Soft ground surface settlement prediction system and method based on multi-scale decomposition optimization

Abstract

The application discloses a soft soil surface subsidence prediction system and method based on multi-scale decomposition optimization, and belongs to the technical field of surface subsidence prediction. The method comprises the following steps: obtaining a soft soil subsidence time series data set and standardizing preprocessing; selecting and removing redundant features; decomposing into multiple IMF components; constructing a time series sample and dividing a training, verification and test set; independently training a Transformer and XGBoost for each IMF component; and adaptively fusing or optimally superimposing to output a final subsidence prediction value. The application constructs a continuous wavelet transform front filter and a self-adaptive noise complete set empirical mode decomposition multi-scale decomposition full component output, a component level Transformer encoder and an XGBoost regressor double model parallel training and self-adaptive weighted fusion nested training architecture, achieves decoupling modeling and complementary advantages of signals in each frequency band, and improves the subsidence prediction precision in the coastal soft soil area.

Description

Technical Field

[0001] This invention relates to the field of ground subsidence prediction technology, and in particular to a soft soil ground subsidence prediction system and method based on multi-scale decomposition and optimization. Background Technology

[0002] Methods for predicting settlement on soft soil surfaces have mainly gone through three technical stages. The first stage, represented by traditional time-series statistical models such as ARIMA and grey models, relies on linear assumptions and data stationarity requirements, making it difficult to capture the nonlinear and non-stationary characteristics of the settlement process. The second stage, using single machine learning and deep learning models, employs models such as LSTM, Transformer, Support Vector Machine, and Random Forest for settlement prediction. Current solutions still primarily use single-model architectures and lack parallel training and adaptive fusion mechanisms for dual models. The third stage, using signal decomposition-assisted prediction, decomposes the settlement signal using wavelet transform or empirical mode decomposition and models it separately. However, high-frequency random terms are directly discarded, failing to establish a complete link between full component retention and component-level modeling.

[0003] In related technologies, the application scenarios of multi-source data fusion frameworks are limited to tunnel construction, and a general data arrangement paradigm and missing value handling system for soft soil settlement monitoring have not been established; the screening process of forward feature selection methods is independent of model training, and no linkage optimization between feature selection and model performance has been formed; wavelet decomposition directly discards random terms; physical constraint construction relies on surrogate models and manual experience thresholds; and the improved Transformer structure is still a single model architecture, without introducing a dual-model parallel training and adaptive weighted fusion mechanism.

[0004] However, existing technologies suffer from problems such as poor data compatibility leading to low sample utilization, single feature selection mode making it difficult to eliminate multicollinearity, fixed denoising strategies with high-frequency components being directly discarded, and rigid model architecture with hyperparameter tuning relying on manual trial and error. These issues make it difficult to meet the engineering needs for refined short- and medium-term prediction of land settlement in coastal soft soil distribution areas. Summary of the Invention

[0005] This application provides a soft soil surface settlement prediction system and method based on multi-scale decomposition optimization, which solves the problems in the prior art such as poor data compatibility leading to low sample utilization, single feature selection mode making it difficult to eliminate multicollinearity, fixed noise reduction strategy with high-frequency components being directly discarded, rigid model architecture and hyperparameter tuning relying on manual trial and error. It realizes decoupled modeling and complementary advantages of signals in different frequency bands, and improves the settlement prediction accuracy of coastal soft soil areas.

[0006] This application provides a method for predicting soft soil surface settlement based on multi-scale decomposition and optimization, including:

[0007] Obtain a standardized soft soil settlement time series dataset, in which the influencing feature independent variables are arranged first and the settlement true value dependent variable is arranged last. Perform missing value removal or cubic spline interpolation filling on the dataset and output a standardized time series sample set.

[0008] Feature selection processing is performed on the standardized time series sample set, and redundant features are removed based on any one of three strategies: Pearson correlation coefficient, tree model feature importance, or intelligent optimization algorithm.

[0009] The original settlement time series data is subjected to continuous wavelet transform (CWT) adaptive threshold denoising. The db4 wavelet basis, 6-level decomposition and soft threshold function are used to filter out high-frequency random interference. The denoised signal is then subjected to complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) to decompose the signal into multiple intrinsic mode function (IMF) components and residual terms, and all components are output.

[0010] Based on the components of each intrinsic mode function and the original multi-feature data, time series samples are constructed according to the historical window length, sliding step size and prediction step size, and the samples are synchronously divided into training set, validation set and test set;

[0011] For each intrinsic mode function component, a Transformer encoder and an XGBoost regressor are trained independently. The Transformer encoder uses a multi-head self-attention mechanism to capture long-range dependencies between time steps, and the XGBoost regressor is trained using a gradient boosting tree framework.

[0012] The error difference between the Transformer encoder and the XGBoost regressor is calculated by means of the validation set mean absolute error. Based on the error threshold, an adaptive decision is made to use either inverse error weighted fusion or single model selection. The prediction results of each component are linearly superimposed to obtain the final settlement prediction value.

[0013] Furthermore, the step of obtaining a standardized soft soil settlement time-series dataset, performing missing value removal or cubic spline interpolation imputation on the dataset, and outputting a standardized time-series sample set specifically includes:

[0014] The data input module reads the ground settlement time series dataset in spreadsheet format, and forces the data column layout rule to place all influencing characteristic independent variables at the beginning and fix the settlement true value dependent variable as the last column, and automatically filters invalid non-numerical columns.

[0015] The system features a switchable cleaning mechanism with two modes: missing value removal and cubic spline interpolation filling. When switched to missing value removal mode, the system calculates the percentage of missing values ​​in a single row of data. If the percentage of missing values ​​in a single row exceeds a preset threshold, the data in that row is removed. When switched to cubic spline interpolation filling mode, the system fills the missing values ​​in the dataset with cubic splines.

[0016] The cleaned dataset is cached in global memory.

[0017] Furthermore, the feature selection process performed on the standardized time-series sample set, based on any one of three strategies—Pearson correlation coefficient, tree model feature importance, or intelligent optimization algorithm—to remove redundant features, specifically includes:

[0018] The feature selection module performs feature selection processing on the standardized time series sample set. The feature selection module is configured with a feature selection master switch. When the feature selection master switch is closed, the original feature dataset is directly output.

[0019] When the feature selection master switch is turned on, feature selection is performed using any of the following three strategies: feature selection based on Pearson correlation coefficient, which removes low-association features based on a preset correlation coefficient threshold; feature selection based on tree model feature importance, which removes low-importance features based on a preset quantile threshold; and feature selection based on intelligent optimization algorithm, which uses a preset feature retention threshold and reuses the built-in optimizer of the model optimization module to dynamically optimize the feature subset.

[0020] Furthermore, the continuous wavelet transform adaptive threshold denoising process performed on the original settlement time series data, employing a db4 wavelet basis, 6-level decomposition, and a soft thresholding function to filter out high-frequency random interference, specifically includes:

[0021] The original settlement time series data is pre-filtered by continuous wavelet transform through a multi-scale decomposition and denoising module. The db4 wavelet basis is used for 6-level decomposition, and adaptive threshold denoising is performed in conjunction with a soft threshold function to filter out high-frequency random interference in the original settlement time series data.

[0022] After the continuous wavelet transform pre-filtering process is completed, the filtered signal is output to the adaptive noise complete set empirical mode decomposition processing flow.

[0023] Furthermore, the adaptive noise complete set empirical mode decomposition process is performed on the denoised signal to decompose the signal into multi-layer intrinsic mode function components and residual terms, and output all components, specifically including:

[0024] An adaptive noise complete set empirical mode decomposition process is performed on the signal after continuous wavelet transform pre-filtering by a multi-scale decomposition noise reduction module. The adaptive noise complete set empirical mode decomposition process has an integration number of 100, a noise figure of 0.2, and a maximum iteration number of 10.

[0025] The filtered signal is adaptively decomposed into multiple intrinsic mode function components and a single residual term. All intrinsic mode function components and residual terms are output to the global buffer without reconstructing each component.

[0026] Furthermore, the step of constructing time-series samples based on each intrinsic mode function component and the original multi-feature data according to the historical window length, sliding step size, and prediction step size, and simultaneously dividing the samples into training set, validation set, and test set, specifically includes:

[0027] The model training module receives all intrinsic mode function components and raw multi-feature data. Based on each intrinsic mode function component and raw multi-feature data, standardized time series samples are constructed according to a unified historical window length, sliding step size and prediction step size. The historical window length, sliding step size and prediction step size can be customized by the user.

[0028] Standardized time-series samples are synchronously divided into training, validation, and test sets according to user-defined proportions. Each intrinsic mode function component shares the same set of original multi-feature data, and each uses the true value of its own intrinsic mode function component as the prediction target.

[0029] Furthermore, the Transformer encoder and XGBoost regressor are trained independently for each intrinsic mode function component. The Transformer encoder uses a multi-head self-attention mechanism to capture long-range dependencies between time steps, and the XGBoost regressor is trained using a gradient boosting tree framework, specifically including:

[0030] The Transformer encoder and XGBoost regressor are trained independently for each intrinsic modality function component through the model training module. The Transformer encoder adopts a two-layer stacked encoder layer structure. Each encoder layer contains a multi-head self-attention submodule and a feedforward network submodule. The number of attention heads in the multi-head self-attention submodule can be customized by the user. The hidden layer dimension of the feedforward network submodule is twice the hidden layer dimension of the model, and the activation function is a modified linear unit.

[0031] The Transformer encoder is trained using the AdamW optimizer with a fixed initial learning rate of 0.001. The loss function is the mean squared error loss, and the full-batch gradient descent method is used. The number of training rounds and batch size can be customized by the user. During training, the validation set loss is monitored and an early stopping mechanism is enabled. When the validation set loss does not decrease for several consecutive rounds, the training is terminated early and the current model parameters are saved.

[0032] The XGBoost regressor is trained using a gradient boosting tree framework. The number of base learners is fixed at 100. The maximum tree depth, learning rate, sample subsampling ratio, and feature subsampling ratio are four optimizable hyperparameters. The L1 regularization coefficient is fixed at 0.1, the L2 regularization coefficient is fixed at 1.0, the minimum leaf node sample weights are fixed at 1, and the objective function is set to the squared regression error.

[0033] Furthermore, the step of calculating the error difference between the Transformer encoder and the XGBoost regressor using the mean absolute error of the validation set, adaptively deciding between inverse error weighted fusion or single-model selection based on the error threshold, and obtaining the final settlement prediction value by linearly superimposing the prediction results of each component, specifically includes:

[0034] The mean absolute error of the Transformer encoder and the mean absolute error of the XGBoost regressor on the validation set are calculated by the model training module. The larger of the two values ​​is used as the denominator and the absolute value of the difference between the two values ​​is used as the numerator to calculate the error difference degree.

[0035] When the error difference is less than the user-defined error threshold, an inverse error weighted fusion strategy is adopted. The fusion weight of the Transformer encoder is the average absolute error of the XGBoost regressor validation set divided by the sum of the average absolute errors of the two validation sets, and the fusion weight of the XGBoost regressor is the average absolute error of the Transformer encoder validation set divided by the sum of the average absolute errors of the two validation sets.

[0036] When the error difference is greater than or equal to the user-defined error threshold, the single model with the smaller mean absolute error of the validation set is selected as the prediction model for that component.

[0037] The prediction results after internal fusion or selection of each intrinsic mode function component are linearly superimposed to obtain the final full-band settlement prediction value.

[0038] This application provides a soft soil surface settlement prediction system based on multi-scale decomposition and optimization, which is used to implement a soft soil surface settlement prediction method based on multi-scale decomposition and optimization, including: a data input module, a feature selection module, a multi-scale decomposition and denoising module, a model training module, a model optimization module, and a result display module.

[0039] The data input module is used to read the ground settlement time series dataset in spreadsheet format, and to enforce the data column arrangement rule that all influencing feature independent variables are arranged at the beginning and the settlement true value dependent variable is fixed as the last column. It automatically filters invalid non-numerical columns, and configures a dual-mode switchable cleaning mechanism of missing value removal and cubic spline interpolation filling. The cleaned dataset is cached in global memory for reuse by various modules.

[0040] The feature selection module has its input end connected to the output end of the data input module. It is configured with a feature selection master switch, which is used to directly output the original feature dataset when the feature selection master switch is closed, and to switch between any one of the feature selection strategies based on Pearson correlation coefficient, feature selection strategies based on tree model feature importance, or feature selection strategies based on intelligent optimization algorithm to remove redundant features when the feature selection master switch is open.

[0041] The multi-scale decomposition denoising module selectively connects its input end to the output end of the data input module or the feature selection module. It is used to perform continuous wavelet transform adaptive threshold denoising on the original settlement time series data. It uses db4 wavelet basis, 6-level decomposition and soft threshold function to filter out high-frequency random interference. It performs adaptive noise complete set empirical mode decomposition on the denoised signal, decomposes the signal into multi-level intrinsic mode function components and residual terms and outputs them completely to the global buffer.

[0042] The model training module has its input end connected to the output end of the multi-scale decomposition and denoising module. It is used to receive all intrinsic mode function components and original multi-feature data, construct time series samples according to historical window length, sliding step size and prediction step size and synchronously divide them into training set, validation set and test set. Transformer encoder and XGBoost regressor are trained independently for each intrinsic mode function component. The error difference between the Transformer encoder and XGBoost regressor is calculated by the mean absolute error of the validation set. Based on the error threshold, an adaptive decision is made to adopt error inverse weighted fusion or single model selection. The prediction results of each component are linearly superimposed to obtain the final settlement prediction value.

[0043] The model optimization module has its input end connected to the intrinsic mode function component training interface of the model training module. It is equipped with an independent optimization switch and has four freely switchable intelligent optimizers: sparrow search algorithm, gray wolf optimization algorithm, lobster optimization algorithm, and leech optimization algorithm. It is used to perform global optimization of the hyperparameters of the XGBoost regressor with the optimization objective of minimizing the mean absolute error of the first intrinsic mode function component validation set. It outputs the optimal hyperparameter combination and stores it in the global cache for reuse by all intrinsic mode function components.

[0044] The results display module, as the system terminal output module, is used to batch calculate five types of evaluation indicators for the training set, validation set, and test set: mean absolute error, mean absolute percentage error, mean square error, root mean square error, and coefficient of determination. It automatically generates a measured-predicted time series comparison curve and a scatter plot, and exports the error index, optimal hyperparameters, fusion weights, and future predicted time series data.

[0045] The soft soil surface settlement prediction method based on multi-scale decomposition optimization provided in this application obtains a standardized soft soil settlement time series dataset. The dataset prioritizes the independent variables of influencing features and prioritizes the dependent variables of true settlement values. Missing value removal or cubic spline interpolation imputation is performed on the dataset to output a standardized time series sample set. Feature selection processing is then performed on the standardized time series sample set, removing redundant features based on Pearson correlation coefficient, tree model feature importance, or intelligent optimization algorithms. Adaptive threshold denoising processing of the original settlement time series data is performed using continuous wavelet transform, employing a db4 wavelet basis, 6-level decomposition, and a soft threshold function to filter out high-frequency random interference. Adaptive noise complete set empirical mode decomposition processing is then performed on the denoised signal, decomposing the signal into multiple intrinsic mode function components and residual terms, outputting all components. Based on each intrinsic mode function component and the original multi-feature data… Time-series samples were constructed based on historical window length, sliding step size, and prediction step size, and these samples were synchronously divided into training, validation, and test sets. A Transformer encoder and an XGBoost regressor were trained independently for each intrinsic mode function component. The Transformer encoder employed a multi-head self-attention mechanism to capture long-range dependencies between time steps, while the XGBoost regressor used a gradient boosting tree framework for training. The error difference between the Transformer encoder and the XGBoost regressor was calculated using the mean absolute error of the validation set. Adaptive decision-making was made based on the error threshold, employing either inverse error weighted fusion or single-model selection. The prediction results of each component were linearly superimposed to obtain the final settlement prediction value. Five evaluation metrics were calculated for the training, validation, and test sets: mean absolute error, mean absolute percentage error, mean square error, root mean square error, and coefficient of determination.

[0046] In this process, by forcibly defining the data column arrangement rules and configuring a dual-mode switching mechanism for missing value removal and cubic spline interpolation filling, standardized import and compatible processing of multi-source heterogeneous sedimentation data were achieved, effectively improving sample utilization. Furthermore, by setting a three-level switchable feature selection strategy, using Pearson correlation coefficient thresholds, tree model feature importance quantiles, or intelligent optimization algorithm feature retention thresholds to automatically remove highly collinear and low-association redundant features, the constraint of multicollinearity among features on the model's generalization ability was suppressed. Furthermore, by constructing a unified denoising decomposition process combining continuous wavelet transform pre-filtering and adaptive noise complete set empirical mode decomposition multi-scale decomposition, and outputting all intrinsic mode function components completely to the downstream modeling stage, the component-level model autonomously distinguishes the effective information and noise of each frequency band signal, avoiding the information loss of traditional single reconstruction sequence denoising. Furthermore, a component-level parallel modeling architecture is used to decouple the modeling of signals in each frequency band. Within each component, a dual-model parallel training of a Transformer encoder and an XGBoost regressor is employed. An adaptive decision fusion strategy based on the validation set error difference and a user-defined error threshold is used to achieve complementary advantages and accurate prediction of signals in each frequency band. Furthermore, multiple built-in intelligent optimization algorithms are used to globally optimize the hyperparameters of the XGBoost regressor with the goal of minimizing the mean absolute error of the validation set. The optimal parameters are automatically reused across all intrinsic mode function components, reducing the time cost of hyperparameter tuning. Attached Figure Description

[0047] Figure 1 A flowchart of a soft soil surface settlement prediction method based on multi-scale decomposition and optimization provided in an embodiment of this application;

[0048] Figure 2 This is a schematic diagram of the structure of the soft soil surface settlement prediction system based on multi-scale decomposition and optimization provided in the embodiments of this application. Detailed Implementation

[0049] This application provides a soft soil surface settlement prediction system and method based on multi-scale decomposition optimization. It solves the problems in the prior art, such as poor data compatibility leading to low sample utilization, single feature selection mode making it difficult to eliminate multicollinearity, fixed noise reduction strategy with direct discarding of high-frequency components, rigid model architecture and hyperparameter tuning relying on manual trial and error. By constructing a nested training architecture with continuous wavelet transform pre-filter and adaptive noise complete set empirical mode decomposition multi-scale decomposition full component output, component-level Transformer encoder and XGBoost regressor dual-model parallel training and adaptive weighted fusion, it realizes decoupled modeling and complementary advantages of signals in each frequency band, and improves the settlement prediction accuracy of coastal soft soil areas.

[0050] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0051] like Figure 1 The diagram shown is a flowchart of a soft soil surface settlement prediction method based on multi-scale decomposition and optimization provided in this application embodiment. This method is applied to a soft soil surface settlement prediction system based on multi-scale decomposition and optimization, and includes the following steps:

[0052] Obtain a standardized soft soil settlement time series dataset, in which the influencing feature independent variables are arranged first and the settlement true value dependent variable is arranged last. Perform missing value removal or cubic spline interpolation filling on the dataset and output a standardized time series sample set.

[0053] Feature selection processing is performed on the standardized time series sample set, and redundant features are removed based on any one of three strategies: Pearson correlation coefficient, tree model feature importance, or intelligent optimization algorithm.

[0054] The original settlement time series data is subjected to continuous wavelet transform adaptive threshold denoising. The db4 wavelet basis, 6-level decomposition and soft threshold function are used to filter out high-frequency random interference. The denoised signal is subjected to adaptive noise complete set empirical mode decomposition. The signal is decomposed into multi-level intrinsic mode function components and residual terms, and all components are output.

[0055] Based on the components of each intrinsic mode function and the original multi-feature data, time series samples are constructed according to the historical window length, sliding step size and prediction step size, and the samples are synchronously divided into training set, validation set and test set;

[0056] For each intrinsic mode function component, a Transformer encoder and an XGBoost regressor are trained independently. The Transformer encoder uses a multi-head self-attention mechanism to capture long-range dependencies between time steps, and the XGBoost regressor is trained using a gradient boosting tree framework.

[0057] The error difference between the Transformer encoder and the XGBoost regressor is calculated by means of the validation set mean absolute error. Based on the error threshold, an adaptive decision is made to use either inverse error weighted fusion or single model selection. The prediction results of each component are linearly superimposed to obtain the final settlement prediction value.

[0058] In this embodiment, the standardized soft soil settlement time-series dataset includes independent variables such as soft soil thickness, water content, void ratio, compression modulus, groundwater level, surface load, rainfall, fault distance, and built-up area density. The dual-mode switchable cleaning mechanism of missing value removal and cubic spline interpolation filling selects the mode based on the actual data quality. When missing data is scattered, interpolation filling is used to retain more samples; when missing data is concentrated in certain time periods, removal is used to ensure data quality.

[0059] In the continuous wavelet transform, the db4 wavelet basis is a wavelet with a vanishing moment order of 4 from the Daubechies wavelet family; the 6-level decomposition decomposes the signal into 6 detail components at different scales and an approximate component; the soft thresholding function sets coefficients below the threshold to zero and shrinks coefficients above the threshold to zero. In the adaptive noise complete ensemble empirical mode decomposition, the mode aliasing problem in traditional empirical mode decomposition is solved by adding pairs of positive and negative white noise multiple times during the decomposition process and taking the ensemble average; the ensemble number of 100 indicates that 100 pairs of positive and negative adaptive noise are added and the ensemble average is taken; the noise figure of 0.2 controls the ratio of the noise amplitude added each time to the standard deviation of the original signal; the maximum number of iterations of 10 controls the upper limit of the screening iterations in the screening process of each intrinsic mode function component.

[0060] The formula for calculating the error difference degree diff is: ;

[0061] in The mean absolute error of the Transformer encoder on the validation set. is the mean absolute error of the XGBoost regressor on the validation set.

[0062] In the error-inverse weighted fusion, the Transformer encoder weights With XGBoost regressor weights The calculation formula is: ; .

[0063] Furthermore, the step of obtaining a standardized soft soil settlement time-series dataset, performing missing value removal or cubic spline interpolation imputation on the dataset, and outputting a standardized time-series sample set specifically includes:

[0064] The data input module reads the ground settlement time series dataset in spreadsheet format, and forces the data column layout rule to place all influencing characteristic independent variables at the beginning and fix the settlement true value dependent variable as the last column, and automatically filters invalid non-numerical columns.

[0065] The system features a switchable cleaning mechanism with two modes: missing value removal and cubic spline interpolation filling. When switched to missing value removal mode, the system calculates the percentage of missing values ​​in a single row of data. If the percentage of missing values ​​in a single row exceeds a preset threshold, the data in that row is removed. When switched to cubic spline interpolation filling mode, the system fills the missing values ​​in the dataset with cubic splines.

[0066] The cleaned dataset is cached in global memory.

[0067] In this embodiment, the data input module is the system's front-end module, responsible for receiving externally imported spreadsheet-formatted settlement monitoring data. The arrangement rule of placing all influencing feature independent variables at the beginning and fixing the true settlement value dependent variable as the last column reflects the meaning and positional relationship of each field in the dataset, enabling the system to automatically identify which columns serve as model input features and which as the prediction target. If the percentage of missing data in a single row exceeds a preset threshold, that row is removed. The preset threshold is set by the user based on data quality; when missing data is concentrated, removing the entire row is more efficient than interpolation. The cubic spline interpolation filling smoothly imputes missing values ​​between adjacent known data points by constructing a cubic polynomial.

[0068] The global memory is a dedicated cache area in the computer's random access memory. The cleaned dataset is stored in this area in tensor or data frame format for the feature selection module, multi-scale decomposition and denoising module, and model training module to read, thus avoiding repeated data reading and cleaning operations by each module.

[0069] Furthermore, the feature selection process performed on the standardized time-series sample set, based on any one of three strategies—Pearson correlation coefficient, tree model feature importance, or intelligent optimization algorithm—to remove redundant features, specifically includes:

[0070] The feature selection module performs feature selection processing on the standardized time series sample set. The feature selection module is configured with a feature selection master switch. When the feature selection master switch is closed, the original feature dataset is directly output.

[0071] When the feature selection master switch is turned on, feature selection is performed using any of the following three strategies: feature selection based on Pearson correlation coefficient, which removes low-association features based on a preset correlation coefficient threshold; feature selection based on tree model feature importance, which removes low-importance features based on a preset quantile threshold; and feature selection based on intelligent optimization algorithm, which uses a preset feature retention threshold and reuses the built-in optimizer of the model optimization module to dynamically optimize the feature subset.

[0072] In this embodiment, the feature selection module is located after the data input module and is used to perform dimensionality reduction processing on the input features. The feature selection master switch allows the user to decide whether to enable feature selection based on the number and significance of the original features. When the number of original features is small or all features have been confirmed to be meaningful, this module is disabled to simplify the process. The feature selection based on the Pearson correlation coefficient calculates the Pearson correlation coefficient between each feature and the settlement amount. Features with |r| below a preset threshold are removed, suitable for screening factors with a linear correlation to sedimentation. The feature selection based on tree model feature importance uses a random forest to measure feature importance based on the reduction in Gini impurity or the cumulative information gain during tree node splits, removing features with importance below a preset quantile. The feature selection based on intelligent optimization algorithms models the feature subset selection problem as a combinatorial optimization problem, using model prediction performance as the fitness function. An optimization algorithm searches for the optimal subset in the feature space, reusing any one of the sparrow search algorithm, gray wolf optimization algorithm, lobster optimization algorithm, or leech optimization algorithm built into the model optimization module.

[0073] Furthermore, the continuous wavelet transform adaptive threshold denoising process performed on the original settlement time series data, employing a db4 wavelet basis, 6-level decomposition, and a soft thresholding function to filter out high-frequency random interference, specifically includes:

[0074] The original settlement time series data is pre-filtered by continuous wavelet transform through a multi-scale decomposition and denoising module. The db4 wavelet basis is used for 6-level decomposition, and adaptive threshold denoising is performed in conjunction with a soft threshold function to filter out high-frequency random interference in the original settlement time series data.

[0075] After the continuous wavelet transform pre-filtering process is completed, the filtered signal is output to the adaptive noise complete set empirical mode decomposition processing flow.

[0076] In this embodiment, the multi-scale decomposition denoising module receives the data output by the feature selection module and is responsible for preprocessing the original settlement time series for denoising. The continuous wavelet transform performs multi-resolution analysis on the signal through scaling and shifting wavelet basis functions, providing local information of the signal in both the time and frequency domains. The 6-layer decomposition corresponds to 6 detail components and one approximate component in different frequency bands, with the frequency band range decreasing progressively for each layer. The adaptive threshold is calculated independently based on the noise level of each layer. The calculation formula is ,in This is an estimate of the noise standard deviation of the detail coefficients for this layer. The signal length. Soft thresholding function. The expression is: .

[0077] Furthermore, the adaptive noise complete set empirical mode decomposition process is performed on the denoised signal to decompose the signal into multi-layer intrinsic mode function components and residual terms, and output all components, specifically including:

[0078] An adaptive noise complete set empirical mode decomposition process is performed on the signal after continuous wavelet transform pre-filtering by a multi-scale decomposition noise reduction module. The adaptive noise complete set empirical mode decomposition process has an integration number of 100, a noise figure of 0.2, and a maximum iteration number of 10.

[0079] The filtered signal is adaptively decomposed into multiple intrinsic mode function components and a single residual term. All intrinsic mode function components and residual terms are output to the global buffer without reconstructing each component.

[0080] In this embodiment, the adaptive noise ensemble empirical mode decomposition (EMD) has advantages in handling nonlinear and non-stationary signals. An ensemble count of 100 means adding 100 pairs of positive and negative adaptive noise to the original signal before performing EMD and taking the ensemble average; a higher ensemble count results in more stable decomposition results. A noise figure of 0.2 controls the ratio of the added noise amplitude to the standard deviation of the original signal, striking a balance between suppressing mode aliasing and maintaining decomposition accuracy. A maximum iteration count of 10 controls the upper limit of the screening iterations for each intrinsic mode function component; the screening process for that component stops when the upper limit is reached. All components are directly output without reconstruction processing, allowing each frequency band signal to enter the downstream modeling stage as an independent time series.

[0081] Furthermore, the step of constructing time-series samples based on each intrinsic mode function component and the original multi-feature data according to the historical window length, sliding step size, and prediction step size, and simultaneously dividing the samples into training set, validation set, and test set, specifically includes:

[0082] The model training module receives all intrinsic mode function components and raw multi-feature data. Based on each intrinsic mode function component and raw multi-feature data, standardized time series samples are constructed according to a unified historical window length, sliding step size and prediction step size. The historical window length, sliding step size and prediction step size can be customized by the user.

[0083] Standardized time-series samples are synchronously divided into training, validation, and test sets according to user-defined proportions. Each intrinsic mode function component shares the same set of original multi-feature data, and each uses the true value of its own intrinsic mode function component as the prediction target.

[0084] In this embodiment, the model training module is the core computing module of the system, responsible for training and predicting the models of each component. The historical window length defines the number of historical observations used in each prediction, with a default value of 10, indicating that data from the previous 10 time steps are used to predict future values. The sliding step size defines the temporal movement interval of training samples, with a default value of 1, indicating that the next training sample is generated by moving one time step backward each time. The prediction step size defines the number of future time steps predicted each time, with a default value of 3, indicating that the settlement prediction values ​​for the next 3 time steps are output at once. The fact that each intrinsic mode function component shares the same set of original multi-feature data means that the original multi-feature data is used as the common input feature matrix for each component model. The prediction target for each component is the actual value sequence of that component. This enables decoupled modeling of signals across different frequency bands. The default partitioning ratio is 7:2:1 (training set: validation set), but users can customize this ratio based on the total amount of data.

[0085] Furthermore, the Transformer encoder and XGBoost regressor are trained independently for each intrinsic mode function component. The Transformer encoder uses a multi-head self-attention mechanism to capture long-range dependencies between time steps, and the XGBoost regressor is trained using a gradient boosting tree framework, specifically including:

[0086] The Transformer encoder and XGBoost regressor are trained independently for each intrinsic modality function component through the model training module. The Transformer encoder adopts a two-layer stacked encoder layer structure. Each encoder layer contains a multi-head self-attention submodule and a feedforward network submodule. The number of attention heads in the multi-head self-attention submodule can be customized by the user. The hidden layer dimension of the feedforward network submodule is twice the hidden layer dimension of the model, and the activation function is a modified linear unit.

[0087] The Transformer encoder is trained using the AdamW optimizer with a fixed initial learning rate of 0.001. The loss function is the mean squared error loss, and the full-batch gradient descent method is used. The number of training rounds and batch size can be customized by the user. During training, the validation set loss is monitored and an early stopping mechanism is enabled. When the validation set loss does not decrease for several consecutive rounds, the training is terminated early and the current model parameters are saved.

[0088] The XGBoost regressor is trained using a gradient boosting tree framework. The number of base learners is fixed at 100. The maximum tree depth, learning rate, sample subsampling ratio, and feature subsampling ratio are four optimizable hyperparameters. The L1 regularization coefficient is fixed at 0.1, the L2 regularization coefficient is fixed at 1.0, the minimum leaf node sample weights are fixed at 1, and the objective function is set to the squared regression error.

[0089] In this embodiment, the two-layer stacked structure of the Transformer encoder progressively extracts high-order features from the temporal data through a multi-layer self-attention mechanism. The multi-head self-attention submodule captures the dependencies between time steps from different subspaces using multiple parallel attention heads. The scaling dot product attention calculation formula is:

[0090] ;

[0091] in , , These are the query matrix, key matrix, and value matrix, respectively. Let be the dimension of the key vector. Multi-head attention will... The outputs of each attention point are concatenated and then linearly transformed to obtain the final output.

[0092] The feedforward network submodule contains two linear transformation layers, with the intermediate hidden layer having a dimension twice that of the model's hidden layers. The activation function is a modified linear unit. The AdamW optimizer introduces weight decay decoupling on top of the Adam optimizer. In the early stopping mechanism, training stops and rolls back to the optimal model parameters when the validation set loss does not decrease within a preset number of rounds. In the XGBoost regressor, the L1 regularization coefficient... The L2 regularization coefficient is used to control model sparsity. Used to control model complexity The minimum sum of sample weights required for a leaf node. Maximum tree depth. Value range [3, 10], learning rate Value range [0.01, 0.3], sample subsampling ratio Value range [0.5, 1.0], feature subsampling ratio The four terms with values ​​in the range [0.5, 1.0] are used as optimizable hyperparameters.

[0093] Furthermore, the step of calculating the error difference between the Transformer encoder and the XGBoost regressor using the mean absolute error of the validation set, adaptively deciding between inverse error weighted fusion or single-model selection based on the error threshold, and obtaining the final settlement prediction value by linearly superimposing the prediction results of each component, specifically includes:

[0094] The mean absolute error of the Transformer encoder and the mean absolute error of the XGBoost regressor on the validation set are calculated by the model training module. The larger of the two values ​​is used as the denominator and the absolute value of the difference between the two values ​​is used as the numerator to calculate the error difference degree.

[0095] When the error difference is less than the user-defined error threshold, an inverse error weighted fusion strategy is adopted. The fusion weight of the Transformer encoder is the average absolute error of the XGBoost regressor validation set divided by the sum of the average absolute errors of the two validation sets, and the fusion weight of the XGBoost regressor is the average absolute error of the Transformer encoder validation set divided by the sum of the average absolute errors of the two validation sets.

[0096] When the error difference is greater than or equal to the user-defined error threshold, the single model with the smaller mean absolute error of the validation set is selected as the prediction model for that component.

[0097] The prediction results after internal fusion or selection of each intrinsic mode function component are linearly superimposed to obtain the final full-band settlement prediction value.

[0098] In this embodiment, the error difference (diff) quantifies the performance difference between the Transformer encoder and the XGBoost regressor on the same component and the same validation set. The value of diff ranges from [0,1]. When diff approaches 0, it indicates that the performance of the two models is similar, and when diff approaches 1, it indicates that the performance difference between the two models is large. The user-defined error threshold... The default value is 0.1, and users can adjust this threshold according to their required prediction accuracy. In the inverse error weighted fusion strategy, the model with smaller error receives a larger fusion weight, and the model with larger error receives a smaller fusion weight, achieving adaptive fusion of the prediction results from the two models. When the performance difference between the two models is large... In this case, a single model with a smaller mean absolute error on the validation set is directly selected to avoid the drag on prediction results caused by a poor-quality model. Prediction results for each intrinsic mode function component. Through linear superposition Restore full-band signal, including This represents the total number of intrinsic mode function components.

[0099] like Figure 2The diagram shown is a structural schematic of the soft soil surface settlement prediction system based on multi-scale decomposition optimization provided in this application embodiment. The soft soil surface settlement prediction system based on multi-scale decomposition optimization provided in this application embodiment includes: a data input module, a feature selection module, a multi-scale decomposition denoising module, a model training module, a model optimization module, and a result display module.

[0100] The data input module is used to read the ground settlement time series dataset in spreadsheet format, and to enforce the data column arrangement rule that all influencing feature independent variables are arranged at the beginning and the settlement true value dependent variable is fixed as the last column. It automatically filters invalid non-numerical columns, and configures a dual-mode switchable cleaning mechanism of missing value removal and cubic spline interpolation filling. The cleaned dataset is cached in global memory for reuse by various modules.

[0101] The feature selection module has its input end connected to the output end of the data input module. It is configured with a feature selection master switch, which is used to directly output the original feature dataset when the feature selection master switch is closed, and to switch between any one of the feature selection strategies based on Pearson correlation coefficient, feature selection strategies based on tree model feature importance, or feature selection strategies based on intelligent optimization algorithm to remove redundant features when the feature selection master switch is open.

[0102] The multi-scale decomposition denoising module selectively connects its input end to the output end of the data input module or the feature selection module. It is used to perform continuous wavelet transform adaptive threshold denoising on the original settlement time series data. It uses db4 wavelet basis, 6-level decomposition and soft threshold function to filter out high-frequency random interference. It performs adaptive noise complete set empirical mode decomposition on the denoised signal, decomposes the signal into multi-level intrinsic mode function components and residual terms and outputs them completely to the global buffer.

[0103] The model training module has its input end connected to the output end of the multi-scale decomposition and denoising module. It is used to receive all intrinsic mode function components and original multi-feature data, construct time series samples according to historical window length, sliding step size and prediction step size and synchronously divide them into training set, validation set and test set. Transformer encoder and XGBoost regressor are trained independently for each intrinsic mode function component. The error difference between the Transformer encoder and XGBoost regressor is calculated by the mean absolute error of the validation set. Based on the error threshold, an adaptive decision is made to adopt error inverse weighted fusion or single model selection. The prediction results of each component are linearly superimposed to obtain the final settlement prediction value.

[0104] The model optimization module has its input end connected to the intrinsic mode function component training interface of the model training module. It is equipped with an independent optimization switch and has four freely switchable intelligent optimizers built in: Sparrow Search Algorithm (SSA), Grey Wolf Optimization Algorithm (GWO), Lobster Optimization Algorithm (COA), and Leech Optimization Algorithm (BSLO). It is used to perform global optimization of the hyperparameters of the XGBoost regressor with the optimization objective of minimizing the mean absolute error of the first intrinsic mode function component validation set, output the optimal hyperparameter combination and store it in the global cache for reuse by all intrinsic mode function components.

[0105] The results display module, as the system terminal output module, is used to batch calculate five types of evaluation indicators for the training set, validation set, and test set: mean absolute error, mean absolute percentage error, mean square error, root mean square error, and coefficient of determination. It automatically generates a measured-predicted time series comparison curve and a scatter plot, and exports the error index, optimal hyperparameters, fusion weights, and future predicted time series data.

[0106] In this embodiment, the data input module is the system's first-end module. Its data column arrangement rules clearly define that all influencing characteristic independent variables are placed at the beginning, while the true value of settlement is fixed at the end, enabling the system to automatically identify the meaning of each column field. In the missing value removal mode, a preset threshold is set so that a row with a missing value ratio exceeding 20% ​​is removed. Cubic spline interpolation filling smoothly imputes missing values ​​between adjacent known data points by constructing a cubic polynomial. The global memory is a dedicated cache area allocated within the computer's random access memory.

[0107] The feature selection module is located after the data input module, and its communication connection method is a data bus or application programming interface call. The main feature selection switch is controlled by the user through a configuration file or graphical interface. In the feature selection strategy based on intelligent optimization algorithms, the optimizer reuses any one of the following algorithms built into the model optimization module: Sparrow Search Algorithm, Gray Wolf Optimization Algorithm, Lobster Optimization Algorithm, and Leech Optimization Algorithm. The preset correlation coefficient threshold is 0.3, the preset quantile threshold is 30%, and the preset feature retention threshold is 20%.

[0108] The multi-scale decomposition and denoising module can selectively connect to the output of either the data input module or the feature selection module. When the feature selection module is off, it directly connects to the data input module; when the feature selection module is on, it connects to the feature selection module. The adaptive noise complete set empirical mode decomposition has an integration count of 100, a noise figure of 0.2, and a maximum iteration count of 10. No reconstruction processing is performed on each component; all intrinsic mode function components and residual terms are output to the global buffer as independent time series.

[0109] The model training module is the core computational module of the system. The default historical window length is 10, the sliding step size is 1, and the prediction step size is 3; users can customize these configurations. The default dataset partitioning ratio is 7 training sets: 2 validation sets: 1 test set. The Transformer encoder uses a two-layer stacked encoder structure. The multi-head self-attention submodule has a default of 2 attention heads. The hidden layer dimension of the feedforward network submodule is twice the hidden layer dimension of the model. The activation function is a modified linear unit. Training is performed using the AdamW optimizer with an initial learning rate of 0.001 and a mean squared error loss function. Error threshold. The default value is 0.1.

[0110] The model optimization module is equipped with an independent optimization switch, allowing the user to decide whether to enable global hyperparameter optimization. The module incorporates four freely switchable intelligent optimizers: Sparrow Search Algorithm (SSA), Grey Wolf Optimization Algorithm (GWO), Lobster Optimization Algorithm (COA), and Leech Optimization Algorithm (BSLO). Users can select any optimization algorithm to perform global optimization based on data size and required optimization accuracy. With the goal of minimizing the mean absolute error of the validation set of the first intrinsic mode function component, global optimization is performed on four hyperparameters of the XGBoost regressor: maximum tree depth, learning rate, sample subsampling ratio, and feature subsampling ratio. The initial population is set to 20 by default, and the number of iterations is set to 30 by default. The population iteration strategy of the selected optimization algorithm updates the individual's position based on the mean absolute error of the validation set corresponding to each hyperparameter combination as the individual's fitness value.

[0111] The built-in plotting function of the results display module uses computer graphics algorithms to implement color density mapping and linear regression fitting in the scatter plot, and automatically labels the sample size, fitting equation, and coefficient of determination. The exported Excel file contains an error index table, an optimal hyperparameter configuration table, a fusion weight table for each component, and a future prediction time series data table.

[0112] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0113] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0114] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0115] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0116] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0117] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting soft soil surface settlement based on multi-scale decomposition and optimization, characterized in that, Includes the following steps: Obtain a standardized soft soil settlement time series dataset, in which the influencing feature independent variables are arranged first and the settlement true value dependent variable is arranged last. Perform missing value removal or cubic spline interpolation filling on the dataset and output a standardized time series sample set. Feature selection processing is performed on the standardized time series sample set, and redundant features are removed based on any one of three strategies: Pearson correlation coefficient, tree model feature importance, or intelligent optimization algorithm. The original settlement time series data is subjected to continuous wavelet transform adaptive threshold denoising. The db4 wavelet basis, 6-level decomposition and soft threshold function are used to filter out high-frequency random interference. The denoised signal is subjected to adaptive noise complete set empirical mode decomposition. The signal is decomposed into multi-level intrinsic mode function components and residual terms, and all components are output. Based on the components of each intrinsic mode function and the original multi-feature data, time series samples are constructed according to the historical window length, sliding step size and prediction step size, and the samples are synchronously divided into training set, validation set and test set; For each intrinsic mode function component, a Transformer encoder and an XGBoost regressor are trained independently. The Transformer encoder uses a multi-head self-attention mechanism to capture long-range dependencies between time steps, and the XGBoost regressor is trained using a gradient boosting tree framework. The error difference between the Transformer encoder and the XGBoost regressor is calculated by means of the validation set mean absolute error. Based on the error threshold, an adaptive decision is made to use either inverse error weighted fusion or single model selection. The prediction results of each component are linearly superimposed to obtain the final settlement prediction value.

2. The method for predicting soft soil surface settlement based on multi-scale decomposition and optimization as described in claim 1, characterized in that, The process of obtaining a standardized soft soil settlement time-series dataset, performing missing value removal or cubic spline interpolation imputation on the dataset, and outputting a standardized time-series sample set specifically includes: The data input module reads the ground settlement time series dataset in spreadsheet format, and forces the data column layout rule to place all influencing characteristic independent variables at the beginning and fix the settlement true value dependent variable as the last column, and automatically filters invalid non-numerical columns. The system features a dual-mode cleaning mechanism that can be switched between missing value removal and cubic spline interpolation filling. When switching to the missing value removal mode, the system calculates the percentage of missing values ​​in a single row of data. If the percentage of missing values ​​in a single row exceeds a preset threshold, the data in that row is removed. When switching to cubic spline interpolation fill mode, cubic spline interpolation fills the missing values ​​in the dataset. The cleaned dataset is cached in global memory.

3. The soft soil surface settlement prediction method based on multi-scale decomposition and optimization as described in claim 1, characterized in that, The feature selection process performed on the standardized time-series sample set, based on any one of three strategies—Pearson correlation coefficient, tree model feature importance, or intelligent optimization algorithm—to remove redundant features, specifically includes: The feature selection module performs feature selection processing on the standardized time series sample set. The feature selection module is configured with a feature selection master switch. When the feature selection master switch is closed, the original feature dataset is directly output. When the feature selection master switch is turned on, feature selection is performed using one of the following three strategies: Feature selection based on Pearson correlation coefficient, using a preset correlation coefficient threshold as the standard to remove low-association features; Feature selection based on the importance of tree model features, using a preset quantile threshold as the standard to remove low-importance features; Based on the feature selection of the intelligent optimization algorithm, the feature subset is dynamically optimized by reusing the built-in optimizer of the model optimization module with the preset feature retention threshold as the standard.

4. The soft soil surface settlement prediction method based on multi-scale decomposition and optimization as described in claim 1, characterized in that, The process of performing continuous wavelet transform adaptive threshold denoising on the original settlement time series data, using a db4 wavelet basis, 6-level decomposition, and a soft thresholding function to filter out high-frequency random interference, specifically includes: The original settlement time series data is pre-filtered by continuous wavelet transform through a multi-scale decomposition and denoising module. The db4 wavelet basis is used for 6-level decomposition, and adaptive threshold denoising is performed in conjunction with a soft threshold function to filter out high-frequency random interference in the original settlement time series data. After the continuous wavelet transform pre-filtering process is completed, the filtered signal is output to the adaptive noise complete set empirical mode decomposition processing flow.

5. The soft soil surface settlement prediction method based on multi-scale decomposition and optimization as described in claim 1, characterized in that, The process of performing adaptive noise complete set empirical mode decomposition on the denoised signal decomposes the signal into multi-layer intrinsic mode function components and residual terms, and outputs all components, specifically including: The signal after continuous wavelet transform pre-filtering is subjected to adaptive noise complete set empirical mode decomposition processing by a multi-scale decomposition noise reduction module. The filtered signal is adaptively decomposed into multiple intrinsic mode function components and a single residual term. All intrinsic mode function components and residual terms are output to the global buffer without reconstructing each component.

6. The method for predicting soft soil surface settlement based on multi-scale decomposition and optimization as described in claim 1, characterized in that, The process involves constructing time-series samples based on the intrinsic mode function components and original multi-feature data, according to the historical window length, sliding step size, and prediction step size. These samples are then synchronously divided into training, validation, and test sets. Specifically, this includes: The model training module receives all intrinsic mode function components and raw multi-feature data. Based on each intrinsic mode function component and raw multi-feature data, standardized time series samples are constructed according to a unified historical window length, sliding step size and prediction step size. The historical window length, sliding step size and prediction step size can be customized by the user. Standardized time-series samples are synchronously divided into training, validation, and test sets according to user-defined proportions. Each intrinsic mode function component shares the same set of original multi-feature data, and each uses the true value of its own intrinsic mode function component as the prediction target.

7. The method for predicting soft soil surface settlement based on multi-scale decomposition and optimization as described in claim 1, characterized in that, The Transformer encoder and XGBoost regressor are trained independently for each intrinsic mode function component. The Transformer encoder uses a multi-head self-attention mechanism to capture long-range dependencies between time steps, and the XGBoost regressor is trained using a gradient boosting tree framework. Specifically, this includes: The Transformer encoder and XGBoost regressor are trained independently for each intrinsic modality function component through the model training module. The Transformer encoder adopts a two-layer stacked encoder layer structure. Each encoder layer contains a multi-head self-attention submodule and a feedforward network submodule. The number of attention heads in the multi-head self-attention submodule can be customized by the user. The hidden layer dimension of the feedforward network submodule is twice the hidden layer dimension of the model, and the activation function is a modified linear unit. The Transformer encoder is trained using the AdamW optimizer; The XGBoost regressor is trained using a gradient boosting tree framework.

8. The method for predicting soft soil surface settlement based on multi-scale decomposition and optimization as described in claim 1, characterized in that, The step involves calculating the error difference between the Transformer encoder and the XGBoost regressor using the mean absolute error of the validation set, adaptively deciding between inverse error weighted fusion or single-model selection based on an error threshold, and linearly superimposing the prediction results of each component to obtain the final settlement prediction value. Specifically, this includes: The mean absolute error of the Transformer encoder and the mean absolute error of the XGBoost regressor on the validation set are calculated by the model training module. The larger of the two values ​​is used as the denominator and the absolute value of the difference between the two values ​​is used as the numerator to calculate the error difference degree. When the error difference is less than the user-defined error threshold, an inverse error weighted fusion strategy is adopted. The fusion weight of the Transformer encoder is the average absolute error of the XGBoost regressor validation set divided by the sum of the average absolute errors of the two validation sets, and the fusion weight of the XGBoost regressor is the average absolute error of the Transformer encoder validation set divided by the sum of the average absolute errors of the two validation sets. When the error difference is greater than or equal to the user-defined error threshold, the single model with the smaller mean absolute error of the validation set is selected as the prediction model for that component. The prediction results after internal fusion or selection of each intrinsic mode function component are linearly superimposed to obtain the final full-band settlement prediction value.

9. A soft soil surface settlement prediction system based on multi-scale decomposition and optimization, used to implement the soft soil surface settlement prediction method based on multi-scale decomposition and optimization as described in any one of claims 1-8, characterized in that, include: The module includes a data input module, a feature selection module, a multi-scale decomposition and denoising module, a model training module, a model optimization module, and a results display module. The data input module is used to read the ground settlement time series dataset in spreadsheet format, and to enforce the data column arrangement rule that all influencing feature independent variables are arranged at the beginning and the settlement true value dependent variable is fixed as the last column. It automatically filters invalid non-numerical columns, and configures a dual-mode switchable cleaning mechanism of missing value removal and cubic spline interpolation filling. The cleaned dataset is cached in global memory for reuse by various modules. The feature selection module has its input end connected to the output end of the data input module. It is configured with a feature selection master switch, which is used to directly output the original feature dataset when the feature selection master switch is closed, and to switch between any one of the feature selection strategies based on Pearson correlation coefficient, feature selection strategies based on tree model feature importance, or feature selection strategies based on intelligent optimization algorithm to remove redundant features when the feature selection master switch is open. The multi-scale decomposition denoising module selectively connects its input end to the output end of the data input module or the feature selection module. It is used to perform continuous wavelet transform adaptive threshold denoising on the original settlement time series data. It uses db4 wavelet basis, 6-level decomposition and soft threshold function to filter out high-frequency random interference. It performs adaptive noise complete set empirical mode decomposition on the denoised signal, decomposes the signal into multi-level intrinsic mode function components and residual terms and outputs them completely to the global buffer. The model training module has its input end connected to the output end of the multi-scale decomposition and denoising module. It is used to receive all intrinsic mode function components and original multi-feature data, construct time series samples according to historical window length, sliding step size and prediction step size and synchronously divide them into training set, validation set and test set. Transformer encoder and XGBoost regressor are trained independently for each intrinsic mode function component. The error difference between the Transformer encoder and XGBoost regressor is calculated by the mean absolute error of the validation set. Based on the error threshold, an adaptive decision is made to adopt error inverse weighted fusion or single model selection. The prediction results of each component are linearly superimposed to obtain the final settlement prediction value. The model optimization module has its input end connected to the intrinsic mode function component training interface of the model training module. It is equipped with an independent optimization switch and has four freely switchable intelligent optimizers: sparrow search algorithm, gray wolf optimization algorithm, lobster optimization algorithm, and leech optimization algorithm. It is used to perform global optimization of the hyperparameters of the XGBoost regressor with the optimization objective of minimizing the mean absolute error of the first intrinsic mode function component validation set. It outputs the optimal hyperparameter combination and stores it in the global cache for reuse by all intrinsic mode function components. The results display module, as the system terminal output module, is used to batch calculate five types of evaluation indicators for the training set, validation set, and test set: mean absolute error, mean absolute percentage error, mean square error, root mean square error, and coefficient of determination. It automatically generates a measured-predicted time series comparison curve and a scatter plot, and exports the error index, optimal hyperparameters, fusion weights, and future predicted time series data.

Images