Roadbed settlement multivariate prediction method fusing related measurement

Abstract

The invention discloses a roadbed settlement multivariate prediction method fusing related measurement, and belongs to the roadbed settlement monitoring and prediction technology. According to the method, multi-block correlation analysis, auto-encoder multi-round denoising game evaluation, covariance matrix and information matrix fusion mapping and multivariate time sequence diagram neural network modeling are comprehensively adopted for on-line monitoring data in the operation and maintenance stage, and multi-point and multi-step combined prediction of roadbed settlement is achieved. The method is clear in structure, adjustable in parameter and easy to integrate with an existing roadbed settlement monitoring system, and has good engineering popularization and application value.

Description

Technical Field

[0001] This invention relates to roadbed settlement monitoring and prediction technology, specifically to a multivariate prediction method for roadbed settlement that integrates relevant metrics. Background Technology

[0002] Subgrade settlement is a key control indicator for the long-term safe operation of infrastructure such as high-speed railways, highways, and urban rail transit. Inaccurate prediction can lead to track irregularities, structural cracking, or even instability and failure. Therefore, how to use monitoring data to predict the evolution trend of subgrade settlement in a timely and accurate manner has always been an important research direction in the field of civil engineering operation and maintenance technology both domestically and internationally.

[0003] Currently, methods for predicting roadbed settlement can be broadly categorized into three types: The first type comprises theoretical and numerical methods, such as consolidation theory, elastoplastic constitutive models, and finite element analysis. These methods can provide overall deformation trends during the design phase, but are highly sensitive to soil parameters, boundary conditions, and load conditions, resulting in significant computational burden. The second type consists of statistical and empirical methods, such as regression analysis, time series models, and grey models. These are suitable for scenarios with small data scales and relatively stable changes, often assuming stationary sequences and simple relationships, thus limiting their ability to characterize abrupt changes and strongly nonlinear behaviors. The third type comprises machine learning and deep learning methods, such as neural networks, support vector machines, and LSTM. These methods have advantages in fitting nonlinearities and have been applied to roadbed or embankment settlement prediction. However, they often model single points or a small number of features, failing to systematically utilize spatial information related to multiple sensors.

[0004] Overall, the methods described above have promoted the development of roadbed settlement prediction technology to some extent. However, under complex working conditions and large-scale online monitoring, the following common problems still exist: First, after denoising the monitoring data, existing studies lack unified and quantifiable evaluation indicators for the denoising effect, making it difficult to objectively select the optimal solution from multiple denoising results. Second, most methods still model each monitoring point individually, with each point being independent. Only in a few studies is simple weighted combination or feature splicing of multi-point data performed. Overall, there is still a lack of a general method that can adaptively characterize the correlation structure of multiple monitoring points based on the statistical characteristics of the monitoring data itself (such as covariance, conditional correlation, etc.). Third, existing time-series prediction models mostly focus on long-term and short-term dependency modeling in the time dimension, lacking sufficient deep coupling with spatial structure, and their engineering practicality still has room for improvement.

[0005] In summary, existing roadbed settlement prediction technologies still have significant shortcomings, including a lack of unified objective indicators for evaluating noise reduction effects, insufficient utilization of spatial information and associated structures from multiple monitoring points, and limited spatiotemporal joint modeling capabilities. With the application of large-scale online monitoring data, there is an urgent need for a new roadbed settlement prediction method that can quantitatively evaluate noise reduction results within a unified framework, adaptively construct multi-monitoring point association structures based on the statistical characteristics of monitoring data, and perform spatiotemporal joint modeling and multi-step prediction on multivariate long-term series data. This would allow for a stable and accurate characterization of roadbed settlement evolution trends under complex working conditions and high-noise backgrounds, providing reliable technical support for operational maintenance decisions and risk warnings. Summary of the Invention

[0006] To address the aforementioned shortcomings in the existing technology, this invention provides a multivariate prediction method for roadbed settlement that integrates relevant metrics.

[0007] To achieve the above-mentioned objectives, the technical solution adopted by this invention is as follows:

[0008] A multivariate prediction method for roadbed settlement integrating relevant metrics includes the following steps:

[0009] S1. Collect settlement data from roadbed monitoring points and preprocess the data to obtain a standardized data matrix;

[0010] S2. Construct stable correlations among multiple monitoring points based on time block correlation statistical analysis;

[0011] S3. Perform multiple rounds of denoising on the data matrix using an autoencoder to obtain multiple sets of denoising results;

[0012] S4. Select the optimal denoised data from multiple denoising results based on relevant game evaluation indicators;

[0013] S5. Construct a graph structure that reflects the spatial correlation of monitoring points by integrating the covariance matrix and the information matrix;

[0014] S6. Construct a multivariate temporal graph neural network on the graph structure, and after training, realize multi-point and multi-step joint prediction of roadbed settlement.

[0015] Furthermore, step S1 specifically includes the following steps:

[0016] S11. Set up N settlement monitoring points at key locations on the roadbed and continuously collect settlement observations according to a fixed sampling cycle;

[0017] S12. Organize the settlement sequences of each monitoring point within the observation time window into a data matrix X∈ Where N is the number of monitoring points and T is the number of time points;

[0018] S13. Perform interpolation to fill in missing values ​​in the original data to provide stable and reliable input data for subsequent related analysis and model training.

[0019] Furthermore, step S2 specifically includes the following steps:

[0020] S21. Divide the preprocessed data matrix X into K consecutive time blocks along the time axis. Each block is... ,satisfy ;

[0021] S22, For each time block Calculate the Pearson correlation coefficient matrix between all pairs of monitoring points. , Let i be the element in the i-th row and j-th column of the matrix. The size of the matrix;

[0022] S23, For any pair of monitoring points Set its correlation coefficients within each time block After removing the maximum and minimum values, calculate the average value to obtain the average correlation coefficient. This yields the average correlation coefficient matrix. , Let i be the element in the i-th row and j-th column of the matrix. is the matrix size.

[0023] Furthermore, S3 specifically includes the following steps:

[0024] S31. Construct an autoencoder neural network containing an encoder and a decoder on the data matrix X. This allows it to learn the main patterns of change in the monitored data and suppress noise. It is trained using the reconstruction error as the loss function to minimize the reconstruction loss.

[0025] ;

[0026] In the formula, Forward propagation volume

[0027] S32. After the autoencoder training converges, M rounds of forward denoising are performed on the data matrix X by changing the input perturbation, dropout mask, or parameter initialization perturbation, resulting in M ​​sets of denoising results. .

[0028] Furthermore, step S4 specifically includes the following steps:

[0029] S41. Denoising results for round t The consistency between the data and the original data is calculated to obtain the first score. :

[0030] ;

[0031] In the formula, This indicates that the matrix can be flattened into a vector. The Pearson correlation coefficient is used. For raw data

[0032] S42, Calculation Overall correlation coefficient matrix And calculate its consistency with the average correlation coefficient matrix A to obtain the second score. :

[0033] ;

[0034] S43. Construct a comprehensive score using the weighting parameter ω, specifically as follows:

[0035] ,

[0036] S44. Calculate the overall score of the denoised data for each round and select the round with the highest overall score. Denoising results As the optimal denoised data.

[0037] Furthermore, S5 specifically includes:

[0038] S51. Optimal denoised data matrix Centralize the data by row and calculate the mean for each monitoring point. And obtain the mean-removed matrix. ,

[0039]

[0040] S52. Calculate the covariance matrix using unbiased estimation of sample covariance, and add regularization terms to the diagonal of the covariance matrix. And calculate the information matrix;

[0041] S43. Calculate and normalize the Frobenius norms of the covariance matrix and the information matrix respectively, and construct the final graph adjacency matrix through weighted fusion and symmetric normalization. The calculation method is as follows:

[0042]

[0043] In the formula, β is the fusion weight. For degree matrix, These are the normalized representations of the norms of the covariance matrix and the information matrix, respectively.

[0044] Furthermore, S65 specifically includes the following steps:

[0045] S61, based on the length of the history window Settlement data is used as input, in the future The step settlement value is used as the output to construct training samples and in the graph structure. Construct a multivariate temporal graph neural network;

[0046] S62. A multivariate time-series graph neural network is trained using a combined loss function. Historical window data from online monitoring is input into the trained multivariate time-series graph neural network, and the future output is obtained. Predicted settlement values ​​at each monitoring point at each time step.

[0047] Furthermore, the combination loss function in S62 is expressed as:

[0048]

[0049] In the formula, y is the true value in the training samples. γ is the network prediction value, and γ is the loss weight coefficient.

[0050] The present invention has the following beneficial effects:

[0051] First, through multi-block correlation analysis and multi-round denoising game evaluation of autoencoders, a set of denoising evaluation theories for engineering monitoring data with multiple monitoring points was established, which maintained key time series evolution characteristics and long-term stable correlation structure.

[0052] Secondly, by constructing a graph structure through the normalized weighted fusion of the covariance matrix and the information matrix, the graph adjacency relationship can simultaneously reflect the overall correlation and conditional correlation, thereby improving the collaborative deformation modeling capability of multiple monitoring points.

[0053] Third, by constructing a multivariate temporal graph neural network on this graph structure, multi-point and multi-step joint prediction is achieved. Compared with traditional single-point models and deep networks that do not consider graph structure, the prediction accuracy and stability are significantly improved in complex working conditions and strong noise backgrounds.

[0054] The method of this invention has a clear structure and adjustable parameters, making it easy to integrate with existing roadbed settlement monitoring systems and possessing good value for engineering promotion and application. Attached Figure Description

[0055] Figure 1 This is a flowchart illustrating the multivariate prediction method for roadbed settlement provided in an embodiment of the present invention. Detailed Implementation

[0056] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0057] like Figure 1 As shown in the figure, the overall process of a multivariate prediction method for roadbed settlement that integrates relevant metrics provided by an embodiment of the present invention includes the following steps:

[0058] S1. Collect settlement data from roadbed monitoring points and preprocess the data to obtain a standardized data matrix;

[0059] S2. Construct stable correlations among multiple monitoring points based on time block correlation statistical analysis;

[0060] S3. Perform multiple rounds of denoising on the data matrix using an autoencoder to obtain multiple sets of denoising results;

[0061] S4. Select the optimal denoised data from multiple denoising results based on relevant game evaluation indicators;

[0062] S5. Construct a graph structure that reflects the spatial correlation of monitoring points by integrating the covariance matrix and the information matrix;

[0063] S6. Construct a multivariate temporal graph neural network on the graph structure, and after training, realize multi-point and multi-step joint prediction of roadbed settlement.

[0064] Specifically, S1 is responsible for collecting raw monitoring data and filling in missing values ​​to form a standardized data matrix; S2 obtains the correlation matrix between multiple monitoring points in the time block dimension; S3 constructs an autoencoder to perform multiple rounds of denoising on the data to obtain multiple sets of denoised data; S4 optimizes the denoised data through relevant game evaluation; S5 fuses the covariance and information matrix to construct a graph structure; S6 builds a spatiotemporal graph neural network on this graph structure to complete the joint prediction of multi-point and multi-step settlement.

[0065] This plan details each implementation step and clarifies its corresponding specific sub-steps, which are described in detail below.

[0066] S1. Settlement monitoring data acquisition and processing

[0067] S11. Install N settlement monitoring points at key locations along the route and laterally on the roadbed. Each monitoring point is equipped with a settlement plate, displacement gauge, or other settlement sensor, forming a sensor set. Settlement observations are collected at fixed sampling intervals (e.g., 1 hour or 1 day) during the operation period to obtain time-series data of length T.

[0068] S12. Place N monitoring points within the observation time window. The settlement values ​​within are organized into a data matrix by rows:

[0069] ,

[0070] Among them, the first OK For the first The sedimentation sequence of each monitoring point at each sampling time, the first... List For a moment Settlement vectors at all monitoring points.

[0071] S13. Perform basic preprocessing on the raw data, mainly by filling in short-term missing data using linear interpolation, spline interpolation, or weighted interpolation based on adjacent monitoring points, to improve the numerical stability of subsequent correlation analysis and deep model training. The preprocessed data is still denoted as matrix X and used as input for step S2.

[0072] S2. Construction of correlation relationships in multivariate time-series data based on time block correlation statistics.

[0073] This embodiment performs statistical analysis on the correlation between multiple monitoring points at the time block scale to obtain the average correlation matrix A, thereby characterizing the stable correlation structure over a longer time scale. The specific steps are as follows:

[0074] S21. Divide the preprocessed data matrix X into K consecutive time blocks along the time axis, with each time block having a length of L time steps, such that:

[0075] ,

[0076] Let the k-th time block be:

[0077] .

[0078] The time block length L can be determined based on the sampling frequency and sedimentation change characteristics, such as "7 consecutive days" or "24 consecutive hours" as a block, to ensure that the amount of data in each block is sufficient for relevant analysis.

[0079] S22, For each time block Calculate the Pearson correlation coefficient matrix between all pairs of monitoring points:

[0080] ,

[0081] in:

[0082] ,

[0083] Let the mean of the i-th monitoring point within time block k be the diagonal element. From this, we can obtain K sets of correlation matrices describing the local correlation structure at different times. It reflects the local correlation structure between multiple points in different time periods.

[0084] S23, for any pair of monitoring points ( , Consider its K correlation coefficients across K time blocks:

[0085]

[0086] To mitigate the impact of local extreme conditions on long-term correlation estimates, the maximum and minimum values ​​are found in this set of K numbers, and these two extreme values ​​are removed from the set. The average correlation coefficient is obtained by averaging the remaining K−2 correlation coefficients.

[0087] ,

[0088] in The set of time block indices retained after removing extreme values ​​is determined by setting the diagonal elements to 1, thus yielding the average correlation coefficient matrix:

[0089] ,

[0090] This matrix effectively weakens the impact of local anomalies on correlation estimation by removing extreme values ​​from the correlation coefficient set of each pair of monitoring points along the time block dimension and then averaging them, reflecting the stable correlation structure among multiple monitoring points over a longer time scale. This provides a statistical basis for the subsequent evaluation of denoising results and the construction of the graph structure.

[0091] S3. Multi-round noise reduction processing based on autoencoder

[0092] This embodiment utilizes an autoencoder to perform nonlinear denoising on multivariate time-series data, and obtains 30 rounds of denoised data.

[0093] S31. Construct an autoencoder neural network containing an encoder and a decoder, and process the data matrix. For nonlinear denoising and reconstruction, let the autoencoder be:

[0094] ,

[0095] in It is a nonlinear mapping consisting of several fully connected layers or convolutional layers, where θ is the network parameter. The reconstruction result is used. The autoencoder is trained using the reconstruction error as the loss function, which can be written as:

[0096] ,

[0097] in Optional weights are used to reduce the influence of obvious outliers or interpolation points. Stochastic gradient descent or Adam algorithm is used to iteratively update the parameter θ until the reconstruction loss converges, resulting in an autoencoder model that can recover the main subsidence evolution features from noisy data.

[0098] S32. After the autoencoder training converges, based on the learned parameters θ, perform M rounds of forward denoising on the same data matrix X by changing the input perturbation, dropout mask, or using slightly different parameter initialization, to obtain multiple sets of denoising results:

[0099] ,

[0100] in This indicates the specific implementation during the t-th round of forward propagation (e.g., different dropout masks), which causes differences in the details of the denoising results in each round.

[0101] S4. Select the optimal denoising result based on relevant game theory evaluation;

[0102] This embodiment introduces a dual correlation index among the results of multiple rounds of denoising to perform game-like optimization, thereby obtaining the optimal denoised data that both preserves the time form and conforms to a stable correlation structure.

[0103] S41. First, define rating 1 (the ability to maintain the original form) as:

[0104] ,

[0105] in ( ) indicates that a matrix is ​​flattened into a vector by rows or columns. ( , ) represents the Pearson correlation coefficient. The larger the score of 1, the more consistent the denoising result is with the original data in terms of overall trend, and the better it can maintain the sedimentation evolution pattern.

[0106] S42. Then, for each round of denoising results Calculate its overall correlation coefficient matrix:

[0107] ,

[0108] in Based on The Pearson correlation coefficient at all T time points. Then... Comparing the score with the averaged correlation matrix A, we define score 2 (consistency with the long-term stable correlation structure) as:

[0109] ,

[0110] The larger the score 2, the more consistent the multi-point correlation structure presented by the denoising results of the t-th round on the overall time scale is with the stable correlation structure A extracted through multi-block analysis.

[0111] S43. Construct a comprehensive score from the two types of scores and select the optimal denoising result. The comprehensive score is defined as: ,

[0112] in The weights are used to adjust the trade-off between "preserving the shape of the time series" and "maintaining the relevant structure".

[0113] S44. Finally, select the round with the highest overall score:

[0114] ,

[0115] Corresponding denoised data This serves as input data for subsequent structural modeling and prediction.

[0116] S5. Construction of a graph structure based on the fusion of covariance and information matrix correlation metrics

[0117] To obtain the optimal denoised data matrix Subsequently, this embodiment constructs a fusion correlation metric graph structure based on covariance and information matrix to provide adjacency relationships for subsequent graph neural networks.

[0118] S51. Optimal denoised data matrix The data is centrally processed row-wise, and the mean value for each monitoring point is calculated.

[0119] ,

[0120] And obtain the mean-reduced matrix:

[0121] ,

[0122] S52. Calculate the covariance matrix using unbiased estimation of sample covariance:

[0123] ,

[0124] Its elemental form is:

[0125] ,

[0126] The matrix Σ describes the overall linear correlation among multiple monitoring points.

[0127] S53. To ensure the invertibility of the covariance matrix and enhance numerical stability, a regularization term is added to the diagonal of the covariance matrix, resulting in:

[0128] ,

[0129] Where I is the N-dimensional identity matrix, λ is the regularization coefficient, and the information matrix is ​​calculated as follows:

[0130] ,

[0131] It is used to reflect the strength of the direct correlation between monitoring points under the condition of other points.

[0132] S54. Calculate the Frobenius norm of the covariance matrix and the information matrix respectively:

[0133] ,

[0134] And perform normalization:

[0135] ,

[0136] To highlight the correlation strength between nodes and avoid interference from diagonal elements, we take the absolute value of the off-diagonal elements and set the diagonal elements to 0, resulting in:

[0137] .

[0138] S55. A weighted superposition method is used to fuse the covariance and information matrix to construct a relevance metric adjacency matrix:

[0139] ,

[0140] Where β represents the weight, a larger β emphasizes the overall covariance relationship, while a smaller β emphasizes the conditional correlation relationship. To control the complexity of the chart, W can be thresholded sparsity. When the value is less than a preset threshold ε, set it to 0, retaining only strongly related edges. Then calculate the degree matrix:

[0141] ,

[0142] Then, symmetric normalization is performed to obtain the final adjacency matrix used for graph neural network computation:

[0143]

[0144] After normalization and weighted fusion, the covariance matrix Σ and the information matrix Ω form a graph structure that reflects both global correlation and conditional correlation. This provides a topological prior for graph neural networks.

[0145] S6. Training and Joint Prediction of Multivariate Temporal Graph Neural Networks

[0146] This embodiment is based on the aforementioned diagram structure. A multivariate temporal graph neural network is constructed to achieve multi-step joint prediction of settlement at multiple monitoring points.

[0147] S61. Construct training samples using a sliding time window. Let the length of the historical window be... The predicted step size is For any starting time Construct the input tensor:

[0148] ,

[0149] Where F is the feature dimension, which can contain only settlement values ​​or superimposed external environment or load features. The corresponding target output is:

[0150] ,

[0151] Indicates the future Settlement values ​​at each monitoring point at each time point.

[0152] S62, in the diagram structure A multivariate temporal graph neural network is constructed. The node feature matrix at each time step t within the historical window is then processed. Perform graph convolution transformation in the spatial dimension.

[0153] In a certain layer l, the node characteristics at time step t are denoted as:

[0154] ,

[0155] Perform graph convolution at each time step:

[0156] ,

[0157] in The spatial convolution weight matrix, This is the bias vector of the spatial convolutional layer. ( ) is a non-linear activation function. This operation weights and aggregates the features of adjacent monitoring points, introducing spatially relevant information.

[0158] Will Viewed as a sequence unfolding over time, the next step is to model it using one-dimensional convolutions or gated recurrent units in the time dimension. For example, using temporal convolutions:

[0159] ,

[0160] in The length of the temporal convolution kernel. The convolution weights for the l-th layer with a time lag of m are... This is the bias vector of the temporal convolutional layer. Temporal convolution can be used to extract the short-term and medium-to-long-term dependencies of settlement evolution over time.

[0161] S63. The above spatial graph convolution and temporal convolution (or GRU) constitute a spatiotemporal block (ST-Block). Using the output of layer l as the input of layer l+1, several ST-Block layers are stacked along the network depth direction to progressively extract higher-level spatiotemporal features. After L layers of ST-Blocks, the top-level spatiotemporal feature representation is obtained. Features from several time steps near the end of the historical window can be extracted, or pooling operations can be used to compress the features of the entire historical window into a set of node representations. Then, a feedforward network (MLP) or deconvolutional structure can be used to map the spatiotemporal features to future multi-step sedimentation values, for example:

[0162] ,

[0163] S64. Define the loss function and complete network training and online prediction applications based on this loss. A combined loss function of weighted average absolute error and mean squared error is used:

[0164] ,

[0165] in and Let be the predicted value and the true value, respectively, and γ∈[0,1] be the weights of MAE and MSE. The network parameters are trained end-to-end on the training set using optimization algorithms such as Adam, and can be minimized on the independent validation set. To select The network layers and hyperparameters such as ω, β, λ, and γ are specified. After training, only the latest historical window data obtained from online monitoring needs to be used. Input a pre-trained network, and you can obtain the future. Settlement prediction results at each monitoring point at each time step By setting different levels of thresholds and early warning rules according to the needs of the project, the timing early warning information of roadbed settlement can be generated.

[0166] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0167] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0168] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0169] Specific embodiments have been used to illustrate the principles and implementation methods of this invention. The descriptions of the embodiments above are only for the purpose of helping to understand the method and core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

[0170] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of this invention.

Claims

1. A multivariate prediction method for roadbed settlement integrating relevant metrics, characterized in that, Includes the following steps: S1. Collect settlement data from roadbed monitoring points and preprocess the data to obtain a standardized data matrix; S2. Construct stable correlations among multiple monitoring points based on time block correlation statistical analysis; S3. Perform multiple rounds of denoising on the data matrix using an autoencoder to obtain multiple sets of denoising results; S4. Select the optimal denoised data from multiple denoising results based on relevant game evaluation indicators; S5. Construct a graph structure that reflects the spatial correlation of monitoring points by integrating the covariance matrix and the information matrix; S6. Construct a multivariate temporal graph neural network on the graph structure, and after training, realize multi-point and multi-step joint prediction of roadbed settlement.

2. The multivariate prediction method for roadbed settlement integrating relevant metrics according to claim 1, characterized in that, S1 specifically includes the following steps: S11. Set up N settlement monitoring points at key locations on the roadbed and continuously collect settlement observations according to a fixed sampling cycle; S12. Organize the settlement sequences of each monitoring point within the observation time window into a data matrix X∈ Where N is the number of monitoring points and T is the number of time points; S13. Perform interpolation to fill in missing values ​​in the original data to provide stable and reliable input data for subsequent related analysis and model training.

3. The multivariate prediction method for roadbed settlement integrating relevant metrics according to claim 1, characterized in that, S2 specifically includes the following steps: S21. Divide the preprocessed data matrix X into K consecutive time blocks along the time axis. Each block is... ,satisfy ; S22, For each time block Calculate the Pearson correlation coefficient matrix between all pairs of monitoring points. , Let i be the element in the i-th row and j-th column of the matrix. The size of the matrix; S23, For any pair of monitoring points Set its correlation coefficients within each time block After removing the maximum and minimum values, calculate the average value to obtain the average correlation coefficient. This yields the average correlation coefficient matrix. , Let i be the element in the i-th row and j-th column of the matrix. is the matrix size.

4. The multivariate prediction method for roadbed settlement integrating relevant metrics according to claim 1, characterized in that, S3 specifically includes the following steps. S31. Construct an autoencoder neural network containing an encoder and a decoder on the data matrix X. This allows it to learn the main patterns of change in the monitored data and suppress noise. It is trained using the reconstruction error as the loss function to minimize the reconstruction loss. ； In the formula, Forward propagation volume S32. After the autoencoder training converges, M rounds of forward denoising are performed on the data matrix X by changing the input perturbation, dropout mask, or parameter initialization perturbation, resulting in M ​​sets of denoising results. .

5. The multivariate prediction method for roadbed settlement integrating relevant metrics according to claim 1, characterized in that, S4 specifically includes the following steps: S41. Denoising results for round t The consistency between the data and the original data is calculated to obtain the first score. : ； In the formula, This indicates that the matrix can be flattened into a vector. The Pearson correlation coefficient is used. For raw data S42, Calculation Overall correlation coefficient matrix And calculate its consistency with the average correlation coefficient matrix A to obtain the second score. : ； S43. Construct a comprehensive score using the weighting parameter ω, specifically as follows: ， S44. Calculate the overall score of the denoised data for each round and select the round with the highest overall score. Denoising results As the optimal denoised data.

6. The multivariate prediction method for roadbed settlement integrating relevant metrics according to claim 5, characterized in that, Specifically, S5 includes: S51. Optimal denoised data matrix Centralize the data by row and calculate the mean for each monitoring point. And obtain the mean-removed matrix. , S52. Calculate the covariance matrix using unbiased estimation of sample covariance, and add regularization terms to the diagonal of the covariance matrix. And calculate the information matrix; S43. Calculate and normalize the Frobenius norms of the covariance matrix and the information matrix respectively, and construct the final graph adjacency matrix through weighted fusion and symmetric normalization. The calculation method is as follows: In the formula, β is the fusion weight. For degree matrix, These are the normalized representations of the norms of the covariance matrix and the information matrix, respectively.

7. The multivariate prediction method for roadbed settlement integrating relevant metrics according to claim 6, characterized in that, S65 specifically includes the following steps: S61, based on the length of the history window Settlement data is used as input, in the future The step settlement value is used as the output to construct training samples and in the graph structure. Construct a multivariate temporal graph neural network; S62. A multivariate time-series graph neural network is trained using a combined loss function. Historical window data from online monitoring is input into the trained multivariate time-series graph neural network, and the future output is obtained. Predicted settlement values ​​at each monitoring point at each time step.

8. The multivariate prediction method for roadbed settlement integrating relevant metrics according to claim 7, characterized in that, The combined loss function in S62 is expressed as follows: In the formula, y is the true value in the training samples. γ is the network prediction value, and γ is the loss weight coefficient.

Images