BiLSTM-XGBoost settlement prediction method based on NGSA-II multi-objective optimization algorithm

Abstract

The invention discloses a BiLSTM-XGBoost settlement prediction method based on an NGSA-II multi-objective optimization algorithm. The BiLSTM-XGBoost settlement prediction method specifically comprises the following steps: preprocessing engineering monitoring acquisition data; carrying out time sequence feature screening by adopting an ADF; generating new features by adopting BiLSTM, and fusing the new features with the static features; performing optimization on the hyper-parameters of the BiLSTM and the XGBoost by using the NSGA-II; on the basis of the optimized hyper-parameters, a fused BiLSTM-XGBoost model is trained, and efficient prediction of sedimentation is achieved; and evaluating and applying the prediction model. Through efficient fusion and multi-objective optimization of multi-source features, dynamic balance of model prediction precision and complexity is realized, the accuracy of settlement prediction is remarkably improved, meanwhile, the training time is shortened, the calculation complexity is reduced, and the method is suitable for real-time prediction and decision support in engineering construction.

Description

Technical Field

[0001] The present invention belongs to the technical field of engineering monitoring, and particularly relates to a BiLSTM-XGBoost settlement prediction method based on the NGSA-II multi-objective optimization algorithm. Background Art

[0002] Accurate prediction of deformation monitoring data during engineering construction is a key link to ensure construction safety and reduce environmental risks. Especially in projects such as underground tunnel excavation, subway construction, and deep foundation pit excavation, the uncertainty of settlement may cause damage to surrounding buildings, surface collapse, and other risks. Therefore, developing a high-precision settlement prediction model based on complex data is an important requirement in engineering construction management.

[0003] Currently, traditional settlement prediction methods have the following limitations: (1) Methods based on empirical formulas. Traditional settlement prediction mainly relies on empirical formulas in soil mechanics and structural mechanics. These methods do not adequately consider the complexity of actual construction scenarios, are difficult to accurately capture the non-linear relationships between features, and have limited capabilities for settlement prediction under multi-factor coupling conditions. (2) Limitations of single-model prediction. With the development of machine learning technology, methods such as support vector machines (SVM), random forests (RF), and single neural network models (such as LSTM) have been widely used in settlement prediction. These methods perform well when dealing with a single type of feature (such as time-series features or static features), but there are certain performance bottlenecks when fusing multi-source heterogeneous features (such as the time-series and static data that coexist during construction). In addition, single models usually rely on manual tuning in parameter selection, making it difficult to obtain the optimal combination of hyperparameters and affecting the prediction accuracy of the model. (3) Lack of modeling methods for multi-objective optimization. Most current prediction methods often only focus on a single objective (such as prediction accuracy), while ignoring the requirements for model complexity and computational efficiency in actual engineering needs. When predicting engineering monitoring data, the prediction model needs to balance high accuracy, low computational complexity, and interpretability in order to be quickly deployed and applied in real time in resource-constrained environments. However, traditional methods cannot balance the optimization requirements among multiple objectives and are difficult to meet the actual needs of complex construction scenarios.

[0004] In summary, there is an urgent need for a BiLSTM-XGBoost settlement prediction method based on the NGSA-II multi-objective optimization algorithm to solve the problems existing in the prior art. Summary of the Invention

[0005] In order to effectively fuse time-series and static features and achieve a dynamic balance between the accuracy and complexity of the model through a multi-objective evolutionary optimization algorithm, the present invention provides a BiLSTM-XGBoost settlement prediction method based on the NGSA-II multi-objective optimization algorithm.

[0006] A BiLSTM-XGBoost settlement prediction method based on the NGSA-II multi-objective optimization algorithm of the present invention includes the following steps:

[0007] Step 1: Preprocessing of data collected from engineering monitoring.

[0008] Step 2: Conduct time series feature screening using ADF (Augmented Dickey-Fuller).

[0009] Step 3: Generate new features using BiLSTM (Bi-directional Long Short-Term Memory) and fuse them with static features.

[0010] Step 4: Use NSGA-II (Non-dominated Sorting Genetic Algorithm II) to optimize the hyperparameters of BiLSTM and XGBoost.

[0011] Step 5: Based on the optimized hyperparameters, train the fused BiLSTM-XGBoost model to achieve efficient prediction of settlement.

[0012] Step 6: Evaluate and apply the prediction model.

[0013] Further, Step 1 is specifically as follows:

[0014] Step 1.1: Filling missing values and normalizing the data collected from monitoring.

[0015] Use the interpolation method to fill the missing values in the monitoring data, and normalize the filled data to eliminate the influence of different dimensions on model training. The formula is as follows:

[0016]

[0017] In the formula, x is the original data, x′ is the normalized data, and min(X) and max(X) are the minimum and maximum values of the data set respectively.

[0018] Step 1.2: Comparative analysis of preprocessing of model input data.

[0019] Conduct feature engineering processing on time series data and static feature data respectively. The feature selection adopts the principal component analysis (PCA) method. The formula is as follows:

[0020] Z = XW (2)

[0021] In the formula, Z is the feature after dimensionality reduction, X is the original feature matrix, and W is the dimensionality reduction matrix.

[0022] Further, Step 2 is specifically as follows:

[0023] Step 2.1: Monitor and extract the temporal characteristics and static characteristics of the collected data: Extract the temporal characteristics and static characteristics from the preprocessed data. The temporal characteristics include, but are not limited to, construction progress and monitoring data, and the static characteristics include, but are not limited to, soil properties and tunneling parameters.

[0024] Step 2.2: Use the ADF test to evaluate the stationarity of the temporal characteristics: Conduct the ADF unit root test on the extracted temporal characteristics to determine their stationarity. The formula is as follows:

[0025] Δy t =α + βt + γy t-1 + δ1Δy t-1 +…+ δ p-1 Δy t-p+1 + ∈ t (3)

[0026] In the formula, y t is the temporal characteristic, Δ is the difference operation, α, β, γ, δ are parameters to be estimated, and ∈ t is the error term.

[0027] t (time variable of the trend term); p (lag order).

[0028] If the test statistic is less than the critical value, it is considered that the temporal characteristic is stationary.

[0029] Furthermore, Step 3 is specifically as follows:

[0030] Step 3.1: Generate new temporal characteristics of the monitored and collected data; construct a Bidirectional Long Short-Term Memory (BiLSTM) model, input the preprocessed temporal data, and simultaneously capture the forward and backward dependencies in the temporal data through the forward and backward LSTM layers to generate a new temporal characteristic representation. The formula is as follows:

[0031]

[0032] In the formula, x t is the missing value, and are the hidden states of the forward and backward LSTMs respectively, and h t is the connection of the bidirectional hidden states.

[0033] Step 3.2: Fuse the temporal characteristics generated by BiLSTM with the static characteristics; fuse the generated new temporal characteristic h t with the static characteristic s to form a comprehensive feature vector F. The fusion method uses concatenation or weighted average. The formula is as follows:

[0034] F = [h t ; s] (7)

[0035] where F is the comprehensive feature vector, h t is the new time series feature, and s is the static feature.

[0036] Furthermore, step 4 is specifically as follows:

[0037] Step 4.1: Define the optimization objective; in the present invention, the optimization objective is to minimize the mean square error (MSE) and the model complexity. The model complexity is determined by the following two factors:

[0038] Number of BiLSTM units (lstm_units): It represents the number of neuron units in the BiLSTM layer.

[0039] Depth of XGBoost tree (xgb_depth): It represents the maximum depth of each tree in the XGBoost model.

[0040] The optimization objective function is defined as:

[0041] Minimize the mean square error (MSE): It measures the error between the model prediction result and the true value.

[0042] Minimize the model complexity: The model complexity is quantified by lstm_units + xgb_depth, where lstm_units represents the number of units in the BiLSTM model and xgb_depth represents the depth of the XGBoost tree. This formula reflects the linear relationship between the model complexity and these two hyperparameters. The formula is as follows:

[0043]

[0044] where y i is the actual value, is the predicted value, and n is the number of samples.

[0045] Step 4.2: Initialize the population; when initializing the population, a group of hyperparameter combinations of BiLSTM and XGBoost are randomly generated as the initial population. Each individual consists of the following hyperparameters:

[0046] BiLSTM hyperparameters: Number of hidden layer units, learning rate, batch size.

[0047] XGBoost hyperparameters: Maximum depth of the tree, learning rate, subsampling rate, number of trees.

[0048] Step 4.3: Perform non-dominated sorting on each individual in the population to form different ranks (Fronts). Non-dominated sorting is used to compare the quality of each solution, ensuring a certain degree of diversity among the solutions in the optimization process. Within each rank, calculate the crowding distance of each individual. The crowding distance measures the density of individuals in the objective space. By maintaining the diversity of the population, the crowding distance helps us avoid premature convergence and ensures the exploration of more potential excellent solutions.

[0049] Step 4.4: Selection, crossover, and mutation operations; based on non-dominated sorting and crowding distance, select high-quality individuals for crossover and mutation operations to generate a new generation of the population.

[0050] Step 4.5: Iterative optimization, repeat Steps 4.3 to 4.4 until a predetermined termination condition is met, such as reaching the maximum number of iterations or the population convergence criterion. In each generation, the quality of the population should gradually improve, thus achieving better results in multi-objective optimization.

[0051] Step 4.6: Select the optimal hyperparameter combination, select the optimal hyperparameter combination from the Pareto front solution set to achieve the balance between the model prediction accuracy and complexity.

[0052] Furthermore, Step 5 is specifically as follows:

[0053] Step 5.1: Train the BiLSTM model using the BiLSTM hyperparameters (such as the number of hidden layer units, learning rate, batch size) optimized by NSGA-II. The BiLSTM model can learn features from time series data and extract deep representations of time series data to capture the temporal relationships in the data.

[0054] Step 5.2: Train the XGBoost model; use the optimized hyperparameters (maximum depth of the tree, learning rate, subsampling rate, number of trees) to train the XGBoost model to model the relationship between static features and the target variable.

[0055] Step 5.3: Fuse features for comprehensive prediction; fuse the features extracted by BiLSTM and XGBoost and input them into the final regression layer to output the settlement prediction value; the formula is as follows:

[0056]

[0057] In the formula, F is the fused feature matrix, is the predicted settlement amount.

[0058] Furthermore, Step 6 is specifically as follows:

[0059] Step 6.1: Model Evaluation: Evaluate the prediction performance of the optimized model on the test set, including prediction accuracy (MSE, MAE, RMSE, R2) and computational complexity (training time, prediction time). The formulas are as follows:

[0060]

[0061] Step 6.2: Application of the Prediction Model: Apply the optimized BiLSTM-XGBoost model to actual engineering monitoring for real-time settlement prediction and decision support. y is the actual observed value.

[0062] The present invention has the following beneficial technical effects compared with the prior art.

[0063] 1. The present invention models time series data through the BiLSTM network to capture the dynamic change characteristics in the time series, and at the same time combines the XGBoost model to accurately model static features (such as soil properties, tunneling parameters), solving the problem of efficient fusion of multi-source heterogeneous features.

[0064] 2. The present invention uses NSGA-II (Non-dominated Sorting Genetic Algorithm II) to optimize the hyperparameters of BiLSTM and XGBoost, taking into account both the model prediction accuracy (minimizing the mean square error MSE) and complexity (improving the operation efficiency while minimizing the model complexity), and realizing flexible adaptation to the requirements of different engineering scenarios.

[0065] 3. Through the collaborative modeling of BiLSTM and XGBoost, the present invention realizes the comprehensive modeling of dynamic influencing factors and static environmental factors during the construction process. The optimized model can significantly improve the accuracy of settlement prediction, and reduce the training time and computational complexity, providing important support for real-time prediction and decision-making in engineering construction.

[0066] 4. The present invention not only overcomes the deficiencies of traditional prediction methods in non-linear modeling and feature fusion, but also effectively improves the balance ability of prediction accuracy and efficiency through multi-objective optimization, providing an efficient and reliable technical means for the safety management of complex engineering construction. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 It is a schematic flow chart of the BiLSTM-XGBoost settlement prediction method based on the NGSA-II multi-objective optimization algorithm of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0068] The present invention will be further described in detail below with reference to the drawings and specific implementation methods.

[0069] The process of a BiLSTM-XGBoost settlement prediction method based on the NGSA-II multi-objective optimization algorithm of the present invention is as follows Figure 1 shown, and specifically includes the following steps:

[0070] Step 1: Preprocessing of data collected from engineering monitoring.

[0071] Step 1.1: Filling missing values and normalizing the data collected from monitoring. For missing values, interpolation based on adjacent data points is used for completion, and the formula is as follows:

[0072]

[0073] In the formula, x t is the missing value, and x t-1 and x t+1 are adjacent non-missing data time points. For static features, if the proportion of missing values is less than 10%, median filling is used; if the proportion of missing values exceeds 40%, abnormal samples are removed.

[0074] Min-Max normalization is used for data scaling to scale all feature values to the interval [0, 1], and the formula is as follows:

[0075]

[0076] Among them, x is the original feature value, x' is the normalized feature value, and min(X) and max(X) are the minimum and maximum values of the entire feature set respectively.

[0077] Step 1.2: Preprocessing of the input data for the comparison and analysis model. For unknown engineering data, first, it is normalized through the data preprocessing module, and then the time series data is segmented using the sliding window method to generate a feature matrix that conforms to the model input format.

[0078] The specific steps are as follows: The time series data is segmented by the sliding window method at a fixed length (such as 10 minutes to 1 hour), and the step size is half of the window length. Feature selection and dimensionality reduction are performed on the data within each time window to form a feature matrix suitable for model input.

[0079] Step 2: Use ADF (Augmented Dickey-Fuller) for time series feature screening.

[0080] Step 2.1: Monitor and extract the time-series characteristics and static characteristics of the collected data. Taking a tunnel excavated by the shield method as an example, various characteristics are extracted from the preprocessed data, including time-series characteristics and static characteristics. The time-series characteristics include chamber pressure, propulsion speed, grouting pressure, cutter head rotation speed, etc., and their sampling frequency is once per minute. The static characteristics include the number of tunneling rings, soil layer depth, soil properties, initial settlement, etc., which are collected once for each construction link.

[0081] Step 2.2: Use the ADF test to evaluate the stationarity of the time-series characteristics; conduct the ADF unit root test on the extracted time-series characteristics to determine their stationarity. The formula is as follows:

[0082] Δy t = α + βt + γy t-1 + δ1Δy t-1 +…+ δ p-1 Δy t-p+1 + ∈ t (3)

[0083] In the formula, y t is the time-series characteristic, Δ is the difference operation, α, β, γ, δ are parameters to be estimated, and ∈ t is the error term. If the test statistic is less than the critical value, the time-series characteristic is considered stationary.

[0084] The screening condition is: retain the characteristics with p-value ≤ 0.5 as the input for time-series modeling. The reason for this threshold selection is that: p-value ≤ 0.05 only screens out completely stationary characteristics, which may ignore some weakly non-stationary characteristics that are meaningful for the model; p-value ≤ 0.5 can capture the influence of weakly non-stationary characteristics on short-term trend changes and retain important time-series information. For the characteristics with p-value > 0.5, the first-order difference method is used for processing. The formula is as follows:

[0085] y′ t = y t - y t-1 (4)

[0086] In the formula: y′ t is the time-series characteristic after differentiation.

[0087] There are a total of 18 original characteristics, namely time, number of tunneling rings, cutter head rotation speed, screw conveyor rotation speed, left upper chamber pressure, left upper grouting pressure, right upper grouting pressure, right lower grouting pressure, left lower grouting pressure, propulsion speed, cutter head torque, screw conveyor torque, total propulsion force, guide horizontal difference, guide vertical difference, foam concentrate, cumulative amount of grouting liquid a in the current ring (L), depth. Among them, there are 8 time-series characteristics, namely cumulative amount of grouting liquid a in the current ring (L), foam concentrate, screw conveyor torque, guide horizontal difference, left upper grouting pressure, left lower grouting pressure, right upper grouting pressure, right lower grouting pressure, etc.

[0088] Step 3: Use BiLSTM (Bi-directional Long Short-Term Memory) to generate new features and fuse them with static features.

[0089] Step 3.1: Generate new temporal features of the monitored acquisition data. It mainly includes three steps: sliding window segmentation, BiLSTM feature extraction, and pooling and feature mapping:

[0090] Sliding window segmentation: Use the sliding window method to divide the temporal features into time windows of a fixed length. The commonly used window length is from 10 minutes to 1 hour, and the step size is half of the window length.

[0091] BiLSTM feature extraction: Build a bidirectional long short-term memory network (BiLSTM) model. Input the preprocessed temporal data, and simultaneously capture the forward and backward dependencies in the temporal data through the forward and backward LSTM layers to generate a new temporal feature representation. The main parameter settings of BiLSTM are as follows: the number of hidden layer units is 64 or 128 (adjusted according to the data volume), the time step is the same as the sliding window size, the activation function is tanh, and the optimizer is Adam (the initial learning rate is 0.001). BiLSTM generates bidirectional hidden states at each time step, and the formula is as follows:

[0092]

[0093] In the formula, and are the hidden states of the forward and backward LSTM respectively, and h t is the concatenation of the bidirectional hidden states.

[0094] Pooling and feature mapping: Perform a pooling operation (such as average pooling) on the temporal feature matrix output by BiLSTM to convert it into a vector of a fixed length. The formula is as follows:

[0095]

[0096] In the formula, h t is the feature value at the t-th time step, and T is the number of time steps. Use a fully connected layer to further reduce the dimension, and the output vector length is the same as the static feature (the set value is 64), and the activation function is selected as ReLU.

[0097] Step 3.2: Fuse the new temporal features and static features. Concatenate the new temporal features extracted from BiLSTM with the static features to form a fused feature matrix. The formula is as follows:

[0098] F = [h t ; s] (8)

[0099] Wherein, F is the comprehensive feature vector, h t is the new time series feature, and s is the static feature.

[0100] Step 4: Use NSGA-II (Non-dominated Sorting Genetic Algorithm II) to optimize the hyperparameters of BiLSTM and XGBoost.

[0101] Step 4.1: Define the optimization objective. In the present invention, the optimization objective is to minimize the mean squared error (MSE) and the model complexity. The model complexity is determined by the following two factors:

[0102] Number of BiLSTM units (lstm_units): It represents the number of neuron units in the BiLSTM layer.

[0103] Depth of XGBoost tree (xgb_depth): It represents the maximum depth of each tree in the XGBoost model.

[0104] The optimization objective function is defined as:

[0105] Minimize the mean squared error (MSE): It measures the error between the model prediction result and the true value.

[0106] Minimize the model complexity: The model complexity is quantified by lstm_units + xgb_depth, where lstm_units represents the number of units in the BiLSTM model and xgb_depth represents the depth of the XGBoost tree. This formula reflects the linear relationship between the model complexity and these two hyperparameters. The formula is as follows:

[0107]

[0108] Wherein, y i is the actual value, is the predicted value, and n is the number of samples. The model complexity is determined by calculating the total number of model parameters.

[0109] Step 4.2: Initialize the population. When initializing the population, a set of hyperparameter combinations of BiLSTM and XGBoost is randomly generated as the initial population. Each individual consists of the following hyperparameters:

[0110] BiLSTM hyperparameters: Number of hidden layer units, learning rate, batch size.

[0111] XGBoost hyperparameters: Maximum depth of the tree, learning rate, subsampling rate, number of trees.

[0112] Each individual in the population represents a set of hyperparameters of BiLSTM and XGBoost, and the selection of these hyperparameters determines the performance and complexity of the model.

[0113] Step 4.3: Perform non-dominated sorting on each individual in the population to form different ranks (Fronts). Non-dominated sorting is used to compare the quality of each solution, ensuring a certain degree of diversity among each solution in the optimization process. Within each rank, calculate the crowding distance of each individual. The crowding distance measures the density of individuals in the objective space. By maintaining the diversity of the population, the crowding distance helps us avoid premature convergence and ensures the exploration of more potential excellent solutions.

[0114] Step 4.4: Selection, crossover, and mutation operations. Based on non-dominated sorting and crowding distance, select high-quality individuals for crossover and mutation operations to generate a new generation of the population. The specific operations are as follows:

[0115] Selection: Adopt the tournament selection method, randomly select several individuals from the current population, and select individuals with higher fitness to enter the next generation.

[0116] Crossover: Randomly select two parent individuals, exchange some hyperparameters, and generate two offspring individuals.

[0117] Mutation: Randomly mutate the hyperparameters of some offspring individuals to increase the diversity of the population.

[0118] Through these operations, the population gradually evolves and continuously approaches the optimal solution or the Pareto front.

[0119] Step 4.5: Iterative optimization. Repeat steps 4.3 to 4.4 until the predetermined termination conditions are met, such as reaching the maximum number of iterations or the population convergence criterion. In each generation, the quality of the population should be gradually improved, thus achieving better results in multi-objective optimization.

[0120] Step 4.6: Select the optimal hyperparameter combination; select the optimal hyperparameter combination from the Pareto front solution set to achieve the balance between the model prediction accuracy and complexity. The specific selection can be based on the actual engineering requirements, and select the hyperparameter combination with the minimum MSE and the model complexity within an acceptable range.

[0121] Step 5: Based on the optimized hyperparameters, train the fused BiLSTM-XGBoost model to achieve efficient prediction of settlement.

[0122] Step 5.1: Use the BiLSTM hyperparameters (such as the number of hidden layer units, learning rate, batch size) optimized by NSGA-II to train the BiLSTM model. The BiLSTM model can learn features from time series data and extract the deep representation of time series data to capture the temporal relationships in the data.

[0123] Step 5.2: Train the XGBoost model; use the optimized XGBoost hyperparameters (such as the maximum depth of the tree, learning rate, subsampling rate, number of trees) to train the XGBoost model and model the relationship between static features and the target variable.

[0124] Step 5.3: Integrate features for comprehensive prediction; integrate the features extracted by BiLSTM and XGBoost and input them into the final regression layer to output the settlement prediction value. The formula is as follows:

[0125]

[0126] In the formula, F is the integrated feature matrix, is the predicted settlement.

[0127] Step 6: Evaluate and apply the prediction model.

[0128] Step 6.1: Model evaluation. Evaluate the prediction performance of the optimized model on the test set, including prediction accuracy (such as MSE, MAE, RMSE, R2) and computational complexity (such as training time, prediction time). The formula is as follows:

[0129]

[0130] Step 6.2: Model application. Apply the optimized BiLSTM-XGBoost model to actual engineering construction for real-time settlement prediction and decision support.

[0131] Example:

[0132] The final settlement data of the longitudinal surface vertical displacement monitored from a certain shield tunnel and the tunneling parameters of the corresponding rings. The input parameters are respectively the tunneling ring number, cutter head rotation speed, screw conveyor rotation speed, left upper soil chamber pressure, left upper grouting pressure, right upper grouting pressure, right lower grouting pressure, left lower grouting pressure, propulsion speed, cutter head torque, screw conveyor torque, total propulsion force, guiding horizontal difference, guiding vertical difference, foam concentrate, cumulative amount of grouting liquid a in the current ring (L), buried depth and other 17 parameters; the output parameter is the final settlement data of the vertical displacement monitoring points. A total of 53 samples of surface settlement datasets are constructed, and each sample consists of 13 input parameters and 1 output parameter. The training set and the test set are divided according to the ratio of 8:2. First, the collected engineering monitoring data is preprocessed, including missing value filling (using the adjacent data interpolation method) and normalization (Min-Max normalization). Then, the time series data is segmented into small blocks of fixed length by the sliding window method, and the time series features are extracted. The stationary features are screened through the ADF unit root test, and the non-stationary features are differenced. Subsequently, a BiLSTM model is constructed to extract the deep features of the time series data, and pooling is performed and static features are fused. The NSGA-II algorithm is used to optimize the hyperparameters of BiLSTM and XGBoost, and the optimization goal is to minimize the MSE and the model complexity. The optimized BiLSTM and XGBoost models are trained, and the settlement prediction is carried out by combining the extracted features. Finally, the MSE, MAE, RMSE and R 2 and other performance indicators are evaluated on the test set (i.e., the data of 10 monitoring points in Table 1).

[0133] The prediction and measured results of the present invention and each comparative scheme are shown in Table 1.

[0134] Table 1 Prediction and Measured Results of Embodiments

[0135]

[0136] The performance evaluation of the present invention and the comparative algorithms is shown in Table 2.

[0137] Table 2 Performance Evaluation Results of the Present Invention and Comparative Algorithms

[0138] Prediction algorithm MSE MAE RMSE R2 NSGA-II-BiLSTM-XGBoost 0.0400 0.2000 0.2000 0.9998 BiLSTM 1.6140 1.2400 1.2704 0.9921 XGBoost 0.1230 0.3300 0.3507 0.9994 BP neural network 2.2570 1.4900 1.5023 0.9889 LSTM 0.6290 0.7700 0.7931 0.9969

[0139] It can be seen from the experimental results that the NSGA-II-BiLSTM-XGBoost model is superior to other models in terms of four indicators: mean absolute error (MAE), mean square error (MSE), root mean square error (RMSE) and regression value (R 2 ). The prediction result of the NSGA-II-BiLSTM-XGBoost model has high reliability and can be used for real-time feedback to adjust construction parameters during shield construction.

[0140] It should be noted that although the above embodiments have specifically illustrated the technical solutions of the present invention, however, the present invention is not limited to the above specific embodiments. For those of ordinary skill in the art, various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and these variations all fall within the protection scope of this application.

Claims

1. A BiLSTM-XGBoost settlement prediction method based on the NGSA-II multi-objective optimization algorithm, characterized in that, It includes the following steps: Step 1: Preprocessing of engineering monitoring collected data; Step 2: Using ADF for time series feature screening; Step 3: Generating new features using BiLSTM and fusing them with static features; Step 4: Using NSGA-II to optimize the hyperparameters of BiLSTM and XGBoost; Step 5: Based on the optimized hyperparameters, training the fused BiLSTM-XGBoost model to achieve efficient prediction of settlement; Step 6: Evaluation and application of the prediction model.

2. The BiLSTM-XGBoost settlement prediction method based on the NGSA-II multi-objective optimization algorithm according to claim 1, characterized in that, The specific content of Step 1 is as follows: Step 1.1: Filling missing values and normalizing the monitoring collected data; The missing values in the monitoring data are filled using the interpolation method, and the filled data is normalized to eliminate the influence of different dimensions on model training. The formula is as follows: In the formula, x is the original data, x′ is the normalized data, and min(X) and max(X) are the minimum and maximum values of the dataset respectively; Step 1.2: Comparative analysis of model input data preprocessing; Feature engineering processing is performed on the time series data and static feature data respectively. The principal component analysis method is used for feature selection. The formula is as follows: Z = XW (2) In the formula, Z is the feature after dimensionality reduction, X is the original feature matrix, and W is the dimensionality reduction matrix.

3. A BiLSTM-XGBoost settlement prediction method based on the NGSA-II multi-objective optimization algorithm according to claim 1, characterized in that, The specific content of Step 2 is as follows: Step 2.1: Extracting time series features and static features of the monitoring collected data: Time series features and static features are extracted from the preprocessed data. Time series features include but are not limited to construction progress and monitoring data, and static features include but are not limited to soil properties and tunneling parameters; Step 2.2: Using ADF test to evaluate the stationarity of time series features: An ADF unit root test is performed on the extracted time series features to determine their stationarity. The formula is as follows: Δy t = α + βt + γy t-1 + δ1Δy t-1 + … + δ p-1 Δy t-p+1 + ∈ t (3) where y t is the time series feature, Δ is the difference operation, α, β, γ, δ are parameters to be estimated, and ∈ t is the error term; If the test statistic is less than the critical value, it is considered that the time series features are stationary.

4. A BiLSTM-XGBoost settlement prediction method based on the NGSA-II multi-objective optimization algorithm according to claim 1, characterized in that, The specific content of Step 3 is as follows: Step 3.1: Generating new time series features of the monitoring collected data; Constructing a bidirectional long short-term memory network BiLSTM model, inputting the preprocessed time series data, and simultaneously capturing the forward and backward dependencies in the time series data through the forward and backward LSTM layers to generate a new time series feature representation. The formula is as follows: where x t is a missing value, and are the hidden states of the forward and backward LSTMs respectively, and h t is the concatenation of the bidirectional hidden states; Step 3.2: Fuse the temporal features generated by BiLSTM with the static features; the generated new temporal feature h t is fused with the static feature s to form a comprehensive feature vector F. The fusion method uses concatenation or weighted average, and the formula is as follows: F = [h t ; s] (7) where F is the comprehensive feature vector, h t is the new time-series feature, and s is the static feature.

5. A BiLSTM-XGBoost settlement prediction method based on the NGSA-II multi-objective optimization algorithm according to claim 1, wherein, The specific content of Step 4 is as follows: Step 4.1: Defining the optimization objective; The optimization objective is to minimize the mean squared error MSE and the model complexity. The model complexity is determined by the following two factors: The number of BiLSTM units lstm_units: Represents the number of neuron units in the BiLSTM layer; The depth of the XGBoost tree xgb_depth: Represents the maximum depth of each tree in the XGBoost model; The optimization objective function is defined as: Minimizing the mean squared error MSE: Measures the error between the model prediction result and the true value; Minimizing the model complexity: The model complexity is quantified by lstm_units + xgb_depth, where lstm_units represents the number of units in the BiLSTM model and xgb_depth represents the depth of the XGBoost tree. The formula is as follows: where y i is the actual value, is the predicted value, and n is the number of samples; Step 4.2: Initializing the population; Randomly generating a set of hyperparameter combinations as the initial population; Step 4.3: Perform non-dominated sorting on each individual in the population to form different levels of Front; non-dominated sorting is used to compare the pros and cons of each solution to ensure that each solution in the optimization process has a certain degree of diversity; within each level, calculate the crowding distance of each individual; crowding distance measures the density of individuals in the target space; by maintaining the diversity of the population, crowding distance helps us avoid early convergence and ensures the exploration of more potential excellent solutions; Step 4.4: Selection, crossover and mutation operations; According to the non-dominated sorting and crowding distance, select high-quality individuals for crossover and mutation operations to generate a new generation of population; Step 4.5: Iterative optimization, repeating steps 4.3 to 4.4 until the predetermined termination condition is met, such as reaching the maximum number of iterations or the population convergence criterion; in each generation, the quality of the population should be gradually improved, thereby achieving better results in multi-objective optimization; Step 4.6: Select the optimal hyperparameter combination. Select the optimal hyperparameter combination from the Pareto frontier solution set to achieve a balance between model prediction accuracy and complexity.

6. A BiLSTM-XGBoost settlement prediction method based on the NGSA-II multi-objective optimization algorithm according to claim 1, characterized in that, The step 5 is specifically as follows: Step 5.1: Use the BiLSTM hyperparameters optimized by NSGA-II to train the BiLSTM model; the BiLSTM model can learn features from time series data and extract deep representations of time series data to capture the time series relationships in the data; Step 5.2: Train the XGBoost model; use the optimized hyperparameters including the maximum depth of the tree, learning rate, subsampling rate, and number of trees to train the XGBoost model to model the relationship between static features and the target variable; Step 5.3: Fusion features for comprehensive prediction; Fusion of features extracted by BiLSTM and XGBoost, input to the final regression layer, output of settlement prediction value; formula is as follows: Where F is the fused feature matrix, is the predicted settlement.

7. A BiLSTM-XGBoost settlement prediction method based on the NGSA-II multi-objective optimization algorithm according to claim 1, wherein The step 6 is specifically as follows: Step 6.1: Model evaluation: Evaluate the prediction performance of the optimized model on the test set, including prediction accuracy: MSE, MAE, RMSE, R2, and computational complexity: training time, prediction time. The formula is as follows: Step 6.2: Application of prediction model: Apply the optimized BiLSTM-XGBoost model to actual engineering monitoring for real-time settlement prediction and decision support.

Images