Rail transit digital twin system and method based on catenary feature data

Abstract

The invention relates to the technical field of rail transit, and discloses a rail transit digital twin system and method based on catenary characteristic data, and the system comprises a data input module which is used for receiving sensor data representing the physical state of a catenary; the model storage module is used for storing a parameterized digital model of the contact network, wherein the parameterized digital model is provided with correctable internal model parameters; the parameter correction module is connected with the data input module and the model storage module; and the model updating module is used for applying the correction value to the parameterized digital model so as to finish synchronous updating of the model. By adopting a data assimilation framework based on an adjoint method, efficient and accurate reverse correction of internal parameters of the model is realized, real-time dynamic synchronization and evolution can be carried out on a large-scale and full-line contact network digital twin system, and the practicability and timeliness of the digital twin system in the field of rail transit are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of rail transit technology, and in particular to a rail transit digital twin system and method based on contact network characteristic data. Background Art

[0002] As critical infrastructure that directly supplies power to trains, the stability and reliability of the rail transit catenary system is directly related to the operational safety and efficiency of the entire rail transit system. In recent years, with the development of digital twin technology, the industry has begun to explore the construction of high-fidelity digital models of the catenary system, hoping to play a key role in design optimization, condition monitoring, fault diagnosis, and predictive maintenance. However, when applying the digital twin concept to complex and large-span catenary systems, existing technologies still have several insurmountable limitations.

[0003] Existing technologies for constructing digital twins of catenary systems generally face a sharp conflict between physical fidelity and computational real-time performance. To accurately replicate the complex nonlinear dynamic behavior of the catenary system, researchers often employ sophisticated finite element models. While these models can achieve high physical fidelity, they are computationally intensive, and a full-line dynamic simulation is often extremely time-consuming. This makes them unsuitable for dynamic digital twin applications that require real-time or near-real-time data interaction with physical entities, and they are typically limited to offline simulation and analysis tools. Furthermore, to meet real-time requirements, some solutions employ significantly simplified analytical or empirical models. However, this simplification often sacrifices key physical properties, resulting in models that fail to accurately reflect the mechanical behavior of the actual catenary system, and their predicted results can deviate significantly from reality over time. Therefore, existing technologies lack a model construction and modification method that balances physical fidelity with computational efficiency to support a truly dynamic catenary digital twin system.

[0004] When using sensor data to modify digital models, existing technologies typically employ crude data fusion strategies that lack a deep understanding of the physical scenario and adaptive adjustment capabilities. A common practice is to incorporate sensor data from all sources, indiscriminately or using fixed static weights, into a unified cost function, performing parameter calibration by minimizing the overall error between the model prediction and all observed data. However, the dominant physical factors of the overhead contact network vary significantly under different operating conditions. For example, at night when no trains are passing, its shape is primarily determined by temperature stress, while when high-speed trains are passing, it manifests primarily as a complex forced vibration response. This existing "one size fits all" correction approach, which attempts to balance errors under different physical mechanisms using a fixed set of weights, often leads to slow convergence of the correction process and even physically unreasonable parameter solutions, failing to specifically address the core model deviation issues under the current operating conditions.

[0005] Furthermore, the vast majority of current catenary digital twin construction and application approaches follow a "static snapshot" model. Initially, the digital model is calibrated through a one-time on-site measurement. Once calibration is complete, the model's inherent physical parameters are fixed. This model ignores the fact that the physical catenary itself is a continuously evolving system, whose material properties, component states, and geometry are subject to continuous and slow drift due to factors such as material aging, component wear, environmental corrosion, and human maintenance interventions over time. A "static" twin, precisely calibrated at the initial moment, will inevitably deviate from its physical counterpart over time, gradually losing its effectiveness as a digital mirror of the real world. Existing technologies generally lack an effective closed-loop mechanism that allows digital twins to continuously and automatically incorporate new observational information to track and synchronize this evolution of physical properties. Summary of the Invention

[0006] In response to the shortcomings of the existing technology, the present invention provides a rail transit digital twin system and method based on contact network characteristic data, which solves the contradiction between the fidelity of the digital twin model and the real-time performance of the calculation in the existing technology, and realizes the long-term, dynamic and high-fidelity synchronization between the digital twin and the physical contact network.

[0007] To achieve the above objectives, the present invention is implemented through the following technical solutions: A rail transit digital twin system based on contact network characteristic data includes:

[0008] A data input module, configured to receive sensor data representing a physical state of the contact network;

[0009] A model storage module, used for storing a parameterized digital model of the contact network having modifiable internal model parameters;

[0010] a parameter correction module, connected to the data input module and the model storage module, configured to: construct a cost function based on the sensor data and the predicted output of the parameterized digital model; and calculate correction values ​​for the internal model parameters by optimizing the cost function;

[0011] The model updating module is used to apply the correction value to the parameterized digital model to complete the synchronous update of the model.

[0012] Preferably, the parameterized digital model is a differentiable reduced-order physical model, the dynamic behavior of which is described by the following set of equations:

[0013]

[0014] Where: x(t) is the state vector of the system at time t; M is the mass matrix of the system; C(θ) is the damping matrix that depends on the internal model parameter θ; K(θ) is the stiffness matrix that depends on the internal model parameter θ; F ext (t) is the external load vector.

[0015] Preferably, the parameter correction module optimizes the cost function J(θ) by calculating the gradient of the cost function with respect to the internal model parameter θ And based on the gradient, the internal model parameters are iteratively updated according to the following rules:

[0016]

[0017] Where: θ k is the internal model parameter at the kth iteration; θ k+1 is the internal model parameter at the k+1th iteration; is the cost function in θ k The gradient at α k is the learning rate for the kth iteration.

[0018] Preferably, the parameter correction module calculates the gradient in the following manner:

[0019] constructing and solving adjoint equations associated with the governing equations of the parameterized digital model to obtain adjoint variables;

[0020] The gradient is calculated analytically using the adjoint variables and the forward solution state of the model.

[0021] Preferably, the cost function J(θ) includes an observation term for measuring the weighted error between the model prediction output and the sensor data, and its expression is:

[0022]

[0023] Where: θ is the internal model parameter; x(θ) is the model state vector determined by the internal model parameter θ; z is the observation vector composed of the sensor data; H is the observation operator that maps the model state vector to the observation space; W is the dynamic weighting matrix; θ b is the background field value of the internal model parameter; R is the regularization matrix;

[0024] The parameter correction module is also used to dynamically generate the dynamic weighting matrix W.

[0025] Preferably, the parameter correction module dynamically generates the dynamic weighting matrix W in the following manner:

[0026] First, identifying the current physical condition of the overhead line;

[0027] Then, different weights are assigned to different types of sensor data according to the physical working conditions.

[0028] Preferably, the physical working conditions include static working conditions and dynamic response working conditions;

[0029] Under the static working condition, increasing the weight of the geometric position sensor data;

[0030] Under the dynamic response working condition, the weight of the dynamic response sensor data is increased.

[0031] Preferably, the internal model parameters are divided into parameter subsets of different spatiotemporal scales.

[0032] Preferably, the parameter correction module adopts a hierarchical correction strategy, and performs corrections on parameter subsets of different spatiotemporal scales using different correction frequencies and data ranges.

[0033] The rail transit digital twin method based on contact network characteristic data includes the following steps:

[0034] S1. receiving sensor data representing the physical state of the contact network;

[0035] S2. constructing a cost function based on the sensor data and a predicted output of a stored parameterized digital model having modifiable internal model parameters;

[0036] S3. Calculating and obtaining correction values ​​for the internal model parameters by optimizing the cost function;

[0037] S4. Apply the correction value to the parameterized digital model to complete the synchronous update of the model.

[0038] The present invention provides a rail transit digital twin system and method based on contact network characteristic data. It has the following beneficial effects:

[0039] 1. This invention constructs a differentiable reduced-order physical model as the computational core of the digital twin and employs a data assimilation framework based on the adjoint method to achieve efficient and accurate reverse correction of the model's intrinsic parameters. This method circumvents the inherent contradiction between the high computational cost of traditional finite element simulation and the insufficient fidelity of simplified models. It enables real-time dynamic synchronization and evolution of large-scale, full-line catenary digital twin systems while ensuring consistency with the key characteristics of the physical entity, significantly improving the practicality and timeliness of digital twin systems in the rail transit sector.

[0040] 2. This invention automatically identifies the different physical conditions of the catenary, such as static or dynamic response, and assigns different weights to different types of sensor data accordingly. This eliminates the need for blind mathematical fitting and instead enables intelligent optimization with the ability to understand the physical scenario. This mechanism ensures that the system always focuses computing resources on the most significant physical contradictions, thereby improving the relevance of the correction process, the speed of convergence, and the physical interpretability of the results, thereby enhancing the digital twin system's adaptability to complex and changing operating environments.

[0041] 3. The digital twin system constructed by this invention can continuously track and compensate for drift in the physical entity's characteristics caused by factors such as material aging, component wear, and environmental changes. This continuous adaptive correction capability ensures that the digital twin maintains long-term, high-fidelity synchronization with the physical entity throughout its entire lifecycle, providing a reliable digital foundation for implementing higher-level applications such as predictive maintenance, fault warning, and condition assessment. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 This is a system architecture diagram of the present invention;

[0043] Figure 2 is a flow chart of the method of the present invention; DETAILED DESCRIPTION

[0044] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0045] Example:

[0046] Please see the attached Figure 1 , an embodiment of the present invention provides a rail transit digital twin system based on contact network characteristic data, including:

[0047] A data input module for receiving sensor data representing the physical state of the contact network;

[0048] In this embodiment, the data input module serves as a bridge connecting the contact network in the physical world and the digital twin system. Its core function is to comprehensively and accurately obtain the multi-dimensional physical state characteristics that characterize the contact network in the real operating environment, and process them into structured data with a unified time and space reference that can be used by subsequent parameter correction modules.

[0049] The data input module is configured to receive and process data from multiple different sensor types, known as heterogeneous sensor data. The integration of this heterogeneous data is crucial because a single type of sensor data can only reflect a specific aspect of the catenary's physical properties. This invention aims to build a digital twin that fully replicates real-world physical behavior, requiring comprehensive observation of the physical entity from multiple dimensions, including geometry, dynamic response, and environmental influences.

[0050] In a preferred embodiment, the sensor data processed by the data input module may specifically include the following categories:

[0051] The first is geometric state sensor data. This type of data is primarily used to accurately describe the static or quasi-static spatial shape of the catenary at a specific moment. Specifically, it can include continuous conductance and pullout values ​​along the line, measured by vehicle- or drone-mounted 3D LiDAR scanning systems or high-resolution industrial vision systems. This geometric data is fundamental to determining whether the catenary meets safe operating procedures and is a key basis for model morphology fitting and calibration under static operating conditions.

[0052] The second is the dynamic response sensor data. This type of data is intended to capture the transient behavior of the contact network when it is subjected to external dynamic excitation (especially when the train pantograph passes). Specifically, it may include the vibration acceleration signal measured by the accelerometer installed at the key stress points of the contact network (such as the suspension point, the root of the arm, the positioning point, etc.), or the structural stress and strain signal measured by the strain gauge. These dynamic data directly reflect the intrinsic dynamic characteristics of the system, such as stiffness and damping, and are the core input for the present invention to correct and identify the model parameters under dynamic working conditions.

[0053] The third is environmental status sensor data. This type of data is used to quantify external environmental factors that affect the physical characteristics of the contact network. Specifically, it can include the ambient temperature measured by temperature sensors deployed along the line, and the wind field information measured by anemometers. Temperature data is crucial for accurately simulating the changes in the contact line's sag and tension caused by thermal expansion and contraction, while wind field data is crucial for calculating the external load vector F. ext (t) Direct input of wind load components.

[0054] After receiving the raw data streams from the aforementioned sensors, the data input module must perform a series of critical preprocessing and alignment operations to ensure data validity and consistency. This preprocessing includes digitally filtering the raw signals to remove high-frequency noise, using statistical methods or preset thresholds to identify and remove obvious outliers or bad pixels caused by communication interruptions, and performing necessary data normalization.

[0055] It is particularly important that the data input module performs spatiotemporal alignment operations. Since different sensors may have independent sampling clocks and coordinate systems, all data must be unified under the same spatiotemporal reference. In the time dimension, preferably, the precise timing signal provided by the Global Positioning System (GPS) can be used as the master clock of the system to synchronize and calibrate the timestamps of all sensor data. In the spatial dimension, the mileage coordinates of the line are used as a unified spatial reference. Through the coordinate transformation algorithm, the measurement data of sensors installed at different physical locations are all mapped to this one-dimensional line coordinate system.

[0056] After completing the above processing flow, the final output of the data input module is a structured, spatiotemporally aligned observation vector z. This vector z not only contains the measurements of all valid sensors at a specific moment or time window, but also contains metadata corresponding to each measurement value, such as the data type (geometric, dynamic, or environmental), the precise spatiotemporal coordinates, and the measurement uncertainty.

[0057] This observation vector z is the fundamental basis for the subsequent technical steps of the present invention. It will be passed directly to the parameter correction module as the "ground truth value" for comparison with the model prediction output H(x(θ)) when constructing the cost function J(θ). At the same time, its accompanying metadata, especially the data type information, is the basis for the subsequent dynamic weighting matrix W to be intelligently constructed according to the physical working conditions. In this way, the data input module ensures that the real-world data on which the entire digital twin system relies is comprehensive, accurate and available, providing solid data support for the effective execution of all subsequent correction and synchronization processes.

[0058] A model storage module, used for storing a parameterized digital model of the contact network with modifiable internal model parameters;

[0059] In this embodiment, the model storage module constitutes the computing core and digital carrier of the digital twin system. Its function is to store and manage a specially designed parameterized digital model that characterizes the inherent physical properties of the catenary. This model is not a static data structure, but rather a dynamically evolving digital object that can be accessed and modified by other modules.

[0060] To simultaneously meet the dual, often conflicting, requirements of the digital twin system for high fidelity of physical behavior and real-time computing performance, the parameterized digital model stored in the model storage module in this embodiment is not a traditional, computationally intensive, full-scale finite element model. Preferably, the model is a differentiable reduced-order physical model. This reduced-order model is constructed by spatially discretizing the continuous catenary structure into a series of key nodes and using a coupled set of ordinary differential equations to describe the dynamic behavior of these nodes. This significantly reduces the computational complexity of the model solution while preserving the key physical properties of the system, laying the foundation for real-time operation of the system.

[0061] Specifically, the dynamic behavior of the reduced-order physical model stored in the model storage module can be accurately described by the following set of second-order ordinary differential equations:

[0062]

[0063] In this equation, each symbol has a clear physical meaning: x(t) is an n×1-dimensional system state vector, whose elements represent the displacement of all n discrete nodes in the model at time t, and is a direct reflection of the spatial form of the model at any time. M is an n×n-dimensional system mass matrix, which characterizes the inertial characteristics of each node and is usually considered to be fixed after the model is established. C(θ) is an n×n-dimensional system damping matrix, which describes the energy dissipation characteristics of the system during movement. K(θ) is an n×n-dimensional system stiffness matrix, which comprehensively reflects the system's elastic recovery ability, geometric nonlinearity, and geometric stiffness determined by tension. F ext (t) is an n×1 dimensional external load vector, which is used to describe all external forces applied to the contact network, such as pantograph lift force, wind load, ice load, etc.

[0064] Crucially, one of the core innovations of this model lies in its parameterized nature. As shown above, the system's damping matrix C(θ) and stiffness matrix K(θ) are both constructed as explicit functions of the internal model parameter vector θ. This p×1-dimensional parameter vector θ is not an abstract mathematical symbol; it encompasses all the modifiable internal model parameters that characterize the intrinsic physical properties of the catenary. These parameters may include, but are not limited to, the equivalent elastic modulus of each contact wire or catenary segment, the Poisson's ratio of the material, the equivalent rotational and translational stiffnesses of the component connections, the system's structural damping coefficient, the material's thermal expansion coefficient, and the baseline tension values ​​of each anchor segment. Parameterizing these physical properties into a vector θ aims to transform the digital model from a fixed, unchanging "digital specimen" into a "digital embryo" whose inherent properties can be quantitatively adjusted. Subsequent parameter correction modules will use this parameter vector θ as their sole correction target.

[0065] Another key technical feature of this model is its differentiability. The entire control equation is constructed to be continuous and mathematically differentiable with respect to the system state x and the internal model parameter vector θ. This property is the fundamental prerequisite for implementing the core technical solution of the present invention. It is precisely because the model is differentiable that the subsequent parameter correction module can adopt an efficient, gradient-based optimization algorithm. Specifically, it allows the gradient of the cost function J with respect to the large-scale parameter vector θ to be analytically calculated by solving the adjoint equation This makes it possible, thus avoiding the "dimensionality disaster" faced by the traditional finite difference method when dealing with high-dimensional parameter problems, and ensuring the efficiency and accuracy of the correction algorithm.

[0066] In a preferred embodiment, to support a more refined hierarchical correction strategy, the parameter vector θ stored in this module can be pre-divided into parameter subsets at different spatiotemporal scales. For example, it can be divided into a global slowly varying parameter subset representing the overall characteristics of the entire line, a regionally varying parameter subset representing the characteristics of a specific anchor section or power supply zone, and a local rapidly varying parameter subset representing the characteristics of a specific suspension or key component. This structured parameter storage approach enables the parameter correction module to perform targeted corrections on different parameter subsets using different correction frequencies and data ranges.

[0067] In summary, the model storage module is not a simple file database but the core of the digital twin system. The parameterized, differentiable reduced-order physical model it stores not only supports the system's real-time operation with its low computational cost, but also provides a complete mathematical interface and physical support for subsequent efficient and accurate model corrections based on the adjoint method, forming the cornerstone of the entire "observe-correct-synchronize" closed-loop process.

[0068] A parameter correction module, connected to the data input module and the model storage module, is used to: construct a cost function based on the sensor data and the predicted output of the parameterized digital model; and calculate the correction value of the internal model parameter by optimizing the cost function;

[0069] In this embodiment, the parameter correction module constitutes the core processing unit and intelligent decision-making center of the present invention. It is configured to conduct two-way information interaction with the data input module and the model storage module. Its fundamental task is to systematically compare the real observation data from the physical world with the virtual prediction output of the digital model, and based on this, calculate a set of accurate internal model parameter correction values ​​that can enable the digital model to approach the physical entity.

[0070] The parameter correction module is essentially a rigorous, physically constrained mathematical optimization process. Its goal is to find an optimal internal model parameter vector θ that minimizes a carefully constructed cost function J(θ) that quantifies the discrepancy between the model and reality. This process can be broken down into the following interrelated, progressive technical steps.

[0071] First, at the beginning of the optimization process, the parameter correction module needs to construct the cost function J(θ). The construction of this cost function is not arbitrary, but deeply integrates physical prior knowledge and mathematical optimization theory. Its complete expression is:

[0072]

[0073] The function consists of two parts with clear physical meanings. The first part is the observation term, which aims to quantify the weighted error between the model prediction and the true observation. In this term, z is the observation vector obtained from the data input module, representing the true state of the physical world; x(θ) is the high-dimensional model state vector calculated by the reduced-order physical model in the model storage module under the current internal model parameters θ. Since the dimension of the model state x(θ) is much higher than the dimension of the sensor observation z, a key observation operator H is introduced. Its function is to map the high-dimensional model state space to the low-dimensional observation space, that is, to extract the predicted value corresponding to the position and type of the sensor measurement point from x(θ) so that it can be directly compared with z.

[0074] What is particularly important is that this observation item includes a dynamic weighting matrix W that is dynamically generated by the parameter correction module itself. The introduction of this matrix is ​​one of the key innovations that distinguishes the present invention from traditional data assimilation methods. Its purpose is to give the optimization process the ability to perceive the physical scene. Specifically, the parameter correction module will first determine the current physical working condition S of the contact network based on the real-time data stream, such as static or dynamic response conditions. Then, based on the identified working conditions, the module will assign different confidence weights to sensor data from different sources. For example, under static conditions, the weight corresponding to the geometric position sensor data is increased, while under dynamic response conditions, the weight corresponding to the vibration sensor data is increased. This dynamic weighting mechanism ensures that the model correction can always focus on the main physical contradictions under the current working conditions, making the optimization process more targeted and physically meaningful.

[0075] The second part of the cost function is the background term or regularization term. This term is introduced to ensure the well-conditioned nature of the optimization problem and the physical rationality of the solution. b Represents the background field or prior value of the parameter, which can usually be the design value or the optimization result of the previous time step; R is the regularization matrix. bPenalizing solutions that are too large effectively prevents the model from producing unrealistic parameter values ​​due to overfitting the observed data.

[0076] After constructing the cost function J(θ), the core task of the parameter correction module is to efficiently solve the optimization problem. Preferably, a gradient-based optimization algorithm is used, the core of which is to calculate the gradient of the cost function J with respect to the large-scale parameter vector θ. Considering that the dimension of the parameter vector θ may be extremely high, the present invention adopts the adjoint method to overcome the efficiency bottleneck faced by traditional gradient calculation methods. The process includes:

[0077] Forward solution: The module first instructs the model storage module to use the current parameters θ k Perform a forward calculation to obtain the model state x(θ k ).

[0078] Adjoint Solver: The module then constructs and solves a linear adjoint equation that is computationally equivalent to the forward model based on the governing equations of the physical model and the expression of the cost function to obtain the adjoint variable λ.

[0079] Gradient analytical calculation: Once the forward state x and the accompanying variable λ are obtained, the complete gradient vector can be calculated at once through an analytical expression The computational cost of this method is basically independent of the number p of parameters to be optimized, thereby achieving efficient solution of large-scale parameter gradients, which is the key to the engineering application of the present invention.

[0080] In obtaining the gradient After that, the parameter correction module can iteratively update the model parameters, and its update rules are:

[0081]

[0082] Where k is the number of iterations, α k The learning rate is used to control the convergence speed and stability. This iterative process will continue until the value of the cost function J converges to an acceptable minimum.

[0083] Furthermore, in a more refined implementation, to further improve correction efficiency and accuracy, the parameter correction module can also implement a hierarchical correction strategy. Specifically, the parameter vector θ stored in the model storage module is divided into subsets of different spatiotemporal scales based on their physical properties. Different correction strategies are then applied to different subsets, thereby achieving an organic combination of macro-calibration and micro-focusing.

[0084] Finally, the parameter correction module will complete the optimized and latest internal model parameter vector θ k+1The corrected value is then output to the model update module. In summary, this module, the system's "intelligent brain," achieves precise, efficient, and intelligent corrections to the intrinsic properties of the digital twin model by constructing a physically meaningful cost function, applying an efficient adjoint method for gradient solving, and incorporating advanced strategies such as dynamic weighting and hierarchical correction.

[0085] The model updating module is used to apply the correction value to the parameterized digital model to complete the synchronous updating of the model.

[0086] In this embodiment, the model update module serves as the final execution link of the closed-loop correction process of the digital twin system of the present invention. Its core function is to materialize the parameter correction value calculated by the parameter correction module, which represents a deeper understanding of the physical world, into an actual change in the intrinsic properties of the parameterized digital model, and ultimately complete the state synchronization between the digital twin and the physical entity at the current moment.

[0087] Specifically, the operation process of the model update module begins when it receives an optimal internal model parameter vector θ output after a complete optimization iteration from the parameter correction module. k+1 This vector is not a meaningless set of numbers, but rather a best estimate of parameters that characterize the intrinsic physical properties of the catenary, such as equivalent stiffness, damping coefficient, and reference tension. It is the latest understanding of the intrinsic properties of the physical entity, derived from integrating the latest sensor observations and undergoing a physically constrained optimization process.

[0088] Upon receiving the parameter vector θ k+1 After that, the model update module performs its first task, which is to apply parameter updates. This process is done by interacting with the model storage module. The module will use the new parameter vector θ received k+1 , to overwrite or replace the old parameter vector θ stored in the model storage module k .

[0089] This parameter update operation will directly and profoundly affect the dynamic behavior of the parameterized digital model. In other words, when the parameter vector θ is updated, the internal element values ​​of the system matrices that depend on this vector and are stored in the model storage module, namely the damping matrix C(θ) and the stiffness matrix K(θ), will also change accordingly. This change is directly reflected in the governing equations of the reduced-order physical model:

[0090]

[0091] In this way, the calculation results of the parameter correction module are seamlessly and deterministically transferred to the core mathematical description of the digital twin, making its physical properties fundamentally closer to the real world.

[0092] However, simply updating the internal parameters θ is not enough to complete a complete synchronization. This is because after the parameters are updated, the original state vector x(θ k ) and its new intrinsic properties θ k+1 Therefore, the model update module also needs to perform its second key task, which is to trigger state synchronization.

[0093] This state synchronization process refers to the application of new model parameters θ k+1 Afterwards, the instruction system re-solves the control equations of the model to calculate the new equilibrium state or dynamic response trajectory x(θ k+1 For example, in static conditions, the module solves the new static equilibrium equation K(θ k+1 )x=F ext To determine the geometry of the contact network under the new stiffness parameters. This step is necessary to ensure that the digital twin is synchronized with the physical entity not only in its "intrinsic genes" (parameters θ) but also in its "external performance" (state x).

[0094] After state synchronization is complete, the digital twin stored in the model storage module has achieved the highest fidelity with the physical entity within the available observation dimensions at that moment. This completes the closed-loop process of "observe-weight-correct-synchronize."

[0095] In summary, the model update module transforms abstract mathematical optimization results into tangible changes in the physical properties and states of the digital twin by applying two key actions: parameter updates and triggering state synchronization, thus completing the closed-loop correction process. Each successful operation of this module signifies that the digital twin has completed a self-improvement and evolution with the help of new external information, and is ready to receive the next batch of sensor data and initiate a new round of dynamic correction cycles. This is the key to the invention's ability to achieve long-term, continuous, and adaptive synchronization between the digital twin and the physical entity.

[0096] Please see the attached Figure 2 Another embodiment of the present invention provides a rail transit digital twin method based on contact network characteristic data, comprising the following steps:

[0097] S1, receiving sensor data representing the physical state of the contact network;

[0098] S2, constructing a cost function based on the sensor data and the predicted output of a stored parameterized digital model with modifiable internal model parameters;

[0099] S3. Calculate and obtain correction values ​​of internal model parameters by optimizing the cost function;

[0100] S4. Apply the correction value to the parameterized digital model to complete the synchronous update of the model.

[0101] The method of this embodiment can be used to execute the above system embodiment. Its principles and technical effects are similar and will not be described in detail here.

[0102] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. The rail transit digital twin system based on contact network characteristic data is characterized by: include: A data input module, configured to receive sensor data representing a physical state of the contact network; A model storage module, used for storing a parameterized digital model of the contact network having modifiable internal model parameters; a parameter correction module, connected to the data input module and the model storage module, configured to: construct a cost function based on the sensor data and the predicted output of the parameterized digital model; and calculate correction values ​​for the internal model parameters by optimizing the cost function; The model updating module is used to apply the correction value to the parameterized digital model to complete the synchronous update of the model.

2. The rail transit digital twin system based on contact network characteristic data according to claim 1 is characterized in that: The parameterized digital model is a differentiable reduced-order physical model, and the dynamic behavior of the model is described by the following set of equations: Where: x(t) is the state vector of the system at time t; M is the mass matrix of the system; C(θ) is the damping matrix that depends on the internal model parameter θ; K(θ) is the stiffness matrix that depends on the internal model parameter θ; F ext (t) is the external load vector.

3. The rail transit digital twin system based on contact network characteristic data according to claim 1 is characterized in that: The parameter modification module optimizes the cost function J(θ) by calculating the gradient of the cost function with respect to the internal model parameter θ And based on the gradient, the internal model parameters are iteratively updated according to the following rules: Where: θ k is the internal model parameter at the kth iteration; θ k+1 is the internal model parameter at the k+1th iteration; is the cost function in θ k The gradient at α k is the learning rate for the kth iteration.

4. The rail transit digital twin system based on contact network characteristic data according to claim 1 is characterized in that: The parameter correction module calculates the gradient in the following way: constructing and solving adjoint equations associated with the governing equations of the parameterized digital model to obtain adjoint variables; The gradient is calculated analytically using the adjoint variables and the forward solution state of the model.

5. The rail transit digital twin system based on contact network characteristic data according to claim 1 is characterized in that: The cost function J(θ) includes an observation term for measuring the weighted error between the model prediction output and the sensor data, and its expression is: Where: θ is the internal model parameter; x(θ) is the model state vector determined by the internal model parameter θ; z is the observation vector composed of the sensor data; H is the observation operator that maps the model state vector to the observation space; W is the dynamic weighting matrix; θ b is the background field value of the internal model parameter; R is the regularization matrix; The parameter correction module is also used to dynamically generate the dynamic weighting matrix W.

6. The rail transit digital twin system based on contact network characteristic data according to claim 5 is characterized in that: The parameter correction module dynamically generates the dynamic weighting matrix W in the following manner: First, identifying the current physical condition of the overhead line; Then, different weights are assigned to different types of sensor data according to the physical working conditions.

7. The rail transit digital twin system based on contact network characteristic data according to claim 6 is characterized in that: The physical working conditions include static working conditions and dynamic response working conditions; Under the static working condition, increasing the weight of the geometric position sensor data; Under the dynamic response working condition, the weight of the dynamic response sensor data is increased.

8. The rail transit digital twin system based on contact network characteristic data according to claim 1 is characterized in that: The internal model parameters are divided into parameter subsets of different spatiotemporal scales.

9. The rail transit digital twin system based on contact network characteristic data according to claim 1 is characterized in that: The parameter correction module adopts a hierarchical correction strategy, and uses different correction frequencies and data ranges to perform corrections on parameter subsets of different spatiotemporal scales.

10. A rail transit digital twin method based on contact network characteristic data, according to any one of claims 1 to 9, characterized in that: The following steps are involved: S1. receiving sensor data representing the physical state of the contact network; S2. constructing a cost function based on the sensor data and a predicted output of a stored parameterized digital model having modifiable internal model parameters; S3. Calculating and obtaining correction values ​​for the internal model parameters by optimizing the cost function; S4. Apply the correction value to the parameterized digital model to complete the synchronous update of the model.

Images