Method and device for determining position of magnetically levitated train in braking process based on digital twinning

Abstract

The invention discloses a position determination method and device in the braking process of a magnetically levitated train based on digital twinning, relates to the technical field of digital twinning of rail transit, and mainly aims to solve the problem that accurate parking positions cannot be guaranteed based on manual observation and operation in the parking process. The method mainly comprises the steps that a linear and nonlinear combined train position forecasting model is constructed, and the output of an unknown nonlinear dynamic system is obtained through forecasting based on a trained online deep learning forecasting model; in response to the train parking preparation instruction, acquiring real-time operation data of the target train; inputting real-time operation data into the linear model, and calculating a linear forecast value; performing prediction processing on the real-time operation data based on the trained online deep learning prediction model to obtain a nonlinear prediction value; and adding the linear forecast value and the nonlinear forecast value to obtain a position forecast value of the target train, and outputting the position forecast value to an interaction terminal. The method is mainly used for determining the train position in the braking process.

Description

Technical Field

[0001] This invention relates to the field of digital twin technology for rail transit, and in particular to a method and apparatus for determining the position during the braking process of a maglev train based on digital twins. Background Technology

[0002] When a maglev train is running at high speed, the onboard speed measurement system (such as a speed measuring device based on eddy current sensors) can provide relatively accurate speed feedback. However, when the train enters the final stage of braking and the speed decreases significantly, this speed measurement method relying on eddy current sensors faces a fundamental challenge. For eddy current sensors, when the train enters the braking phase and the speed decreases, the amount of eddy current generated decreases, the signal is weak, and the signal-to-noise ratio is low, resulting in a large error in the eddy current sensor at low speeds. At this time, if the automatic driving system (ATO) uses this inaccurate speed signal to calculate the train's position and control braking, a deviation between the train's position and its actual position is inevitable.

[0003] Given the significant errors in automatic driving systems during the stopping phase, current stopping processes primarily rely on manual intervention. The driver observes the distance to reference points on the platform and manually operates the gear shift lever to brake the train and control its speed, ensuring accurate stopping position. However, this method of manually observing reference points has high uncertainty and unreliability, making it impossible to guarantee accurate stopping position. Summary of the Invention

[0004] In view of this, the present invention provides a method and device for determining the position during the braking process of a maglev train based on digital twins. The main purpose is to solve the problem that the accuracy of the stopping position cannot be guaranteed by relying on manual observation and operation during the stopping process.

[0005] According to one aspect of the present invention, a method for determining the position during the braking process of a maglev train based on digital twins is provided, comprising: A train position prediction model combining linear and nonlinear methods is constructed, wherein the train position prediction model includes an offline identified linear model and an unknown nonlinear dynamic system, and the output of the unknown nonlinear dynamic system is predicted based on a trained online deep learning prediction model. In response to a train stopping preparation command, acquire real-time operating data of the target train; The real-time running data is input into the linear model to calculate the linear forecast value; and the real-time running data is then processed for prediction based on the trained online deep learning forecast model to obtain the nonlinear forecast value. The linear forecast value and the nonlinear forecast value are added together to obtain the target train's position forecast value, which is then output to the interactive terminal to assist in stopping at the target parking point.

[0006] Furthermore, the construction of the train position prediction model combining linear and nonlinear methods includes: A train position control closed-loop system equation is constructed, and based on the equation, the train position dynamic system is decomposed into a linear part and a nonlinear part. The nonlinear part includes multiple nonlinear fluctuation terms characterizing uncertainties in the train control system. Based on historical train operation data, the parameter vector of the linear part is identified using a least squares algorithm to obtain a linear model. An unknown nonlinear dynamic system is constructed based on the identification error of the parameter vector and the unknown nonlinear terms. A deep learning model for predicting the unknown nonlinear dynamic system is constructed, and the deep learning model is trained based on historical train position time-series data to obtain a trained online deep learning prediction model. A train position prediction model combining linear and nonlinear elements is constructed based on the linear model, the unknown nonlinear dynamic system, and the trained online deep learning prediction model.

[0007] Furthermore, the process of constructing the equations for the train position control closed-loop system includes: Construct a position dynamic model that includes a linear model and unknown nonlinear terms, wherein the linear model includes control level positions and driving permissions; A control law is constructed that includes a dynamic error proportional term and a feedforward compensation term. The dynamic error proportional term is dynamically adjusted based on the deviation between the real-time position of the train and the target position. The feedforward compensation term is calculated based on multiple train operation state variables, wherein the train operation state variables include at least one of the following: level, driving permission, most stringent target point position, most stringent target point speed, train speed, reference speed value, temporary speed limit, recommended speed, emergency braking speed, maximum speed limit, alarm speed, and gradient. The output of the control law is applied to the position dynamic model to obtain the closed-loop system equations for train position control.

[0008] Furthermore, the process of identifying the parameter vector of the linear component based on historical train operation data using a least squares algorithm to obtain a linear model includes: The input matrix and output matrix of train operation control are constructed based on historical operation data at continuous time points. The input matrix includes the train position, target stopping point, control level position and train operation state variables at different historical time points. Based on the input matrix and the output matrix, construct a parameter identification equation with the parameter vector of the linear part as the identification object; The parameter identification equation is solved using the least squares algorithm to obtain an estimate of the parameter vector, and the parameter vector of the linear part is updated with the estimate to obtain a linear model.

[0009] Furthermore, the historical train position time series data includes historical train operation data and historical position measurement values ​​at consecutive time points; The construction and training process of the online deep learning prediction model includes: A deep learning model containing a multi-layer recurrent neural network is constructed, wherein the multi-layer recurrent neural network adopts an architecture including an input gate, a forget gate, an output gate, and a state memory unit, for extracting nonlinear features of train position time series data; Calculate the historical linear output of the linear model corresponding to the historical train operation data at each moment, as well as the difference between the historical position measurement value and the historical linear output, and use the difference as the sample label of the unknown nonlinear dynamic system to construct the training sample set of the initial online deep learning prediction model; The deep learning model is trained using the training sample set to learn the changing patterns of the unknown nonlinear dynamic system, resulting in a well-trained online deep learning prediction model.

[0010] Furthermore, the method is implemented based on an edge-cloud-edge architecture. The acquisition of real-time operation data of the target train includes: collecting real-time operation data based on edge sensors deployed on the target train and uploading it to the real-time operation database on the cloud side; and downloading the real-time operation data of the target train from the real-time operation database on the cloud side. Before inputting the real-time running data into the linear model, the method further includes: Linear models are built on the cloud side, and online deep learning prediction models are trained online. The edge loads the linear model and the trained online deep learning prediction model from the cloud to calculate the location prediction value on the edge based on real-time running data using the linear model and the trained online deep learning prediction model. The online deep learning prediction model is trained online on the cloud side based on network parameter self-correction mechanism and network structure self-correction mechanism.

[0011] Furthermore, the method also includes: On the cloud side, network parameter self-calibration and network structure self-calibration are performed on the online deep learning prediction model based on real-time updated train operation data to obtain a network structure self-calibration model and a network parameter self-calibration model. If the first error of the network parameter self-calibration model is found to be greater than the first preset error threshold and the second error of the network structure self-calibration model is less than the first preset error threshold, the network structure of the network parameter self-calibration model is corrected based on the network structure of the network structure self-calibration model, so as to continue online training based on the network parameter self-calibration model with the updated network structure. If the third error of the online deep learning prediction model on the edge is greater than the second preset error threshold, and the first error of the network parameter self-calibration model is less than the second preset error threshold, the weights and biases of the online deep learning prediction model on the edge are corrected based on the weights and biases of the network parameter self-calibration model to obtain an updated online deep learning prediction model.

[0012] According to another aspect of the present invention, a position determination device for the braking process of a magnetic levitation train based on digital twins is provided, comprising: A construction module is used to construct a train position prediction model that combines linear and nonlinear features. The train position prediction model includes an offline identified linear model and an unknown nonlinear dynamic system. The output of the unknown nonlinear dynamic system is predicted based on a trained online deep learning prediction model. The acquisition module is used to acquire real-time operating data of the target train in response to the train stopping preparation command; The calculation module is used to input the real-time running data into the linear model to calculate the linear forecast value; and to perform prediction processing on the real-time running data based on the trained online deep learning forecast model to obtain the nonlinear forecast value. The output module is used to add the linear forecast value and the nonlinear forecast value to obtain the target train's position forecast value, and output the position forecast value to the interactive terminal to assist in stopping at the target parking point.

[0013] Furthermore, the building module includes: The first construction unit is used to construct the train position control closed-loop system equations and decompose the train position dynamic system into linear and nonlinear parts based on the train position control closed-loop system equations, wherein the nonlinear part includes multiple nonlinear fluctuation terms used to characterize the uncertainties in the train control system; the identification unit is used to identify the parameter vector of the linear part using the least squares algorithm based on historical train operation data to obtain a linear model; the second construction unit is used to construct an unknown nonlinear dynamic system based on the identification error of the parameter vectors and the unknown nonlinear terms; the training unit is used to construct a deep learning model for predicting the unknown nonlinear dynamic system and train the deep learning model based on historical train position time series data to obtain a trained online deep learning prediction model; The third building unit is used to construct a train position prediction model combining linear and nonlinear elements based on the linear model, the unknown nonlinear dynamic system, and the trained online deep learning prediction model.

[0014] Furthermore, in a specific application scenario, the first construction unit is specifically used to construct a position dynamic model that includes a linear model and unknown nonlinear terms, wherein the linear model includes control level positions and driving permissions; A control law is constructed that includes a dynamic error proportional term and a feedforward compensation term. The dynamic error proportional term is dynamically adjusted based on the deviation between the real-time position of the train and the target position. The feedforward compensation term is calculated based on multiple train operation state variables, wherein the train operation state variables include at least one of the following: level, driving permission, most stringent target point position, most stringent target point speed, train speed, reference speed value, temporary speed limit, recommended speed, emergency braking speed, maximum speed limit, alarm speed, and gradient. The output of the control law is applied to the position dynamic model to obtain the closed-loop system equations for train position control.

[0015] Furthermore, in specific application scenarios, the identification unit is specifically used to construct the input matrix and output matrix of train operation control based on historical operation data at continuous time points. The input matrix includes the train position, target stopping point, control level position, and train operation state variables at different historical time points. Based on the input matrix and the output matrix, construct a parameter identification equation with the parameter vector of the linear part as the identification object; The parameter identification equation is solved using the least squares algorithm to obtain an estimate of the parameter vector, and the parameter vector of the linear part is updated with the estimate to obtain a linear model.

[0016] Furthermore, in specific application scenarios, the training unit is specifically used to construct a deep learning model containing a multi-layer recurrent neural network, wherein the multi-layer recurrent neural network adopts an architecture including an input gate, a forget gate, an output gate, and a state memory unit, and is used to extract nonlinear features of train position time series data. Calculate the historical linear output of the linear model corresponding to the historical train operation data at each moment, as well as the difference between the historical position measurement value and the historical linear output, and use the difference as the sample label of the unknown nonlinear dynamic system to construct the training sample set of the initial online deep learning prediction model; The deep learning model is trained using the training sample set to learn the changing patterns of the unknown nonlinear dynamic system, resulting in a well-trained online deep learning prediction model.

[0017] Furthermore, the device also includes: The end-side acquisition module is used to collect real-time operation data based on the end-side sensors deployed on the target train and upload it to the real-time operation database on the cloud side. The side-side data acquisition module is used to download the real-time operation data of the target train from the cloud-side real-time operating database; The cloud-side construction module is used to build linear models on the cloud side and train online deep learning prediction models. The edge computing module is used to load the linear model and the trained online deep learning prediction model from the cloud side, so as to calculate the location prediction value on the edge based on real-time running data, using the linear model and the trained online deep learning prediction model. The online deep learning prediction model is trained online on the cloud side based on network parameter self-correction mechanism and network structure self-correction mechanism.

[0018] Furthermore, the device also includes: The self-calibration module is used on the cloud side to perform network parameter self-calibration and network structure self-calibration on the online deep learning prediction model based on real-time updated train operation data, so as to obtain a network structure self-calibration model and a network parameter self-calibration model. The cloud-side network structure self-calibration module is used to correct the network structure of the network parameter self-calibration model based on the network structure of the network structure self-calibration model when the first error of the network parameter self-calibration model is greater than the first preset error threshold and the second error of the network structure self-calibration model is less than the first preset error threshold, so as to continue online training based on the network parameter self-calibration model with the updated network structure. The cloud-side network parameter self-calibration module is used to correct the weights and biases of the edge-side online deep learning prediction model based on the weights and biases of the network parameter self-calibration model when the third error of the edge-side online deep learning prediction model is greater than the second preset error threshold, and the first error of the network parameter self-calibration model is less than the second preset error threshold, thereby obtaining an updated online deep learning prediction model.

[0019] According to another aspect of the present invention, a storage medium is provided, wherein at least one executable instruction is stored therein, the executable instruction causing a processor to perform an operation corresponding to the above-described position determination method for the braking process of a magnetic levitation train based on digital twins.

[0020] According to another aspect of the present invention, a terminal is provided, comprising: a processor, a memory, a communication interface, and a communication bus, wherein the processor, the memory, and the communication interface communicate with each other through the communication bus; The memory is used to store at least one executable instruction, which causes the processor to perform the operation corresponding to the above-described position determination method for the braking process of a maglev train based on digital twins.

[0021] By employing the above-described technical solutions, the technical solutions provided by the embodiments of the present invention have at least the following advantages: This invention provides a method and apparatus for determining the position during the braking process of a maglev train based on digital twins. The embodiments of this invention construct a train position prediction model combining linear and nonlinear methods. This model includes an offline-identified linear model and an unknown nonlinear dynamic system. The output of the unknown nonlinear dynamic system is predicted based on a trained online deep learning prediction model. In response to a train stopping preparation command, real-time operating data of the target train is acquired. The real-time operating data is input into the linear model to calculate a linear prediction value. The real-time operating data is then processed using the trained online deep learning prediction model to obtain a nonlinear prediction value. The linear and nonlinear prediction values ​​are added to obtain the predicted position value of the target train, which is then output to an interactive terminal to assist in stopping at the target stopping point. By constructing a combined prediction architecture of a linear model and a nonlinear deep learning model, the system dynamics are described more completely, thereby significantly reducing prediction errors. Simultaneously, it provides more advanced and reliable state perception information for the train's automatic driving system, thus greatly improving the success rate of accurate one-time positioning and stopping.

[0022] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention, it can be implemented according to the contents of the specification. Furthermore, in order to make the above and other objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0023] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 A flowchart of a method for determining the position during braking of a magnetic levitation train based on digital twins, provided by an embodiment of the present invention, is shown. Figure 2 A flowchart of another method for determining the position during the braking process of a magnetic levitation train based on digital twins, provided by an embodiment of the present invention, is shown. Figure 3 This diagram illustrates the structure of an edge-cloud collaborative parking location determination system according to an embodiment of the present invention. Figure 4 A flowchart of a network structure search provided by an embodiment of the present invention is shown; Figure 5 This invention provides a curve comparing the actual and predicted train position values ​​according to an embodiment of the invention. Figure 6A block diagram of a position determination device for the braking process of a magnetic levitation train based on digital twins, provided by an embodiment of the present invention, is shown. Figure 7 A schematic diagram of the structure of a terminal provided in an embodiment of the present invention is shown. Detailed Implementation

[0024] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0025] To address the issue that existing maglev train braking processes, which rely on manual observation and operation, cannot guarantee accurate stopping positions, this invention provides a digital twin-based method for determining the stopping position of a maglev train during braking. Figure 1 As shown, the method includes: 101. Construct a train position prediction model combining linear and nonlinear methods.

[0026] In this embodiment of the invention, to accurately describe the complex dynamics of train operation, a combined approach of linear sub-models and nonlinear systems is used to construct a train position prediction model. Specifically, the train position prediction model includes an offline-identified linear model, an unknown nonlinear dynamic system, and a trained online deep learning prediction model for predicting the output of the unknown nonlinear dynamic system. The linear model captures the main part of the train position dynamics that can be described by linear theory, and its parameters remain fixed during the prediction phase, ensuring the model's basic stability and computational efficiency. The online deep learning prediction model is used to model and predict nonlinear fluctuations caused by unknown disturbances, unmodeled dynamics, and changes in complex operating conditions. Combining the two retains the interpretability and robustness of the linear model while enhancing the ability to characterize complex, time-varying systems through the nonlinear model, thus forming a highly adaptable composite prediction model.

[0027] 102. In response to the train stopping preparation command, obtain the real-time operation data of the target train.

[0028] In this embodiment of the invention, the train position prediction process is triggered by a train stopping preparation command. This command can be issued by the Automatic Train Operation (ATO) system or by the dispatch center when the train enters the station stopping phase. Real-time operational data includes control level positions, as well as train operation status variables such as driving permission, maximum target point position, maximum target point speed, train speed, reference speed value, temporary speed limit, recommended speed, emergency braking speed, maximum speed limit, alarm speed, track section, and gradient. Real-time operational data can be collected in real time through onboard sensors, trackside equipment, and the control center. 103. Input the real-time running data into the linear model to calculate the linear forecast value; and perform prediction processing on the real-time running data based on the trained online deep learning forecast model to obtain the nonlinear forecast value.

[0029] In this embodiment of the invention, the predicted position of the target train employs a strategy of parallel processing and feature separation. Real-time operational data vectors are organized according to a predetermined format and input into the model. Based on its offline identified parameter estimates, the model performs a forward computation and outputs a linear prediction of the train's position at a future time, representing the position prediction under the assumption of ideal linear dynamics. Furthermore, an online deep learning prediction model is used to extract train operational state features from the input data and predict the dynamic deviations caused by uncertainties, i.e., nonlinear prediction values.

[0030] 104. Add the linear forecast value and the nonlinear forecast value to obtain the target train's position forecast value, and output the position forecast value to the interactive terminal to assist in stopping at the target parking point.

[0031] In this embodiment of the invention, the linear and nonlinear forecast values ​​calculated in the previous step are algebraically added together, and the sum is the target train's position forecast value, which integrates linear trends and nonlinear fluctuations. This forecast value is then sent to an interactive terminal in real time. The interactive terminal can be a human-machine interface in the train driver's cab, a monitoring screen for a ground dispatcher, etc. Through visualization, this forecast value can provide operators with advanced and accurate predictive information about the train's future position, thereby assisting them to adjust traction or braking strategies earlier and more smoothly, thus stopping the train more accurately at the target stopping point.

[0032] In one embodiment of the present invention, for further illustration and limitation, such as Figure 2 As shown, the step of constructing a train position prediction model combining linear and nonlinear methods includes: 201. Construct the train position control closed-loop system equations, and decompose the train position dynamic system into linear and nonlinear parts based on the train position control closed-loop system equations.

[0033] 202. Based on historical train operation data, the parameter vector of the linear part is identified by the least squares algorithm to obtain the linear model.

[0034] 203. Construct an unknown nonlinear dynamic system based on the identification error of the parameter vector and the unknown nonlinear term.

[0035] 204. Construct a deep learning model for predicting the unknown nonlinear dynamic system, and train the deep learning model based on historical train position time series data to obtain a trained online deep learning prediction model.

[0036] 205. Construct a train position prediction model combining linear and nonlinear elements based on the linear model, the unknown nonlinear dynamic system, and the trained online deep learning prediction model.

[0037] The nonlinear component includes multiple nonlinear fluctuation terms characterizing uncertainties in the train control system. In this embodiment, a closed-loop system equation for train position control is constructed from a system equation reflecting the closed-loop dynamic characteristics of the train under the action of the automatic driving controller. This equation typically contains a linear equation and a lumped unknown term. Through design, the equation is modified to use a linear time-invariant model (i.e., with respect to the shift operator). The part described by the polynomial is explicitly separated and defined as the linear part; while all remaining uncertainties, unmodeled dynamics, time-varying characteristics, and complex disturbances are uniformly classified into the nonlinear part. The nonlinear part here is specifically manifested as multiple nonlinear fluctuation terms, which together characterize the complex dynamics caused by uncertainties such as modeling errors, internal variable coupling of the control system, and external unknown disturbances.

[0038] Furthermore, based on a large amount of historical operational data, a system identification method is used to parameterize the linear component. Specifically, historical data (including train position, control inputs such as level position, and feedforward variables such as train operation permits) are organized into input-output data pairs to construct a least-squares estimation problem. By minimizing the sum of squared errors between the model output and the actual data, the optimal parameter vector in the linear model can be estimated. After identification, this parameter vector is fixed, forming a linear model. Since the identification result of the linear sub-model is based on the estimated parameter values ​​and has a certain bias, the identification error of the linear sub-model (i.e., the difference between the actual output and the estimated output of the linear model) is combined with the unknown nonlinear terms in the original equation to jointly define a new synthetic system, namely, the unknown nonlinear dynamic system.

[0039] To predict the aforementioned unknown nonlinear dynamic system, this application introduces a deep learning model. Considering the temporal dependence and nonlinear characteristics of this dynamic system, a recurrent neural network architecture such as a Long Short-Term Memory network or a gated recurrent unit is preferred for constructing the prediction model. First, a series of "true values" of the unknown nonlinear dynamic system are calculated as training labels according to the definition in the previous step. The historical input data (including running state variables) corresponding to these labels constitute the training samples. The deep learning model is trained using these training samples, enabling it to learn the complex mapping relationship from the input time series to the unknown nonlinear dynamic output. The trained model constitutes the online deep learning prediction model. Finally, the two sub-models are integrated in parallel and summed. The combined model fully leverages the advantages of the linear model's clear structure and good extrapolation, as well as the deep learning model's ability to approximate complex nonlinear functions, thereby achieving high-precision and robust prediction of the stopping position of the maglev train.

[0040] In one embodiment of the present invention, for further explanation and limitation, the process of constructing the train position control closed-loop system equations includes: Construct a positional dynamic model that includes a linear model and unknown nonlinear terms; Construct a control law that includes a dynamic error proportional term and a feedforward compensation term; The output of the control law is applied to the position dynamic model to obtain the closed-loop system equations for train position control.

[0041] In this embodiment of the invention, the operation control process of a maglev train is analyzed to determine the variables required for establishing a dynamic position model of the maglev train, namely, the digital twin modeling variables for the maglev train's parking process, including train position, target parking point, level, and train operation status variables. The train operation status variables specifically include at least one of the following: driving permission, most stringent target point position, most stringent target point speed, train speed, reference speed value, temporary speed limit, recommended speed, emergency braking speed, maximum speed limit, alarm speed, and gradient.

[0042] Under the control of the Automatic Train Operation (ATO) system, the maglev train's position... At their target parking spot The surrounding fluctuations can be described using a linear model and unknown nonlinear terms to represent the position dynamics of the maglev train. The position dynamics model is expressed as: (1); in, This indicates the class at time k. This indicates the driving permission at time k. This represents the unknown nonlinear term at time k. Represents a linear model, where, , and For about polynomial, , , , , , For parameters.

[0043] Rule-based reasoning-based ATO (Automatic Train Control) systems employ variable proportional control with feedforward, meaning the control is based on the train's position. With the target parking spot The error changes the proportional controller parameters, and according to ~ Variables are fed forward for compensation, thereby controlling the position at the control level. This is used to control the train's position. The control law for this process is expressed as: (2); in, This represents the scaling factor at time k. This represents the feedforward compensation coefficient of the j-th variable at time k. This represents the value of the j-th train's operating state variable at time k. The dynamic error proportional term is dynamically adjusted based on the deviation between the train's real-time position and the target position, corresponding to the value in the formula. The feedforward compensation term is calculated based on multiple train operating state variables, corresponding to the formula in... Table 1 below shows the correspondence between variables and symbols in the digital twin modeling of the position during the stopping process of a maglev train. This represents the feedforward compensation coefficient of the train operation state variable corresponding to the corresponding symbol.

[0044] Table 1: Variables for Digital Twin Modeling of Position During Maglev Train Parking Process

[0045] Substituting equation (2) into equation (1), we obtain the equation for the train position control closed-loop system as follows: (3); By rearranging equation (3), we can obtain: (4); Where b is a parameter, using constants P and Replace the time-varying parameters in equation (4) respectively and Equation (4) can be expressed as: (5); Where c is a parameter, for The unknown nonlinear term of the fusion error, which consists of the modeling error and the fusion error, can be expressed as: (6).

[0046] Equation (5) can be used to express the train position closed-loop system equation as a linear model. and unknown nonlinear terms The sum of (7); Among them, the input variables of the linear model and parameter vector They are respectively: (8); (9); in, .

[0047] In one embodiment of the present invention, for further explanation and limitation, based on historical train operation data, the parameter vector of the linear component is identified using a least squares algorithm to obtain a linear model, including: The input and output matrices for train operation control are constructed based on historical operation data at continuous time points. Based on the input matrix and the output matrix, construct a parameter identification equation with the parameter vector of the linear part as the identification object; The parameter identification equation is solved using the least squares algorithm to obtain an estimate of the parameter vector, and the parameter vector of the linear part is updated with the estimate to obtain a linear model.

[0048] In this embodiment of the invention, the input matrix includes the train position, target stopping point, control level position, and train operation state variables at different historical times. The input and output matrices are constructed using M sets of actual operating data of the maglev train. The input matrix is ​​represented as follows: (10); The output matrix includes the train position at the next time step corresponding to the time step in the input matrix, and is represented as: (11). The parameter vector of the linear part is identified by the least squares method. The parameter identification equation constructed based on the above two matrices is expressed as: (12); The objective function of the least squares algorithm is: (13); in, The parameter vector can be obtained using the least squares algorithm. The estimated value , is represented as: (14).

[0049] In one embodiment of the present invention, for further explanation and limitation, the historical train position time series data includes historical train operation data and historical position measurement values ​​at continuous time points; The construction and training process of the online deep learning prediction model includes: A deep learning model containing a multi-layer recurrent neural network is constructed, wherein the multi-layer recurrent neural network adopts an architecture including an input gate, a forget gate, an output gate, and a state memory unit, for extracting nonlinear features of train position time series data; Calculate the historical linear output of the linear model corresponding to the historical train operation data at each moment, as well as the difference between the historical position measurement value and the historical linear output, and use the difference as the sample label of the unknown nonlinear dynamic system to construct the training sample set of the initial online deep learning prediction model; The deep learning model is trained using the training sample set to learn the changing patterns of the unknown nonlinear dynamic system, resulting in a well-trained online deep learning prediction model.

[0050] In this embodiment of the invention, after obtaining the estimated value of the parameter vector, the train position model can be represented as an identification model by combining the estimated value of formula (14) with the train position closed-loop system equation of formula (7). The sum of the unknown nonlinear dynamic system v(k) is expressed as: (15); in, It is caused by the unknown nonlinear term of the fusion error. and identification error The unknown nonlinear dynamic system is represented by the following table: (16); From equation (15), the initial train position prediction model can be obtained, expressed as: (17); where the unknown nonlinear dynamic system v(k+1) in this initial train position prediction model is unknown. Therefore, a deep learning model is needed to predict the unknown nonlinear dynamic system. The prediction result is the output of the unknown nonlinear dynamic system. Substituting it back into formula (17) will give the predicted train position value. This deep learning model is the online deep learning prediction model.

[0051] Taking LSTM (Long Short-Term Memory) as an example, the construction and training process of an online deep learning prediction model is explained: The training labels for the LSTM network are: (18); Using neural networks and input data Predicting an unknown nonlinear dynamic system can be expressed as: (19); in, The input data represents the nonlinear forecast value, where n represents the order of the unknown nonlinear dynamic system. for: (20).

[0052] Train operation data is input into the hidden layer, which is then processed by... l The network is constructed using a layer LSTM network, which will be the first layer LSTM network. ( Output of the 1st layer LSTM network Input to the l Layer LSTM network, the first l Output of the layered LSTM network This refers to the nonlinear features extracted by a nonlinear neural network. A fully connected layer consists of Dense layers, with the [number of layers] being the [number of layers]. l Output of the layered LSTM network As input, the Dense layer... The output of the LSTM network is obtained through calculation. This is the predicted value for an unknown nonlinear dynamic system, expressed as: (21); among which, Indicates the weights of the output layer. This indicates the bias of the output layer.

[0053] In one embodiment of the present invention, for further explanation and limitation, the method is implemented based on an edge-cloud-edge architecture, and the acquisition of real-time operation data of the target train includes: collecting real-time operation data based on edge sensors deployed on the target train and uploading it to the real-time operation database on the cloud side; and the edge side downloading the real-time operation data of the target train from the real-time operation database on the cloud side. Before inputting the real-time running data into the linear model, the method further includes: Linear models are built on the cloud side, and online deep learning prediction models are trained online. The edge loads the linear model and the trained online deep learning prediction model from the cloud to calculate the location prediction value on the edge based on real-time running data.

[0054] In this embodiment of the invention, the train position prediction process is completed through collaboration between the edge, cloud, and terminal. The edge-side data acquisition and transmission device transmits data generated during the operation of the maglev train to the cloud-side data server and artificial intelligence computing platform in real time. The cloud-side data server receives the maglev train operation data from the edge side in real time, processes and stores the data, and the artificial intelligence computing platform uses the data stored in real time on the data server to train the cloud-side online deep learning prediction model. The online deep learning prediction model is trained online on the cloud side based on network parameter self-correction and network structure self-correction mechanisms. The edge side receives the train position prediction model downloaded from the cloud side, i.e., the digital twin model of the operation and control process, and performs real-time prediction of the train position based on this model. Specifically, the edge-side online deep learning prediction model performs real-time prediction of the unknown nonlinear dynamic system without online training. The edge-side linear model reads the input data from the cloud-side database in real time to calculate the linear prediction value of the linear model, and adds it to the nonlinear prediction value of the unknown nonlinear dynamic system to obtain the predicted train position value.

[0055] In one embodiment of the present invention, for further explanation and limitation, the method further includes: On the cloud side, network parameter self-calibration and network structure self-calibration are performed on the online deep learning prediction model based on real-time updated train operation data to obtain a network structure self-calibration model and a network parameter self-calibration model. If the first error of the network parameter self-calibration model is found to be greater than the first preset error threshold and the second error of the network structure self-calibration model is less than the first preset error threshold, the network structure of the network parameter self-calibration model is corrected based on the network structure of the network structure self-calibration model, so as to continue online training based on the network parameter self-calibration model with the updated network structure. If the third error of the online deep learning prediction model on the edge is greater than the second preset error threshold, and the first error of the network parameter self-calibration model is less than the second preset error threshold, the weights and biases of the online deep learning prediction model on the edge are corrected based on the weights and biases of the network parameter self-calibration model to obtain an updated online deep learning prediction model.

[0056] In embodiments of the present invention, such as Figure 3 As shown, the cloud-based AI computing platform processes and stores the data collected from the edge, trains the network structure self-calibration model online based on the data updated in real time in the database, and trains the weights and biases of the network parameter self-calibration model online.

[0057] A network parameter self-calibration model is constructed using an LSTM network. Output of the network parameter self-calibration model Represented as: (twenty two).

[0058] During training, mean squared error (MSE) was used as the loss function, and the Adam (Adaptive Moment Estimation) optimization algorithm was employed to train the network weights and biases in real time. The loss function represents the weights and biases in an LSTM network. Represented as: (23); among which, These are the predicted values ​​from the cloud-side network parameter self-calibration model. B represents the training labels, and B represents the training batch size.

[0059] The gradient gt in the t-th iteration is expressed as: (twenty four); Exponential moving average of gradient Exponential moving average of gradient squared They are respectively: (25); among which, This represents the first-order moment attenuation coefficient. This represents the second-order moment attenuation coefficient.

[0060] Deviation correction for exponential moving average of gradient Deviation correction of the exponential moving average of the gradient squared They are respectively: (26); Network weights and bias parameters during the t-th iteration The update formula is: (27); in, For learning rate, The minimum value greater than 0 is used to ensure that the denominator is not 0.

[0061] Network weights and bias parameters during the t-th iteration The update formula is: (28); in, For learning rate, The minimum value greater than 0 is used to ensure that the denominator is not zero. When new data is obtained, it is used to form new training data and the above training process is repeated to train the network's weights and bias parameters online, so that the network learns new characteristics that appear in different working conditions, thereby improving the generalization ability of the cloud-side network parameter self-calibration model.

[0062] With the operation of maglev trains, the original network structure may not be able to guarantee high accuracy under the latest operating conditions. Therefore, it is necessary to establish a cloud-based network structure self-calibration model to search the network structure using real-time updated data. The search ranges for the number of LSTM neurons n, the number of nodes h, and the number of layers l are set to [n...]. min n max ]、[h min h max ] and [l min , l max The search step sizes are ns, hs, and ls, respectively. The network structure is a self-correcting model. Output for (29).

[0063] Furthermore, the root mean square error (RMSE) is used to evaluate the self-calibration model of the network structure. The accuracy is evaluated using the following formula: (30); where N is the amount of data.

[0064] Since grid search is time-consuming, a greedy algorithm is used to search the neural network structure. The search algorithm process is as follows: Figure 4 As shown. First, load the model training data and set the network structure search range. The number of neurons n in the LSTM network should not exceed the upper limit of the search range. n max In the case of step size n s Deep learning models with different numbers of neurons are trained sequentially and their values ​​are calculated. ,if Less than the minimum value RMSE min If the current network structure is saved, then proceed to the next round of searching for the number of neurons. Less than the minimum value RMSE minIf the search for the number of neurons n is not completed, the search proceeds directly to the next round of searching for the number of neurons h and the number of layers l, ultimately yielding the locally optimal network structure. After one round of network structure search is completed, the latest industrial data is used as new training data, and the above network structure search process is repeated to learn new characteristics that emerge in the industrial process. This allows the network structure to adaptively adjust with changes in operating conditions, ensuring that the prediction model still has a relatively optimal network structure under the latest operating conditions.

[0065] Based on the self-calibrating extrema of the network structure, the root mean square error of the network parameter self-calibration model is calculated in real time. When the cloud-side network parameter self-calibration model The error is greater than the first preset error threshold, and the cloud-side network structure self-correction model... If the error is less than the first preset error threshold, the network structure of the network structure self-calibration model is used to correct the network structure of the network parameter self-calibration model. The cloud side then continues online training using the updated network parameter self-calibration model. Based on the network parameter self-calibration mechanism, the accuracy of the edge-side online deep learning prediction model is monitored online, and the root mean square error of the online deep learning prediction model is calculated in real time. When the side-by-side online deep learning prediction model The error is greater than the second preset error threshold, and the cloud-side network parameter self-calibration model... When the error is less than the second preset error threshold, the weights and biases of the cloud-side network parameter self-calibration model are used to correct the weights and biases of the edge-side online deep learning prediction model. The edge-side online deep learning prediction model updated with the weights and biases continues to perform online prediction, thereby ensuring the prediction accuracy of the stopping position of the maglev train on the edge side.

[0066] Because the cloud-side network parameter self-calibration model uses real-time data for online training, it can update model parameters in real time according to changes in train operating conditions. Therefore, using the cloud-side network parameter self-calibration model to calibrate the edge-side online deep learning prediction model can ensure high-precision predictions at the edge. Simultaneously, deploying the computationally intensive model self-calibration process on the cloud reduces the computational requirements on the edge. Even when the edge lacks the computational power required for model self-updates, the robustness of the edge-side online deep learning prediction model and the accuracy of the prediction values ​​can still be guaranteed.

[0067] The above method has been verified in practice. 50,000 sets of actual operating data from maglev trains were selected as the test set. The root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), goodness of fit (R²), and rising trend accuracy (TPR) were used to evaluate the experimental results. Table 2 shows the experimental performance of the maglev train position side-side online prediction, and Table 3 shows the experimental performance of the maglev train position cloud-side online prediction. Compared with the actual values, the RMSE of the side-side online train position prediction model is 3388, and the MAE is 3. 171. The MAPE is 0.00072%, TPR is 97.89%, and R² is 99.90%. The RMSE of the cloud-based online train position prediction model is 1052, MAE is 624, MAPE is 0.00062%, TPR is 98.29%, and R² is 99.99%. The errors between the predicted and actual values ​​of both the edge and cloud sides are very small. The goodness of fit and the accuracy of the upward trend are close to the ideal value of 1. This indicates that the model has high accuracy and can fit the data well and make accurate trend predictions. It is suitable for online prediction of the position of maglev trains.

[0068] Table 2. Experimental Performance of Online Position Prediction for Maglev Trains

[0069] Table 3. Experimental Performance of Online Cloud-Based Prediction of Maglev Train Location

[0070] To present the experimental results more clearly and intuitively, six different neural network structures were randomly selected. For each architecture, predictions were performed using both cloud-side and edge-side models, and each neural network structure was tested based on 200 sets of data. With 14 neurons (n), 240 hidden layer nodes (h), and 5 neural network layers (l), the comparison curve between the actual and predicted train positions is shown below. Figure 5 As shown.

[0071] This invention provides a method for determining the position during the braking process of a maglev train based on digital twins. The embodiment of this invention constructs a train position prediction model combining linear and nonlinear methods. This model includes an offline-identified linear model and an unknown nonlinear dynamic system. The output of the unknown nonlinear dynamic system is predicted based on a trained online deep learning prediction model. In response to a train stopping preparation command, real-time operating data of the target train is acquired. This real-time operating data is input into the linear model to calculate a linear prediction value. The real-time operating data is then processed using the trained online deep learning prediction model to obtain a nonlinear prediction value. Finally, the linear and nonlinear prediction values ​​are added to obtain the predicted position value of the target train, which is then output to an interactive terminal to assist in stopping at the target stopping point. By constructing a combined prediction architecture of a linear model and a nonlinear deep learning model, the system dynamics are described more completely, thus significantly reducing prediction errors. Simultaneously, it provides more advanced and reliable state perception information for the train's automatic driving system, thereby greatly improving the success rate of accurate one-time positioning and stopping.

[0072] Furthermore, as a response to the above Figure 1 The implementation of the method shown in this invention provides a position determination device for the braking process of a magnetic levitation train based on digital twins. Figure 6 As shown, the device includes: Module 31 is used to construct a train position prediction model that combines linear and nonlinear features. The train position prediction model includes an offline identified linear model and an unknown nonlinear dynamic system. The output of the unknown nonlinear dynamic system is predicted based on a trained online deep learning prediction model. The acquisition module 32 is used to acquire real-time operating data of the target train in response to the train stopping preparation command; The calculation module 33 is used to input the real-time running data into the linear model to calculate the linear forecast value; and to perform prediction processing on the real-time running data based on the trained online deep learning forecast model to obtain the nonlinear forecast value. The output module 34 is used to add the linear forecast value and the nonlinear forecast value to obtain the target train's position forecast value, and output the position forecast value to the interactive terminal to assist in stopping at the target parking point.

[0073] Furthermore, the building module 31 includes: The first construction unit is used to construct the train position control closed-loop system equations and decompose the train position dynamic system into linear and nonlinear parts based on the train position control closed-loop system equations, wherein the nonlinear part includes multiple nonlinear fluctuation terms used to characterize the uncertainties in the train control system; the identification unit is used to identify the parameter vector of the linear part using the least squares algorithm based on historical train operation data to obtain a linear model; the second construction unit is used to construct an unknown nonlinear dynamic system based on the identification error of the parameter vectors and the unknown nonlinear terms; the training unit is used to construct a deep learning model for predicting the unknown nonlinear dynamic system and train the deep learning model based on historical train position time series data to obtain a trained online deep learning prediction model; The third building unit is used to construct a train position prediction model combining linear and nonlinear elements based on the linear model, the unknown nonlinear dynamic system, and the trained online deep learning prediction model.

[0074] Furthermore, in a specific application scenario, the first construction unit is specifically used to construct a position dynamic model that includes a linear model and unknown nonlinear terms, wherein the linear model includes control level positions and driving permissions; A control law is constructed that includes a dynamic error proportional term and a feedforward compensation term. The dynamic error proportional term is dynamically adjusted based on the deviation between the real-time position of the train and the target position. The feedforward compensation term is calculated based on multiple train operation state variables, wherein the train operation state variables include at least one of the following: level, driving permission, most stringent target point position, most stringent target point speed, train speed, reference speed value, temporary speed limit, recommended speed, emergency braking speed, maximum speed limit, alarm speed, and gradient. The output of the control law is applied to the position dynamic model to obtain the closed-loop system equations for train position control.

[0075] Furthermore, in specific application scenarios, the identification unit is specifically used to construct the input matrix and output matrix of train operation control based on historical operation data at continuous time points. The input matrix includes the train position, target stopping point, control level position, and train operation state variables at different historical time points. Based on the input matrix and the output matrix, construct a parameter identification equation with the parameter vector of the linear part as the identification object; The parameter identification equation is solved using the least squares algorithm to obtain an estimate of the parameter vector, and the parameter vector of the linear part is updated with the estimate to obtain a linear model.

[0076] Furthermore, in specific application scenarios, the training unit is specifically used to construct a deep learning model containing a multi-layer recurrent neural network, wherein the multi-layer recurrent neural network adopts an architecture including an input gate, a forget gate, an output gate, and a state memory unit, and is used to extract nonlinear features of train position time series data. Calculate the historical linear output of the linear model corresponding to the historical train operation data at each moment, as well as the difference between the historical position measurement value and the historical linear output, and use the difference as the sample label of the unknown nonlinear dynamic system to construct the training sample set of the initial online deep learning prediction model; The deep learning model is trained using the training sample set to learn the changing patterns of the unknown nonlinear dynamic system, resulting in a well-trained online deep learning prediction model.

[0077] Furthermore, the device also includes: The end-side acquisition module is used to collect real-time operation data based on the end-side sensors deployed on the target train and upload it to the real-time operation database on the cloud side. The side-side data acquisition module is used to download the real-time operation data of the target train from the cloud-side real-time operating database; The cloud-side construction module is used to build linear models on the cloud side and train online deep learning prediction models. The edge computing module is used to load the linear model and the trained online deep learning prediction model from the cloud side, so as to calculate the location prediction value on the edge based on real-time running data, using the linear model and the trained online deep learning prediction model. The online deep learning prediction model is trained online on the cloud side based on network parameter self-correction mechanism and network structure self-correction mechanism.

[0078] Furthermore, the device also includes: The self-calibration module is used on the cloud side to perform network parameter self-calibration and network structure self-calibration on the online deep learning prediction model based on real-time updated train operation data, so as to obtain a network structure self-calibration model and a network parameter self-calibration model. The cloud-side network structure self-calibration module is used to correct the network structure of the network parameter self-calibration model based on the network structure of the network structure self-calibration model when the first error of the network parameter self-calibration model is greater than the first preset error threshold and the second error of the network structure self-calibration model is less than the first preset error threshold, so as to continue online training based on the network parameter self-calibration model with the updated network structure. The cloud-side network parameter self-calibration module is used to correct the weights and biases of the edge-side online deep learning prediction model based on the weights and biases of the network parameter self-calibration model when the third error of the edge-side online deep learning prediction model is greater than the second preset error threshold, and the first error of the network parameter self-calibration model is less than the second preset error threshold, thereby obtaining an updated online deep learning prediction model.

[0079] This invention provides a position determination device for the braking process of a maglev train based on digital twins. The embodiment of this invention constructs a train position prediction model combining linear and nonlinear methods. This model includes an offline-identified linear model and an unknown nonlinear dynamic system. The output of the unknown nonlinear dynamic system is predicted based on a trained online deep learning prediction model. In response to a train stopping preparation command, real-time operating data of the target train is acquired. This real-time operating data is input into the linear model to calculate a linear prediction value. The real-time operating data is then processed using the trained online deep learning prediction model to obtain a nonlinear prediction value. The linear and nonlinear prediction values ​​are added to obtain the target train's position prediction value, which is then output to an interactive terminal to assist in stopping at the target stopping point. By constructing a combined prediction architecture of a linear model and a nonlinear deep learning model, the system dynamics are described more completely, significantly reducing prediction errors. Simultaneously, it provides more advanced and reliable state perception information for the train's automatic driving system, thereby greatly improving the success rate of accurate one-time positioning and stopping.

[0080] According to one embodiment of the present invention, a storage medium is provided, the storage medium storing at least one executable instruction, the computer-executable instruction being able to execute the position determination method for the braking process of a magnetic levitation train based on digital twins in any of the above method embodiments.

[0081] Figure 7 The diagram shows a structural schematic of a terminal according to an embodiment of the present invention. The specific implementation of the present invention does not limit the specific implementation of the terminal.

[0082] like Figure 7 As shown, the terminal may include: a processor 402, a communication interface 404, a memory 406, and a communication bus 408.

[0083] The processor 402, communication interface 404, and memory 406 communicate with each other via communication bus 408.

[0084] Communication interface 404 is used for network communication with other devices such as clients or other servers.

[0085] The processor 402 is used to execute program 410, specifically to execute the relevant steps in the above embodiment of the position determination method for the braking process of a maglev train based on digital twins.

[0086] Specifically, program 410 may include program code that includes computer operation instructions.

[0087] Processor 402 may be a central processing unit (CPU), a specific integrated circuit (ASIC), or one or more integrated circuits configured to implement embodiments of the present invention. The terminal may include one or more processors of the same type, such as one or more CPUs; or it may include processors of different types, such as one or more CPUs and one or more ASICs.

[0088] Memory 406 is used to store program 410. Memory 406 may include high-speed RAM memory, and may also include non-volatile memory, such as at least one disk storage device.

[0089] Specifically, program 410 can be used to cause processor 402 to perform the following operations: A train position prediction model combining linear and nonlinear methods is constructed, wherein the train position prediction model includes an offline identified linear model and an unknown nonlinear dynamic system, and the output of the unknown nonlinear dynamic system is predicted based on a trained online deep learning prediction model. In response to a train stopping preparation command, acquire real-time operating data of the target train; The real-time running data is input into the linear model to calculate the linear forecast value; and the real-time running data is then processed for prediction based on the trained online deep learning forecast model to obtain the nonlinear forecast value. The linear forecast value and the nonlinear forecast value are added together to obtain the target train's position forecast value, which is then output to the interactive terminal to assist in stopping at the target parking point.

[0090] It is obvious to those skilled in the art that the modules or steps of the present invention described above can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. Optionally, they can be implemented using computer-executable program code, thereby storing them in a storage device for execution by a computing device. In some cases, the steps shown or described can be performed in a different order than those presented herein, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, the present invention is not limited to any particular combination of hardware and software.

[0091] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for determining the position during the braking process of a maglev train based on digital twins, characterized in that, include: A train position prediction model combining linear and nonlinear methods is constructed, wherein the train position prediction model includes an offline identified linear model and an unknown nonlinear dynamic system, and the output of the unknown nonlinear dynamic system is predicted based on a trained online deep learning prediction model. In response to a train stopping preparation command, acquire real-time operating data of the target train; The real-time running data is input into the linear model to calculate the linear forecast value; and the real-time running data is then processed for prediction based on the trained online deep learning forecast model to obtain the nonlinear forecast value. The linear forecast value and the nonlinear forecast value are added together to obtain the target train's position forecast value, which is then output to the interactive terminal to assist in stopping at the target parking point.

2. The method according to claim 1, characterized in that, The construction of the train position prediction model combining linear and nonlinear methods includes: A train position control closed-loop system equation is constructed, and based on the equation, the train position dynamic system is decomposed into a linear part and a nonlinear part. The nonlinear part includes multiple nonlinear fluctuation terms characterizing uncertainties in the train control system. Based on historical train operation data, the parameter vector of the linear part is identified using a least squares algorithm to obtain a linear model. An unknown nonlinear dynamic system is constructed based on the identification error of the parameter vector and the unknown nonlinear terms. A deep learning model for predicting the unknown nonlinear dynamic system is constructed, and the deep learning model is trained based on historical train position time-series data to obtain a trained online deep learning prediction model. A train position prediction model combining linear and nonlinear elements is constructed based on the linear model, the unknown nonlinear dynamic system, and the trained online deep learning prediction model.

3. The method according to claim 2, characterized in that, The process of constructing the equations for the train position control closed-loop system includes: Construct a position dynamic model that includes a linear model and unknown nonlinear terms, wherein the linear model includes control level positions and driving permissions; A control law is constructed that includes a dynamic error proportional term and a feedforward compensation term. The dynamic error proportional term is dynamically adjusted based on the deviation between the real-time position of the train and the target position. The feedforward compensation term is calculated based on multiple train operation state variables, wherein the train operation state variables include at least one of the following: level, driving permission, most stringent target point position, most stringent target point speed, train speed, reference speed value, temporary speed limit, recommended speed, emergency braking speed, maximum speed limit, alarm speed, and gradient. The output of the control law is applied to the position dynamic model to obtain the closed-loop system equations for train position control.

4. The method according to claim 2, characterized in that, The method of identifying the parameter vector of the linear component based on historical train operation data and using the least squares algorithm to obtain a linear model includes: The input matrix and output matrix of train operation control are constructed based on historical operation data at continuous time points. The input matrix includes the train position, target stopping point, control level position and train operation state variables at different historical time points. Based on the input matrix and the output matrix, construct a parameter identification equation with the parameter vector of the linear part as the identification object; The parameter identification equation is solved using the least squares algorithm to obtain an estimate of the parameter vector, and the parameter vector of the linear part is updated with the estimate to obtain a linear model.

5. The method according to claim 2, characterized in that, The historical train position time series data includes continuous historical train operation data and historical position measurement values; The construction and training process of the online deep learning prediction model includes: A deep learning model containing a multi-layer recurrent neural network is constructed, wherein the multi-layer recurrent neural network adopts an architecture including an input gate, a forget gate, an output gate, and a state memory unit, for extracting nonlinear features of train position time series data; Calculate the historical linear output of the linear model corresponding to the historical train operation data at each moment, as well as the difference between the historical position measurement value and the historical linear output, and use the difference as the sample label of the unknown nonlinear dynamic system to construct the training sample set of the initial online deep learning prediction model; The deep learning model is trained using the training sample set to learn the changing patterns of the unknown nonlinear dynamic system, resulting in a well-trained online deep learning prediction model.

6. The method according to any one of claims 1-5, characterized in that, The method is implemented based on an edge-cloud-edge architecture. The acquisition of real-time operation data of the target train includes: collecting real-time operation data based on edge sensors deployed on the target train and uploading it to the real-time operation database on the cloud side; and downloading the real-time operation data of the target train from the real-time operation database on the cloud side. Before inputting the real-time running data into the linear model, the method further includes: Linear models are built on the cloud side, and online deep learning prediction models are trained online. The edge loads the linear model and the trained online deep learning prediction model from the cloud to calculate the location prediction value on the edge based on real-time running data using the linear model and the trained online deep learning prediction model. The online deep learning prediction model is trained online on the cloud side based on network parameter self-correction mechanism and network structure self-correction mechanism.

7. The method according to claim 1, characterized in that, The method further includes: On the cloud side, network parameter self-calibration and network structure self-calibration are performed on the online deep learning prediction model based on real-time updated train operation data to obtain a network structure self-calibration model and a network parameter self-calibration model. If the first error of the network parameter self-calibration model is found to be greater than the first preset error threshold and the second error of the network structure self-calibration model is less than the first preset error threshold, the network structure of the network parameter self-calibration model is corrected based on the network structure of the network structure self-calibration model, so as to continue online training based on the network parameter self-calibration model with the updated network structure. If the third error of the online deep learning prediction model on the edge is greater than the second preset error threshold, and the first error of the network parameter self-calibration model is less than the second preset error threshold, the weights and biases of the online deep learning prediction model on the edge are corrected based on the weights and biases of the network parameter self-calibration model to obtain an updated online deep learning prediction model.

8. A position determination device for the braking process of a magnetic levitation train based on digital twins, characterized in that, include: A construction module is used to construct a train position prediction model that combines linear and nonlinear features. The train position prediction model includes an offline identified linear model and an unknown nonlinear dynamic system. The output of the unknown nonlinear dynamic system is predicted based on a trained online deep learning prediction model. The acquisition module is used to acquire real-time operating data of the target train in response to the train stopping preparation command; The calculation module is used to input the real-time running data into the linear model to calculate the linear forecast value; and to perform prediction processing on the real-time running data based on the trained online deep learning forecast model to obtain the nonlinear forecast value. The output module is used to add the linear forecast value and the nonlinear forecast value to obtain the target train's position forecast value, and output the position forecast value to the interactive terminal to assist in stopping at the target parking point.

9. A storage medium, characterized in that, The storage medium stores at least one executable instruction that causes the processor to perform the operation corresponding to the position determination method for the braking process of a magnetic levitation train based on digital twins as described in any one of claims 1-7.

10. A terminal, characterized in that, include: The processor, memory, communication interface, and communication bus are provided, wherein the processor, memory, and communication interface communicate with each other via the communication bus. The memory is used to store at least one executable instruction that causes the processor to perform the operation corresponding to the position determination method for the braking process of a magnetic levitation train based on digital twins as described in any one of claims 1-7.

Images