Data-model hybrid driven high temporal-spatial resolution deformation data prediction method

Abstract

The embodiment of the invention provides a data-model hybrid-driven high-temporal-spatial-resolution deformation data prediction method, which belongs to the technical field of measurement, and specifically comprises the following steps: processing radar image data to obtain earth surface sight historical deformation data; the projected GNSS deformation observation value carries out adjustment on the interpolated historical deformation data; taking a Kalman filtering model as a basic model, and dynamically updating a state transition matrix A, a process noise covariance matrix Q and an observation noise covariance matrix R through a space-time diagram convolutional network model; taking the adjusted historical deformation data as a training sample, and estimating an optimal hyper-parameter of the time sequence prediction model by using a hyper-parameter search method; and inputting time needing to be predicted into the target prediction model in combination with subsequently acquired GNSS deformation data to obtain earth surface deformation data at future moments. Through the scheme of the invention, the future trend can be predicted by considering the space-time correlation of the historical deformation of the earth surface at the same time, and the prediction efficiency, accuracy and adaptability are improved.

Description

Technical Field

[0001] The embodiments of the present invention relate to the field of measurement technologies, and in particular, to a method for predicting deformation data with high spatio-temporal resolution driven by a data-model hybrid. Background Art

[0002] Land subsidence refers to an environmental phenomenon in which the regional ground elevation decreases due to the compression of the soil layer on the earth's surface under the combined action of natural and human factors, also known as ground settlement or land subsidence. In recent decades, with the rapid increase in population and the acceleration of urbanization, the problem of land subsidence has become increasingly severe, posing a significant threat to human social activities and infrastructure. Therefore, implementing large-scale land deformation prediction work has important guiding significance for earthquake prevention and disaster reduction, ensuring infrastructure safety, rational resource development, and scientific urban development planning.

[0003] Currently, two main categories of methods are mainly used in the field of land deformation prediction. The first category is traditional physical models or time series mathematical models (such as Kalman filtering) based on physical parameters. Among them, the reliability of the method based on physical models depends on the degree of description of the deformation mechanism by the model, and its reliability is not high when facing complex deformation scenarios. Time series mathematical models such as Kalman filtering predict the future deformation trend through the historical evolution process of land deformation, and its dependence on the physical deformation mechanism is weaker than that of physical models. However, traditional Kalman filtering assumes that key matrices (such as the state transition matrix, process noise matrix, etc.) are all fixed values, which leads to poor accuracy when facing complex non-linear time series deformation processes.

[0004] Another category of methods is deep learning models based on data-driven. Deep learning models based on data-driven, including time series prediction models such as Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and one-dimensional Convolutional Neural Network (CNN), effectively solve the prediction reliability problem in large-scale land subsidence prediction due to the difficulty of accurately estimating model parameters or the scarcity of relevant data in traditional methods with their highly non-linear modeling capabilities. Nevertheless, the efficient operation of deep learning models highly depends on a large amount of high-quality training data, and the training process takes a long time, which to a certain extent hinders the application prospect of this model in land deformation areas with high real-time requirements and less available data. In this scenario, there is an urgent need for a method for predicting deformation data with high spatio-temporal resolution that has high prediction efficiency, accuracy, and adaptability. Summary of the Invention

[0005] In view of this, an embodiment of the present invention provides a high spatio-temporal resolution prediction method for surface deformation driven by a data-model hybrid, so as to overcome the problems that existing time series models such as Kalman filter are difficult to predict highly nonlinear deformation, and data-driven deep learning methods highly rely on a large number of training samples and have low migration ability, etc.

[0006] An embodiment of the present invention provides a high spatio-temporal resolution prediction method for surface deformation driven by a data-model hybrid, including:

[0007] Step 1, obtain multi-temporal repeat-pass synthetic aperture radar image data within the monitoring period of the target area, use the time series InSAR technology to process the radar image data to obtain the historical deformation data in the satellite line-of-sight direction of the surface, and use spline curve interpolation for the historical deformation data;

[0008] Step 2, collect the surface GNSS three-dimensional deformation data within the same period, project it onto the satellite line-of-sight direction, and use the projected GNSS deformation observation values to adjust the interpolated historical deformation data;

[0009] Step 3, take the Kalman filter model as the basic model, and dynamically update the state transition matrix A, the process noise covariance matrix Q, and the observation noise covariance matrix R in it through a spatio-temporal graph convolutional network model, so as to form a data-model hybrid driven time series prediction model;

[0010] Step 4, use the adjusted historical deformation data as training samples, and use the hyperparameter search method to estimate the best hyperparameters of the time series prediction model to obtain the target prediction model;

[0011] Step 5, combine the subsequently collected GNSS deformation data, input the time to be predicted into the target prediction model, and obtain the surface deformation data at future moments.

[0012] According to a specific implementation manner of an embodiment of the present invention, the time series prediction model includes a state prediction equation and a state update equation.

[0013] According to a specific implementation manner of an embodiment of the present invention, the state prediction equation is

[0014]

[0015]

[0016] f = STGNN F

[0017] where f represents the state transition equation, represents the state prediction value at the current moment t, represents the output result of the Kalman filter state at the previous moment, and f′ represents the Jacobian matrix equation, which is the first-order derivative result of f and is used to estimate the current state covariance matrix Adopt STGNN F The model models f, Q t represents the real-time value of the process noise covariance matrix Q, STGCN Q is to predict the real-time Q t prediction model of

[0018] According to a specific implementation manner of an embodiment of the present invention, the state update equation is

[0019]

[0020] where Z t represents the observed data at the current moment, K t represents the Kalman filter gain at the current moment, I represents the unit weight matrix, R t represents the real-time value of the observation noise covariance matrix R, STGCN R is to predict the real-time R t prediction model of

[0021] A data-model hybrid-driven high spatio-temporal resolution prediction method for surface deformation in an embodiment of the present invention includes: Step 1, obtaining multi-temporal repeat-pass synthetic aperture radar image data within the monitoring period of the target area, using the time series InSAR technology to process the radar image data to obtain the historical deformation data in the satellite line-of-sight direction, and interpolating the historical deformation data using a spline curve; Step 2, collecting the surface GNSS three-dimensional deformation data within the same period, and projecting it onto the satellite line-of-sight direction, and using the projected GNSS deformation observation values to adjust the interpolated historical deformation data; Step 3, using the Kalman filter model as the basic model, dynamically updating the state transition matrix A, the process noise covariance matrix Q, and the observation noise covariance matrix R through a spatio-temporal graph convolutional network model, so as to form a data-model hybrid-driven time series prediction model; Step 4, using the adjusted historical deformation data as the training sample, using the hyperparameter search method to estimate the optimal hyperparameters of the time series prediction model to obtain the target prediction model; Step 5, combining the subsequent collected GNSS deformation data, inputting the time to be predicted into the target prediction model to obtain the surface deformation data at the future moment.

[0022] The beneficial effect of the embodiment of the present invention is that through the solution of the present invention, in the case where there are only GNSS high-time-resolution data and InSAR high-space-resolution data, and high spatio-temporal resolution deformation data needs to be obtained during the monitoring process, the STGCN_KF method can be used to fuse the two existing data to obtain the high spatio-temporal resolution deformation data required during the monitoring process, providing an accurate basis for surface deformation monitoring. Brief Description of the Drawings

[0023] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0024] Figure 1 It is a schematic flowchart of a high spatio-temporal resolution prediction method for surface deformation driven by data-model hybrid provided by an embodiment of the present invention;

[0025] Figure 2 It is an internal structure diagram of an STGCN_KF model provided by an embodiment of the present invention;

[0026] Figure 3 It is a schematic diagram of data division of a data sliding window provided by an embodiment of the present invention. Detailed Embodiments

[0027] The embodiments of the present invention will be described in detail below with reference to the drawings.

[0028] The following illustrates the embodiments of the present invention through specific examples. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Obviously, the described embodiments are only some embodiments of the present invention, not all of them. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, without conflict, the following embodiments and the features in the embodiments can be combined with each other. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts belong to the scope of protection of the present invention.

[0029] It should be noted that the following describes various aspects of the embodiments within the scope of the appended claims. It should be obvious that the aspects described herein can be embodied in a wide variety of forms, and any specific structure and / or function described herein is illustrative only. Based on the present invention, those skilled in the art should understand that one aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number of aspects described herein can be used to implement the device and / or practice the method. In addition, this device can be implemented and this method can be practiced using other structures and / or functions in addition to one or more of the aspects described herein.

[0030] It should also be noted that the illustrations provided in the following embodiments only schematically illustrate the basic concept of the present invention. The diagrams only show the components related to the present invention, rather than being drawn according to the number, shape, and size of the components in actual implementation. The type, quantity, and proportion of each component in actual implementation can be arbitrarily changed, and the component layout type may also be more complex.

[0031] In addition, in the following description, specific details are provided to facilitate a thorough understanding of the examples. However, those skilled in the art will understand that the described aspects can be practiced without these specific details.

[0032] An embodiment of the present invention provides a data-model hybrid-driven high spatio-temporal resolution prediction method for surface deformation, which can be applied to the surface settlement prediction process in a geomonitoring scenario.

[0033] See Figure 1 , which is a schematic flowchart of a data-model hybrid-driven high spatio-temporal resolution prediction method for surface deformation provided by an embodiment of the present invention. As Figure 1 shown, the method mainly includes the following steps:

[0034] Step 1, obtain multi-temporal repeat-pass synthetic aperture radar image data within the monitoring period of the target area, use the time series InSAR technology to process the radar image data to obtain the historical deformation data in the satellite line-of-sight direction of the surface, and interpolate the historical deformation data using a spline curve;

[0035] In specific implementation, currently, the means for obtaining surface deformation mainly include GNSS technology and InSAR technology. GNSS technology can obtain high-time-resolution data but cannot obtain high-spatial-resolution data. InSAR technology can provide high-spatial-resolution data but cannot provide high-time-resolution data. High spatio-temporal resolution deformation data of the surface cannot be obtained by either of the above two methods. Obtain multi-temporal synthetic aperture radar image data within the specified time period of the target area. The radar observation data used in this case is Sentinel-1A ascending orbit image data (a synthetic aperture radar observation satellite designed and launched by the European Space Agency, which can obtain image data with a short revisit period, good interference quality, and accurate orbital attitude). Use the small baseline subset (SBAS) method to process the image data to obtain the InSAR deformation data in the LOS direction; perform non-linear interpolation on the InSAR data, and the interpolation method is spline curve fitting, and interpolate it into time data with a time interval of 1 day to increase the number of InSAR historical samples.

[0036] Step 2, collect the surface GNSS three-dimensional deformation data within the same period, project it onto the satellite line-of-sight direction, and adjust the interpolated historical deformation data with the projected GNSS deformation observation values;

[0037] In specific implementation, first, obtain the ground GNSS observation data in the same coverage time period, obtain the deformation data results in three directions of GNSS, and use the moving average window method to denoise the numerical values in three directions of GNSS; second, project the GNSS data onto the line-of-sight direction based on the three-dimensional deformation to line-of-sight direction equation, compare the GNSS data in the line-of-sight direction with the InSAR data, and judge whether the deformation trends of the GNSS data and the InSAR data are the same, so as to judge whether the InSAR result is accurate; third, with the projected line-of-sight direction GNSS data as a constraint, adjust the interpolated InSAR data by indirect adjustment to improve the accuracy of the InSAR historical samples.

[0038] The three-dimensional to line-of-sight direction equation is as follows:

[0039] Zj = Zncos(α - 270°) + Zecos(360° - α) = -Znsin(α) + Zecos(α)

[0040] Z l = Zucos(θ) - Zjsin(θ)

[0041] Zl = Zucos(θ) + Znsin(α)sin(θ) - Zecos(α)sin(θ)

[0042] where θ is the radar incident angle, α is the radar course incident angle, Z n is the deformation in the north direction, Z e is the deformation in the east direction, Z u is the deformation in the vertical direction.

[0043] Step 3: Taking the Kalman filter model as the basic model, dynamically update the state transition matrix A, the process noise covariance matrix Q, and the observation noise covariance matrix R in it through the spatio-temporal graph convolutional network model, so as to form a data-model hybrid-driven time series prediction model;

[0044] In specific implementation, the Kalman filter (KF) is an efficient recursive algorithm first proposed by Rudolph in 1960 for estimating the state of a dynamic system from a series of observation data in a noisy environment. The basic equations of KF are as follows:

[0045] y t = Ay t-1 + w, w ~ N(0, Q)

[0046] Z t = Hy t + v, v ~ N(0, R)

[0047] The Kalman filter is divided into two main stages, namely the prediction stage and the update stage. The equations in the prediction stage first use the state estimate value at the previous moment Predict the state value at the current moment jointly with the state transition equation A Utilize the state covariance matrix at the previous moment And the process noise covariance matrix Q to estimate the current state covariance matrix The equations in the prediction stage are as follows:

[0048]

[0049]

[0050] The core equation in the KF update stage utilizes the observation value Z at the current moment t To optimize and correct the predicted value in the prediction stage Obtain a more accurate state estimate In this stage, first calculate the current Kalman filter gain value K according to the observation equation H, the estimated current covariance matrix And the observation noise covariance matrix R t , and then update to obtain the accurate Kalman filter output value And the covariance matrix The equations in the KF update stage are as follows:

[0051]

[0052] In the traditional Kalman filter formula, the A matrix in the prediction equation usually describes linear state transition. The process noise covariance matrix Q and the observation noise covariance matrix R are often regarded as known and remain unchanged during the iterative update process. The above parameters reflect the accuracy of state estimation and the uncertainty of the system. However, thanks to the dynamic spatio-temporal modeling ability of the spatio-temporal graph convolutional network (STGCN), it can effectively model the non-linear dynamic spatio-temporal parameters in the Kalman filter. Therefore, based on the traditional Kalman filter, the STGCN_KF model formed by combining the STGCN model can accurately capture the non-linear characteristics of the dynamic Kalman filter system and provide flexible and accurate state estimation.

[0053] y t =f(y t-1 ) + w, w ~ N(0, Q)

[0054] Z t =y t + v, v ~ N(0, R)

[0055] The above formula is the STGCN_KF equation, which is used to depict the spatio-temporal evolution law of the surface deformation state. Figure 2This is the specific implementation process of STGCN_KF. Compared with the traditional Kalman filter model, in the STGCN-KF model, the state transition matrix A, the process noise covariance matrix Q, and the observation noise covariance matrix R are dynamically modeled using the STGCN spatio-temporal model, and the observation matrix H is set as the identity matrix I to better adapt to the dynamic characteristics of surface deformation.

[0056]

[0057] f = STGNN F

[0058]

[0059] The above formula is the state prediction equation of the STGCN_KF model. Given that the state transition equation f is highly nonlinear and its direct mathematical expression is difficult to determine, a corresponding STGCN F model f is constructed to replace the linear state transition matrix A, and the output result of the Kalman filter state at the previous moment is used as the input to predict the state prediction value at the current moment The first-order derivative term f′ in the second formula, that is, the Jacobian matrix equation, is used to estimate the current state covariance matrix reflecting the uncertainty change in the state transition process. This parameter is dynamically spatio-temporally varying in time series and needs to be calculated and updated in real time; in addition, the process noise covariance Q is also dynamically spatio-temporally varying, so a STGCN Q model is introduced to predict the real-time Q value, using the state prediction data at the current moment as the input data of the STGCN Q model, and performing exponentiation and diagonalization operations on the output result. Specifically, the exponentiation result is used as the diagonal elements of the diagonal matrix, and then this diagonal matrix is used as the process noise covariance Q.

[0060]

[0061] The above formula is the state update stage of the STGCN_KF model. In this stage, except that the state noise covariance R value is dynamically estimated and Z t is obtained from external observations, other parameters are automatically updated with the system. To obtain the dynamically changing observation noise covariance R, a STGCN R network is also introduced, using the observation data Z at the current moment t as the input data of the model, and performing exponentiation and diagonalization operations on the prediction result of the STGCN Q model in the same way as the output result processing method of the STGCN R model. By constructing STGCN Q and STGCN RThe network can learn the variation rules of the Q and R matrices under different observation conditions, achieve accurate prediction of Q and R, and improve the robustness of the dynamic Kalman filter model.

[0062] The real-time observation data Z in the state update stage t is indispensable. However, in the process of fusing GNSS data and InSAR data by the STGCN_KF network, the challenge is that GNSS technology can only monitor sparse settlement points, and the spatial resolution is much lower than that of InSAR technology, unable to meet the requirement of providing the existing observation data Z for all image pixel points. t To overcome this difficulty, a traditional indirect adjustment method is introduced to generate the observation data. Specifically, using the GNSS data at each moment as a conditional constraint, the state prediction value of the STGCN F model at the corresponding moment is adjusted to generate the observation data Z at each moment. t In this way, the adjusted observation data Z t can be accurately input into the Kalman filter network to realize the effective update of the state matrix and the covariance matrix .

[0063] Step 4: Using the adjusted historical deformation data as training samples, estimate the best hyperparameters of the time series prediction model by the hyperparameter search method to obtain the target prediction model.

[0064] In specific implementation, after the model is constructed, using the adjusted InSAR deformation interpolation data as training samples, carry out model data preprocessing, that is, training sample division and standardization. First, extract the time series data of each pixel point in the image, take the extracted image pixel points as sample data, and divide the data set into training set and test set data, where the training set: test set = 8:2; to significantly improve the convergence speed of the deep learning algorithm and optimize the training efficiency and performance of the model, perform pixel-by-pixel normalization processing on the InSAR deformation interpolation data in the time series. The normalization method is the maximum-minimum normalization, where Z max is the maximum value of the pixel point time series deformation, Z min is the minimum value of the pixel point time series deformation, Z original is the pixel deformation value, Z normal is the normalized deformation data, and the data range will be compressed within [0,1]; use the sliding window to divide the training window, and use the sliding window method to divide the windows of GNSS data and the adjusted InSAR deformation interpolation data respectively. Let the sliding window length be L + 1, as Figure 3 shown, where the data t1,..., t L is used as the training input data, and the data t2,..., t L+1For true experimental verification data, that is, labeled data, this processing method can not only expand the scale of the data set, but also enable the model to learn more continuous and comprehensive time series features. The maximum-minimum normalization equation is as follows:

[0065]

[0066] To accelerate training and improve the generalization ability of the model, the Adam optimization algorithm is used to update the network weights. The Adam algorithm combines the advantages of RMSProp (Root Mean Square Propagation) and Momentum, and has characteristics such as adaptive learning rate, fast convergence speed, strong adaptability to sparse matrices, simple implementation and high computational efficiency. Secondly, to prevent the model from overfitting during training, the strategy of randomly discarding neurons is applied, and the L2 regularization and learning rate decay strategies are introduced. In addition, a custom model loss function is used to quantify the model error, where yi is the real data, is the prediction result of the STGCN_KF model, is STGCN F prediction result; finally, to search for the STGCN_KF model parameters, that is, to search for STGCN F 、STGCN Q 、STGCN R model parameters and learning rate, number of iterations, etc., a hyperparameter search method is used to search for the optimal parameter configuration. This method can not only optimize the model performance, but also improve the generalization ability and stability of the model. In this example, after hyperparameter search, the following parameter combination is adopted: the model batch size is 512, and the number of neurons in STGCN F 、STGCN Q 、STGCN R models are all 32, the number of network layers is all 1 layer, the random neuron discard probability is 0.1, the learning rate is set to 0.001, and the number of iterations is 200 times. A fully connected layer is set after each STGCN model, and the weights are initialized by the Xavier method, and the bias term is initialized to 0. The custom loss function is as follows:

[0067]

[0068] Input the training data into the STGCN_KF model, and use the above parameters to complete the model training to obtain the trained target prediction model.

[0069] Step 5, combine the subsequent GNSS deformation data collected, input the time to be predicted into the target prediction model, and obtain the deformation data of the ground at future moments.

[0070] In specific implementation, the InSAR deformation data and GNSS deformation data are obtained in real time and input pixel by pixel into the STGCN_KF model to recursively obtain the deformation data of the ground surface at future moments. Through these prediction results, the deformation trend of the ground surface in the future period can be understood more deeply, providing an important basis for subsequent monitoring and engineering decision-making.

[0071] To reflect the advantages of the method provided by this application, the STGCN_KF method proposed in this application is compared with several other ground deformation interpolation methods in terms of accuracy. RMSE and MAE are used as the accuracy evaluation results, as shown in Table 1.

[0072] It can be seen from Table 1 that the method provided by this application can greatly improve the accuracy of terrain estimation compared with the existing methods. Its RMSE (Root Mean Squared Error) is 1.77 mm and the MAE value is 1.15 mm. The accuracy is much higher than several other prediction methods. Therefore, the STGCN_KF model provided by this application can be used to fuse GNSS data and InSAR data as a method to obtain high spatio-temporal resolution data of the ground surface.

[0073] Table 1

[0074]

[0075]

[0076] A high spatio-temporal resolution prediction method for surface deformation driven by data-model hybrid. First, obtain multi-temporal repeat-pass synthetic aperture radar image data within the monitoring period of the target area, process the radar image data using time-series InSAR technology to obtain historical surface line-of-sight deformation data, and use spline curve interpolation for InSAR deformation data to increase the number of InSAR historical samples. Second, collect surface GNSS three-dimensional deformation data within the same period, project the three-dimensional deformation onto the satellite line-of-sight direction, and use the projected GNSS deformation observations to adjust the interpolated InSAR deformation data to improve the accuracy of InSAR historical samples. Third, use the Kalman filter (KF) as the basic model, where the state transition matrix A, the process noise covariance matrix Q, and the observation noise covariance matrix R are regarded as variable and are dynamically updated through the spatio-temporal graph convolutional network (STGCN) model to form a data-model hybrid-driven time-series prediction model. Fourth, use the adjusted InSAR deformation interpolation data as training samples, and use the hyperparameter search method to estimate the optimal hyperparameters of the STGCN_KF model to obtain the target prediction model; Fifth, combine the subsequent collected GNSS deformation data, input the time to be predicted into the target prediction model, and obtain the surface deformation data at future times. This method can use the STGCN_KF method to fuse two existing data to obtain the high spatio-temporal resolution deformation data required during the monitoring process when there are only GNSS high-time-resolution data and InSAR high-space-resolution data, and high spatio-temporal resolution deformation data is required during the monitoring process, providing an accurate basis for surface deformation monitoring.

[0077] It should be understood that each part of the present invention can be implemented by hardware, software, firmware, or a combination thereof.

[0078] The above is only a specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by those skilled in the art within the technical scope disclosed by the present invention should be covered by the protection scope of the present invention. Therefore, the protection scope of the present invention should be subject to the protection scope of the claims.

Claims

1. A data-model hybrid driven high spatiotemporal resolution deformation data prediction method, characterized in that: include: Step 1: obtain multiple periods of heavy-track synthetic aperture radar image data in the target area during the monitoring period, use time-series InSAR technology to process the radar image data to obtain historical deformation data of the surface line of sight, and use spline curves to interpolate the historical deformation data; Step 2: Collect GNSS three-dimensional deformation data of the surface in the same period, project it to the satellite line of sight, and adjust the projected GNSS deformation observations to the interpolated historical deformation data; Step 3: Taking the Kalman filter model as the basic model, the state transfer matrix A, the process noise covariance matrix Q and the observation noise covariance matrix R are dynamically updated through the spatiotemporal graph convolutional network model to form a data-model hybrid driven time series prediction model; Step 4, using the adjusted historical deformation data as training samples, using the hyperparameter search method to estimate the optimal hyperparameters of the time series prediction model, and obtaining the target prediction model; Step 5: Combined with the GNSS deformation data collected subsequently, the time to be predicted is input into the target prediction model to obtain the deformation data of the surface at the future moment.

2. The method according to claim 1, characterized in that The time series prediction model includes a state prediction equation and a state update equation.

3. The method according to claim 2, characterized in that The state prediction equation is: f=STGNN A Where f represents the state transfer equation, represents the state prediction value at the current time t, represents the Kalman filter state output result at the previous moment, f′ represents the Jacobian matrix equation, which is the first-order derivative result of f and is used to estimate the current state covariance matrix The STGCN model is used to model f, Q t represents the real-time value of the process noise covariance matrix Q, STGCN Q Indicates the predicted real-time Q t prediction model.

4. The method according to claim 3, characterized in that The state update equation is Among them, Z t represents the observed data at the current moment, K t represents the Kalman filter gain at the current moment, I represents the unit weight matrix, R t represents the real-time value of the observation noise covariance matrix R, STGCN R Represents the predicted real-time R t prediction model.

Images