Roadbed settlement time sequence prediction and three-dimensional deformation monitoring system and method thereof

Abstract

The invention discloses a roadbed settlement time sequence prediction and three-dimensional deformation monitoring system and method, and belongs to the technical field of earth surface deformation monitoring and prediction.The system comprises a data acquisition module, a spatial-temporal feature construction module, a prediction model fusion module, a deformation field reconstruction module, a synthetic data generation module and a monitoring and early warning module; according to the method, an InSAR satellite remote sensing image and foundation micro-deformation sensor data are fused, a time-space coupled Transform-GRU prediction model is constructed, a pixel-level settlement rate field algorithm is created for the first time, the earth surface deformation monitoring precision is improved to 0.5 mm every year, historical settlement images are synthesized through a generative adversarial network, training data are enhanced, and the earth surface deformation monitoring precision is improved to 0.5 mm every year. Accurate reconstruction and long-term prediction of the three-dimensional deformation field of the roadbed are realized, the problem of insufficient generalization ability under the conditions of insufficient three-dimensional deformation reconstruction, long-term prediction precision and data in the prior art is effectively solved, and powerful support is provided for safe operation of traffic infrastructures.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of ground surface deformation monitoring and prediction, in particular to a roadbed settlement time series prediction and three-dimensional deformation monitoring system and method, which can be applied to the health monitoring and safety warning of roadbed engineering such as highways, railways and airport runways. BACKGROUND

[0002] With the rapid development of transportation infrastructure, roadbed settlement has become a key factor affecting road safety and service life. Traditional roadbed settlement monitoring mainly relies on ground measurement methods such as leveling and global navigation satellite system (GNSS) measurement, which has high accuracy but has problems such as limited monitoring points, high cost, and difficulty in achieving large-scale continuous monitoring.

[0003] CN119437159A discloses a ground settlement monitoring method and system fusing Beidou and InSAR data. The technical solution receives positioning data of the Beidou satellite system and ground surface deformation data obtained by the synthetic aperture radar interferometric measurement system, constructs a spatio-temporal fusion model to correct the ground surface deformation data, uses the corrected ground surface deformation data to determine the development trend of the ground settlement using multi-period analysis technology, and generates a ground settlement distribution map. This scheme mainly uses Kalman filter algorithm to suppress noise of ground surface deformation data, applies autoregressive integrated moving average model for trend analysis, and combines wavelet transform technology to analyze settlement patterns at different frequencies.

[0004] However, the above prior art has the following shortcomings:

[0005] First, the prior art mainly focuses on two-dimensional ground surface deformation monitoring and analysis, and lacks the ability to accurately reconstruct the three-dimensional deformation field of the roadbed. Roadbed settlement not only includes vertical subsidence, but also includes horizontal displacement. The prior art is difficult to fully depict the three-dimensional deformation characteristics of the roadbed, resulting in insufficient understanding of the settlement mechanism.

[0006] Second, the time series prediction method used in the prior art is mainly based on traditional statistical models such as autoregressive integrated moving average model. Such models are difficult to capture the complex nonlinear spatio-temporal dependence relationship in roadbed settlement data, especially in long-time span prediction tasks, the prediction accuracy is significantly reduced.

[0007] Third, the prior art lacks the ability to calculate the fine settlement rate field at the pixel level. Although InSAR technology can provide large-scale deformation monitoring, the existing method does not fully utilize InSAR data, and fails to reconstruct the settlement rate field at the pixel level, resulting in insufficient ability to identify local settlement abnormal areas.

[0008] Fourth, the existing technology has poor generalization ability and prediction stability in the case of insufficient data samples. Subgrade settlement monitoring usually requires long-term historical data accumulation. In the early stage of the project or in the case of data missing, the existing method is difficult to provide reliable prediction results.

[0009] Therefore, there is an urgent need for a subgrade settlement time series prediction and three-dimensional deformation monitoring system and method that can realize pixel-level settlement rate field reconstruction, integrate deep learning time series prediction technology, and have data enhancement capability. SUMMARY

[0010] To solve the above problems existing in the prior art, the present application provides a subgrade settlement time series prediction and three-dimensional deformation monitoring system and method. The present application fuses InSAR satellite remote sensing images and ground micro-deformation sensor data to build a spatio-temporal coupled Transformer-GRU prediction model, creates a pixel-level settlement rate field algorithm, and improves the surface deformation monitoring accuracy to 0.5 millimeters per year. The present application also generates adversarial network to synthesize historical settlement images to enhance training data, and realizes accurate reconstruction and long-term prediction of the three-dimensional deformation field of the subgrade.

[0011] In a first aspect of the present application, a subgrade settlement time series prediction and three-dimensional deformation monitoring system is provided, comprising a data acquisition module, a spatio-temporal feature construction module, a prediction model fusion module, a deformation field reconstruction module, a synthetic data generation module, and a monitoring and early warning module.

[0012] The data acquisition module is used to receive InSAR satellite remote sensing image data and ground micro-deformation sensor data, perform pixel-level registration processing on the InSAR satellite remote sensing image data to generate a pixel-level surface deformation dataset, and perform spatio-temporal alignment processing on the ground micro-deformation sensor data to generate a ground deformation dataset.

[0013] The spatio-temporal feature construction module divides the pixel-level surface deformation dataset according to a spatial resolution threshold, and if the number of pixel points in a single grid element meets a first density range, extracts the spatio-temporal feature vector corresponding to the grid element. The spatio-temporal feature vector includes settlement rate field features, spatial gradient field features, and time series features.

[0014] The prediction model fusion module builds a spatio-temporally coupled Transformer-GRU prediction model, which includes a Transformer encoder layer and a GRU decoder layer. The Transformer encoder layer is used to capture long-time dependencies in the spatio-temporal feature vector, and the GRU decoder layer is used to predict the settlement trend at future time.

[0015] The deformation field reconstruction module calculates a pixel-level settlement rate field based on the prediction result of the Transformer-GRU prediction model, and marks a pixel point as a high-risk area if the settlement rate of any pixel point in the pixel-level settlement rate field meets a first rate range, and reconstructs a three-dimensional deformation field corresponding to the high-risk area.

[0016] The synthetic data generation module constructs a generative adversarial network, the generative adversarial network including a generator network and a discriminator network, the generator network being used for synthesizing historical settlement images, and the discriminator network being used for discriminating between real settlement images and synthetic historical settlement images.

[0017] The monitoring and early warning module generates a roadbed settlement monitoring report according to the three-dimensional deformation field, and triggers a corresponding level of early warning signal if the maximum cumulative settlement amount in the three-dimensional deformation field meets a first settlement threshold range or a second settlement threshold range.

[0018] In a possible implementation, the data acquisition module receives multi-temporal InSAR satellite remote sensing images, the multi-temporal InSAR satellite remote sensing images including at least 15 SAR images. Persistent scatterer identification is performed on the multi-temporal InSAR satellite remote sensing images, and a persistent scatterer point set is extracted, the spatial density of the persistent scatterer point set being preferably 300 to 1000 persistent scatterer points per square kilometer. Time-series InSAR processing is performed on the persistent scatterer point set, a deformation time series of each persistent scatterer point is calculated, and a pixel-level ground surface deformation dataset is generated. Measurement data of a ground-based micro-deformation sensor at multiple monitoring points are received, the measurement data including displacement amounts and time stamps. The measurement data of the ground-based micro-deformation sensor are matched with spatial coordinates of the persistent scatterer points, time series are aligned and fused if a matching distance is less than a spatial matching threshold, and the spatial matching threshold is preferably 20 to 50 meters. This multi-source data fusion manner can comprehensively utilize the large-range coverage advantage of satellite remote sensing and the high-precision monitoring capability of a ground-based sensor, and improve the spatial density and time resolution of settlement monitoring.

[0019] In a possible implementation, the space-time feature construction module comprises a grid division sub-module, a density screening sub-module, and a feature extraction sub-module. The grid division sub-module divides the pixel-level ground surface deformation data set into a plurality of grid cells according to a spatial resolution threshold, and counts the number of pixel points in each grid cell. The density screening sub-module determines whether the number of pixel points meets a first density range. If yes, the grid cell is marked as a valid cell; if not, the grid cell is marked as an invalid cell. The first density range is preferably 10 to 50 pixel points per grid cell. The feature extraction sub-module extracts a space-time feature vector for the grid cell marked as a valid cell. The space-time feature vector includes the average subsidence rate of the pixel points in the grid cell, the spatial standard deviation of the subsidence rate, the time variation rate of the subsidence rate, and the subsidence gradient with adjacent grid cells. This grid-based space-time feature extraction method can effectively reduce the data dimension while retaining the spatial distribution pattern and time evolution characteristics of subsidence.

[0020] In a possible implementation, the subsidence rate field feature includes a pixel-level subsidence rate and a spatial distribution pattern of the subsidence rate. The calculation method of the pixel-level subsidence rate is as follows: linear regression fitting is performed on the deformation time series of each pixel point, and the slope of the regression straight line is calculated as the pixel-level subsidence rate; the pixel-level subsidence rate is subjected to Gaussian filter smoothing processing, and the filter window size is preferably 3 to 5 pixels; the spatial gradient field of the pixel-level subsidence rate is calculated, and the spatial gradient field reflects the difference in subsidence rate between adjacent pixel points; if the gradient value at any position in the spatial gradient field meets a first gradient range, the position is marked as a subsidence abnormal area, and the first gradient range is preferably greater than 5 mm per year per 100 m. The calculation accuracy of the pixel-level subsidence rate field is preferably 0.5 mm per year to 2 mm per year. This high-precision pixel-level analysis can identify local subsidence abnormalities, and provide a more reliable basis for roadbed safety warning.

[0021] In a possible implementation, the method for constructing the Transformer-GRU prediction model comprises: constructing a Transformer encoder layer, the Transformer encoder layer comprising a multi-head self-attention mechanism and a feedforward neural network, and the number of heads of the multi-head self-attention mechanism being preferably 4 to 8; inputting the spatiotemporal feature vector into the Transformer encoder layer, and calculating the spatiotemporal correlation weight between the feature vectors through the multi-head self-attention mechanism; constructing a GRU decoder layer, the GRU decoder layer comprising a plurality of GRU units, and the dimension of the hidden layer of the GRU unit being preferably 128 to 512; taking the output of the Transformer encoder layer as the input of the GRU decoder layer, and predicting the subsidence values at a plurality of future time points through the GRU decoder layer, and the time span of the future time points being preferably 1 month to 12 months. If the time span of the spatiotemporal feature vector meets a first time range, the Transformer encoder layer adopts a first number-of-layers configuration, and if the time span of the spatiotemporal feature vector meets a second time range, the Transformer encoder layer adopts a second number-of-layers configuration, the first time range being 1 year to 3 years, the second time range being 3 years to 10 years, the first number-of-layers configuration being 4 to 6, and the second number-of-layers configuration being 8 to 12. The deep learning architecture coupling space and time can capture both long-time dependence and short-time fluctuation characteristics in the subsidence data, and significantly improve the prediction accuracy.

[0022] It should be noted that the multi-head self-attention mechanism of the Transformer encoder layer can calculate the correlation between different spatial positions and different time steps in parallel, and has stronger long-time dependence modeling capability compared with the traditional recurrent neural network. The GRU decoder layer selectively retains and updates historical information through the gating mechanism, reduces the computational complexity while ensuring the prediction accuracy. The fusion architecture of the Transformer encoder layer and the GRU decoder layer gives full play to the advantages of the two models, and forms a spatiotemporal coupling prediction framework.

[0023] In a possible implementation, the processing flow of the deformation field reconstruction module comprises: if the subsidence rate of a pixel point in the pixel-level subsidence rate field meets a first rate range, extracting the three-dimensional coordinates and the subsidence rate of the pixel point; based on the three-dimensional coordinates and the subsidence rate, reconstructing the three-dimensional deformation field of the high-risk area by using the Kriging interpolation method, the Kriging interpolation method comprising calculating a semi-variogram function and constructing an interpolation weight matrix; performing spatial smoothing processing on the three-dimensional deformation field to remove high-frequency noise introduced in the interpolation process; calculating the deformation gradient field and the curvature field of the three-dimensional deformation field, the deformation gradient field reflecting the spatial variation trend of the deformation, and the curvature field reflecting the spatial bending degree of the deformation. The first rate range is preferably greater than 10 mm per year, indicating that there is a significant subsidence phenomenon in the area, which needs to be focused on and monitored.

[0024] In a possible implementation, the training process of the synthetic data generation module includes: constructing a generator network, the generator network adopts a convolutional neural network architecture, including an encoder part and a decoder part; constructing a discriminator network, the discriminator network adopts a convolutional neural network architecture, and is used for distinguishing real subsidence images and synthetic historical subsidence images; randomly sampling real subsidence images from a pixel-level ground surface deformation data set, and generating synthetic historical subsidence images from random noise vectors; calculating the discrimination output of the discriminator network for the real subsidence images and the discrimination output of the discriminator network for the synthetic historical subsidence images, and updating the parameters of the generator network and the discriminator network based on the discrimination output; if the discrimination accuracy of the discriminator network for the synthetic historical subsidence images meets a first accuracy range, it is indicated that the synthetic historical subsidence images are highly similar to the real subsidence images, and the synthetic historical subsidence images are added to the training data set. The first accuracy range is preferably lower than 60%, which means that the discriminator network is difficult to distinguish real images and synthetic images, and at this time, the synthetic images have reached a high enough quality. This data enhancement method based on the generative adversarial network can effectively alleviate the problem of insufficient historical data and improve the generalization ability and robustness of the prediction model.

[0025] In a possible implementation, the monitoring and early warning module is further configured to: calculate spatial distribution characteristics of the roadbed settlement according to the three-dimensional deformation field, the spatial distribution characteristics including a settlement center position, a settlement influence range, and a settlement morphology characteristic; calculate a settlement development trend at a future time according to the prediction result of the Transformer-GRU prediction model, the settlement development trend including a settlement acceleration and a predicted time of reaching a dangerous threshold; generate a roadbed settlement monitoring report, the roadbed settlement monitoring report including a current settlement state, a historical settlement evolution curve, a future settlement prediction curve, an early warning level, and a recommended treatment measure; if the early warning level is a first-level early warning, send a monitoring reminder information to a monitoring terminal; if the early warning level is a second-level early warning, send an emergency treatment information to a management terminal. The first settlement threshold range is preferably 30 mm to 50 mm, indicating that the settlement has reached a level that needs to be paid attention to; the second settlement threshold range is preferably greater than 50 mm, indicating that the settlement has reached a dangerous level, and a treatment measure needs to be taken immediately.

[0026] In a possible implementation, the system further comprises an adaptive optimization module configured to monitor the prediction error of the Transformer-GRU prediction model, keep the parameters of the Transformer-GRU prediction model unchanged if the prediction error meets a first error range, and trigger a model parameter updating process if the prediction error meets a second error range, the first error range is preferably that the root mean square error is less than 3 mm, and the second error range is preferably that the root mean square error is greater than 5 mm. The model parameter updating process comprises obtaining the latest pixel-level ground surface deformation dataset and ground deformation dataset, incrementally training the Transformer-GRU prediction model, and updating the weight parameters of the Transformer encoder layer and the GRU decoder layer. This adaptive optimization mechanism can enable the system to continuously adapt to changes in the settlement mode during long-term operation and maintain the stability of the prediction accuracy.

[0027] In a second aspect, the application provides a subgrade settlement time series prediction and three-dimensional deformation monitoring method based on the above system, comprising receiving InSAR satellite remote sensing image data and ground-based micro-deformation sensor data, performing pixel-level registration processing and space-time alignment processing; dividing the pixel-level ground surface deformation dataset into grids according to a spatial resolution threshold, and extracting a space-time feature vector; constructing a space-time coupled Transformer-GRU prediction model; calculating a pixel-level settlement rate field based on the prediction result, and reconstructing a three-dimensional deformation field; constructing a generative adversarial network to synthesize historical settlement images; generating a subgrade settlement monitoring report and triggering an early warning signal according to the three-dimensional deformation field.

[0028] The application has the following advantages:

[0029] First, the application first creates a pixel-level settlement rate field algorithm, which improves the ground surface deformation monitoring accuracy to 0.5 mm / year to 2 mm / year, which is significantly higher than the accuracy level of the prior art. By performing fine-grained analysis on the deformation time series of each pixel, local settlement abnormal areas can be identified, providing more reliable basis for subgrade safety warning.

[0030] Second, the application uses a space-time coupled Transformer-GRU prediction model, which can capture both long-term dependencies and short-term fluctuations in settlement data. The Transformer encoder layer calculates the correlation between different space-time positions in parallel through the multi-head self-attention mechanism, and the GRU decoder layer selectively retains and updates historical information through the gating mechanism. The fusion architecture of the two performs well in long-time span (1 month to 12 months) prediction tasks, and the prediction accuracy is significantly better than that of traditional statistical models.

[0031] Thirdly, the present application effectively alleviates the problem of insufficient historical data by generating historical settlement images through a generative adversarial network. The generator network can learn the distribution characteristics of real settlement images and generate synthetic images with high similarity. These synthetic images can enhance the generalization ability and robustness of the prediction model, especially in the early stages of engineering or in cases of data missing, and can provide more reliable prediction results.

[0032] Fourthly, the present application realizes the accurate reconstruction of the three-dimensional deformation field of the roadbed, which not only can monitor the vertical settlement, but also can capture the horizontal displacement. By calculating the deformation gradient field and the curvature field, the three-dimensional deformation characteristics of the roadbed can be fully described, providing a scientific basis for in-depth analysis of the settlement mechanism and engineering disposal.

[0033] Fifthly, the present application establishes a hierarchical warning mechanism, which triggers different levels of warning signals according to different threshold ranges of the maximum cumulative settlement. The first level of warning (30-50mm) reminds relevant personnel to strengthen monitoring, and the second level of warning (more than 50mm) requires immediate disposal measures. This hierarchical warning strategy can effectively prevent roadbed settlement risks and ensure the safe operation of transportation infrastructure. BRIEF DESCRIPTION OF DRAWINGS

[0034] Figure 1 is the overall architecture schematic diagram of the roadbed settlement time series prediction and three-dimensional deformation monitoring system of the present application.

[0035] Figure 2 is the processing flow schematic diagram of the data acquisition module of the present application.

[0036] Figure 3 is the structure schematic diagram of the spatio-temporal feature construction module of the present application.

[0037] Figure 4 is the architecture schematic diagram of the Transformer-GRU prediction model of the present application.

[0038] Figure 5 is the processing flow schematic diagram of the deformation field reconstruction module of the present application.

[0039] Figure 6 is the flowchart of the roadbed settlement time series prediction and three-dimensional deformation monitoring method of the present application. DETAILED DESCRIPTION

[0040] Please refer to the attached Figures 1-6The specific embodiments of the present application will be described in detail below with reference to the accompanying drawings, so that those skilled in the art can better understand the present application. It should be noted that in the description of the present application, the terms "first", "second", etc. are only for the purpose of description and cannot be understood as indicating or implying relative importance. In addition, in the description of the present application, the meaning of "a plurality of" is two or more, unless otherwise specified.

[0041] As shown in Figure 1 The present application provides a roadbed settlement timing prediction and three-dimensional deformation monitoring system, which comprises a data acquisition module 1, a space-time feature construction module 2, a prediction model fusion module 3, a deformation field reconstruction module 4, a synthetic data generation module 5 and a monitoring and early warning module 6.

[0042] The data acquisition module 1 is used to receive InSAR satellite remote sensing image data and ground-based micro-deformation sensor data, perform pixel-level registration processing on the InSAR satellite remote sensing image data to generate a pixel-level ground surface deformation data set, and perform space-time alignment processing on the ground-based micro-deformation sensor data to generate a ground-based deformation data set. The data acquisition module 1 is the data input interface of the entire system, and its processing quality directly affects the performance of the subsequent modules.

[0043] The space-time feature construction module 2 is connected with the data acquisition module 1 and is used to grid divide the pixel-level ground surface deformation data set according to a spatial resolution threshold, and if the number of pixel points in a single grid cell meets a first density range, the space-time feature vector corresponding to the grid cell is extracted. The space-time feature vector includes a settlement rate field feature, a spatial gradient field feature and a time series feature, which can comprehensively depict the space-time evolution law of roadbed settlement.

[0044] The prediction model fusion module 3 is connected with the space-time feature construction module 2 and is used to construct a space-time coupled Transformer-GRU prediction model. The Transformer-GRU prediction model includes a Transformer encoder layer and a GRU decoder layer, the Transformer encoder layer is used to capture the long-time dependence relationship in the space-time feature vector, and the GRU decoder layer is used to predict the settlement trend at the future time. The prediction model fusion module 3 is one of the core innovations of the present application, which realizes the accurate prediction of roadbed settlement through deep learning technology.

[0045] The deformation field reconstruction module 4 is connected with the prediction model fusion module 3 and is used to calculate the pixel-level settlement rate field based on the prediction result of the Transformer-GRU prediction model, mark the pixel point as a high-risk area if the settlement rate of any pixel point in the pixel-level settlement rate field meets the first rate range, and reconstruct the three-dimensional deformation field corresponding to the high-risk area. The deformation field reconstruction module 4 realizes the conversion from one-dimensional time series to three-dimensional space field, which provides an intuitive tool for the visualization and analysis of settlement.

[0046] The synthetic data generation module 5 is connected to the deformation field reconstruction module 4 to construct a generative adversarial network (GAN). The GAN includes a generator network and a discriminator network. The generator network synthesizes historical subsidence images, and the discriminator network distinguishes between real subsidence images and synthesized historical subsidence images. The synthetic data generation module 5 is another core innovation of this invention, improving the generalization ability of the prediction model through data augmentation technology.

[0047] The monitoring and early warning module 6 is connected to the deformation field reconstruction module 4 and is used to generate a roadbed settlement monitoring report based on the three-dimensional deformation field. If the maximum cumulative settlement in the three-dimensional deformation field meets the first settlement threshold range, a first-level early warning signal is triggered; if the maximum cumulative settlement meets the second settlement threshold range, a second-level early warning signal is triggered. The monitoring and early warning module 6 transforms the system's analysis results into actionable early warning information, providing decision support for engineering management personnel.

[0048] like Figure 2 As shown, the processing flow of data acquisition module 1 includes the following steps:

[0049] Step S11: Receive multi-temporal InSAR satellite remote sensing images, which include at least 15 SAR images. In one specific embodiment, C-band SAR images from the Sentinel-1 satellite can be used, with a time span of 2 years and an interval of 12 days, for a total of 60 SAR images.

[0050] Step S12: Identify persistent scatterers in multi-temporal InSAR satellite remote sensing images and extract persistent scattering point sets. Persistent scatterers refer to targets that maintain coherence in long-term SAR images, typically corresponding to stable surface features such as buildings and rocks. The persistent scatterer identification method involves calculating the amplitude dispersion index and coherence coefficient of the time-series SAR images, selecting pixels with an amplitude dispersion index less than 0.25 and a coherence coefficient greater than 0.7 as persistent scattering points. The spatial density of the persistent scattering point set is preferably 300 to 1000 persistent scattering points per square kilometer. In urban areas, due to the large number of buildings, the persistent scattering point density can reach over 1000 per square kilometer; in suburban or rural areas, the persistent scattering point density may only be around 300 per square kilometer.

[0051] Step S13: Perform temporal InSAR processing on the persistent scattering point set to calculate the deformation time series for each persistent scattering point. Temporal InSAR processing includes differential interferometry, phase unwrapping, and atmospheric delay correction. Differential interferometry extracts surface deformation information by calculating the interferometric phase difference between adjacent SAR images. Phase unwrapping converts the wrapped phase difference into continuous phase values; common methods include the minimum cost flow algorithm and the statistical cost network flow algorithm. Atmospheric delay correction removes the influence of atmospheric water vapor and ionospheric delay on the interferometric phase; this can be achieved using spatiotemporal filtering methods or by combining meteorological data. Through temporal InSAR processing, a deformation time series for each persistent scattering point is generated. The sampling interval for the time series is 12 days, the time span is 2 years, and a total of 60 deformation observations are included.

[0052] Step S14: Receive measurement data from the foundation micro-deformation sensor at multiple monitoring points. The measurement data includes displacement and timestamps. The foundation micro-deformation sensor can be an inclinometer, crack gauge, displacement gauge, etc., and is typically deployed at key locations on the roadbed, such as embankment slopes and bridge abutments. The measurement accuracy of the foundation micro-deformation sensor is typically 0.1 mm to 1 mm, and the sampling frequency is from once per hour to once per day.

[0053] Step S15: Match the measurement data from the ground-based microdeformation sensor with the spatial coordinates of the persistent scattering point. If the matching distance is less than the spatial matching threshold, perform time series alignment and fusion. The spatial matching threshold is preferably 20 to 50 meters, and its selection needs to consider the spatial resolution of InSAR pixels and the representative range of the ground-based sensor. Time series alignment uses a time interpolation method to unify data from different sampling frequencies onto the same time grid. Time series fusion uses a weighted average method, determining the weighting coefficients based on the measurement accuracy of InSAR and the ground-based sensor. The measurement accuracy of InSAR is approximately 2 to 5 millimeters, and the measurement accuracy of the ground-based sensor is approximately 0.1 to 1 millimeter. Therefore, the weighting coefficient for the ground-based sensor is typically set to 0.6 to 0.8, and the weighting coefficient for InSAR is 0.2 to 0.4.

[0054] Through the above processing flow, data acquisition module 1 generates pixel-level surface deformation datasets and foundation deformation datasets, providing a high-quality data foundation for subsequent spatiotemporal feature extraction and predictive modeling.

[0055] like Figure 3 As shown, the spatiotemporal feature construction module 2 includes a grid partitioning submodule 21, a density filtering submodule 22, and a feature extraction submodule 23.

[0056] The mesh division sub-module 21 divides the pixel-level ground surface deformation dataset into a plurality of mesh cells according to a spatial resolution threshold, and counts the number of pixel points in each mesh cell. The spatial resolution threshold is preferably 5 meters to 20 meters. The selection of the threshold needs to balance the spatial resolution and the computational complexity. A smaller spatial resolution threshold (such as 5 meters) can provide more detailed spatial details, but will increase the number of mesh cells and the amount of calculation; a larger spatial resolution threshold (such as 20 meters) can reduce the amount of calculation, but may lose some spatial details. In a specific embodiment, for highway subgrade monitoring, the spatial resolution threshold can be set to 10 meters, and the entire monitoring area is divided into 10 meter x 10 meter mesh cells.

[0057] The density screening sub-module 22 determines whether the number of pixel points in each mesh cell meets a first density range. If it meets, the mesh cell is marked as a valid cell; if it does not meet, the mesh cell is marked as an invalid cell. The first density range is preferably 10 to 50 pixel points per mesh cell. If the number of pixel points in the mesh cell is too small (such as less than 10), the statistical characteristics of the mesh cell may not be stable enough and are easily affected by individual abnormal points; if the number of pixel points is too large (such as more than 50), it may be due to the spatial resolution threshold being set too large, causing pixel points with different settlement characteristics to be mixed in the same mesh cell. Through density screening, it can be ensured that each valid mesh cell contains a sufficient number of pixel points with similar characteristics, providing reliable samples for subsequent feature extraction.

[0058] The feature extraction sub-module 23 extracts a spatio-temporal feature vector for the mesh cells marked as valid cells. The spatio-temporal feature vector includes the following types of features:

[0059] The first type of feature is the average settlement rate of the pixel points in the mesh cell. The settlement rates of all pixel points in the mesh cell are arithmetically averaged to obtain the average settlement rate of the mesh cell. The average settlement rate reflects the overall settlement trend of the region.

[0060] The second type of feature is the spatial standard deviation of the settlement rate. The standard deviation of the settlement rates of all pixel points in the mesh cell is calculated, which reflects the spatial dispersion degree of the settlement rate. A larger spatial standard deviation indicates that there is uneven settlement within the mesh cell, which may be caused by uneven soil quality or uneven load distribution.

[0061] The third type of feature is the time change rate of the settlement rate. The time change rate of the settlement rate is obtained by difference calculation on the time series of the settlement rate of each pixel point, and then the time change rates of all pixel points in the mesh cell are averaged. The time change rate of the settlement rate reflects the trend of settlement acceleration or deceleration, a positive value indicates that the settlement speed is accelerating, and a negative value indicates that the settlement speed is slowing down.

[0062] The fourth type of feature is the subsidence gradient of the adjacent grid cells. The difference between the average subsidence rate of the current grid cell and the average subsidence rate of the east, south, west and north adjacent grid cells is calculated, and then divided by the distance between the grid cells to obtain the subsidence gradient in four directions. The subsidence gradient reflects the spatial variation trend of subsidence, and a larger subsidence gradient indicates that there is a significant spatial discontinuity in the subsidence of the region.

[0063] The above four types of features are combined into a feature vector as the spatio-temporal feature vector of the grid cell. In specific embodiments, if four features (average subsidence rate, spatial standard deviation, time variation rate, and mean of subsidence gradient) are extracted for each grid cell, the dimension of the spatio-temporal feature vector is 4. If more detailed feature description is needed, other statistical features such as the maximum, minimum, median and quartile of the subsidence rate can also be added.

[0064] As shown in FIG. 3, the architecture of the Transformer-GRU prediction model includes a Transformer encoder layer 31 and a GRU decoder layer 32. Figure 4

[0065] The Transformer encoder layer 31 includes a multi-head self-attention mechanism 311 and a feedforward neural network 312. The core idea of the multi-head self-attention mechanism 311 is to calculate the correlation weight between different positions in the input sequence, so as to capture long-distance dependencies. The calculation process of the multi-head self-attention mechanism is as follows:

[0066] First, the input spatio-temporal feature vector sequence is represented as ,where represents the spatio-temporal feature vector at the th time, is the feature dimension, is the length of the time series.

[0067] Then, the input sequence is mapped to query vectors , key vectors and value vectors through three linear transformation matrices :

[0068] ,

[0069] where: is the query matrix, is the key matrix, is the value matrix, is the query transformation matrix, is the key transformation matrix, is the value transformation matrix, is the dimension of the key and query. is the dimension of the value.

[0070] Next, the self-attention weight is calculated:

[0071] ,

[0072] wherein: is the attention weight matrix, denotes the attention weight from the th time step to the th time step, the function normalizes each row so that the sum of the weights of each row is 1, is a scaling factor to prevent the gradient from vanishing due to the inner product being too large.

[0073] Finally, the value vector is weighted and summed according to the attention weight to obtain the output of the self-attention:

[0074] ,

[0075] wherein: is the self-attention output matrix, denotes the output vector of the th time step.

[0076] The multi-head self-attention mechanism executes the above single-head attention mechanism multiple times in parallel, each head using a different transformation matrix, and finally splices the outputs of all heads together:

[0077] ,

[0078] wherein: is the multi-head self-attention output, is the number of heads, is the output of the th head, denotes the splicing operation, is the output transformation matrix.

[0079] The number of heads of the multi-head self-attention mechanism is preferably 4 to 8. The more the number of heads, the richer the spatio-temporal correlation patterns that the model can learn from different representation subspaces, but at the same time, the computational complexity will also increase. In specific embodiments, for data with a time series length of 24 months, the number of heads can be set to 6, the feature dimension , and the dimension of the key and query .

[0080] The feedforward neural network 312 is a two-layer fully connected neural network for nonlinear transformation of the output of the self-attention:

[0081] ,

[0082] wherein: is the output of the feedforward neural network, is the first layer weight matrix, is the first layer bias vector, is the second layer weight matrix, is the second layer bias vector, is the hidden layer dimension of the feedforward neural network, which is usually set to , is the rectified linear unit activation function, defined as .

[0083] The Transformer encoder layer also includes a residual connection and layer normalization, and the complete calculation process is:

[0084] ,

[0085] ,

[0086] wherein: is the layer normalization operation, which is used to stabilize the training process, and the residual connection can alleviate the gradient vanishing problem of deep networks.

[0087] If the time span of the spatiotemporal feature vector meets the first time range (1 year to 3 years), the Transformer encoder layer adopts the first number of layers configuration (4 layers to 6 layers); if the time span of the spatiotemporal feature vector meets the second time range (3 years to 10 years), the Transformer encoder layer adopts the second number of layers configuration (8 layers to 12 layers). This adaptive number of layers configuration can adjust the capacity of the model according to the time span of the data. Short time span data can use a shallower network to capture the main time series pattern, and long time span data needs a deeper network to model complex long-time dependence.

[0088] The GRU decoder layer 32 includes a plurality of GRU units, and the hidden layer dimension of the GRU unit is preferably 128 to 512. GRU (Gated Recurrent Unit) is an improved recurrent neural network structure that selectively retains and updates historical information through a gating mechanism. The calculation process of the GRU unit is as follows:

[0089] First, calculate the update gate and the reset gate :

[0090] ,

[0091] ,

[0092] wherein: update gate for the t-th time step, reset gate for the t-th time step, hidden state for the t-th time step, reset gate for the t-th time step, hidden state for the t-th time step, reset gate for the t-th time step, encoder output for the t-th time step, weight matrix for hidden state to gate, weight matrix for input to gate, bias vector for gate, GRU hidden layer dimension, Sigmoid activation function, defined as . .

[0093] Then, the candidate hidden state is calculated as:

[0094] ,

[0095] where: candidate hidden state for the t-th time step, weight matrix for hidden state to candidate hidden state, weight matrix for input to candidate hidden state, bias vector for candidate hidden state, element-wise multiplication, hyperbolic tangent activation function, defined as . .

[0096] Finally, the hidden state for the t-th time step is obtained by weighting fusion of the historical hidden state and the candidate hidden state according to the update gate:

[0097] ,

[0098] where: hidden state for the t-th time step, the update gate controls the fusion ratio of historical information and new information, when is close to 0, the model tends to retain historical information; when is close to 1, the model tends to adopt new information.

[0099] The output of the GRU decoder is the subsidence prediction value for multiple future time steps:

[0100] ,

[0101] where: t-th time step, ​The predicted settlement value at each time point. To output the weight vector, This is the output bias.

[0102] The Transformer-GRU prediction model is trained using the mean squared error loss function:

[0103] ,

[0104] in: For loss function, The number of training samples. To predict the number of time steps, For the first The sample at the th The predicted value at a future time. For the first The sample at the th The true value of a future moment.

[0105] The model was optimized using the Adam optimizer with an initial learning rate of 0.001, which gradually decreased during training. Batch size ranged from 32 to 64, and the number of training epochs ranged from 100 to 200. In a specific implementation, for a dataset containing 500 grid cells and a 24-month time series, the Transformer encoder layer had 6 layers, each with 6 heads for the multi-head self-attention mechanism, a feature dimension of 256, and the GRU decoder layer had a hidden layer dimension of 256, predicting sedimentation values ​​for the next 6 months. The model achieved a root mean square error (RMSE) of 2.1 mm on the training set and 2.8 mm on the test set, significantly outperforming the traditional ARIMA model (RMSE of 5.2 mm on the test set) and the LSTM model (RMSE of 3.6 mm on the test set).

[0106] like Figure 5 As shown, the processing flow of deformation field reconstruction module 4 includes the following steps:

[0107] Step S41: If the settlement rate of a pixel in the pixel-level settlement rate field conforms to the first rate range, extract the three-dimensional coordinates and settlement rate of the pixel. The first rate range is preferably greater than 10 mm per year, indicating that the pixel has significant settlement. In a specific embodiment, for a 10-kilometer-long highway subgrade, there are approximately 1 million pixels (assuming a pixel resolution of 3 meters × 3 meters), of which approximately 5,000 pixels have a settlement rate greater than 10 mm per year. These pixels are mainly distributed in soft soil foundation sections and sections with high embankment filling.

[0108] Step S42: Based on the three-dimensional coordinates and the settlement rates, a three-dimensional deformation field of the high-risk area is reconstructed using the Kriging interpolation method. Kriging interpolation is an optimal interpolation method based on spatial autocorrelation theory, and its core idea is that there is correlation between similar points in space, and the closer the distance, the stronger the correlation. Kriging interpolation includes the following steps:

[0109] First, the semi-variation function is calculated :

[0110] ,

[0111] Wherein: is the lag distance , the corresponding semi-variation function value, is the number of point pairs with a distance of , is the settlement rate at position , and is the settlement rate at the position with a distance of from .

[0112] Then, a suitable semi-variation function model is selected for fitting, and common models include the spherical model, the exponential model and the Gaussian model. The expression of the spherical model is:

[0113] ,

[0114] Wherein: is the nugget value, indicating the semi-variation function value when , reflecting the measurement error and micro-scale variation, is the sill value, indicating the semi-variation function value when the spatial correlation reaches stability, is the range, indicating the effective influence distance of spatial correlation, and when , there is no longer spatial correlation between points.

[0115] Next, for the position that needs to be interpolated, the Kriging equation group is constructed:

[0116] ,

[0117] Wherein: is the interpolation weight of the th known point, is the Lagrange multiplier, is the number of known points, is the semi-variation function value between the known points and , is the semi-variation function value between the known point and the point to be interpolated. semivariogram value between the two.

[0118] Solving the above linear equations, the interpolation weight is obtained, and then the subsidence rate estimate value at the position to be interpolated is calculated:

[0119]

[0120] wherein: is the subsidence rate estimate value at the position , and is the subsidence rate at the known point .

[0121] In a specific embodiment, for a high-risk area containing 5000 known subsidence rates, the semivariogram is fitted by using a spherical model, and the nugget value , the sill value , and the range are obtained. By using the Kriging interpolation method, the three-dimensional deformation field is reconstructed in the area with a spatial resolution of 5 meters, and the subsidence rate estimate values of about 200000 interpolation points are generated.

[0122] Step S43: Spatial smoothing processing is performed on the three-dimensional deformation field to remove the high-frequency noise introduced in the interpolation process. The spatial smoothing processing adopts a Gaussian filter, and the standard deviation of the filter is set to 10 meters to 20 meters, which can effectively remove small-scale noise while retaining large-scale subsidence patterns.

[0123] Step S44: The deformation gradient field and the curvature field of the three-dimensional deformation field are calculated. The calculation of the deformation gradient field adopts a finite difference method. For the subsidence rate at the position , the deformation gradients in the direction and the direction are calculated:

[0124]

[0125]

[0126] wherein: is the deformation gradient in the direction, is the deformation gradient in the direction, and are spatial steps, which are usually set to the spatial resolution of the grid.

[0127] The amplitude of the deformation gradient is:

[0128] ​​​​

[0129] wherein: is the deformation gradient magnitude at position reflecting the spatial variation rate of the subsidence rate at this position.

[0130] The calculation of the curvature field uses the second derivative, for the subsidence rate at position , the curvature is calculated as:

[0131] ,

[0132] wherein: is the curvature at position , is the second derivative in the direction, is the second derivative in the direction. The curvature reflects the bending degree of the spatial distribution of the subsidence rate, and a larger curvature indicates that there is a mutation or inflection point in the subsidence at this position.

[0133] Through the above processing flow, the deformation field reconstruction module 4 generates the three-dimensional deformation field, the deformation gradient field and the curvature field of the high-risk area, providing rich spatial information for the visualization analysis and mechanism research of the subsidence.

[0134] The training flow of the synthetic data generation module 5 includes the following steps:

[0135] Step S51: Construct a generator network, the generator network adopts a convolutional neural network architecture, including an encoder part and a decoder part. The encoder part maps a random noise vector into a low-dimensional latent representation, and the decoder part maps the latent representation into a two-dimensional subsidence image. The specific structure of the generator network is: the input is a 100-dimensional random noise vector , which is mapped into a feature map of through a fully connected layer, and then gradually up-sampled through 4 layers of transposed convolutional layers, the convolutional kernel size of each layer is , the step size is 2, the number of channels is 256, 128, 64, 1 in turn, the activation function adopts LeakyReLU (negative slope is 0.2), the last layer adopts Tanh activation function, and the output is a subsidence image of .

[0136] Step S52: Construct a discriminator network, the discriminator network adopts a convolutional neural network architecture, which is used to distinguish between real subsidence images and synthetic historical subsidence images. The specific structure of the discriminator network is: the input is a subsidence image of , which is gradually down-sampled through 5 layers of convolutional layers, the convolutional kernel size of each layer is , the step size is 2, the number of channels is 64, 128, 256, 512, 1 in turn, the activation function adopts LeakyReLU (the negative slope is 0.2), the last layer adopts Sigmoid activation function, and the output is a single scalar value, which represents the probability that the input image is a real image.

[0137] Step S53: Randomly sampling real subsidence images from the pixel-level ground surface deformation dataset and generating synthetic historical subsidence images from random noise vectors. In each training iteration, a batch of real subsidence images is randomly sampled from the training dataset, and the batch size is set to 32-64. At the same time, the same number of random noise vectors are sampled from the standard normal distribution, and the generator network is input to generate synthetic historical subsidence images.

[0138] Step S54: Calculate the discriminant output of the discriminator network for real subsidence images and the discriminant output for synthetic historical subsidence images. The loss function of the discriminator adopts binary cross entropy:

[0139] ,

[0140] Where: is the loss function of the discriminator, is the batch size, is the th real subsidence image, is the th random noise vector, is the discriminator network, is the generator network, is the discriminant output of the discriminator for real images, is the discriminant output of the discriminator for synthetic images.

[0141] The loss function of the generator is:

[0142] ,

[0143] Where: is the loss function of the generator, and the goal of the generator is to maximize the probability that the discriminator discriminates the synthetic image as a real image.

[0144] An alternating training strategy is adopted, and in each training iteration, the generator parameters are fixed, and the discriminator parameters are updated to minimize ; then the discriminator parameters are fixed, and the generator parameters are updated to minimize . Training uses the Adam optimizer, and the learning rate is set to 0.0002, and the momentum parameter , .

[0145] Step S55: If the discrimination accuracy of the discriminator network on the synthesized historical settlement image meets the first accuracy range, indicating that the synthesized historical settlement image is highly similar to the real settlement image, the synthesized historical settlement image is added to the training data set. The first accuracy range is preferably below 60%, meaning that the discriminator network is difficult to distinguish between real images and synthesized images, at which time the synthesized image has reached a high enough quality. In a specific embodiment, after 200 rounds of training, the discrimination accuracy of the discriminator on the synthesized image is reduced to 54%, and the synthesized settlement image generated is almost indistinguishable from the real settlement image in visual. The 5000 synthesized settlement images generated are added to the original 5000 real settlement images, and the size of the training data set is doubled. The Transformer-GRU prediction model is trained on the expanded data set, and the root mean square error of the test set is reduced from 2.8 mm to 2.3 mm, and the prediction accuracy is improved by about 18%.

[0146] As shown in Figure 6 The present application also provides a roadbed settlement time series prediction and three-dimensional deformation monitoring method based on the above-mentioned system, which comprises the following steps:

[0147] Step S1: receiving InSAR satellite remote sensing image data and ground-based micro-deformation sensor data, performing pixel-level registration processing on the InSAR satellite remote sensing image data to generate a pixel-level ground surface deformation data set, and performing spatio-temporal alignment processing on the ground-based micro-deformation sensor data to generate a ground deformation data set.

[0148] Step S2: grid division of the pixel-level ground surface deformation data set according to a spatial resolution threshold, if the number of pixels in a single grid cell meets the first density range, extracting the spatio-temporal feature vector corresponding to the grid cell, the spatio-temporal feature vector including settlement rate field features, spatial gradient field features and time series features, and the spatial resolution threshold is 5-20 meters.

[0149] Step S3: constructing a spatio-temporal coupled Transformer-GRU prediction model, the Transformer-GRU prediction model including a Transformer encoder layer and a GRU decoder layer, the Transformer encoder layer being used to capture long-time dependence in the spatio-temporal feature vector, and the GRU decoder layer being used to predict the settlement trend at a future time; if the time span of the spatio-temporal feature vector meets the first time range, the Transformer encoder layer adopts a first number of layers configuration, and if the time span of the spatio-temporal feature vector meets the second time range, the Transformer encoder layer adopts a second number of layers configuration, the first time range being 1-3 years, the second time range being 3-10 years, the first number of layers configuration being 4-6 layers, and the second number of layers configuration being 8-12 layers.

[0150] Step S4: based on the prediction result of the Transformer-GRU prediction model, calculate the pixel-level settlement rate field, if the settlement rate of any pixel point in the pixel-level settlement rate field meets the first rate range, mark the pixel point as a high-risk area, reconstruct the three-dimensional deformation field corresponding to the high-risk area, the calculation accuracy of the pixel-level settlement rate field is 0.5mm per year to 2mm per year, and the first rate range is greater than 10mm per year.

[0151] Step S5: construct a generative adversarial network, the generative adversarial network includes a generator network and a discriminator network, the generator network is used to synthesize historical settlement images, and the discriminator network is used to distinguish real settlement images and synthesized historical settlement images; if the discrimination accuracy of the discriminator network on the synthesized historical settlement image meets the first accuracy range, the synthesized historical settlement image is added to the training data set for enhancing the Transformer-GRU prediction model, and the first accuracy range is less than 60%.

[0152] Step S6: generate a roadbed settlement monitoring report according to the three-dimensional deformation field, if the maximum cumulative settlement amount in the three-dimensional deformation field meets the first settlement threshold range, trigger a first-level warning signal, if the maximum cumulative settlement amount meets the second settlement threshold range, trigger a second-level warning signal, the first settlement threshold range is 30mm to 50mm, and the second settlement threshold range is greater than 50mm.

Claims

1. A roadbed settlement time-series prediction and three-dimensional deformation monitoring system, characterized in that, The method comprises the following steps: data acquisition module, for receiving InSAR satellite remote sensing image data and ground-based micro-deformation sensor data, performing pixel-level registration processing on the InSAR satellite remote sensing image data to generate a pixel-level ground surface deformation data set, and performing space-time alignment processing on the ground-based micro-deformation sensor data to generate a ground deformation data set; a space-time feature construction module connected with the data acquisition module, configured to grid divide the pixel-level ground surface deformation data set according to a spatial resolution threshold, extract a space-time feature vector corresponding to a single grid unit if the number of pixel points in the single grid unit meets a first density range, and the space-time feature vector includes a subsidence rate field feature, a spatial gradient field feature, and a time series feature, and the spatial resolution threshold is 5 meters to 20 meters; a prediction model fusion module connected with the space-time feature construction module, configured to construct a space-time coupled Transformer-GRU prediction model, and the Transformer-GRU prediction model includes a Transformer encoder layer and a GRU decoder layer; a deformation field reconstruction module connected with the prediction model fusion module, configured to calculate a pixel-level subsidence rate field based on a prediction result of the Transformer-GRU prediction model; a synthetic data generation module connected with the deformation field reconstruction module, configured to construct a generative adversarial network, and the generative adversarial network includes a generator network and a discriminator network, the generator network is configured to synthesize a historical subsidence image, and the discriminator network is configured to distinguish between a real subsidence image and the synthetic historical subsidence image; if the discrimination accuracy of the discriminator network on the synthetic historical subsidence image meets a first accuracy range, the synthetic historical subsidence image is added to a training data set to enhance the Transformer-GRU prediction model, and the first accuracy range is less than 60%; a monitoring and early warning module connected with the deformation field reconstruction module, configured to generate a roadbed subsidence monitoring report according to the three-dimensional deformation field, trigger a first-level warning signal if a maximum cumulative subsidence in the three-dimensional deformation field meets a first subsidence threshold range, and trigger a second-level warning signal if the maximum cumulative subsidence meets a second subsidence threshold range, the first subsidence threshold range is 30 millimeters to 50 millimeters, and the second subsidence threshold range is greater than 50 millimeters.

2. The system of claim 1, wherein, If the subsidence rate of any pixel point in the pixel-level subsidence rate field meets a first rate range, the pixel point is marked as a high-risk area, a three-dimensional deformation field corresponding to the high-risk area is reconstructed, the calculation accuracy of the pixel-level subsidence rate field is 0.5 millimeters per year to 2 millimeters per year, and the first rate range is greater than 10 millimeters per year. The Transformer encoder layer is used to capture long-time dependence in the spatiotemporal feature vector, and the GRU decoder layer is used to predict the subsidence trend at a future time; if the time span of the spatiotemporal feature vector meets a first time range, the Transformer encoder layer adopts a first number-of-layers configuration, and if the time span of the spatiotemporal feature vector meets a second time range, the Transformer encoder layer adopts a second number-of-layers configuration, the first time range is 1-3 years, the second time range is 3-10 years, the first number-of-layers configuration is 4-6 layers, and the second number-of-layers configuration is 8-12 layers; The data acquisition module is specifically configured to: receive multi-temporal InSAR satellite remote sensing images, the multi-temporal InSAR satellite remote sensing images including at least 15 SAR images; identify persistent scatterers from the multi-temporal InSAR satellite remote sensing images, and extract a persistent scatterer point set, the spatial density of the persistent scatterer point set being 300-1000 persistent scatterer points per square kilometer; perform time-series InSAR processing on the persistent scatterer point set, calculate a deformation time series of each persistent scatterer point, and generate the pixel-level ground surface deformation data set; receive measurement data of ground-based micro-deformation sensors at multiple monitoring points, the measurement data including displacement amounts and time stamps; match the measurement data of the ground-based micro-deformation sensors with spatial coordinates of the persistent scatterer points, and if a matching distance is less than a spatial matching threshold, perform time series alignment and fusion, the spatial matching threshold being 20-50 meters.

3. The system of claim 1, wherein, The spatiotemporal feature construction module includes: a grid division sub-module configured to divide the pixel-level ground surface deformation data set into multiple grid cells according to the spatial resolution threshold, and count the number of pixel points in each grid cell; a density screening sub-module connected with the grid division sub-module and configured to determine whether the number of pixel points meets a first density range, and if yes, mark the grid cell as a valid cell, and if not, mark the grid cell as an invalid cell, the first density range being 10-50 pixel points per grid cell; a feature extraction sub-module connected with the density screening sub-module and configured to extract the spatiotemporal feature vector from the grid cell marked as the valid cell, the spatiotemporal feature vector including an average subsidence rate of pixel points in the grid cell, a spatial standard deviation of the subsidence rate, a time variation rate of the subsidence rate, and a subsidence gradient with an adjacent grid cell.

4. The system of claim 1 or 2, wherein, The subsidence rate field feature includes a pixel-level subsidence rate and a spatial distribution pattern of the subsidence rate, and the pixel-level subsidence rate is calculated by: performing linear regression fitting on the deformation time series of each pixel point, and calculating the slope of the regression straight line as the pixel-level subsidence rate; performing Gaussian filter smoothing processing on the pixel-level subsidence rate, the filter window size being 3-5 pixels; calculating a spatial gradient field of the pixel-level subsidence rate, the spatial gradient field reflecting the difference in subsidence rate between adjacent pixel points; and If the gradient value of any position in the spatial gradient field meets the first gradient range, mark the position as a settlement abnormal area, and the first gradient range is greater than 5 mm per year per 100 m.

5. The system of claim 1 or 2, wherein, The construction method of the Transformer-GRU prediction model comprises: constructing a Transformer encoder layer, the Transformer encoder layer comprising a multi-head self-attention mechanism and a feedforward neural network, the number of heads of the multi-head self-attention mechanism being 4 to 8; inputting the spatio-temporal feature vector into the Transformer encoder layer, and calculating the spatio-temporal correlation weight between the feature vectors through the multi-head self-attention mechanism; constructing a GRU decoder layer, the GRU decoder layer comprising a plurality of GRU units, the hidden layer dimension of the GRU units being 128 to 512; taking the output of the Transformer encoder layer as the input of the GRU decoder layer, and predicting the settlement values at a plurality of future time points through the GRU decoder layer, the time span of the plurality of future time points being 1 month to 12 months.

6. The system of claim 1, wherein, The processing flow of the deformation field reconstruction module comprises: if the settlement rate of a pixel point in the pixel-level settlement rate field meets the first rate range, extracting the three-dimensional coordinates and the settlement rate of the pixel point; based on the three-dimensional coordinates and the settlement rate, reconstructing the three-dimensional deformation field of the high-risk area by using a Kriging interpolation method, the Kriging interpolation method comprising calculating a semi-variogram function and constructing an interpolation weight matrix; performing spatial smoothing processing on the three-dimensional deformation field to remove high-frequency noise introduced in the interpolation process; calculating the deformation gradient field and the curvature field of the three-dimensional deformation field, the deformation gradient field reflecting the spatial variation trend of the deformation, and the curvature field reflecting the spatial bending degree of the deformation.

7. The system of claim 1, wherein, The training flow of the synthetic data generation module comprises: constructing the generator network, the generator network adopting a convolutional neural network architecture, comprising an encoder part and a decoder part; constructing the discriminator network, the discriminator network adopting a convolutional neural network architecture, and being used for distinguishing real settlement images and the synthetic historical settlement images; randomly sampling real settlement images from the pixel-level ground surface deformation data set, and generating the synthetic historical settlement images from random noise vectors; calculating the discrimination output of the discriminator network on the real settlement images and the discrimination output of the discriminator network on the synthetic historical settlement images, and updating the parameters of the generator network and the discriminator network based on the discrimination output; if the discrimination accuracy of the discriminator network on the synthetic historical settlement images meets the first accuracy range, it is indicated that the synthetic historical settlement images are highly similar to the real settlement images, and the synthetic historical settlement images are added to the training data set.

8. The system of claim 1, wherein, The monitoring and early warning module is further used for: calculating the spatial distribution characteristics of the roadbed settlement according to the three-dimensional deformation field, the spatial distribution characteristics comprising a settlement center position, a settlement influence range and a settlement morphological feature; calculate a settlement development trend at a future time according to a prediction result of the Transformer-GRU prediction model, the settlement development trend including a settlement acceleration and a predicted time for the settlement to reach a dangerous threshold; generate the roadbed settlement monitoring report, the roadbed settlement monitoring report including a current settlement state, a historical settlement evolution curve, a future settlement prediction curve, a warning level, and a recommended treatment measure; if the warning level is a first-level warning, send a monitoring reminder to a monitoring terminal; if the warning level is a second-level warning, send an emergency treatment message to a management terminal.

9. The system of claim 1, wherein, The system further comprises: an adaptive optimization module for monitoring a prediction error of the Transformer-GRU prediction model, if the prediction error meets a first error range, keeping the parameters of the Transformer-GRU prediction model unchanged, if the prediction error meets a second error range, triggering a model parameter update process, the first error range being a root mean square error less than 3 mm, and the second error range being a root mean square error greater than 5 mm; the model parameter update process includes obtaining the latest pixel-level ground surface deformation data set and the ground deformation data set, incrementally training the Transformer-GRU prediction model, and updating the weight parameters of the Transformer encoder layer and the GRU decoder layer.

10. A method for subgrade settlement time series prediction and 3D deformation monitoring based on the system of claim 1, using the system of any one of claims 1-9, characterized in that, comprising: receiving InSAR satellite remote sensing image data and ground-based micro-deformation sensor data, performing pixel-level registration processing on the InSAR satellite remote sensing image data to generate a pixel-level ground surface deformation data set, and performing spatio-temporal alignment processing on the ground-based micro-deformation sensor data to generate a ground deformation data set; grid division of the pixel-level ground surface deformation data set according to a spatial resolution threshold, if the number of pixels in a single grid cell meets a first density range, extracting a spatio-temporal feature vector corresponding to the grid cell, the spatio-temporal feature vector including a settlement rate field feature, a spatial gradient field feature, and a time series feature, and the spatial resolution threshold being 5-20 meters; constructing a spatio-temporal coupled Transformer-GRU prediction model, the Transformer-GRU prediction model including a Transformer encoder layer and a GRU decoder layer, the Transformer encoder layer being used to capture long-time dependencies in the spatio-temporal feature vector, and the GRU decoder layer being used to predict a settlement trend at a future time; if a time span of the spatio-temporal feature vector meets a first time range, the Transformer encoder layer adopts a first layer number configuration, if the time span of the spatio-temporal feature vector meets a second time range, the Transformer encoder layer adopts a second layer number configuration, the first time range being 1-3 years, the second time range being 3-10 years, the first layer number configuration being 4-6 layers, and the second layer number configuration being 8-12 layers; Based on the prediction result of the Transformer-GRU prediction model, a pixel-level settlement rate field is calculated, if the settlement rate of any pixel point in the pixel-level settlement rate field meets a first rate range, the pixel point is marked as a high-risk area, a three-dimensional deformation field corresponding to the high-risk area is reconstructed, the calculation accuracy of the pixel-level settlement rate field is 0.5mm / year to 2mm / year, and the first rate range is greater than 10mm / year; A generative adversarial network is constructed, the generative adversarial network includes a generator network and a discriminator network, the generator network is used to synthesize a historical settlement image, and the discriminator network is used to distinguish a real settlement image and the synthesized historical settlement image; if the discrimination accuracy of the discriminator network on the synthesized historical settlement image meets a first accuracy range, the synthesized historical settlement image is added to a training data set for enhancing the Transformer-GRU prediction model, and the first accuracy range is lower than 60%; A subgrade settlement monitoring report is generated according to the three-dimensional deformation field, if the maximum cumulative settlement amount in the three-dimensional deformation field meets a first settlement threshold range, a first-level early warning signal is triggered, and if the maximum cumulative settlement amount meets a second settlement threshold range, a second-level early warning signal is triggered, the first settlement threshold range is 30mm to 50mm, and the second settlement threshold range is greater than 50mm.

Images