW method flow is as follows: Calculate the current parameters in the t-th iteration. gradient : ; Update first-order moment estimate : ,in The attenuation rate is estimated by the first moment; Update the second moment estimate : ,in The attenuation rate is estimated by the second moment; Bias correction for calculating first-order moment estimates : ; Bias correction for calculating second-order moment estimates : ; By calculating the bias correction, these estimates can be effectively adjusted, reducing the bias caused by inaccurate estimates of the first and second moments in the early stages of optimization, and making them more accurately reflect the true gradient information. Update parameters : ,in: The learning rate; This is the weight decay coefficient; It is a very small constant used to prevent the denominator from being zero; Through the above mechanism, AdamW can efficiently guide the model to converge quickly in multi-stage settlement data, improve the prediction effect, and at the same time suppress the risks of overfitting and underfitting.

[0018] S52: Although AdamW can achieve fast convergence, it may still get stuck in local optima when dealing with nonlinear time series data with multiple local extrema. Therefore, based on AdamW, a simulated annealing mechanism is integrated to further improve the global optimization capability of the model. The simulated annealing algorithm guides the optimization process from random search to gradual stabilization by controlling the gradual decay of the "temperature" parameter T. The simulated annealing algorithm consists of three stages: heating stage, isothermal stage, and cooling stage. S521: Heating stage, i.e. algorithm initialization, setting the initial temperature and initial solution; S522: The isothermal stage involves multiple iterations at the same temperature, applying a temperature-related random perturbation to the current solution vector using the following formula: Where: T is the simulated temperature, t is the initial temperature; controls the cooling progress; This is the current solution vector. The new solution after applying the perturbation; r is a random number uniformly distributed in [-1,1]. The step size vector has the same length as the solution vector; If the new solution is better than the current solution, then accept it; if the new solution is worse, then accept it based on probability. Where: E is the objective function value; The difference between the objective function of the old and new solutions; S523: After a certain number of iterations, the temperature T gradually decreases (cooling phase): ,in: The simulated temperature at the current time step; The simulated temperature of the previous time step; This is the temperature decrease coefficient, typically taken as 0.95-0.99; S524: Repeat the isothermal and cooling phases until the temperature drops to the minimum value. The simulated annealing mechanism enables the model to escape local optima during training, thereby exploring the solution space more comprehensively and improving the stability and generalization performance of settlement prediction.

[0019] As a further technical solution of the present invention, in S6, experimental training is first performed, and the range of values ​​of each hyperparameter is adjusted one by one using the controlled variable method. The influence of the fitting image of the training set on the model performance is analyzed, thereby narrowing the parameter search space. Then, by designing cross-experiments, different combinations of hyperparameters are compared to evaluate their fitting on the training set and test set. The mean squared error (MSE) is used as a metric to record the model performance corresponding to each combination of hyperparameters in detail, and finally the optimal hyperparameter setting scheme is determined.

[0020] As a further technical solution of the present invention, in S7, the cumulative settlement-time curve of the building predicted by the model is compared with the actual monitoring curve to intuitively show the deviation distribution between the predicted value and the true value; at the same time, by drawing multiple evaluation index images, including but not limited to root mean square error (RMSE) and mean square error (MSE), the performance and convergence trend of the model at different training stages are clearly reflected.

[0021] A time-series prediction system for settlement of adjacent buildings during deep foundation pit excavation is provided to implement a method for time-series prediction of settlement of adjacent buildings during deep foundation pit excavation. The system includes the following modules: The data acquisition module is used to acquire raw monitoring data on the settlement of buildings on site; The preprocessing module is used to preprocess the acquired raw monitoring data; The model building module is used to build recurrent neural networks; The analysis module is used to determine the model's input and output parameters through principal component analysis; The optimization module is used to optimize recurrent neural networks based on the optimizer. The parameter optimization module is used to optimize hyperparameters based on the optimization results. The prediction module is used to perform time-series prediction of sedimentation using an optimized recurrent neural network.

[0022] A computer-readable storage medium storing a plurality of instructions adapted for loading and execution by a processor of a terminal device of a method for time-series prediction of settlement of adjacent buildings during deep foundation pit excavation.

[0023] A terminal device includes a processor and a computer-readable storage medium, the processor being configured to implement various instructions; the computer-readable storage medium being configured to store multiple instructions adapted for loading and execution by the processor to provide a method for time-series prediction of settlement of adjacent buildings during deep foundation pit excavation.

[0024] The beneficial effects of this invention are as follows: 1. By introducing recurrent neural networks (RNN), long short-term memory networks (LSTM), and locally connected networks, a settlement prediction model is constructed to achieve dynamic modeling and accurate prediction of the settlement trend of buildings during each stage of foundation pit excavation.

[0025] 2. This method combines historical settlement monitoring data with multi-layer soil excavation information, and enhances the model's adaptability to complex working conditions by introducing multi-source input parameters.

[0026] 3. Through systematic data preprocessing and optimization strategies, high-precision time-series prediction of building settlement during future soil excavation stages is achieved, providing solid data support for safety control and risk assessment during foundation pit construction.

[0027] 4. It can provide a scientific and reliable technical basis for safety risk assessment and settlement control during the construction phase, and then be used to guide the optimization design of the foundation pit excavation scheme and the dynamic adjustment of the construction process, improve the safety and scientific nature of the project implementation process, and reduce the risk of damage to the surrounding building structure or even accidents caused by uneven settlement. Attached Figure Description

[0028] Figure 1 This is an overall flowchart of a method for predicting the time series settlement of adjacent buildings during deep foundation pit excavation, as proposed in this invention. Figure 2 This is a schematic diagram of the recurrent neural network method used in the time-series prediction method for settlement of adjacent buildings during deep foundation pit excavation proposed in this invention. Detailed Implementation

[0029] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0030] Please see the appendix Figure 1 A method for time-series prediction of settlement of adjacent buildings during deep foundation pit excavation includes the following specific steps: S1: Obtain raw monitoring data on the settlement of buildings on site; The aim is to monitor the settlement changes of adjacent buildings at each stage of the foundation pit excavation process in real time, providing accurate and reliable basic data support for subsequent model construction and prediction. The main methods include, but are not limited to, deploying high-precision monitoring points at key locations and employing an automated hydrostatic leveling system to continuously and quantitatively collect time-series information on the settlement process, capturing the settlement data of buildings during foundation pit excavation. After collection, the settlement monitoring data is systematically organized and archived, laying a solid foundation for data preprocessing and model training.

[0031] S2: Perform data preprocessing on the acquired raw monitoring data; Before building the model, the settlement data of the buildings collected on site needs to be systematically and comprehensively preprocessed to ensure data quality and the effectiveness of model training.

[0032] First, the raw data undergoes an integrity check to identify and remove obviously missing or abnormal monitoring point data. Then, the data for each monitoring point is organized into a "cumulative settlement-time" curve and normalized to the [-1,0] interval. All normalized settlement curves are plotted on the same coordinate graph to visually identify and remove curves with significantly abnormal settlement trends. The mean error (CMSE) between the remaining curves and the average curve is calculated, and the maximum error value is compared with a preset judgment threshold ε. When the maximum error is less than ε, the remaining curves are considered to have consistent trends and learning value, and the data cleaning process ends. If the error exceeds the threshold, the curve with the largest error is removed, and the above judgment is repeated until the consistency standard is met.

[0033] After completing the overall data screening, outliers inevitably remain within each curve. Therefore, a statistically based "three-times root mean square method" is used for smoothing. This method calculates the root mean square value of each curve to identify and correct points with excessive deviations, effectively filtering out outliers and improving the stability and continuity of the data, thereby providing reliable data support for subsequent model training.

[0034] Missing value imputation methods include, but are not limited to, linear interpolation, spline interpolation, or imputation using data from nearby time points; outlier removal methods include, but are not limited to, first using statistical methods to remove obvious outliers, and then combining them with rules such as setting thresholds for further fine-tuning to improve overall data quality and model training effect; outlier removal methods include, but are not limited to, identifying outliers through statistical analysis (such as the 3σ principle) or time series stationarity detection methods, and correcting or removing them according to specific circumstances.

[0035] S3: Determine input and output parameters based on preprocessed monitoring data; The aim is to clarify the key input variables and target output results required for settlement prediction models, ensuring that the models can fully reflect the core factors affecting building settlement, thereby achieving scientific modeling, dynamic prediction, and reasonable assessment of the settlement process. Appropriate selection of input and output parameters is fundamental to improving the model's prediction accuracy and generalization ability.

[0036] By systematically analyzing the main engineering and geological factors affecting building settlement, soil parameters, groundwater level, and excavation parameters are selected as input parameters for the model. The output variable of the model is the cumulative settlement of the building, that is, the total settlement of the building since excavation is predicted at each time step. The above multi-source information is encoded as a time series input so that the RNN model can effectively learn the dynamic relationship between it and settlement changes. Soil parameters include but are not limited to soil weight, internal friction angle, and cohesion. Soil excavation parameters include but are not limited to excavation location and excavation volume.

[0037] S31: The main reason for settlement of adjacent buildings during foundation pit excavation is that the excavation of the soil in the foundation pit removes the earth pressure acting on the support structure, causing lateral movement of the support structure and thus redistributing the stress in the foundation soil of the building; according to Rankine's earth pressure theory, the formula for active earth pressure intensity is: in, This refers to the active earth pressure intensity. The soil weight is the density, reflecting the soil's self-weight stress. In the formula: The weight of the soil. For the volume of soil; This refers to the depth of the soil. The internal friction angle reflects the shear strength of the soil. Cohesion is the force of reaction between particles; in this process, soil parameters related to earth pressure are key factors affecting the settlement process, and are therefore selected as input parameters for the model. S32: Changes in groundwater level, especially a decrease in groundwater level within the bearing layer of a building, can increase the additional stress on the soil, thereby affecting building settlement. Therefore, monitoring the depth of groundwater at different times is used as an input parameter for the model. For example, a water level gauge can be used to periodically measure the depth of groundwater relative to the ground surface to obtain dynamic data on groundwater level changes. S33: During the excavation of the foundation pit, the soil unloading process over time is the direct cause of settlement. Therefore, the soil excavation parameters that change over time should also be used as model input parameters to reflect the impact of soil unloading on settlement during the excavation process. The excavation location is represented by coordinates or relative distance to determine the relative position of the excavation area and the building. For example, with a reference point of the building as the origin, a coordinate system is established, and the center coordinates of the excavation area are (x, y). The excavation location can then be represented as the relative position of this coordinate system with the building reference point. The excavation volume is the amount of earthwork excavated from the foundation pit, which can be calculated based on the shape and depth of the excavation area. For example, for a rectangular foundation pit, the excavation volume... long Width depth; S34: Regarding the support structure, commonly used rigid support structures (such as pile foundations and diaphragm walls) are constructed before the foundation pit is excavated and are factors that do not change over time. For internal supports or anchor cables added during the foundation pit excavation, the prestress applied will cause a slight rebound in the settlement of the building. The duration is short and can be regarded as a one-time rebound amount, used to correct the model prediction results, rather than being directly used as model input. The self-weight of the building, the foundation type, and its relative geometric relationship with the foundation pit remain unchanged during the excavation process and are not used as input parameters of the model. S35: The cumulative settlement S of the building is the model output parameter, representing the total settlement of the building at a certain time step since excavation. It can be calculated by measuring the settlement at different monitoring points on the building and taking the average or weighted average. For example, if multiple settlement monitoring points are set up on the building, the settlement at each monitoring point is... The cumulative settlement (Average method), or (Weighted average method) (Assign weights to each monitoring item).

[0038] S4: Construct a recurrent neural network (RNN) and divide it into training and test sets; Based on the construction of recurrent neural networks (RNN), long short-term memory network (LSTM) units are introduced to enhance the ability to remember long-term dependencies, and local connection subnetworks are added to capture spatial local features, forming a hybrid neural network model that integrates RNN, LSTM and local connection networks. S41: Input layer: Receives preprocessed building settlement-related data, including soil parameters (soil weight). internal friction angle Cohesion ), groundwater level And excavation parameters (excavation location coordinates) Data such as excavation volume (V) are arranged and combined according to time series to form an input vector sequence. , where t represents the time step; S42: RNN layer: Consists of several RNN units, each of which receives the hidden state from the previous time step. and the input at the current moment Calculate the hidden state at the current time step. : ,in: For example, the sigmoid function; Let be the weight matrix input to the hidden layer, with dimension . Input size Hidden layer size The weight matrix from hidden layer to hidden layer has a dimension of . Hidden layer size Hidden layer size Here is the hidden layer bias vector, with dimension . Hidden layer size ; S43: LSTM Layer: An LSTM layer is added on top of the RNN layer. An LSTM unit contains a forget gate, an input gate, and an output gate. The forget gate determines how much past information needs to be retained, the input gate controls the degree to which current input information is written, and the output gate determines the amount of output information. The specific calculation is as follows: S431: The Gate of Oblivion : ,in: It is the weight matrix of the forget gate; It is the bias vector of the forget gate; This is the sigmoid function, with an output range between 0 and 1, representing the degree of forgetting; S432: Input Gate : It is the weight matrix of the input gate; It is the bias vector of the input gate, used to control the degree of update of the input information; S433: Candidate memory cell status : ,in: It is the weight matrix of the candidate memory cell states. It is a bias vector. The function compresses the input to a range between -1 and 1; S434: Memory cell status update: ,in: It is the state of memory cells from the previous moment; It is the updated state of the memory cells at the current moment; Indicates element-level multiplication; S435: Output Gate : It is the weight matrix of the output gate. It is the bias vector of the output gate, which controls the amount of output information; S436: Output in Hidden State : The hidden state integrates the memory cell state and the control information of the output gate; S44: Local Connectivity Subnetwork: The outputs of the RNN and LSTM layers are concatenated and then input into the local connectivity subnetwork. The local connectivity subnetwork consists of several fully connected layers. The neurons in each fully connected layer are only connected to a subset of neurons in the previous layer, used to extract local features; the first fully connected layer... The calculation formula is: ,in: This represents the output of the RNN layer. This represents the output of the LSTM layer. It is the weight matrix of the first fully connected layer. It is the bias vector; the calculation of subsequent fully connected layers is similar, performing feature extraction and nonlinear transformation layer by layer; S45: Multi-feature fusion module: fuses the local features extracted by the local connectivity sub-network with the temporal features extracted by the RNN and LSTM layers using a weighted summation method. ,in: This represents the final model output; This represents the feature output of a locally connected subnetwork; This represents the temporal feature output of RNN and LSTM layers; and These are weighting coefficients used to balance the importance of local features and temporal features, and are learned through training data. S46: Attention Mechanism: An attention mechanism is added after the multi-feature fusion module to weight the fused features and highlight important feature information; attention weights. The calculation formula is: ,in: It is the attention weight matrix. It is a bias term, which uses the softmax function to normalize the weights to between 0 and 1; the weighted features Represented as: This allows the model to focus more on feature information related to settlement prediction; S47: Output Layer: Receives the processed feature information and outputs the predicted cumulative settlement of the building. The calculation formula is: ,in: It is the weight matrix of the output layer. It is a bias term; Through the above steps, the model can combine historical settlement effects with parameters such as the current groundwater level and excavation location to dynamically predict the cumulative settlement at each time point, thus achieving time-series simulation of building settlement throughout the entire foundation pit excavation process.

[0039] This circular connection mechanism enables the network to transmit and accumulate information over time, flexibly process dynamically changing input data, and capture the dynamic characteristics of the input data that evolve over time, thereby achieving accurate prediction of the settlement of adjacent buildings during the foundation pit excavation process.

[0040] By dividing all settlement monitoring data into training and testing sets in chronological order, the first 70% of the data is used for model training, allowing the model to fully learn the historical settlement change patterns, while the last 30% of the data is used as the testing set to test the model's predictive ability on unseen data.

[0041] This step aims to ensure that the RNN can effectively learn settlement patterns from historical data and has good generalization ability on unknown data, thereby enabling the prediction of future settlement.

[0042] S5: Select an appropriate optimization algorithm to optimize the recurrent neural network (RNN); Employing appropriate optimization algorithms can not only accelerate model convergence, shorten training time, and improve resource utilization, but also significantly enhance the model's prediction accuracy and generalization ability, avoid getting trapped in local optima, and reduce the risk of overfitting or underfitting. A hybrid optimization algorithm combining AdamW and simulated annealing is selected. This method combines AdamW's adaptive gradient adjustment capability with the global search characteristics of simulated annealing, aiming to improve the model's global optimization capability and nonlinear sequence modeling capability while maintaining stable model convergence, thereby improving the accuracy of settlement prediction.

[0043] S51: First, AdamW is used for basic optimization. AdamW is an improved adaptive gradient optimization algorithm that can more reasonably reduce weights while maintaining the ability to estimate first and second moments. In each training iteration, AdamW is used to adaptively adjust the learning rate and weight reduction to achieve efficient convergence. The specific AdamW method flow is as follows: Calculate the current parameters in the t-th iteration. gradient : ; Update first-order moment estimate : ,in The attenuation rate is estimated by the first moment; Update the second moment estimate : ,in The attenuation rate is estimated by the second moment; Bias correction for calculating first-order moment estimates : ; Bias correction for calculating second-order moment estimates : ; By calculating the bias correction, these estimates can be effectively adjusted, reducing the bias caused by inaccurate estimates of the first and second moments in the early stages of optimization, and making them more accurately reflect the true gradient information. Update parameters : ,in: The learning rate; This is the weight decay coefficient; It is a very small constant used to prevent the denominator from being zero; Through the above mechanism, AdamW can efficiently guide the model to converge quickly in multi-stage settlement data, improve the prediction effect, and at the same time suppress the risks of overfitting and underfitting.

[0044] S52: Although AdamW can achieve fast convergence, it may still get stuck in local optima when dealing with nonlinear time series data with multiple local extrema. Therefore, based on AdamW, a simulated annealing mechanism is integrated to further improve the global optimization capability of the model. The simulated annealing algorithm guides the optimization process from random search to gradual stabilization by controlling the gradual decay of the "temperature" parameter T. The simulated annealing algorithm consists of three stages: heating stage, isothermal stage, and cooling stage. S521: Heating stage, i.e. algorithm initialization, setting the initial temperature and initial solution; S522: The isothermal stage involves multiple iterations at the same temperature, applying a temperature-related random perturbation to the current solution vector using the following formula: Where: T is the simulated temperature, t is the initial temperature; controls the cooling progress; This is the current solution vector. The new solution after applying the perturbation; r is a random number uniformly distributed in [-1,1]. The step size vector has the same length as the solution vector; If the new solution is better than the current solution, then accept it; if the new solution is worse, then accept it based on probability. Where: E is the objective function value; The difference between the objective function of the old and new solutions; S523: After a certain number of iterations, the temperature T gradually decreases (cooling phase): ,in: The simulated temperature at the current time step; The simulated temperature of the previous time step; This is the temperature decrease coefficient, typically taken as 0.95-0.99; S524: Repeat the isothermal and cooling phases until the temperature drops to the minimum value. The simulated annealing mechanism enables the model to escape local optima during training, thereby exploring the solution space more comprehensively and improving the stability and generalization performance of settlement prediction.

[0045] This hybrid optimization algorithm, which integrates AdamW and simulated annealing, not only makes full use of the efficient optimization path brought by gradient information in the algorithm structure, but also introduces a probability-based global search capability through simulated annealing. The two complement each other, enabling the model to have both fast and stable convergence characteristics during training, and to escape local optima traps, thereby improving the overall optimization effect.

[0046] S6: Perform hyperparameter optimization based on the optimization results; Aimed at improving the training efficiency and predictive performance of models, this approach scientifically selects and adjusts key hyperparameters to enable RNNs to achieve optimal learning outcomes during training. Hyperparameters include, but are not limited to, the learning rate. The control parameters (β1, β2) of the AdamW optimization algorithm, the control parameters (T, θ) of the simulated annealing optimizer, and the batch size Number of neurons in the hidden layer Weight decay coefficient Number of layers in a recurrent neural network .

[0047] In the initial training phase, the model undergoes experimental training, and the range of values ​​for each hyperparameter is adjusted one by one using the controlled variable method. The impact of the adjusted values ​​on the model performance is analyzed by combining the fitted images of the training set, thereby narrowing the parameter search space. Subsequently, cross-experiments are designed to compare different combinations of hyperparameters, evaluate their fitting performance on the training and test sets, and use mean squared error (MSE) as a metric to record the model performance corresponding to each hyperparameter combination in detail, ultimately determining the optimal hyperparameter setting scheme.

[0048] S61: Controlled variable method for adjusting hyperparameters: First, determine an initial combination of hyperparameters, keep other hyperparameters unchanged, adjust the values ​​of each hyperparameter in turn, conduct experimental training, obtain the fitted image of the training set, observe the changes in the loss value of the model at different times, and determine their respective candidate ranges. S62: Cross-validation experiment: S621: Split the dataset: Divide the training set into a certain proportion (e.g., 5-fold cross-validation) There are 3 subsets, each of roughly the same size, and the data distribution of each subset is guaranteed to be representative. S622: Design Hyperparameter Combinations: Within the candidate range determined by the control variable method, select different hyperparameter combinations and design a total of... A variety of hyperparameter combinations; S623: Cross-validation process: Training and validation: For each combination of hyperparameters, sequentially... One subset is used as training data, and the remaining one subset is used as validation data. The training and validation are repeated. Each training iteration sets the model parameters according to the current hyperparameter combination and calculates the loss values ​​on the training and validation sets. Record the loss: After each training and validation cycle, record the corresponding training loss. ) and validation loss ( ); S624: Calculate the average loss: For each combination of hyperparameters, calculate the average loss... Average training loss under cross-validation ( ) and average validation loss ( ): , ; S63: Evaluate and select the optimal hyperparameters: S631: Calculate the mean squared error (MSE): training set : test set : ,in: The number of samples in the training set. The number of samples in the test set. This represents the actual settlement. To predict settlement, for each combination of hyperparameters, the MSE value is calculated on the training and test sets respectively. S632: Comparison and Recording: Record in detail the average training loss, average validation loss, and training set data for each hyperparameter combination. Test set By comparing these performance metrics, we can find the hyperparameter combination that achieves a smaller MSE on both the training and test sets. At the same time, we should avoid overfitting (i.e., a smaller MSE on the training set but a larger MSE on the test set). S633: Determine the optimal combination: ultimately select the combination that exhibits the best overall performance in cross-validation (e.g., minimum validation loss and minimum test set value). The optimal combination of hyperparameters is used for subsequent model prediction.

[0049] This method can not only significantly improve the convergence speed and stability of the model, but also enhance its adaptability to complex time series data, laying the foundation for improving the overall prediction accuracy of the model.

[0050] S7: Comparison and analysis of model prediction results with measured data and verification of evaluation indicators; During the hyperparameter optimization process, it is also necessary to compare and analyze the model's prediction results with actual monitoring data to further verify the model's fitting accuracy and predictive ability. Specifically, the cumulative settlement-time curve of the building predicted by the model is compared with the actual monitoring curve to visually show the distribution of the deviation between the predicted and actual values. At the same time, by plotting multiple evaluation index graphs, including but not limited to root mean square error (RMSE) and mean square error (MSE), the model's performance and convergence trend at different training stages are clearly reflected. This process not only helps to verify the rationality of the hyperparameter settings but also comprehensively evaluates the model's fitting ability and stability, thus providing a reliable basis for finally determining the optimal hyperparameter combination.

[0051] Figure 2The method shown is a flowchart of the RNN algorithm, where is the input of the first time step, is the input of the second time step, and is the input data of the t-th time step. is the output of the first time step, is the output of the second time step, and is the output data of the t-th time step. is the input of the current time step and all information from previous time steps captured by the hidden layer. is the initial hidden state, is the hidden state of the first time step, is the hidden state of the second time step, and so on, is the hidden state of the t-th time step, which can capture the input of the current time step and all information from previous time steps.

[0052] S8: Use the optimized recurrent neural network to predict the time series of sedimentation.

[0053] A time-series prediction system for settlement of adjacent buildings during deep foundation pit excavation is provided to implement a method for time-series prediction of settlement of adjacent buildings during deep foundation pit excavation. The system includes the following modules: The data acquisition module is used to acquire raw monitoring data on the settlement of buildings on site; The preprocessing module is used to preprocess the acquired raw monitoring data; The model building module is used to build recurrent neural networks; The analysis module is used to determine the model's input and output parameters through principal component analysis; The optimization module is used to optimize recurrent neural networks based on the optimizer. The parameter optimization module is used to optimize hyperparameters based on the optimization results. The prediction module is used to perform time-series prediction of sedimentation using an optimized recurrent neural network.

[0054] A computer-readable storage medium storing multiple instructions adapted for loading and execution by a processor of a terminal device: a method for time-series prediction of settlement of adjacent buildings during deep foundation pit excavation.

[0055] A terminal device includes a processor and a computer-readable storage medium. The processor implements various instructions; the computer-readable storage medium stores multiple instructions adapted for loading and execution by the processor. This relates to a method for time-series prediction of settlement of adjacent buildings during deep foundation pit excavation. As can be seen from the above description, the above embodiments of the present invention achieve the following technical effects: by introducing recurrent neural networks (RNN), long short-term memory networks (LSTM) and local connection networks to construct a settlement prediction model, dynamic modeling and accurate prediction of the settlement trend of buildings during the excavation of foundation pits at various stages can be realized.

[0056] This method combines historical settlement monitoring data with multi-layer soil excavation information, and enhances the model's adaptability to complex working conditions by introducing multi-source input parameters.

[0057] Through systematic data preprocessing and optimization strategies, high-precision time-series prediction of building settlement during future soil excavation stages is achieved, providing solid data support for safety control and risk assessment during foundation pit construction.

[0058] It can provide a scientific and reliable technical basis for safety risk assessment and settlement control during the construction phase, and then be used to guide the optimization design of the foundation pit excavation scheme and the dynamic adjustment of the construction process, improve the safety and scientific nature of the project implementation process, and reduce the risk of damage to the surrounding building structure or even accidents caused by uneven settlement.

[0059] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of the invention (including the claims) is limited to these examples; within the framework of the invention, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of the different aspects of the invention as described above, which are not provided in the details for the sake of brevity.

[0060] This invention is intended to cover all such substitutions, modifications, and variations falling within the broad scope of the claims. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A method for time-series prediction of settlement of adjacent buildings during deep foundation pit excavation, characterized in that, The specific steps include the following: S1: Obtain raw monitoring data on the settlement of buildings on site; S2: Perform data preprocessing on the acquired raw monitoring data; S3: Determine input and output parameters based on preprocessed monitoring data; S4: Construct a recurrent neural network and divide it into training and test sets; S5: Select a suitable optimization algorithm to optimize the recurrent neural network; S6: Perform hyperparameter optimization based on the optimization results; S7: Comparison and analysis of model prediction results with measured data and verification of evaluation indicators; S8: Use the optimized recurrent neural network to predict the time series of sedimentation.

2. The method for predicting the settlement of adjacent buildings during deep foundation pit excavation according to claim 1, characterized in that, In step S2, data preprocessing includes performing integrity checks on the original monitoring data, identifying and removing missing data and abnormal monitoring point data, converting the monitoring data into a normalized cumulative settlement-time curve, using the mean error CMSE and a preset threshold for consistency judgment, gradually filtering out data with learning value, and then using the three-times root mean square method to smooth local outliers in the curve, effectively removing outliers.

3. The method for predicting the settlement of adjacent buildings during deep foundation pit excavation according to claim 1, characterized in that, In S3, the main engineering and geological factors affecting building settlement are systematically analyzed, and soil parameters, groundwater level and excavation parameters are selected as input parameters of the model; the output parameter of the model is the cumulative settlement of the building, that is, the total settlement of the building since excavation is predicted at each time step; the soil parameters include but are not limited to soil weight, internal friction angle and cohesion; the excavation parameters include but are not limited to excavation location and excavation volume.

4. The method for predicting the settlement of adjacent buildings during deep foundation pit excavation according to claim 1, characterized in that, In step S4, based on the construction of a recurrent neural network, a long short-term memory network unit is introduced to enhance the ability to remember long-term dependencies, and a locally connected subnetwork is added to capture local spatial features, forming a hybrid neural network model that integrates RNN, LSTM and locally connected networks. The first 70% of the data is used for model training to enable the model to fully learn the historical subsidence change patterns, and the last 30% of the data is used as a test set to test the model's predictive ability on unseen data.

5. The method for predicting the settlement of adjacent buildings during deep foundation pit excavation according to claim 1, characterized in that, In S5, a hybrid optimization algorithm based on AdamW and simulated annealing mechanism is used to optimize network parameters.

6. The method for predicting the settlement of adjacent buildings during deep foundation pit excavation according to claim 1, characterized in that, In step S6, experimental training is first conducted, and the range of values ​​for each hyperparameter is adjusted one by one using the controlled variable method. The impact of the hyperparameters on the model performance is analyzed by combining the fitted images of the training set, thereby narrowing the parameter search space. Subsequently, cross-experiments are designed to compare different combinations of hyperparameters, evaluate their fitting performance on the training and test sets, and use the mean squared error as a metric to record the model performance corresponding to each combination of hyperparameters in detail. Finally, the optimal hyperparameter setting scheme is determined.

7. The method for predicting the settlement of adjacent buildings during deep foundation pit excavation according to claim 1, characterized in that, In step S7, the cumulative settlement-time curve of the building predicted by the model is compared with the actual monitoring curve to intuitively show the deviation distribution between the predicted value and the true value; at the same time, by drawing multiple evaluation index images, including but not limited to root mean square error and mean square error, the performance and convergence trend of the model at different training stages are clearly reflected.

8. A time-series prediction system for settlement of adjacent buildings during deep foundation pit excavation, used to implement the time-series prediction method for settlement of adjacent buildings during deep foundation pit excavation as described in any one of claims 1-7, characterized in that, Includes the following modules: The data acquisition module is used to acquire raw monitoring data on the settlement of buildings on site; The preprocessing module is used to preprocess the acquired raw monitoring data; The model building module is used to build recurrent neural networks; The analysis module is used to determine the model's input and output parameters through principal component analysis; The optimization module is used to optimize recurrent neural networks based on the optimizer. The parameter optimization module is used to optimize hyperparameters based on the optimization results. The prediction module is used to perform time-series prediction of sedimentation using an optimized recurrent neural network.

9. A computer-readable storage medium storing a plurality of instructions, characterized in that, The instructions are adapted to be loaded and executed by the processor of a terminal device as described in any one of claims 1-7, which is a method for predicting the settlement of adjacent buildings during deep foundation pit excavation.

10. A terminal device, comprising a processor and a computer-readable storage medium, wherein the processor is configured to implement instructions; and the computer-readable storage medium is configured to store multiple instructions, characterized in that, The instructions are adapted to be loaded by a processor and executed as described in any one of claims 1-7, a method for predicting the settlement of adjacent buildings during deep foundation pit excavation.

Images