Soft soil area roadbed settlement prediction system and method

Abstract

The invention relates to the field of geotechnical engineering, in particular to a soft soil area roadbed settlement prediction system and method which are applied to soft soil foundation roadbed settlement monitoring and prediction of traffic infrastructures such as highways and railways. The system comprises five modules of data acquisition, preprocessing, model construction, prediction analysis and feedback optimization. The data acquisition module is responsible for collecting roadbed parameters and settlement historical data; the preprocessing module carries out abnormal value elimination and normalization processing; the model building module is used for building a prediction model based on a gradient lifting decision tree algorithm, and the prediction analysis module is used for carrying out settlement prediction by utilizing the processed data and the model; the feedback optimization module compares a prediction result with actually measured data and optimizes the model; compared with a traditional method, the prediction precision is remarkably improved, and the average error is reduced by more than 40%; in addition, the system has strong environmental adaptability through multi-model parallel prediction and AI-driven parameter adaptive correction, and can automatically adjust a prediction strategy to adapt to different geological conditions.

Description

Technical Field

[0001] The present invention relates to the field of geotechnical engineering, and particularly to a subgrade settlement prediction system and method in soft soil areas, which are applied to the settlement monitoring and prediction of soft soil subgrades of transportation infrastructure such as highways and railways. Background Art

[0002] Soft soil areas have characteristics such as high compressibility, low bearing capacity, and deformation sensitivity, which make the subgrades built in such areas often face serious settlement problems. Accurately predicting the subgrade settlement in soft soil areas is of great significance for engineering design, construction control, and later maintenance.

[0003] Traditional subgrade settlement prediction methods mainly rely on empirical formulas or simple numerical simulations. These methods usually have the following problems: First, traditional methods are difficult to effectively handle the complex geological conditions in soft soil areas, resulting in a large deviation between the prediction results and the actual situation; Second, the existing prediction models have insufficient adaptability and cannot dynamically adjust the prediction parameters according to real-time monitoring data; Third, traditional methods lack comprehensive consideration of the coupled effects of multiple parameters and are difficult to accurately reflect the complex mechanism of subgrade settlement; Finally, there is a lack of an effective feedback adjustment mechanism and the prediction model cannot be optimized according to the measured data.

[0004] With the development of artificial intelligence technology, it has become possible to apply advanced machine learning algorithms to the field of subgrade settlement prediction. However, there is still a lack of a dedicated prediction system for the particularity of soft soil areas and a comprehensive solution that integrates physical models and data-driven models. In view of the above problems, it is urgent to develop a high-precision prediction system and method specifically for subgrade settlement in soft soil areas. Summary of the Invention

[0005] The purpose of the present invention is to provide a subgrade settlement prediction system and method in soft soil areas, aiming to solve the problems of insufficient prediction accuracy, poor model adaptability, and lack of real-time feedback optimization mechanism in the prior art.

[0006] The present invention realizes high-precision prediction of subgrade settlement in soft soil areas by establishing a complete technical system of data collection, preprocessing, model construction, prediction analysis, and feedback optimization, and adopting the gradient boosting decision tree algorithm combined with physical model fusion technology. The system has key technologies such as multi-dimensional data collection, intelligent data processing, adaptive model construction, multi-model parallel prediction, and parameter correction based on artificial intelligence, forming a closed-loop optimization architecture of "collection - processing - modeling - prediction - feedback".

[0007] The present invention provides a subgrade settlement prediction system in soft soil areas, including:

[0008] A data collection module, which is used to collect subgrade parameters and settlement historical data in soft soil areas;

[0009] A data preprocessing module, communicatively connected to the data acquisition module, for removing outliers and normalizing the collected data;

[0010] A model construction module, communicatively connected to the data preprocessing module, for constructing a roadbed settlement prediction model based on the gradient boosting decision tree algorithm;

[0011] A prediction and analysis module, communicatively connected to the model construction module, for performing settlement prediction according to the processed data and the constructed model;

[0012] A feedback and optimization module, communicatively connected to the prediction and analysis module, for comparing the prediction results with the measured data and optimizing the prediction model.

[0013] Preferably, the data acquisition module includes:

[0014] A roadbed parameter acquisition unit, for acquiring the filling height, soil type and design type of the roadbed in the soft soil area;

[0015] A geological exploration data acquisition unit, for acquiring the soil layer depth and soil layer longitudinal section diagram in the soft soil area;

[0016] A settlement history data acquisition unit, for acquiring the measured roadbed settlement values at different time nodes;

[0017] A real-time monitoring unit, for performing real-time monitoring of the roadbed settlement and acquiring the current settlement data.

[0018] Preferably, the data preprocessing module includes:

[0019] An outlier processing unit, for identifying and removing outliers, missing values and duplicate values in the data;

[0020] A data conversion unit, for converting the preprocessed data into an ordered list form;

[0021] A normalization processing unit, for mapping the data to the interval [0, 1] by using the maximum-minimum normalization method;

[0022] A correlation analysis unit, for analyzing the linear correlation coefficient between the data, and establishing a relationship between the parameters when the correlation coefficient is greater than a preset threshold.

[0023] Preferably, the model construction module includes:

[0024] A feature extraction unit, for extracting the input features required for the prediction model from the preprocessed data;

[0025] A GBDT algorithm unit, for training weak classifiers iteratively and constructing a gradient boosting decision tree model;

[0026] A tree structure optimization unit for controlling the model complexity through the tree structure complexity parameter γ and the leaf node penalty parameter λ;

[0027] A physical model fusion unit for fusing physical test results and numerical simulation results to enhance the physical meaning of the model.

[0028] Preferably, the GBDT algorithm unit is specifically used for:

[0029] Initialize the model parameters, let t = 1;

[0030] Train a weak classifier for each sample point i where the weak classifier determines the optimal splitting point by calculating the average loss of the samples in each leaf region;

[0031] Calculate the total error and update the tree structure complexity parameter γ and the leaf node penalty parameter λ;

[0032] Optimize the model through the iterative process extended by the Taylor formula;

[0033] Judge whether the termination condition is satisfied. If satisfied, output the final model H(x).

[0034] Preferably, the prediction analysis module includes:

[0035] An input processing unit for converting the parameters of the subgrade to be predicted into the input format required by the model;

[0036] A multi-model prediction unit for obtaining multiple sets of prediction results through parallel calculations of multiple prediction models;

[0037] A result evaluation unit for evaluating the errors of each prediction result and selecting the result with the smallest error as the final prediction value;

[0038] A settlement curve generation unit for generating a relationship curve of subgrade settlement varying with time according to the final prediction value.

[0039] Preferably, the feedback optimization module includes:

[0040] An error calculation unit for calculating the error between the predicted value and the measured value;

[0041] An AI parameter correction unit for automatically correcting the parameters based on the artificial intelligence algorithm when the error is greater than the preset threshold;

[0042] A historical data correction unit for correcting the model input parameters based on the error between the historical prediction data and the measured data;

[0043] A model update unit for optimizing and updating the prediction model according to the correction result.

[0044] Preferably, the AI ​​parameter correction unit is specifically used for:

[0045] Establish a mapping relationship between prediction parameters and errors;

[0046] Identify the key parameters that have the greatest impact on forecast results;

[0047] Dynamically adjust the weights of key parameters based on real-time settlement data;

[0048] The corrected parameters are fed back to the model building module for model optimization.

[0049] Preferably, it further comprises a modeling environment module, which is in communication connection with the model building module and is used for:

[0050] Storing multiple sets of modeling environment information, including modeling environment data, modeling environment analysis data, and modeling environment determination data;

[0051] Select the optimal modeling environment through similarity comparison analysis method;

[0052] Among them, the similarity comparison and analysis method includes calculating the similarity coefficient and correlation between the modeling environment data information and the modeling environment information, and selecting the optimal solution based on the weighted results.

[0053] A method for predicting roadbed settlement in soft soil areas comprises the following steps:

[0054] S1. Collect roadbed parameters and settlement history data in soft soil areas;

[0055] S2, removing outliers and normalizing the collected data;

[0056] S3, constructing a roadbed settlement prediction model based on the gradient boosting decision tree algorithm;

[0057] S4. Predict settlement based on the processed data and the constructed model;

[0058] S5, comparing the predicted results with the measured data;

[0059] S6. When the prediction error is greater than a preset threshold, the model parameters are corrected based on the artificial intelligence algorithm;

[0060] S7. updating the prediction model according to the correction result;

[0061] S8, re-predicting the settlement using the updated model until the prediction error is less than a preset threshold;

[0062] S9. Output the final roadbed settlement prediction results, including the settlement amount and the relationship curve between settlement and time.

[0063] The beneficial effects of the present invention are as follows: First, compared with traditional prediction methods, the prediction accuracy is significantly improved, and the average prediction error is reduced by more than 40%; Second, through parallel prediction of multiple models and AI-driven parameter adaptive correction, the system has strong environmental adaptability and can automatically adjust the prediction strategy according to different geological conditions; Third, a closed-loop feedback mechanism based on real-time monitoring data is established, which can continuously optimize the prediction model and improve the reliability of long-term prediction; Finally, the special algorithms and processing procedures designed for the special characteristics of the soft soil area significantly enhance the application effect under complex geological conditions. Brief Description of the Drawings

[0064] Figure 1 It is the overall architecture diagram of the subgrade settlement prediction system in the soft soil area of the present invention;

[0065] Figure 2 It is the structural schematic diagram of the data acquisition module of the present invention;

[0066] Figure 3 It is the functional flow chart of the data preprocessing module of the present invention;

[0067] Figure 4 It is the structural schematic diagram of the model construction module of the present invention;

[0068] Figure 5 It is the training flow chart of the GBDT algorithm of the present invention;

[0069] Figure 6 It is the functional structure diagram of the prediction and analysis module of the present invention;

[0070] Figure 7 It is the working flow chart of the feedback optimization module of the present invention;

[0071] Figure 8 It is the functional structure diagram of the modeling environment module of the present invention;

[0072] Figure 9 It is the flow chart of the subgrade settlement prediction method in the soft soil area of the present invention;

[0073] Figure 10 It is the comparison schematic diagram of the prediction result and the measured data of the present invention;

[0074] Figure 11 It is the architecture diagram of the multi-scale spatio-temporal feature extraction and fusion technology of the present invention;

[0075] Figure 12 It is the working principle diagram of the spatio-temporal self-attention mechanism of the present invention. Detailed Embodiments

[0076] Please refer to the attached Figures 1-12 drawings, and the present invention will be further described in detail below in conjunction with the drawings and specific embodiments.

[0077] As Figure 1 shown, a subgrade settlement prediction system provided by the present invention mainly includes: a data acquisition module 1, a data preprocessing module 2, a model construction module 3, a prediction analysis module 4, a feedback optimization module 5, and a modeling environment module 6. Each module forms a complete workflow through communication connection to achieve accurate prediction of subgrade settlement in soft soil areas.

[0078] As Figure 2 shown, the data acquisition module 1 includes a subgrade parameter acquisition unit 11, a geological exploration data acquisition unit 12, a settlement history data acquisition unit 13, and a real-time monitoring unit 14.

[0079] The subgrade parameter acquisition unit 11 is used to acquire basic parameters such as the filling height, soil type, and design type of the subgrade in the soft soil area. In practical applications, the filling height is usually measured by a total station with an accuracy of up to ±1 mm; the soil type is determined through on-site sampling and laboratory tests, mainly including indicators such as water content, density, liquid-plastic limit, etc.; the design type records the structural form of the subgrade, such as ordinary embankment, bridgehead transition section, soil replacement subgrade, etc.

[0080] The geological exploration data acquisition unit 12 is used to acquire the soil layer depth and soil layer longitudinal section diagram in the soft soil area. Preferably, the system uses the method of multi-point drilling combined with ground penetrating radar to obtain geological exploration data. The drilling point spacing is usually set to 50 - 100 m, and it can be appropriately densified or reduced according to the complexity of the geological conditions. The soil layer longitudinal section diagram is generated by professional geological modeling software and is used to visually reflect the underground soil layer distribution.

[0081] The settlement history data acquisition unit 13 is used to acquire the measured values of subgrade settlement at different time nodes. In an embodiment of the present invention, a settlement observation point is buried every 10 - 20 m along the cross-section on the subgrade surface, and the elevation changes of each observation point are measured regularly (usually once a week, which can be densified to once a day in the initial stage of construction) to generate a settlement-time curve.

[0082] The real-time monitoring unit 14 is used to monitor the subgrade settlement in real time and obtain the current settlement data. This unit uses an automated settlement monitoring system, including equipment such as precise level gauges, pressure sensors, and data collectors, which can achieve continuous 24-hour monitoring. The monitoring data is uploaded to the system database in real time through wireless transmission.

[0083] As Figure 3 shown, the data preprocessing module 2 includes an outlier processing unit 21, a data conversion unit 22, a normalization processing unit 23, and a correlation analysis unit 24.

[0084] The outlier processing unit 21 is used to identify and eliminate outliers, missing values, and duplicate values in the data. The present invention adopts an outlier detection method based on statistical principles. When the data deviates from the mean by more than 3 times the standard deviation, it is determined as an outlier and eliminated. For missing values, different processing strategies are adopted according to the data density: when the number of missing points is less than 5% of the total data volume, interpolation method is used to fill; when the number of missing points is relatively large, time series prediction method is used for reasonable estimation.

[0085] The data conversion unit 22 is used to convert the preprocessed data into an ordered list form. Specifically, this unit arranges the collected multi-dimensional data in chronological order to form a standardized two-dimensional table, where the rows represent different observation points, the columns represent different time nodes, and the cell values are the corresponding settlement amounts. This standardized data structure facilitates subsequent model training and prediction.

[0086] The normalization processing unit 23 is used to map the data to the interval [0, 1] by using the maximum-minimum normalization method. The normalization process uses the following formula:

[0087] ,

[0088] where, is the normalized value, is the original value, is the minimum value in the dataset, is the maximum value in the dataset. Normalization processing can eliminate the influence of different dimensions of different indicators and improve the model training efficiency and prediction accuracy.

[0089] The correlation analysis unit 24 is used to analyze the linear correlation coefficient between data, and establish a relationship between parameters when the correlation coefficient is greater than the preset threshold. The present invention adopts the Pearson correlation coefficient to evaluate the linear correlation between variables:

[0090] ,

[0091] where, is the correlation coefficient, and are the observed values of two variables, and are their respective means, is the sample size. According to experimental verification, when , it can be considered that there is a strong linear correlation between the two variables, and at this time, a linear relationship between parameters is established; when , the original data is directly used for subsequent processing.

[0092] Such as Figure 4As shown in the figure, the model construction module 3 includes a feature extraction unit 31, a GBDT algorithm unit 32, a tree structure optimization unit 33, and a physical model fusion unit 34.

[0093] The feature extraction unit 31 is used to extract the input features required for the prediction model from the preprocessed data. In the present invention, by combining expert knowledge and feature importance analysis, the most critical features for settlement prediction are screened out, including but not limited to: soft soil layer thickness, water content, compression index, subgrade filling height, filling rate, preloading time, etc. Preferably, the feature extraction process also includes feature combination and interaction feature generation to capture the non-linear relationship between different features.

[0094] The GBDT algorithm unit 32 is used to train weak classifiers iteratively and construct a gradient boosting decision tree model. Gradient Boosting Decision Tree (GBDT) is a powerful ensemble learning algorithm that combines multiple weak learners (usually decision trees) into a strong learner to achieve high-precision prediction. The GBDT algorithm adopted in the present invention has the following process:

[0095] 1) Initialize the model parameters, let , the initial model , where is the loss function, usually the mean square error.

[0096] 2) For each sample point train the weak classifier . First, calculate the negative gradient:

[0097] ,

[0098] 3) Fit a regression tree to obtain the leaf node region .

[0099] 4) For each leaf node region calculate the optimal prediction value:

[0100] ,

[0101] 5) Update the model:

[0102] ,

[0103] where, is the learning rate, usually set between 0.1 - 0.3, is the indicator function.

[0104] 6) Iterate steps 2 - 5 until the termination condition is reached (such as the maximum number of iterations or the error is less than the threshold).

[0105] The tree structure optimization unit 33 is used to control the model complexity through the tree structure complexity parameter γ and the leaf node penalty parameter λ. In practical applications, the maximum depth of the tree is usually set to 3-7. A shallower tree helps prevent overfitting; the minimum number of samples in a leaf node is set to 1%-5% of the total number of samples to ensure that each leaf node has sufficient sample support. The complexity parameter γ and the penalty parameter λ are determined by cross-validation to obtain the optimal values, and the typical values are 0.1-0.5 and 1-10 respectively.

[0106] The physical model fusion unit 34 is used to fuse the physical test results and the numerical simulation results to enhance the physical meaning of the model. The present invention uses a physical constraint method to guide the training of the data-driven model. Specifically, the theoretical settlement is calculated using classical theories in soil mechanics (such as Terzaghi's one-dimensional consolidation theory) as prior knowledge for model training; at the same time, numerical simulations are carried out through finite element software such as ANSYS, and the simulation results are used as auxiliary training data to improve the performance of the model in data-sparse regions.

[0107] As Figure 11 shown, the present invention also innovatively proposes a multi-scale spatio-temporal feature extraction and fusion technology, which is the core innovation of the feature extraction unit 31. This technology aims to capture the complex characteristics of subgrade settlement in soft soil areas from both time and space dimensions. Through multi-scale analysis and feature fusion, the accuracy and generalization ability of the prediction model are significantly improved.

[0108] The multi-scale spatio-temporal feature extraction and fusion technology includes the following core steps:

[0109] 1) Time-scale decomposition: The settlement time series is decomposed into components at three scales: short-term fluctuations, medium-term trends, and long-term evolutions. The present invention uses wavelet decomposition to achieve time-scale decomposition, and its mathematical expression is:

[0110] ,

[0111] where is the original time series, is the wavelet basis function, is the scaling function, is the detail coefficient, is the approximation coefficient, represents the decomposition level, represents the displacement parameter. The present invention preferably uses the db4 wavelet with the decomposition level J = 3, which can effectively separate the settlement characteristics at different time scales.

[0112] 2) Spatial scale feature extraction: Settlement features are extracted from three spatial scales: points, lines, and surfaces. At the point scale, the settlement characteristics of individual monitoring points are concerned; at the line scale, the settlement distribution law along the cross-section or longitudinal section is concerned; at the surface scale, the settlement spatial distribution pattern of the entire subgrade area is concerned. The present invention uses spatial interpolation and tensor decomposition methods to construct multi-scale spatial features:

[0113] Point scale features: Directly use the settlement data of the monitoring points and their first-order derivatives (settlement rate) and second-order derivatives (settlement acceleration).

[0114] Line scale features: Construct the settlement distribution function along the line profile through B-spline interpolation:

[0115] ,

[0116] where is the settlement amount at position x on the cross-section, is the k-th order B-spline basis function, is the control point. In the present invention, k = 3, that is, cubic B-spline is used to ensure the smoothness of the curve.

[0117] Surface scale features: Use the tensor decomposition method to extract the main patterns from the two-dimensional settlement distribution:

[0118] ,

[0119] where, is the settlement data tensor, and are the component vectors of the r-th CP (CANDECOMP / PARAFAC) decomposition respectively, 0 represents the outer product operation, is the decomposition rank (usually selected as 3 - 5).

[0120] 3) Spatiotemporal feature fusion: Fuse the time and space features of different scales through adaptive weights. The present invention uses a feature fusion method based on the attention mechanism to automatically learn the importance weights of different features:

[0121] ,

[0122] where, is the fused feature vector, is the i-th time scale feature, is the 1st spatial scale feature and are the attention weights of the corresponding features respectively, is the number of time scale features, is the number of spatial scale features. The attention weights are calculated as follows:

[0123] ,

[0124] ,

[0125] Among them, and are learnable parameter vectors.

[0126] The core advantages of the multi-scale spatio-temporal feature extraction and fusion technology are as follows: First, it can capture the settlement evolution laws at different time scales, including short-term fluctuations (such as the influence of day-night temperature difference), medium-term trends (such as seasonal changes), and long-term evolution (such as the consolidation process); Second, it can comprehensively consider the settlement characteristics at three spatial scales of points, lines, and surfaces to establish a more comprehensive spatial distribution model; Finally, through the attention mechanism, it realizes the adaptive fusion of features and automatically adjusts the importance weights of each feature according to different engineering scenarios and geological conditions.

[0127] The experimental results show that after adopting the multi-scale spatio-temporal feature extraction and fusion technology, the prediction accuracy of the prediction model at different time periods (construction period, initial consolidation period, long-term consolidation period) has been significantly improved, and the average prediction error has been further reduced by 18.5% compared with the traditional method. Especially, the advantage is more obvious in the long-term settlement prediction.

[0128] As Figure 12 shown, in order to further improve the effective utilization of multi-scale spatio-temporal features, the present invention innovatively introduces a spatio-temporal self-attention mechanism as the core technological innovation point of the prediction analysis module 4. This mechanism can automatically discover and utilize the long-term and short-term time dependence relationships and the near and far spatial correlation relationships in the settlement data, and significantly enhance the model's understanding ability of complex settlement patterns.

[0129] The core algorithm of the spatio-temporal self-attention mechanism is as follows:

[0130] 1) Temporal self-attention calculation: Capturing the dependence relationships between different time points

[0131] First, map the time series features into three representations: query, key, and value:

[0132] ,

[0133] ,

[0134] ,

[0135] Among them, is the time series feature matrix, and are learnable mapping matrices.

[0136] Then, calculate the attention weight matrix:

[0137] ,

[0138] where, is the feature dimension, which is used to scale the dot product to avoid the problem of vanishing gradients.

[0139] Finally, calculate the weighted features:

[0140] ,

[0141] where, is the feature representation after adding the temporal self-attention information.

[0142] 2) Spatial self-attention calculation: Capture the spatial correlation between different monitoring points

[0143] Similar to the temporal self-attention, the spatial self-attention mechanism is implemented through the following steps:

[0144] ,

[0145] ,

[0146] ,

[0147] ,

[0148] ,

[0149] where, is the spatial feature matrix, is the spatial distance matrix. Introducing can incorporate the geographical distance information into the attention calculation. The calculation formula of

[0150] ,

[0151] where, is the geographical distance between monitoring points i and j, is the learnable distance attenuation coefficient, usually initialized to 0.1.

[0152] 3) Spatiotemporal attention fusion: Integrate the attention information in the temporal and spatial dimensions

[0153] The present invention uses a gating mechanism to fuse the temporal and spatial attention features:

[0154] ,

[0155] ,

[0156] Among them, is the sigmoid activation function, represents the concatenation of temporal and spatial features, and are learnable parameters, represents element-wise multiplication, is the finally fused spatio-temporal feature.

[0157] 4) Multi-head attention: To improve the model's expressive ability, the present invention adopts a multi-head attention mechanism, that is, h independent attentions are calculated simultaneously, and then the results are combined:

[0158] ,

[0159] Among them, is the output of the i-th attention head, is the output mapping matrix, is the number of attention heads, which is set to 8 in the present invention.

[0160] The core advantages of the spatio-temporal self-attention mechanism are as follows: First, it can capture long-term dependence relationships and avoid the problem of gradient disappearance in traditional recurrent neural networks; second, it can automatically discover and utilize the spatial correlations between different monitoring points and effectively integrate regional settlement information; finally, through the gated fusion mechanism, it can adaptively integrate temporal and spatial information and adjust the importance weights of the two according to specific scenarios.

[0161] In practical applications, the spatio-temporal self-attention mechanism is usually used in combination with the GBDT algorithm to form a two-stage prediction framework: in the first stage, the spatio-temporal self-attention mechanism is used to extract high-order spatio-temporal features; in the second stage, these features are used as the input of the GBDT model to perform the final settlement prediction. This combination takes advantage of the complementary advantages of the two algorithms: the spatio-temporal self-attention mechanism is good at capturing complex spatio-temporal dependence relationships, while GBDT is good at dealing with non-linear mappings and heterogeneous features.

[0162] Experimental results show that after introducing the spatio-temporal self-attention mechanism, the model has significantly improved in terms of prediction accuracy and robustness: the average prediction error is further reduced by 15.3%; the ability to identify irregular settlement patterns (such as sudden settlement, stage settlement) is significantly enhanced; the anti-interference ability of the model to interference factors (such as weather changes, temporary loads) is also significantly improved.

[0163] As Figure 6 shown, the prediction analysis module 4 includes an input processing unit 41, a multi-model prediction unit 42, a result evaluation unit 43, and a settlement curve generation unit 44.

[0164] The input processing unit 41 is used to convert the parameters of the subgrade to be predicted into the input format required by the model. This unit first checks the input parameters to ensure data integrity and consistency; then applies the same preprocessing and normalization methods as the training data to ensure that the input features are in the same format as expected by the model; finally, fills in the missing features with reasonable default values or estimated values.

[0165] The multi-model prediction unit 42 is used to obtain multiple sets of prediction results through parallel calculations by multiple prediction models. The present invention adopts the idea of model integration and uses GBDT models with different parameters and structures for prediction at the same time. Preferably, 5-10 differentiated models are constructed, which are different in terms of training data subsets, feature selection, tree depth, etc., so as to capture different aspects of the data.

[0166] The result evaluation unit 43 is used to evaluate the errors of each prediction result and select the result with the smallest error as the final prediction value. The evaluation process is based on the historical prediction performance and comprehensively combines the prediction results of each model by means of weighted voting. The weight calculation formula is:

[0167] ,

[0168] where is the weight of the i-th model, is the historical error of this model on the validation set, is the total number of models. In this way, the models with better prediction performance obtain higher weights, improving the overall prediction accuracy.

[0169] The settlement curve generation unit 44 is used to generate a relationship curve of subgrade settlement changing with time according to the final prediction value. This unit adopts the method of numerical fitting and fits the settlement-time relationship through a hyperbolic function or an exponential function:

[0170] ,

[0171] or

[0172] ,

[0173] where is the settlement amount at time t, is the final settlement amount, and are fitting parameters. In this way, the system can not only predict the settlement amount at a specific time point, but also predict the settlement development trend, providing more comprehensive support for engineering decisions.

[0174] Such as Figure 7As shown, the feedback optimization module 5 includes an error calculation unit 51, an AI parameter correction unit 52, a historical data correction unit 53, and a model update unit 54.

[0175] The error calculation unit 51 is used to calculate the error between the predicted value and the measured value. The present invention uses two metrics, the mean absolute error (MAE) and the root mean square error (RMSE), to evaluate the prediction accuracy:

[0176] ,

[0177] ,

[0178] where, is the measured value, is the predicted value, is the number of samples. When the MAE or RMSE exceeds a preset threshold (usually 10% of the measured value), the model optimization process is triggered.

[0179] The AI parameter correction unit 52 is used to automatically correct the parameters based on the artificial intelligence algorithm when the error is greater than the preset threshold. This unit uses the Bayesian optimization algorithm to automatically adjust the model parameters. This algorithm efficiently explores the parameter space by establishing a probability model (Gaussian process) between the parameters and the model performance to find the optimal parameter combination. The key parameters include the learning rate, the maximum depth of the tree, the number of samples in the minimum leaf node, etc.

[0180] The historical data correction unit 53 is used to correct the model input parameters based on the error between the historical prediction data and the measured data. This unit establishes an error compensation mechanism and systematically adjusts the new prediction results by analyzing the pattern of the historical prediction error. Specifically, a mapping relationship between the error and the input features is established:

[0181] ,

[0182] where, is the prediction error, is the input feature. In this way, the system can learn and compensate for the systematic error of the model and improve the prediction accuracy.

[0183] The model update unit 54 is used to optimize and update the prediction model according to the correction results. This unit adopts an incremental learning strategy to integrate the new data and correction results into the existing model without retraining the entire model. Preferably, the model update adopts a sliding window mechanism, which pays more attention to the influence of the recent data, enabling the model to adapt to the changes in geological conditions and settlement laws.

[0184] Specific implementation of the AI parameter correction unit 52

[0185] The AI parameter correction unit 52 is specifically configured to establish a mapping relationship between prediction parameters and errors; identify the key parameters that have the greatest impact on the prediction results; dynamically adjust the weights of the key parameters according to real-time settlement data; and feedback the corrected parameters to the model construction module for model optimization.

[0186] In practical applications, the AI parameter correction unit 52 first determines the parameters that have the greatest impact on the prediction results through sensitivity analysis. Specifically, by using the method of controlling variables, the influence degree of each parameter on the prediction results within a certain range of changes is analyzed, and they are sorted according to the influence degree. Usually, the thickness of the soft soil layer, the compression index, and the load intensity are the three parameters that have the most significant influence on settlement prediction.

[0187] The parameter correction uses the gradient descent method to minimize the prediction error:

[0188] ,

[0189] where, is the model parameter vector, is the learning rate (usually set to 0.01 - 0.05), is the loss function, is the gradient of the loss function with respect to the parameters. To prevent overfitting, a regularization term is also introduced:

[0190] ,

[0191] where, is the regularization coefficient, usually set to 0.001 - 0.01.

[0192] To improve the efficiency of parameter correction, this unit also adopts an adaptive learning rate strategy. When the parameter update directions are the same for several consecutive times, the learning rate is increased; when the parameter update directions change frequently, the learning rate is decreased. This strategy significantly improves the convergence speed and stability of parameter optimization.

[0193] As Figure 8 shown, the present invention further includes a modeling environment module 6, which is communicatively connected to the model construction module 3 and is used to store multiple sets of modeling environment information, including modeling environment data, modeling environment analysis data, and modeling environment determination data; and selects the optimal modeling environment through a similarity comparison and analysis method.

[0194] The key function of the modeling environment module 6 is to select the most suitable modeling environment for the current prediction task from multiple candidate environments through a similarity comparison and analysis method. The similarity comparison uses the following formula:

[0195] ,

[0196] where, is the result of the similarity comparison and analysis, It is the number of the modeling environment data information, It is the number of the modeling environment information, It represents the modeling environment data information and the modeling environment information of the similarity coefficient, It represents their correlation, is the similarity weight (usually 0.6), is the correlation weight (usually 0.4).

[0197] The similarity coefficient The calculation formula is:

[0198] ,

[0199] where, and are respectively the maximum value and the minimum value of the modeling environment data information , and are respectively the maximum value and the minimum value of the modeling environment information .

[0200] The correlation The calculation formula is:

[0201] ,

[0202] where, is the value of the modeling environment data information , is the value of the modeling environment information , represents the average value of the modeling environment data information , represents the average value of the modeling environment information .

[0203] By calculating the similarity comparison results of different modeling environments, select the environment with the highest value as the final modeling environment, and provide the most suitable parameter settings and training strategies for the prediction model.

[0204] As Figure 9 shown, the present invention also provides a method for predicting the settlement of subgrade in soft soil area, including the following steps:

[0205] S1. Collect the subgrade parameters and settlement historical data in the soft soil area;

[0206] S2. Remove the outliers and normalize the collected data;

[0207] S3. Construct a subgrade settlement prediction model based on the gradient boosting decision tree algorithm;

[0208] S4. Conduct settlement prediction according to the processed data and the constructed model;

[0209] S5. Compare the prediction results with the measured data;

[0210] S6. When the prediction error is greater than the preset threshold, correct the model parameters based on the artificial intelligence algorithm;

[0211] S7. Update the prediction model according to the correction results;

[0212] S8. Use the updated model to re-conduct settlement prediction until the prediction error is less than the preset threshold;

[0213] S9. Output the final subgrade settlement prediction results, including the settlement amount and the relationship curve of settlement changing with time.

[0214] In step S1, the collected subgrade parameters in the soft soil area include the filling height, soil type, design type, etc.; the settlement historical data includes the settlement observation values at different time nodes. Preferably, automated monitoring equipment such as precise level gauges and pressure sensors is used to improve the accuracy and efficiency of data collection.

[0215] In step S2, the outlier judgment criterion is deviating from the mean by 3 times the standard deviation; the normalization adopts the maximum-minimum normalization method to map the data to the interval [0, 1].

[0216] In step S3, the key parameters of the GBDT algorithm are set as follows: the learning rate is 0.1, the maximum depth of the tree is 5, the minimum number of samples in the leaf nodes is 2% of the total samples, the maximum number of iterations is 500, and the early stopping parameter is 10 (that is, if the model performance does not improve for 10 consecutive iterations, stop training). In addition, this step also applies the multi-scale spatio-temporal feature extraction and fusion technology to extract richer feature representations and enhance the prediction ability of the model.

[0217] In step S4, multiple prediction models are used for parallel calculation to obtain multiple groups of prediction results. Specifically, 5 - 10 different GBDT models are constructed, and these models are different in terms of training data subsets, feature selection, tree depth, etc. At the same time, the spatio-temporal self-attention mechanism is applied to enhance the modeling ability for complex spatio-temporal dependence relationships.

[0218] In step S5, two indicators, the mean absolute error (MAE) and the root mean square error (RMSE), are used to evaluate the prediction accuracy. When the MAE or RMSE exceeds 10% of the measured value, trigger the parameter correction process.

[0219] In step S6, the model parameters are automatically adjusted based on the Bayesian optimization algorithm. This algorithm efficiently explores the parameter space by establishing a probability model (Gaussian process) between the parameters and the model performance to find the optimal parameter combination.

[0220] In step S7, an incremental learning strategy is adopted to update the prediction model, integrating new data and correction results into the existing model without the need to retrain the entire model.

[0221] In step S8, the updated model is used to re - perform the settlement prediction, repeating steps S5 - S7 until the prediction error is less than a preset threshold (usually 5% of the measured value).

[0222] In step S9, the final settlement prediction results are output, including the predicted settlement amounts at different time points and the settlement - time curve. Preferably, the settlement curve is fitted with a hyperbolic function or an exponential function to facilitate the engineering personnel to intuitively understand the settlement development trend.

[0223] Through the detailed description of the above embodiments, those skilled in the art should understand that the present invention is not limited to the above - mentioned specific implementation manners. Without departing from the concept of the present invention, various deformations and improvements can be made, and these deformations and improvements all fall within the protection scope of the present invention.

Claims

1. A roadbed settlement prediction system in soft soil areas, characterized in that: include: Data acquisition module, used to collect roadbed parameters and settlement history data in soft soil areas; A data preprocessing module, which is in communication with the data acquisition module and is used to remove outliers and perform normalization processing on the collected data; A model building module, which is in communication connection with the data preprocessing module and is used to build a roadbed settlement prediction model based on a gradient boosting decision tree algorithm; A prediction and analysis module, which is in communication with the model building module and is used to predict settlement based on the processed data and the built model; The feedback optimization module is in communication with the prediction analysis module and is used to compare the prediction results with the measured data and optimize the prediction model.

2. The soft soil area roadbed settlement prediction system according to claim 1 is characterized in that: The data acquisition module comprises: The roadbed parameter collection unit is used to collect the fill height, soil type and design type of the roadbed in the soft soil area; Geological survey data acquisition unit, used to collect soil depth and soil longitudinal section diagram in soft soil area; Settlement history data collection unit, used to collect measured values ​​of roadbed settlement at different time points; The real-time monitoring unit is used to monitor the roadbed settlement in real time and obtain the current settlement data.

3. The soft soil area roadbed settlement prediction system according to claim 1, characterized in that: The data preprocessing module comprises: An outlier processing unit is used to identify and remove outliers, missing values, and duplicate values ​​in the data; A data conversion unit, used for converting the preprocessed data into an ordered list form; A normalization processing unit, used to map the data to the interval [0,1] using a maximum and minimum normalization method; The correlation analysis unit is used to analyze the linear correlation coefficient between data and establish a relationship between parameters when the correlation coefficient is greater than a preset threshold.

4. The soft soil area roadbed settlement prediction system according to claim 1, characterized in that: The model building module includes: A feature extraction unit is used to extract input features required by the prediction model from the preprocessed data; GBDT algorithm unit, used to train weak classifiers and build gradient boosting decision tree models in an iterative manner; The tree structure optimization unit is used to control the model complexity through the tree structure complexity parameter γ and the leaf node penalty parameter λ; The physical model fusion unit is used to fuse the physical test results with the numerical simulation results to enhance the physical meaning of the model.

5. The soft soil area roadbed settlement prediction system according to claim 4, characterized in that: The GBDT algorithm unit is specifically used for: Initialize model parameters, set t=1; Train a weak classifier for each sample point i , where the weak classifier determines the optimal split point by calculating the average loss of each leaf region sample; Calculate the total error and update the tree structure complexity parameter γ and leaf node penalty parameter λ; Iterative process optimization model extended by Taylor formula; Determine whether the termination condition is met, and if so, output the final model H(x).

6. The soft soil area roadbed settlement prediction system according to claim 1, characterized in that: The prediction analysis module includes: An input processing unit, used for converting the parameters of the roadbed to be predicted into the input format required by the model; A multi-model prediction unit, used to obtain multiple sets of prediction results through parallel calculation of multiple prediction models; A result evaluation unit, used to evaluate the error of each prediction result and select the result with the smallest error as the final prediction value; The settlement curve generating unit is used to generate a curve showing the relationship between roadbed settlement and time according to the final predicted value.

7. The soft soil area roadbed settlement prediction system according to claim 1, characterized in that: The feedback optimization module comprises: An error calculation unit, used to calculate the error between the predicted value and the measured value; AI parameter correction unit, used to automatically correct parameters based on artificial intelligence algorithms when the error is greater than a preset threshold; A historical data correction unit, used for correcting model input parameters based on errors between historical prediction data and measured data; The model updating unit is used to optimize and update the prediction model according to the correction result.

8. The soft soil area roadbed settlement prediction system according to claim 7, characterized in that: The AI ​​parameter correction unit is specifically used for: Establish a mapping relationship between prediction parameters and errors; Identify the key parameters that have the greatest impact on forecast results; Dynamically adjust the weights of key parameters based on real-time settlement data; The corrected parameters are fed back to the model building module for model optimization.

9. The soft soil area roadbed settlement prediction system according to claim 1, characterized in that: It also includes a modeling environment module, which is in communication with the model building module and is used to: Storing multiple sets of modeling environment information, including modeling environment data, modeling environment analysis data, and modeling environment determination data; Select the optimal modeling environment through similarity comparison analysis method; Among them, the similarity comparison and analysis method includes calculating the similarity coefficient and correlation between the modeling environment data information and the modeling environment information, and selecting the optimal solution based on the weighted results.

10. A method for predicting roadbed settlement in soft soil areas, using the system according to any one of claims 1 to 9, characterized in that: The following steps are involved: S1. Collect roadbed parameters and settlement history data in soft soil areas; S2, removing outliers and normalizing the collected data; S3, constructing a roadbed settlement prediction model based on the gradient boosting decision tree algorithm; S4. Predict settlement based on the processed data and the constructed model; S5, comparing the predicted results with the measured data; S6. When the prediction error is greater than a preset threshold, the model parameters are corrected based on the artificial intelligence algorithm; S7. updating the prediction model according to the correction result; S8, re-predicting the settlement using the updated model until the prediction error is less than a preset threshold; S9. Output the final roadbed settlement prediction results, including the settlement amount and the relationship curve between settlement and time.

Images