Railway line settlement space-time prediction method and device

Abstract

The present application provides a kind of railway along line settlement space-time prediction method and device.The prediction method includes: based on the monitoring data obtained by monitoring along railway, unstable target area is identified;Obtain the time series settlement data of each monitoring point in unstable target area, and utilize optimal denoising model to carry out denoising processing, obtain the time series settlement data of each monitoring point after denoising processing;Based on the distance between each two monitoring points in unstable target area and the distance threshold value of pre-set, obtain adjacency matrix and degree matrix;Based on the time series settlement data of each monitoring point after denoising processing, adjacency matrix and degree matrix, utilize GCN-LSTM prediction model, adopt rolling prediction method to carry out lead prediction to the time series settlement data of each monitoring point, obtain the lead prediction set of each monitoring point.The present application is by constructing the GCN-LSTM prediction model of "space-time" coupling feature to carry out lead prediction to the time series settlement data of monitoring point, with the advantages of high prediction accuracy.

Description

Technical Field

[0001] This invention belongs to the field of geological disaster prediction technology, specifically relating to a method and device for predicting the spatiotemporal settlement along a railway line. Background Technology

[0002] As a vital national transportation infrastructure, the safe operation of railways is directly related to the national economy and the safety of people's lives and property. In recent years, affected by factors such as frequent extreme weather events, geological environmental changes, and intensified human engineering activities, disasters along railway lines have become more frequent. In particular, geological disasters such as roadbed settlement, karst collapse, and slope instability pose a serious threat to the smoothness of railway lines and traffic safety. Traditional monitoring methods mainly rely on manual inspections and local fixed-point monitoring, which have problems such as limited coverage, insufficient real-time performance, and delayed hazard identification, making it difficult to promptly detect potential risk areas along long linear railway lines.

[0003] Furthermore, surface subsidence, as a core indicator reflecting the evolution of a disaster body, has dual value for railway safety operation and maintenance through trend prediction: on the one hand, it can assess overall stability through historical and current subsidence data; on the other hand, it can predict potential future risks based on advanced subsidence predictions. Current mainstream prediction methods mostly focus on feature mining of single-point time-series subsidence data. While these methods can capture the temporal evolution patterns of monitoring points, they neglect the inherent spatial correlation characteristics of InSAR subsidence fields. The lack of spatiotemporal heterogeneity will cause prediction results to deviate from the actual subsidence field distribution pattern. Summary of the Invention

[0004] The purpose of this invention is to capture the spatial neighborhood correlation of the settlement field by introducing a graph convolutional network (GCN), extract the settlement time series features by coupling the optimal time series prediction model, and construct a GCN-LSTM prediction model with "spatial-temporal" coupled features to make advance predictions on the time series settlement data of monitoring points, which has the advantage of high prediction accuracy.

[0005] In a first aspect, the present invention provides a method for spatiotemporal prediction of settlement along a railway line. The prediction method includes the following steps: Step S10: Identifying unstable target areas based on monitoring data obtained from monitoring along the railway line; Step S20: Obtaining time-series settlement data for each monitoring point within the unstable target area, and performing noise reduction processing using an optimal noise reduction model to obtain noise-reduced time-series settlement data for each monitoring point; Step S30: Obtaining an adjacency matrix based on a preset distance threshold and the distance between every two monitoring points within the unstable target area, and constructing a degree matrix based on the adjacency relationships of the adjacency matrix; Step S40: Based on the noise-reduced time-series settlement data for each monitoring point, the adjacency matrix, and the degree matrix, using a GCN-LSTM prediction model and a rolling prediction method to perform advance prediction of the time-series settlement data for each monitoring point, obtaining an advance prediction set for each monitoring point. The GCN-LSTM prediction model has a corresponding relationship with the preset distance threshold and is a trained model. The advance prediction set includes multiple predicted settlement data that correspond to the prediction time.

[0006] In one specific implementation, the monitoring data includes TQI values ​​collected based on vehicle platforms, settlement data obtained based on aerospace platforms, image data collected by drones based on aviation platforms, inspection data obtained by manual inspections based on ground platforms, and geophysical detection data.

[0007] In one specific implementation, step S10 includes: acquiring the TQI value along the railway line using a track inspection vehicle on a vehicle platform, and determining the over-limit section based on a preset TQI value, wherein the TQI value of the over-limit section is greater than or equal to the preset TQI value; acquiring time-series settlement data and annual average settlement rate along the railway line using SBAS-InSAR technology on a space platform, and dividing the railway line into multiple settlement target areas and multiple non-settlement target areas based on the time-series settlement data and annual average settlement rate; identifying unstable target areas and target areas to be verified by combining the over-limit sections determined by the vehicle platform and the settlement target areas and non-settlement target areas determined by the space platform; identifying the settlement target area as an unstable target area if the over-limit section is located within the settlement target area; and identifying the over-limit sections within the non-settlement target areas and the settlement target areas not including the over-limit sections as areas to be verified.

[0008] In one specific implementation, step S10 further includes: acquiring image data of the area to be verified based on an aerial platform and acquiring inspection data of the area to be verified based on manual inspection on a ground platform; further identifying the area to be verified based on the image data and the inspection data; identifying the area to be verified as an unstable target area when the area to be verified has both adjacent construction disturbances and subsidence points; identifying the area to be verified as a further verification area when the area to be verified has adjacent construction disturbances but no subsidence points or has subsidence points but no adjacent construction disturbances; and verifying the further verification area based on geophysical exploration methods.

[0009] In one specific implementation, the optimal denoising model is a wavelet transform model. The method for obtaining the optimal denoising model in step S20 includes: acquiring time-series settlement data of multiple target monitoring points, all of which are located within an unstable target region; performing denoising processing on the time-series settlement data of each target monitoring point using both a wavelet transform model and a Gaussian filtering model, obtaining a first denoising result corresponding to the wavelet transform model and a second denoising result corresponding to the Gaussian filtering model, wherein both the first and second denoising results include the denoised time-series settlement data of each target monitoring point; calculating a denoising accuracy evaluation index based on the time-series settlement data of each target monitoring point, the first denoising result, and the second denoising result, obtaining a first denoising accuracy evaluation index corresponding to the wavelet transform model and a second denoising accuracy evaluation index corresponding to the Gaussian filtering model, wherein the denoising accuracy evaluation index includes the average signal-to-noise ratio and the average root mean square error of multiple target monitoring points; comparing the first denoising accuracy evaluation index and the second denoising accuracy evaluation index to obtain a first comparison result, and determining the wavelet transform model as the optimal denoising model based on the first comparison result.

[0010] In one specific implementation, the preset distance threshold is a predetermined optimal distance threshold. The prediction method further includes a step after step S20 and before step S30, which involves obtaining the optimal distance threshold and the GCN-LSTM prediction model, including: Step (1) Based on the time-series sedimentation data after noise reduction at each monitoring point, the input and output features of the time series prediction are constructed using the sliding window method, and a time feature dataset consisting of multiple data pairs is obtained, where each data pair includes input and output features with a preset time series length; Step (2) The time feature dataset is split into a first training set and a first test set; Step (3) The three time-series prediction models are first trained based on the first training set, and then the three trained time-series prediction models are evaluated based on the first test set to obtain the optimal time-series prediction model, where the three time-series prediction models include a Bayesian ridge regression model, a support vector regression model, and a long short-term memory network model, and the optimal time-series prediction model is a long short-term memory network model; Step (4) The GCN-LSTM prediction model is constructed by fusing the long short-term memory network model with the GCN model. Model testing; Step (5): Set multiple distance thresholds, and construct multiple spatial feature datasets corresponding to multiple distance thresholds one-to-one based on each distance threshold and the distance between every two monitoring points in the unstable target area. Each spatial feature dataset includes an adjacency matrix and a degree matrix corresponding to the distance threshold; Step (6): Based on the time feature dataset and multiple spatial feature datasets, construct multiple second training sets and multiple second test sets corresponding to multiple distance thresholds one-to-one. Each second training set includes the first training set and the spatial feature dataset, and the spatial feature datasets in different second training sets are different. Each second test set includes the first test set and the spatial feature dataset, and the spatial feature datasets in different second test sets are different. The second training set and the second test set including the same spatial feature dataset have a corresponding relationship; Step (7): First, train the GCN-LSTM prediction model using multiple second training sets respectively, and then evaluate the trained GCN-LSTM prediction model using multiple second test sets respectively to obtain the optimal distance threshold and the GCN-LSTM prediction model corresponding to the optimal distance threshold.

[0011] In one specific implementation, step (3) includes: inputting the first training set into the Bayesian Ridge Regression Model, the Support Vector Regression Model, and the Long Short-Term Memory Network Model for training, and obtaining the trained Bayesian Ridge Regression Model, the Support Vector Regression Model, and the Long Short-Term Memory Network Model; inputting the first test set into the trained Bayesian Ridge Regression Model, the Support Vector Regression Model, and the Long Short-Term Memory Network Model for prediction, and obtaining three sets of first prediction results corresponding one-to-one with the three time-series prediction models, each set of first prediction results including multiple first prediction settlement data corresponding one-to-one with multiple input features in the first test set; calculating the prediction accuracy evaluation index of each time-series prediction model based on multiple output features in the first test set and the three sets of first prediction results, and comparing them to obtain a second comparison result, wherein the prediction accuracy evaluation index includes the average of the mean absolute error of multiple monitoring points, the average of the mean square error, and the average of the goodness of fit; and determining the optimal time-series prediction model as the Long Short-Term Memory Network Model based on the second comparison result.

[0012] In one specific implementation, step (7) includes: inputting multiple second training sets into a GCN-LSTM prediction model for training, obtaining multiple trained GCN-LSTM prediction models, wherein the multiple trained GCN-LSTM prediction models have a one-to-one correspondence with multiple distance thresholds; inputting multiple second test sets into the corresponding trained GCN-LSTM prediction models for prediction, obtaining multiple sets of second prediction results that correspond one-to-one with multiple distance thresholds, wherein each set of second prediction results includes multiple second prediction settlement data that correspond one-to-one with multiple input features in the first test set; calculating the prediction accuracy evaluation index of each trained GCN-LSTM prediction model based on multiple output features in the first test set and multiple sets of second prediction results, and comparing them to obtain a third comparison result, wherein the prediction accuracy evaluation index includes the average of the mean absolute error, the average of the mean square error, and the average of the goodness of fit of multiple monitoring points; and obtaining the optimal distance threshold and the GCN-LSTM prediction model corresponding to the optimal distance threshold based on the third comparison result.

[0013] In one specific implementation, the preset time series length is N. The time series length of the rolling prediction method is set according to the preset time series length. The step of using the rolling prediction method to make any advance prediction for any monitoring point among multiple monitoring points includes: step (a) obtaining target time series settlement data for the prediction of the Mth settlement data, wherein the target time series settlement data consists of the latest N settlement data, and the N settlement data included in the target time series settlement data are measured settlement data and / or predicted settlement data; step (b) using the target time series settlement data, the adjacency matrix and the degree matrix as input, and using the GCN-LSTM prediction model to obtain the predicted settlement data for the Mth prediction.

[0014] Secondly, the present invention provides a spatiotemporal prediction device for settlement along a railway line. The prediction device includes: an identification module for identifying an unstable target area based on monitoring data obtained from monitoring along the railway line; a noise reduction module for acquiring time-series settlement data of each monitoring point within the unstable target area and performing noise reduction processing using an optimal noise reduction model to obtain noise-reduced time-series settlement data for each monitoring point; a spatial feature construction module for constructing an adjacency matrix based on a preset distance threshold and the distance between every two monitoring points within the unstable target area, and constructing a degree matrix based on the adjacency relationships of the adjacency matrix; and a prediction module for using a GCN-LSTM prediction model and a rolling prediction method to predict the time-series settlement data of each monitoring point based on the noise-reduced time-series settlement data of each monitoring point, the adjacency matrix, and the degree matrix, to obtain a prediction set for each monitoring point. The GCN-LSTM prediction model corresponds to the preset distance threshold and is a trained model. The prediction set includes multiple predicted settlement data corresponding to the prediction time.

[0015] The beneficial effects of the present invention include at least the following:

[0016] The present invention provides a method for predicting the spatiotemporal settlement along railway lines. First, it identifies unstable target areas. Then, based on the temporal settlement data of monitoring points within the unstable target areas and the spatial feature data between multiple monitoring points constructed based on preset distance thresholds, it utilizes a GCN-LSTM prediction model and a rolling prediction method to predict the temporal settlement data of each monitoring point in advance. Thus, the present invention introduces a Graph Convolutional Network (GCN) to capture the spatial neighborhood associations of the settlement field, couples it with the optimal temporal prediction model LSTM to extract temporal settlement features, and constructs a GCN-LSTM prediction model with "spatial-temporal" coupled features to predict the temporal settlement data of monitoring points in advance, achieving high prediction accuracy. Furthermore, the present invention primarily targets the advance prediction of settlement data at monitoring points in unstable target areas, solving the technical problem of difficult monitoring of unstable areas along long linear railways, and enabling early prediction of key areas, resulting in high monitoring efficiency and low monitoring costs.

[0017] Second, this invention constructs a multi-dimensional and three-dimensional early identification system for unstable target areas along railway lines. By integrating track inspection vehicle detection, large-scale aerospace monitoring and ground-based fine exploration technologies, it establishes hierarchical identification principles, thereby breaking through the limitations of traditional methods and improving the timeliness and accuracy of identifying unstable target areas along long railway lines.

[0018] In addition to the objectives, features, and advantages described above, the present invention has other objectives, features, and advantages. The invention will now be described in further detail with reference to the figures. Attached Figure Description

[0019] Figure 1 A schematic flowchart of a method for predicting spatiotemporal settlement along a railway line according to an embodiment of the present invention;

[0020] Figure 2 for Figure 1 A schematic diagram of the identification process for unstable target areas along the railway line corresponding to step S10 is shown.

[0021] Figure 3 This is a flowchart of the spatiotemporal prediction method for railway settlement corresponding to Embodiment 1 of the present invention;

[0022] Figure 4 This is a monitoring result diagram based on a vehicle platform and a space platform in Embodiment 1 of the present invention, wherein, Figure 4 (a) in the figure shows the relationship between the settlement rate obtained from the aerial platform and the monitoring areas along the railway. Figure 4 (b) in the figure shows the relationship between settlement data obtained from an aerial platform and the location along the railway line. Figure 4 (c) in the figure represents the TQI value along the railway line obtained based on the vehicle platform;

[0023] Figure 5 The image data acquired based on the aviation platform in Embodiment 1 of the present invention;

[0024] Figure 6 This is an adjacency graph of nine monitoring points constructed according to different distance thresholds in Embodiment 1 of the present invention, wherein... Figure 6 The distance threshold corresponding to (a) is 600m. Figure 6 The distance threshold corresponding to (b) is 800m. Figure 6 The distance threshold corresponding to (c) is 1000m. Figure 6 The distance threshold corresponding to (d) in the figure is 1200m. Figure 6 The distance threshold corresponding to (e) is 1400m. Figure 6 The distance threshold corresponding to (f) in the figure is 1600m;

[0025] Figure 7 This is the mean absolute error of 9 monitoring points obtained when using the GCN-LSTM prediction model corresponding to different distance thresholds (600m, 800m, 1000m, 1200m, 1400m) for prediction in Embodiment 1 of the present invention.

[0026] Figure 8 The mean square error of the nine monitoring points was obtained when the GCN-LSTM prediction model corresponding to different distance thresholds (600m, 800m, 1000m, 1200m, 1400m) was used for prediction in Embodiment 1 of the present invention.

[0027] Figure 9 This is the result of the goodness of fit of the nine monitoring points obtained when using the GCN-LSTM prediction model corresponding to different distance thresholds (600m, 800m, 1000m, 1200m, 1400m) for prediction in Embodiment 1 of the present invention.

[0028] Figure 10 The comparison results are as follows: the mean absolute error of the 9 monitoring points corresponding to the GCN-LSTM prediction model in Embodiment 1 of the present invention and the mean absolute error of the 9 monitoring points corresponding to the LSTM time series prediction model.

[0029] Figure 11 The comparison results of the mean square error of the 9 monitoring points corresponding to the GCN-LSTM prediction model and the mean square error of the 9 monitoring points corresponding to the LSTM time series prediction model in Embodiment 1 of the present invention;

[0030] Figure 12 The comparison results of the goodness of fit of the 9 monitoring points corresponding to the GCN-LSTM prediction model and the goodness of fit of the 9 monitoring points corresponding to the LSTM time series prediction model in Embodiment 1 of the present invention;

[0031] Figure 13 This is a time-series settlement data map drawn based on the measured dataset and the advanced prediction dataset of 9 monitoring points in Embodiment 1 of the present invention.

[0032] Figure 14 A block diagram of a railway line settlement spatiotemporal prediction device provided in another embodiment of the present invention. Detailed implementation method:

[0033] The embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0034] To address the shortcomings of existing prediction methods that primarily focus on feature mining of single-point time-series settlement data while neglecting the inherent spatial correlation characteristics of InSAR settlement fields, this invention introduces a Graph Convolutional Network (GCN) to capture the spatial neighborhood correlation of settlement fields. This GCN is coupled with an optimal time-series prediction model to extract settlement time-series features, constructing a GCN-LSTM prediction model with "spatial-temporal" coupled features to predict the time-series settlement data of monitoring points in advance, offering the advantage of high prediction accuracy. Furthermore, this invention addresses the limitations of long linear railway lines, such as limited coverage and insufficient real-time performance due to manual inspections and insufficient monitoring points. To address the problem of delayed hazard identification, this invention constructs a multi-dimensional, three-dimensional early identification system for unstable target areas along railway lines. By integrating track inspection vehicle detection, large-scale aerospace monitoring, and fine-grained ground exploration technologies, it establishes hierarchical identification principles, thereby overcoming the limitations of traditional methods and improving the timeliness and accuracy of unstable target area identification for long railway lines. On the one hand, it can reduce monitoring costs, and on the other hand, it can enable early prediction of settlement data at various monitoring points in key areas (unstable target areas). This allows for early warning and prevention of settlement risks in unstable target areas based on the predicted settlement data.

[0035] It should be noted that the time-series settlement data in this invention consists of several settlement data that correspond to time characteristics in sequence (in chronological order), and the settlement data is the cumulative settlement amount, such as settlement data A1 on January 1, 2023, settlement data A2 on March 1, 2023, settlement data A3 on May 1, 2023, etc.

[0036] Please see Figure 1 According to a first aspect of the present invention, the present invention provides a method for spatiotemporal prediction of settlement along a railway line, the prediction method comprising the following steps:

[0037] Step S10: Based on the monitoring data obtained from monitoring along the railway line, identify unstable target areas.

[0038] In this invention, multiple methods are used to monitor the railway line, and unstable target areas are identified based on the monitoring data obtained by these multiple methods.

[0039] In one optional implementation, the monitoring data includes TQI values ​​collected based on a vehicle platform, time-series settlement results obtained based on a space platform, image data collected by an unmanned aerial vehicle (UAV) based on an aviation platform, inspection data obtained by manual inspection based on a ground-based platform, and geophysical detection results.

[0040] Specifically, the vehicle platform collects TQI values ​​through a track inspection vehicle; the space platform uses SBAS-InSAR technology to obtain time-series settlement results along the railway line, mainly based on Sentinel-1A imagery, 30m spatial resolution SRTM DEM data released by NASA, and Precise Orbit Ephemerides (POD) data from the Sentinel-1A satellite provided by the European Space Agency; the airborne platform collects image data through unmanned aerial vehicles; the ground-based platform is inspected manually; and geophysical exploration methods include multi-source frequency domain seismic wave method and micro-motion detection method.

[0041] Please refer to the following: Figure 2 Step S10 includes:

[0042] Step (1): Use the track inspection vehicle of the vehicle platform to obtain the TQI value along the railway line, and determine the over-limit section based on the preset TQI value. The TQI value of the over-limit section is greater than or equal to the preset TQI value.

[0043] In this invention, the preset TQI value is 15, which is determined based on existing grading standards.

[0044] Step (2): Based on the aerospace platform, use SBAS-InSAR technology to obtain the time-series settlement data and annual average settlement rate along the railway line, and divide the railway line into multiple settlement target areas and multiple non-settlement target areas based on the time-series settlement data and annual average settlement rate.

[0045] In this invention, the sedimentation target area has a significant sedimentation funnel compared to other areas. This can be understood as the maximum absolute value of the sedimentation data in the sedimentation target area being significantly greater than the absolute value of the sedimentation data in other areas. Distinguishing between the sedimentation target area and non-settlement target area is an existing technology, and those skilled in the art can make this judgment based on experience or changes in sedimentation data.

[0046] In this invention, the core calculation formula for SBAS-InSAR technology is:

[0047]

[0048] In the formula Let (x, r) be the interference phase of pixel (x, r) in the j-th interferogram. t B t AThe phase at time (x, r); The phase of the radar line-of-sight is deformed. These are the topographic phase, orbital error, atmospheric delay phase, and noise in the interferometry.

[0049] Step (3): Combining the over-limit section determined by the vehicle platform and the settlement target area and non-settlement target area determined by the aviation platform, identify unstable target areas and target areas to be verified. If the over-limit section is located within the settlement target area, the settlement target area is identified as an unstable target area. Over-limit sections within the non-settlement target area and settlement target areas that do not include over-limit sections are identified as areas to be verified.

[0050] In this invention, railway segments along the railway line that do not belong to the unstable target area or the area to be verified are all considered safe areas.

[0051] In step (3), when identifying unstable target areas, both TQI value and settlement characteristics are considered. If there are obvious settlement characteristics and the TQI value is greater than the preset TQI value (i.e., both limiting conditions are met), it is considered an unstable target area that needs to be monitored closely. The future time series settlement data of the unstable target area needs to be predicted so as to provide early warning and prevention. If only the TQI value is greater than the preset TQI value or the settlement characteristics meet the preset conditions, it is considered an area to be verified and further monitoring data needs to be obtained for further judgment.

[0052] For easier understanding, please refer to the examples provided. Figure 4 ,from Figure 4 It can be seen that the TQI value exceeding the limit section K2261+200 is located at the settlement target area B, so the settlement target area is an unstable target area. Since the settlement target areas A and C and the mileages K2250.6 and K2250.8 do not simultaneously meet the two limiting conditions, they are considered to be areas to be verified.

[0053] Step (4): Obtain image data of the area to be verified based on the aerial platform and inspection data of the area to be verified based on the ground platform. Further identify the area to be verified based on the image data and the inspection data. When the area to be verified has both adjacent construction disturbances and collapse points, the area to be verified is identified as an unstable target area. When the area to be verified has adjacent construction disturbances but no collapse points or has collapse points but no adjacent construction disturbances, the area to be verified is identified as a further verification area.

[0054] In this invention, image data of the area to be verified is captured by a drone to determine whether there is any adjacent construction disturbance in the area to be verified, and manual inspection is used to determine whether there is any subsidence point in the area to be verified. If both adjacent construction disturbance and subsidence point exist, the area to be verified is identified as an unstable target area. If the determination result is that neither adjacent construction disturbance nor subsidence point exists, the area to be verified is identified as a safe area. If the determination result is that either adjacent construction disturbance exists but no subsidence point exists, or a subsidence point exists but no adjacent construction disturbance exists, the area to be verified is identified as a further verification area.

[0055] This invention combines drone aerial photography and manual inspection to determine whether there is adjacent construction disturbance and subsidence points in the area to be verified. The collected data is highly reliable and can further improve the accuracy of unstable target area identification.

[0056] Step (5) Verify the further verification area using geophysical methods.

[0057] In this invention, geophysical methods are used to detect the further verification area, and the stability of the area is determined based on the detection results. If the area is unstable, it is an unstable target area.

[0058] In this invention, unstable target areas are identified through multiple methods, and then the temporal settlement data of monitoring points in the unstable target areas are predicted in advance. On the one hand, this can reduce monitoring costs, and on the other hand, it can enable timely early warning by focusing on monitoring unstable target areas.

[0059] Step S20: Obtain the time-series settlement data of each monitoring point in the unstable target area, and perform noise reduction processing using the optimal noise reduction model to obtain the noise-reduced time-series settlement data of each monitoring point.

[0060] In this invention, SBAS-InSAR technology is used to obtain time-series settlement data for each monitoring point within the unstable target area.

[0061] In this invention, the optimal noise reduction model is a wavelet transform model.

[0062] In an optional implementation, the method for obtaining the optimal noise reduction model in step S20 includes:

[0063] Step (a): Obtain time-series settlement data of multiple target monitoring points, all of which are located within an unstable target area.

[0064] When selecting the optimal noise reduction model, time-series sedimentation data from some or all monitoring points within the unstable target area can be used as the data to be processed.

[0065] For ease of understanding, let’s take an example. Suppose that the number of unstable target areas determined in step S10 is 3, and each unstable target area includes 5 monitoring points. In step (a), 3 or other numbers of monitoring points can be selected as target monitoring points in one unstable target area, or 1 or more monitoring points can be selected as target monitoring points in each unstable target area.

[0066] Step (b): Denoise the time-series settlement data of each target monitoring point using wavelet transform model and Gaussian filter model respectively, and obtain the first denoising result corresponding to the wavelet transform model and the second denoising result corresponding to the Gaussian filter model. Both the first denoising result and the second denoising result include the denoised time-series settlement data of each target monitoring point.

[0067] In this invention, the wavelet transform model is abbreviated as WT (Wavelet Transform) model; the Gaussian filtering model is abbreviated as GF (Gaussian Filtering) model.

[0068] Step (c): Calculate the noise reduction accuracy evaluation index based on the time-series settlement data of each target monitoring point, the first noise reduction result, and the second noise reduction result. Obtain the first noise reduction accuracy evaluation index corresponding to the wavelet transform model and the second noise reduction accuracy evaluation index corresponding to the Gaussian filter model. The noise reduction accuracy evaluation index includes the average signal-to-noise ratio and the average root mean square error of multiple target monitoring points.

[0069] In this invention, the calculation of signal-to-noise ratio (SNR) and root mean square error (RMSE) are existing techniques and will not be described in detail here.

[0070] It should be noted that when there are multiple target monitoring points, each target monitoring point will have a corresponding signal-to-noise ratio and root mean square error when calculating the noise reduction accuracy evaluation index. When evaluating the model, the arithmetic mean of the signal-to-noise ratio and root mean square error of all target monitoring points is used.

[0071] For ease of understanding, let's take an example. Assume there are 5 target monitoring points. The average signal-to-noise ratio is the arithmetic mean of the signal-to-noise ratios of the 5 monitoring points, and the average root mean square error is the arithmetic mean of the root mean square errors of the 5 monitoring points.

[0072] Step (d): Compare the first noise reduction accuracy evaluation index and the second model evaluation index to obtain the first comparison result, and determine the wavelet transform model as the optimal noise reduction model based on the first comparison result.

[0073] In this invention, the first comparison result includes a signal-to-noise ratio (SNR) of the wavelet transform model being greater than that of the Gaussian filter model, and a mean square error (MSE) of the Gaussian filter model being greater than that of the wavelet transform model.

[0074] Step S30: Based on the preset distance threshold and the distance between every two monitoring points in the unstable target area, obtain the adjacency matrix, and construct the degree matrix based on the adjacency relationship of the adjacency matrix.

[0075] In this invention, the predetermined optimal distance threshold is 800m.

[0076] In this invention, the construction of the adjacency matrix and degree matrix are existing technologies. The innovation of this invention lies in the consideration of spatial feature data when predicting the settlement data of monitoring points.

[0077] Construct an adjacency matrix A based on the distances between monitoring points, where i, j, k are monitoring points, and e is the distance between them. i，j The weight of the edge linking monitoring point i and monitoring point j is calculated based on the geographical location (coordinates) of the monitoring points, using the following formula:

[0078]

[0079] In the formula, d i,j H is the distance between the two monitoring points; H is the distance threshold.

[0080] As can be seen from the weight calculation formula above, when the distance between two monitoring points exceeds the distance threshold, their weight value is 0. This means that only the weight of the edge between two monitoring points whose distance is less than the distance threshold H is actually calculated. Thus, the adjacency matrix and degree matrix constructed based on different distance thresholds are different, resulting in a correspondence between the adjacency matrix and degree matrix in the spatial feature data and the distance threshold.

[0081] Step S40: Based on the denoised temporal settlement data of each monitoring point and the spatial feature data, the GCN-LSTM prediction model is used to perform advance prediction on the temporal settlement data of each monitoring point using the rolling prediction method, and the advance prediction set of each monitoring point is obtained. The GCN-LSTM prediction model has a corresponding relationship with the preset distance threshold and is a trained model. The advance prediction set includes multiple predicted settlement data that correspond to the prediction time.

[0082] In this invention, a rolling forecasting method is used to construct the input and output features for time series forecasting. When the time series length is 5, the input feature in the first data pair is L. t-4 L t-3 L t-2 L t-1 L t The output feature is L t+1 For the a-th data pair, the input feature is L t+a-5 L t+a-4 Lt+a-3 L t+a-2 L t+a-1 The output feature is L t+a The total settlement data for group a is output.

[0083] It should be noted that the rolling prediction method used in this invention is existing technology and will not be described in detail here. For details, please refer to the "Method for Monitoring and Predicting Deformation Along Railway Lines in Karst Areas Based on Temporal InSAR and Machine Learning" published in the Journal of Central South University in March 2025.

[0084] It is understood that the time interval between two adjacent predicted settlement data in the advanced prediction set is the same as the time interval between two adjacent settlement data in the time series settlement data.

[0085] To make it easier to understand, let's take an example. Suppose that the interval between two adjacent settlement data in the time series settlement data is 12 days, that is, settlement data is obtained once every 12 days; the advance prediction time is 600 days, then the advance prediction set includes 50 predicted settlement data, and the interval between two adjacent predicted settlement data is 12 days.

[0086] In one alternative implementation, the preset distance threshold is a pre-determined optimal distance threshold.

[0087] In an optional implementation, the prediction method further includes a step of obtaining the optimal distance threshold and the GCN-LSTM prediction model, located after step S20 and before step S30, comprising:

[0088] Step (1): Based on the time-series settlement data after noise reduction at each monitoring point, the sliding window method is used to construct the input and output features for time series prediction, and a time feature dataset consisting of multiple data pairs is obtained. Each data pair includes input and output features with a preset time series length.

[0089] In this invention, a sliding window method is used to construct the input and output features for time series prediction. When the time series length is 5, in the first data pair, the input features are L1, L2, L3, L4, and L5, and the output feature is L6; for the t-5 data pairs, the input feature is L... t-5 L t-4 L t-3 L t-2 L t-1 The output feature is L t This is to complete the construction of the time feature dataset.

[0090] It should be noted that the sliding window method for constructing input and output features for time series prediction is an existing technology and will not be elaborated upon here. For details, please refer to the article "A Method for Monitoring and Predicting Deformation Along Railway Lines in Karst Areas Based on Temporal InSAR and Machine Learning" published in the Journal of Central South University in March 2025.

[0091] Step (2): Split the time feature dataset into a first training set and a first test set.

[0092] In this invention, the time feature dataset is split into a first training set and a first test set in an 8:2 ratio.

[0093] Step (3): First, train the three time series prediction models based on the first training set, and then evaluate the trained three time series prediction models based on the first test set to obtain the optimal time series prediction model. The three time series prediction models include the Bayesian Ridge Regression Model, the Support Vector Regression Model, and the Long Short-Term Memory Network Model. The optimal time series prediction model is the Long Short-Term Memory Network Model.

[0094] In this invention, the Bayesian Ridge Regression model is abbreviated as Bayesian Ridge Regression model; the Support Vector Regression model is abbreviated as SVR (Support Vector Regression) model; and the Long Short-Term Memory network model is abbreviated as LSTM (Long Short-Term Memory) model.

[0095] This step includes:

[0096] Step 3.1: Input the first training set into the Bayesian Ridge Regression Model, Support Vector Regression Model, and Long Short-Term Memory Network Model respectively for training, and obtain the trained Bayesian Ridge Regression Model, Support Vector Regression Model, and Long Short-Term Memory Network Model.

[0097] Step 3.2: Input the first test set into the trained Bayesian ridge regression model, support vector regression model and long short-term memory network model respectively for prediction, and obtain three sets of first prediction results corresponding one-to-one with the three time series prediction models. Each set of first prediction results includes multiple first prediction settlement data corresponding one-to-one with multiple input features in the first test set.

[0098] In this invention, the first predicted settlement data is obtained by taking input features as input, and there is a corresponding relationship between the two.

[0099] To make it easier to understand, let's take an example. Suppose there are 200 input features. Then each first prediction result includes 200 first prediction settlement data based on the 200 input features.

[0100] Step 3.3: Based on multiple output features and three sets of first prediction results in the first test set, calculate the prediction accuracy evaluation index of each time series prediction model and compare them to obtain a second comparison result. The evaluation index includes the average of the mean absolute error, the average of the mean square error, and the average of the goodness of fit of multiple monitoring points.

[0101] In this embodiment, multiple monitoring points refer to all monitoring points. In other embodiments, multiple monitoring points may refer to only some monitoring points, for example, removing the monitoring points corresponding to the minimum mean absolute error and the maximum mean absolute error.

[0102] The steps for obtaining the prediction accuracy evaluation index for any one of the three time series prediction models include:

[0103] Step (a): Obtain the first prediction result corresponding to the time series prediction model.

[0104] Step (b): Based on the first prediction result corresponding to the time series prediction model and multiple output features in the first test set, obtain the mean absolute error, mean square error and goodness of fit for each monitoring point according to the calculation formulas for mean absolute error, mean square error and goodness of fit.

[0105] The calculation method for the mean absolute error, mean square error, and goodness of fit of any monitoring point is as follows: Based on the first prediction result corresponding to the time series prediction model, obtain multiple first predicted settlement data that correspond to each input feature of the monitoring point. Use the output feature corresponding to the input feature in the first test set as the true value, and use the first predicted settlement data corresponding to the input feature as the predicted value. Obtain the mean absolute error, mean square error, and goodness of fit of the monitoring point according to the calculation formulas for mean absolute error, mean square error, and goodness of fit.

[0106] Step (c): Based on the calculation formula of the arithmetic mean, obtain the average of the mean absolute error, the average of the mean square error, and the average of the goodness of fit of multiple monitoring points.

[0107] Step (d) uses the average of the mean absolute error, the average of the mean squared error, and the average of the goodness of fit calculated in step (c) as the evaluation index of the prediction accuracy of the time series prediction model.

[0108] Step 3.4: Based on the second comparison result, the optimal time series prediction model is determined to be the Long Short-Term Memory Network model.

[0109] In this invention, the second comparison result includes the mean absolute error and mean squared error of the long short-term memory network model, the support vector regression model, and the Bayesian ridge regression model increasing sequentially, and the goodness of fit decreasing sequentially.

[0110] Step (4): Based on the long short-term memory network model and the GCN model, the GCN-LSTM prediction model is constructed.

[0111] GCN (Graph Convolutional Network) is a deep learning model based on graph structure that learns graph data features by aggregating information from neighboring nodes.

[0112] Step (5): Set multiple distance thresholds and construct multiple spatial feature datasets corresponding to the multiple distance thresholds based on each distance threshold and the distance between every two monitoring points in the unstable target area. Each spatial feature dataset includes an adjacency matrix and a degree matrix corresponding to the distance threshold.

[0113] In this invention, the distance threshold is set in relation to the minimum and maximum distances between two monitoring points, as well as the settlement monitoring spacing specified in the Railway Subgrade Design Code.

[0114] In this invention, multiple distance thresholds are set as 600m, 800m, 1000m, 1200m, 1400m and 1600m, and the optimal distance threshold is determined to be 800m.

[0115] In this invention, the method for constructing a spatial feature dataset corresponding to any distance threshold includes: constructing an adjacency matrix based on the distance threshold and the distance between every two monitoring points in the unstable target area; and constructing a degree matrix based on the adjacency relationship of the adjacency matrix.

[0116] Step (6): Based on the time feature dataset and multiple spatial feature datasets, construct multiple second training sets and multiple second test sets that correspond one-to-one with multiple distance thresholds. Each second training set includes the first training set and the spatial feature dataset, and the spatial feature datasets in different second training sets are different. Each second test set includes the first test set and the spatial feature dataset, and the spatial feature datasets in different second test sets are different. Second training sets and second test sets with the same spatial feature dataset have a corresponding relationship.

[0117] For ease of understanding, let’s take an example. Suppose that a total of three distance thresholds are set, namely the first distance threshold, the second distance threshold and the third distance threshold. Based on step (5), the first adjacency matrix and the first degree matrix corresponding to the first distance threshold, the second adjacency matrix and the second degree matrix corresponding to the second distance threshold, and the third adjacency matrix and the third degree matrix corresponding to the third distance threshold are constructed.

[0118] Correspondingly, the number of the second training set and the second test set obtained in step (6) is also 3. The second training set 01 corresponding to the first distance threshold is composed of the first training set, the first adjacency matrix and the first degree matrix. The second test set 01 corresponding to the first distance threshold is composed of the first test set, the first adjacency matrix and the first degree matrix. The second training set 01 and the second test set 01 have a corresponding relationship. The second training set 02 corresponding to the second distance threshold is composed of the first training set, the second adjacency matrix and the second degree matrix. The second test set 02 corresponding to the second distance threshold is composed of the first test set, the second adjacency matrix and the second degree matrix. The second training set 02 and the second test set 02 have a corresponding relationship. The second training set 03 corresponding to the third distance threshold is composed of the first training set, the third adjacency matrix and the third degree matrix. The second test set 03 corresponding to the third distance threshold is composed of the first test set, the third adjacency matrix and the third degree matrix. The third training set 03 and the third test set 03 have a corresponding relationship.

[0119] Step (7): First, train the GCN-LSTM prediction model using multiple second training sets, and then evaluate the trained GCN-LSTM prediction model using multiple second test sets to obtain the optimal distance threshold and the GCN-LSTM prediction model corresponding to the optimal distance threshold.

[0120] This step includes:

[0121] Step 7.1: Input multiple sets of the second training set into the GCN-LSTM prediction model for training to obtain multiple trained GCN-LSTM prediction models. The multiple trained GCN-LSTM prediction models have a one-to-one correspondence with multiple distance thresholds.

[0122] Step 7.2: Input multiple second test sets into the pre-trained GCN-LSTM prediction model with corresponding relationships to make predictions, and obtain multiple sets of second prediction results that correspond one-to-one with multiple distance thresholds. Each set of second prediction results includes multiple second prediction settlement data that correspond one-to-one with multiple input features in the first test set.

[0123] In step 7.2, multiple second test sets are input into pre-trained GCN-LSTM prediction models with corresponding relationships for prediction. This can be understood as follows: second test set 01 is input into the GCN-LSTM prediction model trained using second training set 01 for prediction; second test set 02 is input into the GCN-LSTM prediction model trained using second training set 02 for prediction; and second test set 03 is input into the GCN-LSTM prediction model trained using second training set 03 for prediction. That is, the spatial feature dataset used for training the GCN-LSTM prediction model is the same as the spatial feature data used for evaluating the model.

[0124] Step 7.3: Based on multiple output features and multiple sets of second prediction results in the first test set, calculate the prediction accuracy evaluation index of each trained GCN-LSTM prediction model, and compare them to obtain a third comparison result. The evaluation index includes the average of the mean absolute error, the average of the mean square error, and the average of the goodness of fit of multiple monitoring points.

[0125] In an embodiment of the present invention, multiple monitoring points refer to all monitoring points. In other embodiments, multiple monitoring points may refer to only some monitoring points, for example, removing the monitoring points corresponding to the minimum mean absolute error and the maximum mean absolute error.

[0126] The steps for obtaining the prediction accuracy evaluation index of the GCN-LSTM prediction model corresponding to any distance threshold include:

[0127] Step (a): Obtain the second prediction result corresponding to the distance threshold.

[0128] To make it easier to understand, let's take an example. Suppose the distance threshold is 800m. This step is to obtain the second prediction result output by the GCN-LSTM prediction model corresponding to the distance threshold of 800m.

[0129] Step (b): Based on the second prediction result corresponding to the distance threshold and multiple output features in the first test set, the mean absolute error, mean square error and goodness of fit corresponding to each monitoring point are obtained according to the calculation formulas for mean absolute error, mean square error and goodness of fit.

[0130] The calculation method for the mean absolute error, mean square error, and goodness of fit for any monitoring point is as follows: Based on the second prediction result corresponding to the distance threshold, obtain multiple second predicted settlement data that correspond to multiple input features of the monitoring point. Use the output feature corresponding to the input feature in the first test set as the true value, and use the second predicted settlement data corresponding to the input feature as the predicted value. Obtain the mean absolute error, mean square error, and goodness of fit of the monitoring point according to the calculation formulas for mean absolute error, mean square error, and goodness of fit.

[0131] Step (c): Based on the calculation formula of the arithmetic mean, obtain the average of the mean absolute error, the average of the mean square error, and the average of the goodness of fit of multiple monitoring points.

[0132] Step (d) uses the average of the mean absolute error, the average of the mean square error, and the average of the goodness of fit calculated in step (c) as the evaluation index of the prediction accuracy of the GCN-LSTM prediction model corresponding to the distance threshold.

[0133] Step 7.4: Based on the third comparison result, obtain the optimal distance threshold and the GCN-LSTM prediction model corresponding to the optimal distance threshold.

[0134] In one optional implementation, the step of obtaining the optimal distance threshold and the GCN-LSTM prediction model further includes comparing the prediction accuracy evaluation index of the GCN-LSTM prediction model corresponding to the optimal distance threshold with the prediction accuracy evaluation index of the LSTM prediction model.

[0135] In one optional implementation, the preset time series length is N, and the time series length of the rolling prediction method is set according to the preset time series length. The step of using the rolling prediction method to make any advance prediction for any one of the multiple monitoring points includes:

[0136] Step (a) Obtain the target time-series settlement data for the prediction of the Mth settlement data, wherein the target time-series settlement data consists of the latest N settlement data.

[0137] When the preset time series length N is 5, the specific steps are as follows: add the settlement data obtained from the M-1th prediction to the target time series settlement data used to predict the M-1th settlement data, delete the settlement data from the M-6th prediction, and use the settlement data from the M-1th, M-2nd, M-3rd, M-4th, and M-5th predictions as the target time series settlement data for predicting the Mth settlement data.

[0138] In one optional implementation, the N settlement data included in the target time series settlement data are measured settlement data and / or predicted settlement data.

[0139] To facilitate understanding, let's illustrate with an example. Assume the time series length N is 5, and the space platform acquires settlement data every 12 days. The latest settlement data acquisition date is June 8, 2025. To predict the settlement data for June 20, 2025, the target time series settlement data includes C corresponding to June 8, 2025. a-1 C corresponding to 2025.05.27 a-2 C corresponding to 2025.05.15 a-3 The corresponding C for 2025.05.03 a-4 The corresponding C for 2025.04.21 a-5 After completing the settlement data prediction for June 20, 2025, to predict the settlement data for July 2, 2025, the target time-series settlement data includes the predicted settlement value for June 20, 2025, and the C value corresponding to June 8, 2025. a-1 C corresponding to 2025.05.27 a-2 C corresponding to 2025.05.15 a-3 The corresponding C for 2025.05.03 a-4 ...and so on. That is, in the first prediction, all five settlement data points are measured settlement data. In the second prediction, the most recent of the five settlement data points is the predicted settlement data, and the other four are measured settlement data. In the third prediction, the two most recent of the five settlement data points are the predicted settlement data, and the other three are measured settlement data. ...In the sixth and subsequent predictions, all five settlement data points are predicted settlement data.

[0140] Step (b): Using the target time-series settlement data, the adjacency matrix, and the degree matrix force as inputs, the predicted settlement data for the Mth prediction is obtained using the GCN-LSTM prediction model.

[0141] Please see Figure 14According to a second aspect of the present invention, the present invention provides a railway line settlement spatiotemporal prediction device 100, the prediction device 100 comprising: an identification module 101, used to identify an unstable target area based on monitoring data obtained from monitoring along the railway line; a noise reduction processing module 102, used to acquire time-series settlement data of each monitoring point within the unstable target area, and perform noise reduction processing using an optimal noise reduction model to obtain noise-reduced time-series settlement data of each monitoring point; and a spatial feature construction module 103, used to construct spatial features based on a predetermined optimal distance threshold and the distance between every two monitoring points within the unstable target area. The distance is used to construct an adjacency matrix, and a degree matrix is ​​constructed based on the adjacency relationships in the adjacency matrix. The prediction module 104 is used to perform advance prediction on the time-series settlement data of each monitoring point using the GCN-LSTM prediction model and the rolling prediction method based on the noise-reduced time-series settlement data of each monitoring point, the adjacency matrix, and the degree matrix, and to obtain the advance prediction set for each monitoring point. The GCN-LSTM prediction model has a corresponding relationship with the optimal distance threshold and is a trained model. The advance prediction set includes multiple predicted settlement data that have a corresponding relationship with the prediction time.

[0142] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the working process of the prediction device described above can be referred to the content of the foregoing method embodiments, and will not be repeated here.

[0143] Example 1

[0144] To better demonstrate the prediction method of this invention, based on a specific conventional railway example, the railway section of the experiment was selected. InSAR subsidence monitoring was mainly carried out based on Sentinel-1A imagery, SRTM DEM data with a spatial resolution of 30m released by NASA, and Precise Orbit Ephemerides (POD) data of the Sentinel-1A satellite provided by the European Space Agency. The spatial baseline threshold (the maximum spatial difference between two SAR images) was set to 4%, and the temporal baseline threshold was set to 90 days.

[0145] Please refer to the following: Figure 3 The implementation steps are as follows:

[0146] 1.1 Identification of Unstable Target Regions

[0147] First, based on the TQI results obtained from the track inspection vehicle, it can be seen that the TQI values ​​from the two inspections show roughly the same trend, revealing three sections exceeding the limit along the railway line: K2250.6, K2250.8, and K2261.2. Figure 4As shown in (c). Simultaneously, a large-scale "general survey" was conducted along a section of the railway using the "space platform" SBAS-InSAR technology, and the settlement rate and cumulative settlement results are shown below. Figure 4 As shown in (a) and (b), the results show that there are significant differential settlement characteristics along the railway line. Among them, the settlement in regions A (K2253-K2256+500 / K2256.5), B (K2258-K2261+500 / K2261.5), and C (K2243-K2247) is more obvious. The absolute values ​​of the cumulative settlement from January 1, 2019 to October 31, 2023 are ranked from largest to smallest as follows: B (-63.8mm) > A (-49.7mm) > C (-35.6mm), with the settlement in region B being the most obvious. Through the integrated analysis of the "vehicle platform" and the "space platform", it was determined that region B is an "unstable target area". That is, the section with the TQI value exceeding the limit, K2261+200, is located in the settlement target area B. Regions A and C, as well as mileages K2250.6 and K2250.8, need further verification to determine whether they are "unstable target areas".

[0148] To further delineate the "unstable target area," drones based on an "aerial platform" conducted aerial photography along the railway line to monitor its external environment. Through this drone aerial photography, it was discovered that there are multiple human-made disturbances adjacent to the existing railway line, such as... Figure 5 As shown in (a) to (f), the results indicate that there are a total of 6 engineering disturbances near the existing line, mainly distributed in areas A and B. No significant engineering disturbances were observed in aerial photographs near area C and at the aforementioned TQI-exceeding sections K2250.6 and K2250.8. The railway mileage adjacent to the construction sites is shown in Table 1, numbered a to f respectively. Most of the construction disturbances are located near settlement funnels and historical subgrade disaster sites. According to manual inspection results, a total of 8 historical subsidence points were found at the railway subgrade, all located in areas A and B, i.e. Figure 4 The specific mileage information of the collapse points T1 to T8 shown in (a) is shown in Table 2. Therefore, area A is also designated as an "unstable target area" through "verification", while area C is a safe area. The results of judging the stability of the area based on the TQI value and multi-source data fusion are shown in Table 3.

[0149] Table 1. Information on the nearest mileage to construction disturbance

[0150]

[0151] Table 2 Information on the Collapse Site

[0152]

[0153] Table 3 Results of the "Unstable Target Region" Identification Process

[0154]

[0155] 1.2 Determination of the Optimal Noise Reduction Model

[0156] The time-series settlement data of settlement funnel points S1-S6 and collapse points T1-T3 within the unstable target area A, obtained through multi-source collaborative monitoring, were input into the WT and GF models for noise reduction. Monitoring points T3, T2, S4, S6, S5, S2, S3, S1, and T1 were renumbered as monitoring points 1-9, respectively. Table 4 shows the accuracy evaluation results for the nine monitoring points. Analysis shows that the average signal-to-noise ratio (SNR) from highest to lowest is: WT (25.372) > GF (24.372); the average root mean square error (RMSE) from highest to lowest is: GF (3.815 mm) > WT (1.576 mm), indicating that the WT model is the optimal noise reduction model.

[0157] Table 4 Comparison of Predictive Performance between WT Model and GF Model

[0158]

[0159] 1.3 Determination of the Optimal Time Series Prediction Model

[0160] (1) Dataset construction and partitioning

[0161] A dataset was constructed based on the time-series settlement data after denoising using the WT model. The dataset was then divided into a training set and a test set in an 8:2 ratio for subsequent predictive analysis.

[0162] (2) Optimal Time Series Prediction Model Selection Results

[0163] The prediction results were evaluated using three metrics: prediction accuracy and error. The results are shown in Table 5. The mean absolute error (MAE), from highest to lowest, is: Bayesian Ridge regression model (1.0667 mm) > Support Vector Regression (SVR) model (0.8854 mm) > Long Short-Term Memory (LSTM) network model (0.5845 mm); the mean squared error (MSE), from highest to lowest, is: Bayesian Ridge (2.0409 mm). 2 SVR (1.3678mm) 2 )>LSTM(0.6877mm 2 ); goodness of fit R 2 The average values ​​from highest to lowest are: LSTM (0.9667) > SVR (0.9326) > Bayesian Ridge (0.9009), indicating that the LSTM model is the optimal time series prediction model.

[0164] Table 5 Comparison of prediction performance of time series prediction models

[0165]

[0166] 1.4 Determination of the Optimal Spatiotemporal Prediction Model

[0167] (1) Constructing a spatiotemporal prediction model

[0168] The GCN-LSTM model is constructed by fusing the optimal time series prediction model LSTM model with the GCN model.

[0169] The input data of the GCN-LSTM model includes temporal feature data, i.e., time-series settlement data, and spatial feature data, and the output data is time-series settlement data.

[0170] (2) Construction of multiple second training sets and multiple second test sets

[0171] The temporal feature data in the second training set is the first training set constructed in step 1.3. The spatial feature data in the second training set consists of an adjacency matrix and a degree matrix that correspond to the distance threshold. The temporal feature data in the second test set is the first test set constructed in step 1.3. The spatial feature data in the second test set consists of an adjacency matrix and a degree matrix that correspond to the distance threshold. The second training set is used to train the spatiotemporal prediction model, and the second test set is used to test the spatiotemporal prediction model. The second training set and the second test set with the same spatial feature data have a corresponding relationship.

[0172] In this embodiment, a total of 6 second training sets and 6 second test sets were constructed, corresponding to distance thresholds of 600m, 800m, 1000m, 1200m, 1400m and 1600m respectively.

[0173] The spatial feature data is obtained as follows:

[0174] Monitoring points 1-9 in unstable target area A were selected to construct spatial feature data and marked on a map. The distance information between each pair of monitoring points was obtained, and their adjacency relationships were constructed based on distance thresholds. The distance range between monitoring points in target area A is [114.9m, 2632.5m]. According to the "Railway Subgrade Design Code" (TB10001-2016), the monitoring spacing for settlement along railway lines should be 50m-100m. Therefore, 600m-1600m was selected as the distance threshold range. Based on this, six sets of experiments were designed and comparative experiments were conducted. The distance thresholds were selected as 600m, 800m, 1000m, 1200m, 1400m, and 1600m, respectively. The graph structure is shown below. Figure 6 As shown in (a) to (f).

[0175] Taking the experimental group with a distance threshold of 600m as an example, Figure 6 (a) It shows that monitoring points 4 and 5, 3 and 6, 6 and 7, and 8 and 9 are all adjacent. A 9×9 adjacency matrix A is constructed based on the distance between nodes, and a degree matrix D is constructed based on the adjacency relationship. It is a 9×9 diagonal matrix, and the values ​​on the diagonal represent the connection density of each node. The expressions for A and D are as follows:

[0176]

[0177] (3) Training of the spatiotemporal model and selection results of the optimal spatiotemporal prediction model

[0178] First, the GCN-LSTM model was trained using six second training sets. Then, the time-series settlement was predicted for monitoring points 1-9 of the trained GCN-LSTM model with corresponding relationships using six second test sets. The evaluation results were calculated based on the predicted settlement data and the measured settlement data, as shown in Figure 7. Figure 9 As shown in Table 6, the prediction results were evaluated using three prediction accuracy and error evaluation indicators. The overall prediction accuracy evaluation results are shown in Table 6. The results show that the experimental group with a distance threshold of 800m had the best prediction performance. Compared with other experimental groups, its prediction error indicators MAE and MSE were the smallest, at 0.4982 and 0.5771 respectively, while the prediction accuracy indicator R... 2 The maximum value is 0.9994. Meanwhile, the spatiotemporal prediction model GCN-LSTM outperforms the LSTM model, for example... Figures 10-12 As shown.

[0179] Table 6 Comparison of prediction results for different distance thresholds

[0180]

[0181] It should be noted that the data of the prediction accuracy evaluation index corresponding to different distance thresholds in Table 6 are the arithmetic mean of the prediction accuracy evaluation index of the nine monitoring points.

[0182] 1.5 Advanced Prediction

[0183] Based on the last data pair in the dataset, using the spatial data features corresponding to the optimal distance threshold of 800m, a pre-trained GCN-LSTM prediction model and a rolling prediction method are used to predict the temporal deformation data of monitoring points 1-9 in advance, obtaining the settlement prediction results for the next two years. Figure 13 As shown, it should be noted that Figure 13 The three graphs in the first row, from left to right, show the projected results for monitoring points 1, 2, and 3 over the next two years. Figure 13The three graphs in the second row, from left to right, show the projected results for monitoring points 4, 5, and 6 over the next two years. Figure 13 The three figures in the second row, from left to right, show the predicted results for monitoring points 7, 8, and 9 for the next two years.

[0184] Figure 13 The data shows that when the GCN-LSTM model predicts future subsidence values ​​with a lead time of about one year, the surface subsidence characteristics of monitoring points 3–8 within the prediction scale range are characterized by gradual and slow subsidence, without reaching a stable state; monitoring point 1 shows fluctuating subsidence; monitoring point 2 shows continuous and slow ground uplift and tends to stabilize; and the surface subsidence process of monitoring point 9 is rapid subsidence followed by slow subsidence, and finally tends to stabilize.

[0185] It should be noted that the model training method used in this invention is an existing technology, and this invention does not improve upon it. The conditions for terminating the model training iteration can include reaching the maximum number of training rounds, the loss decrease being less than a threshold, and other user-defined conditions. These conditions can be used individually or in combination to ensure that the training is sufficient without overfitting, while also taking into account the efficiency of computational resource utilization.

[0186] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions and substitutions can be made without departing from the inventive concept, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A railway line settlement space-time prediction method, characterized in that, The prediction method comprises the following steps: Step S10, identifying an unstable target area based on monitoring data obtained by monitoring along the railway; Step S20, obtaining time series settlement data of each monitoring point in the unstable target area, and performing denoising processing on the time series settlement data of each monitoring point by using an optimal denoising model to obtain denoising-processed time series settlement data of each monitoring point; Step S30, constructing an adjacency matrix based on a preset distance threshold and distances between each two monitoring points in the unstable target area, and constructing a degree matrix based on an adjacency relationship of the adjacency matrix; Step S40, based on the denoising-processed time series settlement data of each monitoring point, the adjacency matrix, and the degree matrix, using a GCN-LSTM prediction model, and using a rolling prediction method to perform lead prediction on the time series settlement data of each monitoring point to obtain a lead prediction set of each monitoring point, the GCN-LSTM prediction model has a corresponding relationship with the preset distance threshold and is a trained model, wherein the lead prediction set comprises a plurality of predicted settlement data corresponding to a prediction time; The preset distance threshold is a pre-determined optimal distance threshold, and the prediction method further comprises a step of obtaining the optimal distance threshold and the GCN-LSTM prediction model after step S20 and before step S30, comprising: Step (1), based on the denoising-processed time series settlement data of each monitoring point, constructing input features and output features of time series prediction by using a sliding window method to obtain a time feature data set comprising a plurality of data pairs, each data pair comprising input features and output features having a preset time series length; Step (2), splitting the time feature data set into a first training set and a first test set; Step (3), first training three time series prediction models based on the first training set, and then evaluating the trained three time series prediction models based on the first test set to obtain an optimal time series prediction model, the three time series prediction models comprising a Bayesian ridge regression model, a support vector regression model, and a long short-term memory network model, and the optimal time series prediction model is the long short-term memory network model; Step (4), fusing a GCN model based on the long short-term memory network model to construct a GCN-LSTM prediction model; Step (5), setting a plurality of distance thresholds, and constructing a plurality of spatial feature data sets corresponding to the plurality of distance thresholds according to the distance threshold and the distance between each two monitoring points in the unstable target area, each spatial feature data set comprising an adjacency matrix and a degree matrix corresponding to the distance threshold; Step (6), based on the time feature data set and a plurality of spatial feature data sets, a plurality of second training sets and a plurality of second test sets corresponding to a plurality of distance thresholds are constructed, each of the second training sets includes the first training set and the spatial feature data set, and the spatial feature data sets in different second training sets are different, each of the second test sets includes the first test set and the spatial feature data set, and the spatial feature data sets in different second test sets are different, the second training set and the second test set including the same spatial feature data set have a corresponding relationship; Step (7), first, the GCN-LSTM prediction model is trained using a plurality of second training sets, and then the trained GCN-LSTM prediction model is evaluated using a plurality of second test sets to obtain the best distance threshold and the GCN-LSTM prediction model corresponding to the best distance threshold.

2. The railway line settlement space-time prediction method according to claim 1, characterized in that, The monitoring data includes TQI values collected based on a vehicle platform, settlement data obtained based on a space platform, image data collected by a drone based on an aerial platform, inspection data obtained by manual inspection based on a ground platform, and detection data of geophysical prospecting.

3. The railway line settlement space-time prediction method according to claim 1, characterized in that, The step S10 comprises: A track inspection vehicle of the vehicle platform is used to obtain TQI values along the railway, and a section exceeding the limit is determined based on a preset TQI value, the TQI value of the section exceeding the limit being greater than or equal to the preset TQI value; Based on the space platform, time-series settlement data and annual average settlement rate along the railway are obtained by using the SBAS-InSAR technology, and the railway is divided into a plurality of settlement target areas and a plurality of non-settlement target areas based on the time-series settlement data and the annual average settlement rate; In combination with the section exceeding the limit determined by the vehicle platform and the settlement target area and the non-settlement target area determined by the aerial platform, an unstable target area and a to-be-verified target area are identified, the settlement target area is identified as the unstable target area when the section exceeding the limit is located in the settlement target area, and the section exceeding the limit in the non-settlement target area and the settlement target area not including the section exceeding the limit are all identified as the to-be-verified area.

4. The railway line settlement space-time prediction method according to claim 3, characterized in that, The step S10 further comprises: Image data of the to-be-verified area is obtained based on the aerial platform, and inspection data of the to-be-verified area is obtained by manual inspection based on the ground platform, and the to-be-verified area is further identified based on the image data and the inspection data, the to-be-verified area is identified as the unstable target area when the to-be-verified area simultaneously exists adjacent construction disturbance and collapse points, and the to-be-verified area is identified as the further verification area when the to-be-verified area exists adjacent construction disturbance and does not exist collapse points or exists collapse points and does not exist adjacent construction disturbance; The further verification area is verified based on geophysical prospecting means.

5. The railway line settlement space-time prediction method according to claim 1, characterized in that, The optimal denoising model is a wavelet transform model, and the method for obtaining the optimal denoising model in the step S20 comprises: Time-series settlement data of a plurality of target monitoring points are obtained, and the plurality of target monitoring points are all located in the unstable target area; The time series subsidence data of each target monitoring point is respectively denoised by using a wavelet transform model and a Gaussian filter model to obtain a first denoising result corresponding to the wavelet transform model and a second denoising result corresponding to the Gaussian filter model, wherein the first denoising result and the second denoising result each include denoised time series subsidence data of each target monitoring point; A denoising precision evaluation index is calculated based on the time series subsidence data of each target monitoring point, the first denoising result and the second denoising result to obtain a first denoising precision evaluation index corresponding to the wavelet transform model and a second denoising precision evaluation index corresponding to the Gaussian filter model, wherein the denoising precision evaluation index includes a signal-to-noise ratio average value and a root mean square error average value of multiple target monitoring points; The first denoising precision evaluation index and the second denoising precision evaluation index are compared to obtain a first comparison result, and the wavelet transform model is determined as the optimal denoising model based on the first comparison result.

6. The railway line settlement space-time prediction method according to claim 1, characterized in that, The step (3) comprises: The first training set is input into a Bayesian ridge regression model, a support vector regression model and a long short-term memory network model respectively for training to obtain trained Bayesian ridge regression model, support vector regression model and long short-term memory network model; The first test set is input into the trained Bayesian ridge regression model, support vector regression model and long short-term memory network model respectively for prediction to obtain three groups of first prediction results corresponding to the three time series prediction models respectively, and each group of first prediction results includes multiple first prediction subsidence data corresponding to multiple input features in the first test set; Based on the multiple output features in the first test set and the three groups of first prediction results, a prediction precision evaluation index of each time series prediction model is calculated respectively and compared to obtain a second comparison result, wherein the prediction precision evaluation index includes an average absolute error average value, a mean square error average value and a goodness of fit average value of multiple monitoring points; Based on the second comparison result, the optimal time series prediction model is determined as the long short-term memory network model.

7. The railway line settlement space-time prediction method according to claim 1, characterized in that, The step (7) comprises: The multiple second training sets are input into the GCN-LSTM prediction model respectively for training to obtain multiple trained GCN-LSTM prediction models, and the multiple trained GCN-LSTM prediction models have a one-to-one correspondence with the multiple distance thresholds; The multiple second test sets are input into the trained GCN-LSTM prediction model having the corresponding relationship respectively for prediction to obtain multiple groups of second prediction results corresponding to the multiple distance thresholds respectively, and each group of second prediction results includes multiple second prediction subsidence data corresponding to the multiple input features in the first test set; Based on the multiple output features in the first test set and the multiple groups of second prediction results, a prediction precision evaluation index of each trained GCN-LSTM prediction model is calculated respectively and compared to obtain a third comparison result, wherein the prediction precision evaluation index includes an average absolute error average value, a mean square error average value and a goodness of fit average value of multiple monitoring points. Based on the third comparison result, an optimal distance threshold value and a GCN-LSTM prediction model corresponding to the optimal distance threshold value are obtained.

8. The railway line settlement space-time prediction method according to claim 1, characterized in that, The preset time sequence length is N, and the time sequence length of the rolling prediction method is set according to the preset time sequence length. The step of using the rolling prediction method to perform any one step of the advance prediction of any one monitoring point in the plurality of monitoring points comprises: Step (a), obtaining target time sequence settlement data for Mth settlement data prediction, the target time sequence settlement data being composed of the latest N settlement data, and the N settlement data included in the target time sequence settlement data being measured settlement data and / or predicted settlement data; Step (b), taking the target time sequence settlement data, the adjacency matrix and the degree matrix as inputs, and using the GCN-LSTM prediction model to obtain predicted settlement data of the Mth prediction.

9. A device for spatio-temporal prediction of settlement along a railway line, characterized in that, The prediction device comprises: The recognition module is configured to recognize unstable target areas based on monitoring data obtained by monitoring along the railway line. The noise reduction processing module is configured to obtain time sequence settlement data of each monitoring point in the unstable target area, and to perform noise reduction processing on the time sequence settlement data of each monitoring point by using an optimal noise reduction model to obtain noise-reduced time sequence settlement data of each monitoring point. The prediction module is configured to use a GCN-LSTM prediction model to perform advance prediction on the time sequence settlement data of each monitoring point by using a rolling prediction method based on the noise-reduced time sequence settlement data of each monitoring point, the adjacency matrix and the degree matrix, to obtain an advance prediction set of each monitoring point, and the GCN-LSTM prediction model has a corresponding relationship with a preset distance threshold value and is a trained model. The prediction device further comprises an acquisition module configured to acquire an optimal distance threshold value and the GCN-LSTM prediction model, which comprises: Based on the noise-reduced time sequence settlement data of each monitoring point, input features and output features of time sequence prediction are constructed by using a sliding window method to obtain a time feature data set composed of a plurality of data pairs, each data pair comprising input features and output features with a preset time sequence length. The time feature data set is split into a first training set and a first test set. Three time sequence prediction models are trained based on the first training set, and the trained three time sequence prediction models are evaluated based on the first test set to obtain an optimal time sequence prediction model, the three time sequence prediction models comprising a Bayesian ridge regression model, a support vector regression model and a long short-term memory network model, and the optimal time sequence prediction model being the long short-term memory network model. A GCN-LSTM prediction model is constructed by fusing a long short-term memory network model with a GCN model. A plurality of distance threshold values are set, and a plurality of spatial feature data sets corresponding to the plurality of distance threshold values are constructed according to each distance threshold value and the distance between each two monitoring points in the unstable target area, each spatial feature data set including an adjacency matrix and a degree matrix corresponding to the distance threshold value; Based on the time feature data set and the plurality of spatial feature data sets, a plurality of second training sets and a plurality of second test sets corresponding to the plurality of distance threshold values are constructed, each second training set including the first training set and a spatial feature data set, and the spatial feature data sets in different second training sets being different, each second test set including the first test set and a spatial feature data set, and the spatial feature data sets in different second test sets being different, the second training set and the second test set including the same spatial feature data set having a corresponding relationship; The GCN-LSTM prediction model is trained using the plurality of second training sets, and the trained GCN-LSTM prediction model is evaluated using the plurality of second test sets, to obtain the best distance threshold value and the GCN-LSTM prediction model corresponding to the best distance threshold value.

Images