Land subsidence prediction method, device and medium

Abstract

The application discloses a land subsidence prediction method, device and medium, and the method comprises the following steps: acquiring a land subsidence time series subsidence amount of a target point; performing EMD decomposition on the land subsidence time series subsidence amount to obtain IMF components and a trend component; establishing a prediction model by using the IMF components and the trend component, predicting each component, and reconstructing a prediction result; comparing the prediction accuracy of subsidence when the number of different IMF components is different, selecting an optimal IMF number, and taking the corresponding prediction result as the final prediction result of the target point. The application can improve the prediction accuracy of land subsidence and provide a basis for analyzing the influence of land subsidence on high-speed rail operation safety.

Description

Technical Field

[0001] This invention belongs to the field of geological safety technology, and more specifically, relates to a method, device and medium for predicting ground subsidence. Background Technology

[0002] High-speed railway safety operation places higher demands on the stability, deformation degree, and track smoothness of roadbeds and bridges. Regional ground subsidence affects the gradient of high-speed railway lines, and the operational safety of high-speed railways is significantly impacted when they cross subsidence funnel areas and subsidence zones. The Beijing-Tianjin-Hebei region, located in the northern part of the North China Plain, has a complex geological structure and scarce water resources. To meet the water demands of urban construction and rapid population growth, long-term over-extraction of groundwater has led to the formation of large-scale groundwater over-extraction funnels (Gong et al., 2017; Guo et al., 2015). Furthermore, the integrated development of the Beijing-Tianjin-Hebei region, the rapid development of high-density urban clusters, and the rapid expansion of high-speed, multi-dimensional transportation networks have resulted in the formation of multiple ground subsidence zones within the region, exhibiting a cross-regional, contiguous distribution with high subsidence rates (Cui and Lei et al., 2018). The total area of ​​ground subsidence in the Beijing-Tianjin-Hebei region reaches 1.4 × 10⁻⁶. 4 km 2 In some areas, the maximum annual subsidence reaches 160 mm. Areas with severe subsidence are mainly distributed in the northern and southeastern parts of the Beijing Plain; Wangqingtuo District of Tianjin; and Baoding in the central part of the Hebei Plain, Hengshui and Cangzhou in the southeast, and Handan in the south. Furthermore, the subsidence exhibits a cross-regional, contiguous characteristic. The development of regional land subsidence will affect the operational safety of key linear projects such as the Beijing-Shanghai, Beijing-Tianjin, Tianjin-Baoding, Shijiazhuang-Jinan high-speed railways, and the South-to-North Water Diversion Project, and constrain regional sustainable development (Gong et al., 2018). The "14th Five-Year Plan for National Economic and Social Development and the Outline of the Long-Term Goals for 2035" points out that the Beijing-Tianjin-Hebei region should be basically built on rails. Against this backdrop, conducting land subsidence prediction along high-speed railway lines and evaluating the impact of regional land subsidence on the operational safety of high-speed railways has significant research value and practical significance for disaster prevention and mitigation of other major linear projects under construction in the Beijing-Tianjin-Hebei region.

[0003] Traditional land subsidence monitoring techniques, such as leveling, GPS, and stratified beacon surveying, have limitations in large-area, high-temporal-resolution measurements (Berardino et al., 2002; Galloway DL, 1998). Interferometric synthetic aperture radar (InSAR), developed in the 20th century, offers advantages such as wide spatial coverage, high accuracy, and low cost, enabling the acquisition of large-area land subsidence information. Currently, InSAR technology has become a commonly used geodetic method for studying surface deformation, especially techniques developed based on this technology such as Persistent Scatterer InSAR (PS-InSAR), Small Baseline Subset InSAR (SBAS-InSAR), and Interferometric Point Target Analysis (IPTA) (Hooper A, 2008; Gao et al., 2016). These technologies have, to some extent, addressed the limitations of conventional synthetic aperture radar differential interferometry (D-InSAR) in spatial and temporal decorrelation, atmospheric error, orbital error, and terrain error removal. PS-InSAR technology identifies points with stable scattering characteristics (Persistent Scatterer, PS) in time series, and after processing the target points, obtains reliable surface deformation estimation results with monitoring accuracy down to the millimeter level (Ferretti et al., 2000).

[0004] In the field of land subsidence prediction research, models are mainly divided into deterministic models and stochastic models. Deterministic models are based on physical mechanisms, estimating subsidence by simulating the physical processes of subsidence occurrence and development (Ye et al., 2005; Luo et al., 2009; Zhu et al., 2020). Deterministic models are mathematical models composed of groundwater flow models, soil mechanics models, and their coupling models. When simulating the deformation process of aquifers and aquifer systems, they require input of numerous physical parameters such as groundwater extraction rates and geological, geotechnical, and hydrogeological conditions. Stochastic models consider the diversity and uncertainty of factors influencing land subsidence. They typically employ mathematical statistical models, analyzing and simulating the inherent relationships and development characteristics of large amounts of historical monitoring data to establish the correlation between changes in single influencing factors and land subsidence, thereby achieving land subsidence prediction (Tomás et al., 2010; Yue et al., 2020; Edalat et al., 2020; Fan et al., 2013). In recent years, big data and artificial intelligence technologies have developed rapidly. Intelligent prediction algorithms are not limited by complex physical parameters such as regional hydrogeology, and can also solve the problem of low efficiency of traditional mathematical and statistical methods. They have been widely used in land subsidence prediction research (Deng et al., 2017; Zhao et al., 2020; Wang et al., 2018; Arabameri et al., 2020).

[0005] Currently, given the complex evolution of ground subsidence along high-speed railways, improving the accuracy of ground subsidence prediction is a pressing technical problem that needs to be solved. Summary of the Invention

[0006] This invention addresses the aforementioned problems in the prior art. Therefore, there is a need for a ground settlement prediction method, apparatus, and medium to improve the accuracy of ground settlement prediction and provide a basis for analyzing the impact of ground settlement on the operational safety of high-speed railways.

[0007] According to a first aspect of the present invention, a method for predicting ground subsidence is provided, characterized in that the method comprises:

[0008] Obtain the temporal ground settlement at the target point;

[0009] The time-series ground subsidence was decomposed using EMD to obtain IMF components and trend components.

[0010] A prediction model is established using the IMF components and trend components. The prediction results are obtained by predicting and reconstructing each component.

[0011] By comparing the prediction accuracy of settlement with different numbers of IMF components, the optimal number of IMF components is selected and the corresponding prediction result is used as the final prediction result for the target point.

[0012] Further, obtaining the time-series settlement amount of the ground subsidence at the target point includes:

[0013] For each sub-strip of radar image data, select one main image and the rest as auxiliary images, and register the main image and the auxiliary images.

[0014] In the registered master and slave images, the amplitude stability coefficient method is used to select target points;

[0015] Phase analysis is performed on the target point, and the phase information contained in the target point is shown in formula (1):

[0016] (1)

[0017] in, It is the phase of surface deformation. For terrain phase, It is a horizontal phase. Due to atmospheric error, Noise phase;

[0018] Based on solving the overall deformation phase components using the nonlinear standard model, an external digital elevation model (DEM) is introduced to remove the terrain phase, orbital data is used to remove the horizon phase, atmospheric phase information is retrieved and removed using an atmospheric phase model, and the deformation phase information of each target point is retrieved using a star topology analysis method. The deformation phase information of each target point includes nonlinear and linear deformation phases.

[0019] Phase unwrapping is performed on the deformation phase information of the target point to obtain the time-series settlement amount of the ground settlement at the target point.

[0020] Furthermore, the time-series ground subsidence is decomposed using formula (2) to obtain the IMF component and trend component:

[0021] (2)

[0022] in The original time-series data uses ground subsidence measurements along each high-speed railway line. For the first One IMF component, This represents the trend.

[0023] Furthermore, the step of establishing a prediction model using the IMF components and trend components, predicting and reconstructing each component to obtain the prediction result includes:

[0024] Select the time series sedimentation of sample points as the dataset , Indicates the monitoring time, sets the ratio of training set to validation set, and includes the previous... The time-series sedimentation data were used to construct the training set. ,Will Time series sedimentation data were used to construct a validation set. , Less than Monitoring time;

[0025] Initialize for the training dataset , The base learner is represented by a decision tree. For loss function, This represents a constant value that minimizes the loss function;

[0026] For the training dataset, construct M decision trees. A tree, for Calculate the value of the negative gradient descent of the loss function. As a model for fitting the next step of the decision tree model For the current decision tree model Approximate residual value : ;

[0027] The final model is obtained through M iterations. ,in It consists of M trees;

[0028] Using the final model To predict the amount of ground subsidence.

[0029] Furthermore, by comparing the RMSE and MAE values ​​for different numbers of IMF components, the accuracy of settlement prediction is improved. The smaller the RMSE and MAE values, the more accurate the model's prediction results.

[0030] Furthermore, the RMSE value is calculated using the following formula:

[0031] (3)

[0032] in Indicates the number of sample points. This indicates the first time series in the validation dataset. The label values ​​of each data point. Indicates the first [item] in the prediction dataset The predicted label value for each data point.

[0033] Furthermore, the RMSE value is calculated using the following formula:

[0034] (4)

[0035] in Indicates the number of sample points. This indicates the first time series in the validation dataset. The label values ​​of each data point. Indicates the first [item] in the prediction dataset The predicted label value for each data point.

[0036] According to a second aspect of the present invention, a ground subsidence prediction device is provided, the device comprising:

[0037] The acquisition module is configured to acquire the time-series ground settlement at the target point.

[0038] The decomposition module is configured to perform EMD decomposition on the time-series ground subsidence to obtain IMF components and trend components.

[0039] The prediction module is configured to establish a prediction model using the IMF components and trend components, predict each component, and reconstruct the prediction results.

[0040] The selection module is configured to compare the prediction accuracy of settlement with different numbers of IMF components, select the optimal number of IMFs, and use the corresponding prediction result as the final prediction result for the target point.

[0041] Furthermore, the acquisition module is further configured as follows:

[0042] The radar image data of different frame numbers are merged, and the merged data is split into several sub-strips according to different data modes;

[0043] For each sub-strip of radar image data, select one main image and the rest as auxiliary images, and register the main image and the auxiliary images.

[0044] In the registered master and slave images, the amplitude stability coefficient method is used to select target points;

[0045] Phase analysis is performed on the target point, and the phase information contained in the target point is shown in formula (1):

[0046] (1)

[0047] in, It is the phase of surface deformation. For terrain phase, It is a horizontal phase. Due to atmospheric error, Noise phase;

[0048] Based on solving the overall deformation phase components using the nonlinear standard model, an external digital elevation model (DEM) is introduced to remove the terrain phase, orbital data is used to remove the horizon phase, atmospheric phase information is retrieved and removed using an atmospheric phase model, and the deformation phase information of each target point is retrieved using a star topology analysis method. The deformation phase information of each target point includes nonlinear and linear deformation phases.

[0049] Phase unwrapping is performed on the deformation phase information of the target point to obtain the time-series settlement amount of the ground settlement at the target point.

[0050] Furthermore, the decomposition module is further configured to perform EMD decomposition on the time-series ground subsidence using formula (2) to obtain the IMF component and the trend component:

[0051] (2)

[0052] in The original time-series data uses ground subsidence measurements along each high-speed railway line. For the first One IMF component, This represents the trend.

[0053] Furthermore, the prediction module is further configured as follows:

[0054] Select the time series sedimentation of sample points as the dataset , Indicates the monitoring time, sets the ratio of training set to validation set, and includes the previous... The time-series sedimentation data were used to construct the training set. ,Will Time series sedimentation data were used to construct a validation set. , Less than Monitoring time;

[0055] Initialize for the training dataset , The base learner is represented by a decision tree. For loss function, This represents a constant value that minimizes the loss function;

[0056] For the training dataset, construct M decision trees. A tree, for Calculate the value of the negative gradient descent of the loss function. As a model for fitting the next step of the decision tree model For the current decision tree model Approximate residual value : ;

[0057] The final model is obtained through M iterations. ,in It consists of M trees;

[0058] Using the final model To predict the amount of ground subsidence.

[0059] According to a third aspect of the present invention, a readable storage medium is provided, characterized in that the readable storage medium stores one or more programs, which can be executed by one or more processors to implement the modeling method and / or the prediction method as described above.

[0060] The present invention has at least the following beneficial effects:

[0061] This invention addresses the complex evolution of ground subsidence along high-speed railways. When predicting ground subsidence along high-speed railways, stochastic models face significant challenges in predicting nonlinear and non-stationary time series data. This invention combines an empirical normal model with multiple stable time series data points. Gradient boosting decision tree models are then used to perform time series predictions for each stable time series data point. The empirical normal decomposition component with the highest prediction accuracy, along with the prediction result at that point, is extracted as the final result. This improves the accuracy of ground subsidence prediction and provides a basis for analyzing the impact of ground subsidence on the operational safety of high-speed railways. Attached Figure Description

[0062] Figure 1 An overview of the research area according to an embodiment of the present invention is shown;

[0063] Figure 2 A flowchart of a ground subsidence prediction method according to an embodiment of the present invention is shown;

[0064] Figure 3 The distribution of average subsidence rates in a typical area of ​​the Beijing-Tianjin-Hebei region from 2016 to 2020 is shown according to an embodiment of the present invention.

[0065] Figure 4 The cumulative settlement and settlement of typical PS points along each high-speed railway line according to embodiments of the present invention are shown.

[0066] Figure 5 The EMD decomposition results of the original time-series settlement at typical PS points along each high-speed railway line according to an embodiment of the present invention are shown.

[0067] Figure 6 This paper presents a comparison of the RMSE index of the prediction results for each high-speed rail line when the optimal IMF component decomposition is performed and when the maximum decomposition is performed, according to an embodiment of the present invention.

[0068] Figure 7 The following is a prediction of land subsidence along a high-speed railway line based on the EMD-GBDT, GBDT, and ARIMA models according to an embodiment of the present invention (the original values ​​cover the period from January 2016 to September 2020, and the EMD-GBDT, GBDT, and ARIMA models cover the period from January 2016 to September 2021). Detailed Implementation

[0069] To enable those skilled in the art to better understand the technical solutions of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. The embodiments of the present invention will be further described in detail below with reference to the accompanying drawings and specific examples, but this is not intended to limit the present invention. If there is no necessary sequential relationship between the various steps described herein, the order in which they are described as examples should not be considered a limitation. Those skilled in the art should understand that the order can be adjusted, as long as it does not disrupt the logical consistency between them and render the entire process impossible.

[0070] This invention provides a method for predicting land subsidence, which is applied to predicting land subsidence information along high-speed railway lines. This embodiment uses the Beijing-Tianjin-Hebei region as the study area, and elaborates on the specific implementation steps and effects of this invention.

[0071] The Beijing-Tianjin-Hebei region is located in the northern part of the North China Plain, with a total area of ​​approximately 218,000 km². 2 The Beijing-Tianjin-Hebei Plain, mainly encompassing Beijing, Tianjin, and parts of Hebei Province, has a population exceeding 100 million. Its geological structure is complex, broadly divided into a piedmont alluvial-lacustrine plain, a central alluvial-lacustrine plain, and an eastern coastal alluvial-marine plain. The region suffers from severe water scarcity, with a significant shortage of surface water forcing cities to rely on groundwater for water supply. Long-term groundwater extraction has led to the formation of multiple groundwater drawdown cones and land subsidence zones.

[0072] As of 2020, the total mileage of high-speed railways in the Beijing-Tianjin-Hebei region reached 2,163 km. The region is dominated by several high-speed railways, including the Beijing-Guangzhou, Beijing-Shanghai, and Beijing-Tianjin lines running east-west, and the Shijiazhuang-Jinan and Tianjin-Baoding lines running north-south. Areas in the Beijing-Tianjin-Hebei region with relatively rapid land subsidence rates are distributed in Beijing's Chaoyang District and Tongzhou District, Langfang, Baoding, Cangzhou, Hengshui, and Xingtai cities in Hebei Province, and Wangqingtuo District in Tianjin (Guo et al., 2020). However, some high-speed railways pass through or are close to subsidence areas, and the development of land subsidence will affect the safe operation of these railways. Therefore, this study selects typical high-speed railway lines in the Beijing-Tianjin-Hebei region, namely the Beijing-Tianjin, Beijing-Shanghai, Beijing-Guangzhou, Tianjin-Baoding, and Shijiazhuang-Jinan high-speed railways (only sections within the Beijing-Tianjin-Hebei region are selected) as the study area to conduct land subsidence monitoring and prediction along these high-speed railways, providing a basis for research on high-speed railway land subsidence hazards.

[0073] This embodiment selects 204 Sentinel-1A (S1A) archived ascent data, with track number 142 and frame numbers 126, 121, and 116, ranging as follows: Figure 1 As shown in Table 1, the time span is from January 14, 2016 to September 1, 2020. During PS-InSAR processing, the S1A data range of Frame 116 only covers the Beijing-Tianjin-Hebei plain area. The digital elevation model (DEM) used for terrain phase removal is the ShuttleRadar Topography Mission (SRTM) data released by NASA, with a resolution of 30 m. To verify the reliability of the PS-InSAR technology, second-order leveling monitoring data from seven countries within the study area were selected for verification. The locations of the leveling points are shown in Table 1. Figure 1 As shown, the monitoring period was from October 2016 to October 2017.

[0074] Table 1. S1A Radar Image Information

[0075]

[0076] This embodiment uses the Beijing-Tianjin, Beijing-Shanghai, Beijing-Guangzhou, Tianjin-Baoding, and Shijiazhuang-Jinan high-speed railways as the study areas. Spatially, sample points are taken at 1 km intervals along the high-speed railway lines. Since the Beijing-Tianjin high-speed railway is relatively short, the distance for taking sample points along the Beijing-Tianjin high-speed railway is 600 m. Temporally, the time series of the original data is from January 2016 to September 2020, with a total of 57 time points. To ensure the consistency of the time interval in the prediction model, the original data is processed into samples on a monthly basis. At the same time, missing months are interpolated to make up for them, and sample points in areas not covered by radar images are removed. The above-processed points are used as sample data for the prediction model to predict ground subsidence along each high-speed railway line. The number of input samples varies for different high-speed railways. The number of sample points for the Beijing-Shanghai, Beijing-Tianjin, Beijing-Guangzhou, Tianjin-Baoding, and Shijiazhuang-Jinan high-speed railways are 300, 190, 229, 161, and 188, respectively.

[0077] The technical process of this ground subsidence prediction method is as follows: Figure 2 As shown, the specific steps are as shown in S100-S300.

[0078] S100 decomposes the time-series subsidence data obtained by PS-InSAR technology into EMD, and during the decomposition, it iterates through the number of IMF components in the EMD.

[0079] Specifically, Permanent Scatterer Interferometry (PS-InSAR) was proposed by Ferretti et al. in 2000 (Ferretti et al., 2000). This method selects target points (PS points) with stable backscattered signals over a long time series. Phase analysis is performed on the PS points, mainly including terrain phase, horizon phase, atmospheric error, and noise error. The interferometric phase of each PS point is composed of the equation shown in Equation 1 (Wegmüller et al., 2004):

[0080] (1)

[0081] in, It is the phase of surface deformation. For terrain phase, It is a horizontal phase. Due to atmospheric error, This represents the noisy phase. In this study, SARPROZ software was used to process the S1A data. The specific workflow is as follows (Perissin et al., 2011):

[0082] 1) Merge the S1A data of the three frames, and then split them into three sub-strips according to different Swath numbers;

[0083] 2) For the S1A data of each sub-strip, select one main image and the others as auxiliary images. In this study, the image on February 8, 2018 was selected as the main image, and the main and auxiliary images were registered.

[0084] 3) Perform differential interferometry on the registered master and slave images to obtain differential interferograms; select PS points in the registered master and slave images using the amplitude stability coefficient method, with a threshold set to 0.75.

[0085] 4) Phase analysis was performed on the PS points. The phase information contained in the PS points is shown in Equation 1. In this study, SRTMDEM data and precise ephemeris data were used to remove topographic phase and horizontal phase information, respectively. First, the overall deformation phase components were solved using the nonlinear standard model in the APS (Atmospheric Phase Screen, APS) module of SARPROZ, and atmospheric phase information was inverted and removed. Then, the PS point deformation analysis module was used to introduce the inverted atmospheric phase, and the deformation phase of each PS point was inverted using the star topology analysis method, including nonlinear and linear deformation phases.

[0086] 5) Phase unwrapping is performed on the deformation phase information of PS points to finally obtain time-series surface deformation monitoring information.

[0087] Empirical Mode Decomposition (EMD) is a signal analysis method proposed by Huang et al. in 1998 (Huang et al., 1998). It plays a significant role in processing non-stationary time series data. This method can decompose fluctuations or trend signals at different scales in the original data stepwise, obtaining a series of local characteristic signals with different time scales, including a trend component and multiple intrinsic mode function (IMF) components. The EMD decomposition expression is shown in the formula... 2 As shown (Rilling et al., 2003), the IMF decomposed from EMD must satisfy the following conditions: the total number of maximum and minimum values ​​in the IMF is equal to or differs from the number of zero points by at most 1; and the average value of the envelope of the maximum and minimum values ​​at any point in the IMF should be equal to 0.

[0088] (2)

[0089] in This paper uses the ground subsidence data along each high-speed railway line as the original time series data. For the first One IMF component, This represents the trend.

[0090] S200 establishes a GBDT prediction model for the decomposed IMF components and trend components, predicts each component, and reconstructs the prediction results.

[0091] The Gradient Boosting Decision Tree (GBDT) model consists of multiple regression trees. It iterates through these trees, reducing the loss function along the gradient direction. The negative gradient of the loss function in the current model is used as an approximation of the residual in the boosting tree algorithm for the regression problem. The regression trees are then fitted, and the results from all trees are summed to make the final decision. For each sample point along each high-speed rail line, the specific steps of GBDT are as follows:

[0092] 1) Select the time series settlement data of sample points as the dataset. , Indicates the monitoring time, sets the ratio of training set to validation set, and includes the previous... The time-series sedimentation data were used to construct the training set. ,Will Time series sedimentation data were used to construct a validation set. , Less than Monitoring time;

[0093] 2) Initialize the training dataset. , The base learner is represented by a decision tree. For loss function, This represents a constant value that minimizes the loss function;

[0094] 3) For the training dataset, M decision trees need to be built. A tree, for Calculate the value of the negative gradient descent of the loss function. (T represents matrix transpose) as the model for fitting the next step of the decision tree model. For the current decision tree model Approximate residual value : ;

[0095] 4) The final model is obtained through M iterations. ,in It consists of M trees;

[0096] 5) Use the trained model To predict the amount of ground subsidence.

[0097] S300 compares the prediction accuracy of ground subsidence with different numbers of IMF components, selects the optimal number of IMF components, and outputs the prediction result as the final prediction result for that point.

[0098] During the simulation, training data and validation data accounted for 70% (first 39 time points) and 30% (last 16 time points) of the total time series data, respectively. The root mean square error (RMSE) and mean absolute error (MAE) were selected as the evaluation metrics for model accuracy. The calculation formulas are as follows: the smaller the RMSE and MAE values, the better the prediction effect of the model.

[0099] (3)

[0100] (4)

[0101] in Indicates the number of sample points. This indicates the first time series in the validation dataset. The label values ​​of each data point. Indicates the first [item] in the prediction dataset The predicted label value for each data point. Indicates the validation dataset The average of the true label values.

[0102] The prediction results are compared with those of the statistical model Autoregressive Integrated Moving Average Model (ARIMA) (Zhang et al., 2019) and the machine learning model GBDT to verify the feasibility of the EMD-GBDT model constructed using the method of this invention.

[0103] Typical ground subsidence characteristics along high-speed railways in the Beijing-Tianjin-Hebei plain area:

[0104] from Figure 3 As shown in the left image, from 2016 to 2020, multiple land subsidence funnels formed in the Beijing-Tianjin-Hebei plain region, exhibiting a continuous pattern, mainly distributed in the central part of the plain in a north-south oriented band. Areas with severe land subsidence were mainly distributed in Chaoyang-Tongzhou District of Beijing, Wuqing District of Tianjin, and parts of Langfang, Bazhou, Baoding, Hengshui, and Xingtai in Hebei Province. From January 2016 to September 2020, within the radar data coverage area, the maximum land subsidence rate reached 132 mm / year, distributed in Gaoyang County, Julu County, and Nangong City of Baoding City.

[0105] Analysis of ground subsidence along high-speed railways reveals that these railways pass through or are close to subsidence zones. The Tianjin-Baoding and Shijiazhuang-Jinan high-speed railways experience significant subsidence. The Tianjin-Baoding high-speed railway traverses the Tianjin Wuqing-Hebei Langfang and Hebei Xiongxian subsidence zones, with a maximum subsidence rate of 82 mm / year occurring 98 km from its starting point. The Shijiazhuang-Jinan high-speed railway experiences substantial overall subsidence, with 67% of its length experiencing a subsidence rate exceeding 30 mm / year. The maximum subsidence rate reaches 77 mm / year, occurring 136 km from its starting point and crossing the Hebei Jingxian subsidence zone. The Beijing-Shanghai, Beijing-Tianjin, and Beijing-Guangzhou high-speed railways are close to the Hebei Dongguang-Jingxian, Beijing Chaoyang-Tongzhou, and Hebei Xushui subsidence zones, with relatively gentler overall subsidence. The maximum subsidence rates are 55 mm / year, 28 mm / year, and 23 mm / year, respectively, located 250 km, 21 km, and 103 km from their starting points.

[0106] To further analyze the temporal evolution characteristics of ground subsidence along the high-speed railway, the PS point (location shown in [reference]) along the high-speed railway, where the maximum subsidence rate occurred, was selected. Figure 3 (Left) is used as a typical point for cumulative settlement and settlement analysis, such as Figure 4 As shown in the cumulative settlement curve, ground subsidence occurred continuously at each typical point from 2016 to 2020, with the maximum cumulative subsidence reaching 326 mm, 384 mm, and 350 mm for the Beijing-Shanghai, Tianjin-Baoding, and Shijiazhuang-Jinan high-speed railways, respectively. Further analysis of the subsidence variation characteristics at typical points is as follows... Figure 4As shown by the blue curve, the degree of fluctuation in the curve represents the unevenness of the temporal changes in land subsidence. It can be observed that between 2016 and 2020, the subsidence at each typical point exhibited fluctuation characteristics. The uneven evolution of temporal subsidence at typical points along the Beijing-Guangzhou High-Speed ​​Railway was slightly weaker, with subsidence fluctuation amplitudes generally within 5 mm. The subsidence fluctuation amplitudes at typical points along the Beijing-Shanghai and Beijing-Tianjin High-Speed ​​Railways were generally around 10 mm, while the uneven evolution of temporal subsidence at typical points along the Tianjin-Baoding and Shijiazhuang-Jinan High-Speed ​​Railways was the strongest, with maximum amplitudes reaching 16 mm and 12 mm respectively, during the periods of July to August 2016 and July to August 2018, respectively. Overall, the high-speed railway lines are significantly affected by the spatiotemporal development of regional land subsidence, necessitating close monitoring of land subsidence information along these lines.

[0107] PS-InSAR monitoring results verification:

[0108] To verify the accuracy of ground subsidence information acquired using PS-InSAR technology, this paper selects seven leveling monitoring points within the study area. The locations of the leveling points are as follows: Figure 1 As shown, the leveling monitoring period was from October 2016 to October 2017. A buffer zone with a radius of 500 meters was established centered on the leveling point. Settlement monitoring points (PS points) acquired using PS-InSAR within the buffer zone were extracted. The average settlement of these PS points from October 2016 to October 2017 was calculated and compared with the leveling monitoring values ​​from the same period. The comparison results are shown in Table 2. It can be observed that the maximum deviation between the PS-InSAR monitoring results and the leveling measurement results was 11 mm, the minimum deviation was 0 mm, and the average deviation was 5 mm. The Pearson correlation coefficient between the two reached 0.99 (95% confidence interval), indicating a strong correlation. This demonstrates the good accuracy of the PS-InSAR monitoring results and supports this study.

[0109] Table 2 Comparison of PS-InSAR monitoring results and leveling measurement results

[0110]

[0111] Settlement prediction along high-speed railways based on the EMD-GBDT model:

[0112] from Figure 4 The subsidence curves reveal that the ground subsidence exhibits a nonlinear evolution over time, with complex data evolution. To accurately predict ground subsidence along high-speed railway lines, the EMD decomposition method is selected to decompose the time-series subsidence, reducing the time-series complexity of the original data. GBDT models are then constructed for the decomposed IMF components and trend term to improve prediction accuracy.

[0113] For each high-speed rail line, taking the EMD decomposition of a PS point as an example, the settlement at that point is obtained using the EMD decomposition principle described in Section 3.2.1, including all IMF components and trend terms. Figure 5 As shown in the figure, the Beijing-Shanghai and Beijing-Tianjin high-speed railways have four IMF components, while the Beijing-Guangzhou, Tianjin-Baoding, and Shijiazhuang-Jinan high-speed railways each have three. With EMD decomposition, the fluctuation of the original time-series settlement decreases. However, due to the different time-series evolution characteristics of ground settlement sample points along each high-speed railway, the number of IMF components obtained from EMD decomposition varies at each point. To ensure that the prediction results for each sample point are optimal, this study performs a comprehensive decomposition at each point, constructs models for each decomposed IMF component, compares the error values ​​predicted by the GBDT model at different numbers of IMF components, and selects the number of IMF components with the optimal error and the prediction result at that time as the final prediction result. A comparison of the prediction errors for each high-speed railway line with optimal IMF component decomposition and maximum decomposition is shown below. Figure 6 As shown in Table 3, for the EMD-GBDT and GBDT models, the parameters used in the base learner and tree branching are mainly considered. The parameters of the model are optimized based on the grid search method, and the final model parameters are shown in Table 3.

[0114] Table 3. List of EMD-GBDT and GBDT model parameters

[0115]

[0116] The accuracy of the EMD-GBDT model was verified by selecting the GBDT and ARIMA models for comparison. The accuracy evaluation results of the three models are shown in Table 4, and the prediction results are as follows: Figure 7 As shown in the table, the optimized EMD-GBDT model has the lowest RMSE and MAE, indicating that the proposed model has the best prediction effect on time-series land subsidence. This further demonstrates that, given the complex nonlinear characteristics of land subsidence over time, predicting the IMF components and trend terms obtained after EMD decomposition can improve the prediction accuracy.

[0117] Table 4 Comparison of accuracy metrics between EMD-GBDT and GBDT / ARIMA models

[0118]

[0119] Note: Bold text indicates the model with the lowest error index. RMSE and MAE values ​​are in millimeters.

[0120] from Figure 7It can be observed that each high-speed railway line still exhibits a slow settlement phenomenon. To assess the operational stability of high-speed railways, the gradient changes along each high-speed railway line were statistically analyzed, and the results are shown in Table 5. The gradient calculation formula is shown in the following formula (Xie et al., 2019).

[0121] (5)

[0122] This represents the difference in settlement (m) between two points along the high-speed railway line. The distance between the two points is represented by the indicative symbol. The results show that from January 2016 to September 2021, the impact of differential ground subsidence on the slope changes along each high-speed railway was within the engineering technical standards (2‰). The slope changes along most high-speed railways were concentrated within 0.02‰, with the Tianjin-Baoding high-speed railway having a slope change impact range of 0.15‰ to 0.1‰ for a length of 3.13 km. In summary, the results indicate that ground subsidence along typical high-speed railways in the Beijing-Tianjin-Hebei region shows a slow but increasing trend. Ground subsidence affects the slope changes along high-speed railways and thus the stability of high-speed railway operations. Therefore, the development of ground subsidence along each high-speed railway should be closely monitored.

[0123] Table 5. Impact of Differential Ground Subsidence Evolution on Gradient Changes of Typical High-Speed ​​Railway Lines in the Beijing-Tianjin-Hebei Region from 2016 to 2021

[0124]

[0125] Note: The symbol "-" indicates that there is no data in this interval.

[0126] Excessive long-term groundwater extraction has led to significant land subsidence in the Beijing-Tianjin-Hebei region. With the integrated development of transportation in the region, the dense rail transit network and the ongoing land subsidence are impacting the stability of high-speed rail operations. This study selected five typical high-speed rail lines within the Beijing-Tianjin-Hebei plain. Based on land subsidence information from typical areas, it explored the evolution of land subsidence along the high-speed rail lines, used an optimized EMD-GBDT model to predict land subsidence along the lines in 2021, and investigated the operational stability of the high-speed rail. The conclusions are as follows:

[0127] 1) From 2016 to 2020, land subsidence in the Beijing-Tianjin-Hebei Plain mainly occurred in the central part, showing a north-south zonal distribution. Multiple land subsidence funnels in the region were connected, with Gaoyang County, Julu County and Nangong City experiencing severe subsidence, with the maximum land subsidence rate reaching 132 mm / year.

[0128] 2) The north-south high-speed railways of Tianjin-Baoding and Shijiazhuang-Jinan in the Beijing-Tianjin-Hebei region have experienced severe subsidence. The Tianjin-Baoding Expressway passes through the ground subsidence funnel between Wuqing, Tianjin and Langfang, Hebei and Xiongxian, Hebei, with the maximum subsidence rate along the line reaching 82 mm / year.

[0129] 3) Ground subsidence exhibits complex fluctuations over time. Decomposing and re-predicting time-series ground subsidence using the EMD-GBDT model can improve prediction accuracy. As of September 2021, ground subsidence has a certain impact on the gradient changes along various high-speed railways, but it is still within the engineering and technical standards.

[0130] In summary, high-speed railways are widely distributed in the Beijing-Tianjin-Hebei Plain, and the continuous development of ground subsidence affects the safe operation of these railways. Analyzing the current status of ground subsidence along high-speed railway lines and predicting it is of great significance for the operational safety of high-speed railways. This paper uses Persistent Scatterer Interferometric Synthetic Aperture Radar (PS-InSAR) technology to process Sentinel-1A data covering the Beijing-Tianjin-Hebei Plain from 2016 to 2020 to obtain ground subsidence information in the Beijing-Tianjin-Hebei Plain and along typical high-speed railway lines. Based on the analysis of the characteristics and patterns of ground subsidence along high-speed railway lines, a joint empirical mode decomposition and gradient boosting decision tree method is proposed to predict ground subsidence along high-speed railway lines. The research results show that some sections of typical high-speed railways in the Beijing-Tianjin-Hebei Plain pass through or are close to ground settlement funnels. Among them, the maximum cumulative settlement along the Beijing-Shanghai, Tianjin-Baoding, and Shijiazhuang-Jinan high-speed railways reached 326 mm, 384 mm, and 350 mm, respectively. The empirical mode decomposition method can reduce the time series complexity of ground settlement. On this basis, when the gradient boosting decision tree model is used for refined settlement prediction, the root mean square error of the model is 0.38 mm to 0.56 mm and 0.23 mm to 0.38 mm, respectively, which improves the prediction accuracy of ground settlement along high-speed railways.

[0131] In addition, it should be noted that the ground subsidence prediction method provided by the present invention includes, but is not limited to, the application in typical high-speed railway lines in the Beijing-Tianjin-Hebei Plain as exemplified above. It can also be applied to the ground subsidence prediction of other linear features. This embodiment is merely an example and is not intended to limit the present invention.

[0132] Furthermore, although exemplary embodiments have been described herein, their scope includes any and all embodiments based on the invention that have equivalent elements, modifications, omissions, combinations (e.g., schemes involving intersections of various embodiments), adaptations, or alterations. Elements in the claims will be interpreted broadly based on the language used in the claims and are not limited to the examples described in this specification or during the implementation of this application, and such examples will be interpreted as non-exclusive. Therefore, this specification and examples are intended to be considered illustrative only, and the true scope and spirit are indicated by the following claims and the full scope of their equivalents.

[0133] The above description is intended to be illustrative and not restrictive. For example, the above examples (or one or more of them) can be used in combination with each other. Other embodiments may be used by those skilled in the art upon reading the above description. Furthermore, in the above detailed description, various features may be grouped together to simplify the invention. This should not be construed as an intention that a feature of an unclaimed invention is necessary for any claim. Rather, the subject matter of the invention may be less than all the features of a particular embodiment of the invention. Thus, the following claims are incorporated herein by reference as examples or embodiments, wherein each claim is independently considered as a separate embodiment, and these embodiments are contemplated as being able to be combined with each other in various combinations or arrangements. The scope of the invention should be determined by reference to the appended claims and the full scope of their equivalents.

Claims

1. A method for predicting ground subsidence, characterized in that, The method includes: Obtain the temporal settlement of the ground at the target point; The time-series ground subsidence was decomposed using EMD to obtain IMF components and trend components. A prediction model is established using the IMF components and trend components. The prediction results are obtained by predicting and reconstructing each component. By comparing the prediction accuracy of settlement with different numbers of IMF components, the optimal number of IMF components is selected and the corresponding prediction result is used as the final prediction result of the target point. The acquisition of the time-series ground settlement at the target point includes: For each sub-strip of radar image data, select one main image and the rest as auxiliary images, and register the main image and the auxiliary images. In the registered master and slave images, the amplitude stability coefficient method is used to select target points; Phase analysis is performed on the target point, and the phase information contained in the target point is shown in formula (1): （1） in, It is the phase of surface deformation. For terrain phase, It is a horizontal phase. Due to atmospheric error, Noise phase; Based on solving the overall deformation phase components using the nonlinear standard model, an external digital elevation model is introduced to remove the terrain phase, orbital data is used to remove the horizon phase, atmospheric phase information is retrieved and removed using the atmospheric phase model, and star topology analysis method is used to retrieve all deformation phase information of each target point. The deformation phase information of each target point includes nonlinear and linear deformation phases. Phase unwrapping is performed on the deformation phase information of the target point to obtain the time-series ground settlement of the target point; The EMD decomposition of the time-series ground subsidence using formula (2) yields the IMF component and trend component: （2） in The original time-series data uses ground subsidence measurements along each high-speed railway line. For the first One IMF component, This is a trend item; The process of establishing a prediction model using the IMF components and trend components, predicting and reconstructing each component to obtain the prediction result includes: Select the time series sedimentation of sample points as the dataset , Indicates the monitoring time, sets the ratio of training set to validation set, and includes the previous... The time-series sedimentation data were used to construct the training set. ,Will Time series sedimentation data were used to construct a validation set. , Less than Monitoring time; Initialize for the training dataset , The base learner is represented by a decision tree. For loss function, This represents a constant value that minimizes the loss function; For the training dataset, construct M decision trees. A tree, for Calculate the value of the negative gradient descent of the loss function. As a reference for fitting the next step of the decision tree model For the current decision tree model Approximate residual value : ; The final model is obtained through M iterations. ,in It consists of M trees; Using the final model To predict the amount of ground subsidence.

2. The ground subsidence prediction method according to claim 1, characterized in that, By comparing the RMSE and MAE values, the accuracy of settlement prediction is improved when the number of IMF components is different. The smaller the RMSE and MAE values, the more accurate the model's prediction results are.

3. The ground subsidence prediction method according to claim 2, characterized in that, The RMSE value is calculated using the following formula: （3） in Indicates the number of sample points. This indicates the first time series in the validation dataset. The label values ​​of each data point. Indicates the first [item] in the prediction dataset The predicted label value for each data point.

4. The ground subsidence prediction method according to claim 3, characterized in that, The RMSE value is calculated using the following formula: （4） in Indicates the number of sample points. This indicates the first time series in the validation dataset. The label values ​​of each data point. Indicates the first [item] in the prediction dataset The predicted label value for each data point.

5. A ground settlement prediction device for implementing the method as described in any one of claims 1 to 4, characterized in that, The device includes: The acquisition module is configured to acquire the time-series ground settlement at the target point. The decomposition module is configured to perform EMD decomposition on the time-series ground subsidence to obtain IMF components and trend components. The prediction module is configured to establish a prediction model using the IMF components and trend components, predict each component, and reconstruct the prediction results. The selection module is configured to compare the prediction accuracy of settlement with different numbers of IMF components, select the optimal number of IMFs, and use the corresponding prediction result as the final prediction result for the target point.

6. A readable storage medium, characterized in that, The readable storage medium stores one or more programs, which can be executed by one or more processors to implement the method as described in any one of claims 1 to 4.

Images