High-speed rail frequency shift automatic block virtual-real fusion intelligent simulation system based on digital twinning

Abstract

The invention relates to the field of intelligent simulation of rail transit, and discloses a high-speed rail frequency shift automatic block virtual-real fusion intelligent simulation system based on digital twinning, which comprises a data acquisition module, a digital twinning modeling module, a dynamic prediction module, a multi-agent cooperative control module, a fault chain evolution module and a simulation verification module. The data acquisition module acquires state parameters in real time; the digital twinning modeling module establishes a virtual-real mapping basis; the dynamic prediction module realizes multi-step state extrapolation; the multi-agent cooperative control module generates a section cooperative strategy conforming to the four-display block logic; the fault chain type evolution module dynamically adjusts the weight of a fault propagation path; and the simulation verification module realizes virtual and real consistency maintenance. The system can complete block equipment full-state sensing, control strategy closed-loop verification, compound fault dynamic deduction and system reliability evaluation, and provides a complete simulation verification system from a physical entity to a digital space for a high-speed rail block system.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of intelligent simulation of rail transit, in particular to a high-speed rail frequency shift automatic block virtual-real fusion intelligent simulation system based on digital twinning. BACKGROUND

[0002] The simulation verification technology of the high-speed rail block system is the core support for ensuring the safety of train operation. The traditional method mainly adopts the mode of combining mechanism model offline simulation with device-level independent test.

[0003] However, the asynchronous updating mechanism of physical devices and virtual models makes it difficult for the simulation system to track the state drift of the devices in real time, especially in the multi-device coupling scenario. The deviation between the model simplification assumption and the real physical constraint may cause the risk of control strategy mismatch. The existing simulation platform mainly relies on pre-defined scripts in fault evolution simulation, and cannot dynamically generate a propagation path that conforms to the actual physical connection relationship between devices, which limits the coverage of emergency plan testing.

[0004] In addition, the fragmented design of control instruction generation and virtual-real consistency verification makes the system lack closed-loop self-correction capability, and it is easy to accumulate model errors in long-term operation.

[0005] These limitations restrict the credibility and practical value of the simulation system in complex operating environments. SUMMARY

[0006] In view of the deficiencies of the prior art, the present application provides a high-speed rail frequency shift automatic block virtual-real fusion intelligent simulation system based on digital twinning, which overcomes the defects of asynchronous virtual-real state, static pre-set fault evolution path and lack of closed-loop fidelity of control strategy in traditional simulation methods, and realizes integrated intelligent simulation of multi-device coupling state dynamic mapping, composite fault causality simulation and control instruction closed-loop verification.

[0007] To achieve the above purpose, the present application realizes the following technical scheme: a high-speed rail frequency shift automatic block virtual-real fusion intelligent simulation system based on digital twinning, comprising:

[0008] A data acquisition module is used to acquire the electrical parameters and logical states of the physical devices in the high-speed rail frequency shift automatic block system in real time, and output the acquired data to the digital twinning modeling module;

[0009] A digital twinning modeling module is used to construct a virtual-real mapping space-time state matrix and embed the physical constraints of the block system, and the output thereof is connected to a dynamic prediction module and a multi-agent collaborative control module;

[0010] A dynamic prediction module is used to dynamically predict the device state based on tensor decomposition method, and input the prediction result to the multi-agent collaborative control module as prior information for collaborative strategy optimization;

[0011] The multi-agent cooperative control module achieves cooperative control of multiple block sections through a distributed optimization algorithm, and the control strategy it generates is simultaneously output to the fault chain evolution module and the simulation verification module.

[0012] The fault chain evolution module generates composite fault scenarios based on the fault propagation graph and dynamic evolution equation. Its output fault path data is fed back to the simulation verification module for logical rationality verification.

[0013] The simulation verification module is used to perform consistency verification of virtual and real data and closed-loop control correction. Its output model correction parameters are fed back into the digital twin modeling module to complete the closed-loop update, and the verified control commands are fed back to the multi-agent collaborative control module to realize strategy iterative optimization.

[0014] Preferably, the data acquisition module includes:

[0015] The frequency shifter parameter acquisition unit is used to measure the carrier frequency, low frequency, and output current.

[0016] The track circuit parameter acquisition unit is used to calculate transmission attenuation and measure the inlet current and outlet current;

[0017] The relay status acquisition unit is used to monitor the contact status and coil voltage.

[0018] Preferably, in the digital twin modeling module:

[0019] The spatiotemporal state matrix Defined as:

[0020]

[0021] Where N is the total number of devices, T is the predefined time window length, and Δt is the sampling interval. α (i) , Let j represent the carrier frequency, transmission attenuation, and coil voltage of device i, respectively, where j = 0, 1, ..., T-1, and t is the current timestamp.

[0022] The physical constraints include:

[0023] Frequency shift signal constraint: carrier frequency f c With low frequency f l The combination must satisfy the preset blocking logic rules;

[0024] Track circuit attenuation constraint: The transmission attenuation α and the segment length L satisfy α=k·L, where k is the material property coefficient.

[0025] Preferably, the step of the dynamic prediction module dynamically predicting the device status includes:

[0026] The spatiotemporal state matrix is ​​expanded into a three-dimensional tensor. Where D is the state parameter dimension, including carrier frequency, transmission attenuation and coil voltage;

[0027] Perform CP decomposition on the three-dimensional tensor:

[0028]

[0029] in, This represents the latent feature vector of the r-th device, reflecting the correlation between devices; This represents the r-th time-series latent feature vector, reflecting the temporal changes in state. Let R represent the latent feature vector of the r-th state, reflecting the multi-parameter coupling relationship; R is the decomposition rank, which is dynamically adjusted according to the physical topology of the block system.

[0030] Predicting the device state at the next moment based on the decomposition results:

[0031]

[0032] Where ∈(t) is the Gaussian noise term.

[0033] Preferably, the multi-agent cooperative control module achieves cooperative control of multiple block sections through a distributed optimization algorithm, including the following steps:

[0034] Each block section is defined as an independent intelligent agent, whose control inputs include frequency shift parameter adjustment amounts, and whose outputs include track circuit current and signal display status.

[0035] Design a distributed optimization objective function, which includes four explicit blocking logic constraints, namely, the reference output is dynamically set according to the number of available blocks ahead.

[0036] The distributed solution is achieved using the alternating direction multiplier method.

[0037] Preferably, the objective function is:

[0038]

[0039] Where N1 is the total number of agents, u i y represents the control input of agent i. i This represents the output of agent i. This represents the reference output of agent i. Let represent the set of segments adjacent to agent i. λ is the index for adjacent segments, and μ is the weighting coefficient.

[0040] Preferably, the fault chain evolution module generates composite fault scenarios based on fault propagation graphs and dynamic evolution equations, including the following steps:

[0041] Construct a directed fault propagation graph G = (V, E), where node v ∈ V represents a device, and edge (u, v) ∈ E represents the path of fault propagation from device u to device v;

[0042] edge weight w uv Initialize based on the physical connection relationships between devices, satisfying... d(u,v) is the physical distance between devices u and v;

[0043] Define a dynamic fault propagation equation to describe the temporal evolution of node fault probability through differential equations;

[0044] The edge weights are dynamically adjusted based on a reinforcement learning strategy, while the causality of fault propagation is constrained.

[0045] Preferably, the dynamic fault propagation equation is:

[0046]

[0047] Where, x v x(t) represents the probability of device v failing at time t. u (t) represents the failure probability of device u at time t, k v This represents the self-recovery coefficient of device v, where Δt is a predefined time step.

[0048] Preferably, the step of dynamically adjusting edge weights based on a reinforcement learning strategy while constraining the causality of fault propagation includes:

[0049] Define the reinforcement learning state space S(t) = {x v (t),w uv (t),E}, which includes the current device failure probability, edge weights and propagation graph structure;

[0050] Define the action space A(t) = {Δw} uv} represents the adjustment amount of the edge weight;

[0051] Design the reward function R(t) = β1·Accuracy(t) - β2·CausalityViolation(t), where:

[0052] Accuracy(t) is the degree of matching between the fault prediction result and the actual fault.

[0053] CausalityViolation(t) counts the number of edge weight adjustments for violating causality.

[0054] β1 and β2 are the fault prediction accuracy weight and causality constraint weight, respectively;

[0055] The edge weights are updated using the policy gradient algorithm, with the following constraints:

[0056] w uv (t+1)=w uv (t)+Δw uv (t)

[0057] and

[0058] (u,v)∈E DAG

[0059] Among them, E DAG Given a directed acyclic graph with edges, backpropagation of edges is prohibited. Weight adjustment.

[0060] Preferably, the simulation verification module includes:

[0061] The forward verification unit is used to verify the consistency of virtual operation control over physical devices;

[0062] The reverse verification unit is used to correct the virtual model based on sudden changes in the physical device state.

[0063] This invention provides a high-speed rail frequency-shift automatic block signaling virtual-real fusion intelligent simulation system based on digital twins.

[0064] It has the following beneficial effects:

[0065] 1. This invention employs a two-way closed-loop mechanism of forward control command verification and reverse model parameter correction to detect and eliminate state deviations between the digital twin model and the physical equipment in real time. The dynamic calibration algorithm based on residual Mahalanobis distance can adaptively adjust the spatiotemporal matrix parameters, ensuring that the virtual model accurately reflects the actual operating state of the equipment and providing a reliable simulation environment for control strategy verification.

[0066] 2. This invention employs a chain-like evolutionary model coupled with directed fault propagation graphs and reinforcement learning, capable of generating complex fault scenarios involving multiple devices and modes. A causal constraint mechanism based on physical topology ensures that fault paths conform to actual logic, supporting comprehensive testing and optimization of emergency plans for train block systems.

[0067] 3. This invention decomposes the global optimization problem into local subproblems and solves them in parallel using the distributed ADMM algorithm. Combined with the adjacency communication topology to limit the information exchange range, it can reduce the communication overhead and computational delay of large-scale block section collaborative control and meet the millisecond-level response requirements of high-speed rail systems.

[0068] 4. The spatiotemporal state matrix of this invention integrates device parameters, temporal evolution and physical coupling relationships, and uses CP tensor decomposition to achieve heterogeneous feature decoupling. This can overcome the information loss problem caused by traditional single-dimensional modeling and provide high-fidelity input data for dynamic prediction.

[0069] 5. This invention provides a complete technical chain from physical equipment data acquisition, digital twin modeling, dynamic control optimization to fault evolution inference, forming a closed-loop verification system driven by both virtual and real factors, which can improve the reliability of the high-speed rail block system throughout its entire life cycle from perception to decision-making to execution to verification. Attached Figure Description

[0070] Figure 1 This is a schematic diagram of the system architecture of the present invention;

[0071] Figure 2 This is a schematic diagram of the CP tensor decomposition process of the dynamic prediction module of the present invention.

[0072] Figure 3 This is a schematic diagram of the ADMM iterative process of the multi-agent cooperative control module of the present invention.

[0073] Figure 4 This is a schematic diagram of the interaction timing of the virtual-real closed-loop verification of the simulation verification module of the present invention. Detailed Implementation

[0074] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0075] Please see the appendix Figure 1 - Appendix Figure 4 This invention provides a high-speed rail frequency-shift automatic block signaling virtual-real fusion intelligent simulation system based on digital twins. The system realizes state prediction, cooperative control and fault evolution simulation of the block system by constructing a real-time interactive closed loop between physical equipment and virtual models.

[0076] like Figure 1 As shown, the high-speed rail frequency-shift automatic block signaling virtual-real fusion intelligent simulation system based on digital twins includes a data acquisition module, a digital twin modeling module, a dynamic prediction module, a multi-agent collaborative control module, a fault chain evolution module, and a simulation verification module.

[0077] The following is a detailed description of each module in the system of this invention, providing a comprehensive explanation of the specific implementation principles, technical details, and processes of each module.

[0078] Data acquisition module

[0079] This invention's data acquisition module utilizes multi-source heterogeneous sensing technology to achieve state perception of physical equipment in a high-speed rail frequency-shift automatic block system. This module constructs a closed-loop data acquisition system, providing real-time input for subsequent digital twin modeling. The specific implementation process is as follows:

[0080] The frequency shifter's parameter acquisition unit extracts signal features based on a high-frequency sampling circuit. For example, the carrier frequency parameter acquisition employs a combination architecture of an anti-aliasing filter and a high-speed AD converter, with a sampling frequency range set to 1-10kHz to accommodate different block systems. For the carrier frequency signal f... c (t) and low-frequency modulation signal f l (t), the spectral energy distribution is calculated using Fast Fourier Transform:

[0081]

[0082] Where s(t) is the time-domain sampled signal, T is the signal period, and f is the analysis frequency. This transformation enables the separation and detection of carrier frequency and low-frequency components, avoiding incorrect parameter extraction caused by harmonic interference.

[0083] Output current parameter I out (t) is measured by a Hall sensor, and its sensitivity coefficient is preferably 0.1V / A.

[0084] The track circuit parameter acquisition unit uses a transmission line theory model to establish attenuation calculation relationships. The input current I is defined. in (t) and output current I out The ratio of (t) is the transmission coefficient:

[0085]

[0086] In the formula, α(t) represents the transmission loss of the track circuit, and its calculation process incorporates a temperature compensation algorithm to eliminate environmental interference. A high-precision current transformer is preferably used for inlet current measurement, with its phase error controlled within ±0.5° to ensure the accuracy of complex impedance calculation.

[0087] The relay status acquisition unit integrates digital quantity detection and analog quantity acquisition functions. Contact status S contact (t) is converted to a TTL level signal through an opto-isolation circuit, with a rise time of less than 10μs to meet real-time requirements. Coil voltage V coil (t) Measurements are performed using a differential amplifier circuit, with a common-mode rejection ratio preferably greater than 80dB. This unit establishes a timing correlation model between contact action and coil excitation:

[0088]

[0089] This model can accurately calibrate the mechanical action delay time T of a relay. delay This provides timing characteristics for fault diagnosis.

[0090] Each acquisition unit achieves data synchronization through a unified clock source, with timestamp deviation controlled within 1ms. Preferably, a distributed synchronization network is established using the IEEE 1588 precision clock protocol, where the master clock node periodically broadcasts synchronization messages to calibrate the clock offset of the slave nodes. This mechanism can eliminate time base jitter in data acquired by multiple devices, ensuring the timing consistency of the spatiotemporal state matrix construction.

[0091] The above implementation method ensures the physical authenticity of the collected data through a multi-level data verification mechanism, and its output standardized data stream provides initial boundary conditions for the digital twin model.

[0092] Digital Twin Modeling Module

[0093] The digital twin modeling module of this invention achieves high-fidelity mapping between virtual and real systems through spatiotemporal state matrix construction and physical constraint embedding. This module transforms multi-source heterogeneous acquired data into structured representations, providing a mathematical model foundation for dynamic prediction and control. The specific implementation process is as follows:

[0094] The spatiotemporal state matrix is ​​constructed to meet the need for heterogeneous data fusion from multiple devices. Since the parameter types and dimensions of frequency shifting devices, track circuits, and relays differ fundamentally, directly mixing and storing them would lead to dimensional chaos in the model. Therefore, a device type block encoding method is adopted to establish device category identifiers in the matrix row dimension, ensuring that parameters of the same type of device are continuously distributed in the row space. The time window length T is set considering the inertial characteristics of device state changes; preferably, the window length is determined based on the signal period (e.g., a 2-second period for a frequency shifting signal) to ensure complete coverage of the dominant frequency component of state fluctuations.

[0095] Define matrix Where N represents the total number of frequency shifting devices, track circuits, and relays in the system, and T is the preset time window length, preferably dynamically adjusted according to the equipment status change cycle. Matrix element s i,j (t) Assign values ​​based on differences in equipment physical characteristics:

[0096]

[0097] In the formula, Δt is the data sampling interval, j = 0, 1, ..., T-1 is the time index offset, and t is the current timestamp. This design can preserve the temporal evolution characteristics of the device state, while avoiding dimensional conflicts between different types of parameters through dimensional separation.

[0098] The embedding of frequency-shift signal combination constraints stems from the mandatory compliance requirements of blocking logic rules. In actual systems, the combination of carrier frequency and low-frequency signals directly determines the signal controller's display state; if the virtual model generates an illegal combination, the simulation results will fail. For the signal combination constraints of frequency-shifting devices, the carrier frequency f is defined as follows: c With low frequency f l Permissible set:

[0099]

[0100] Among the set elements (f) c,k ,f l,k The corresponding four-display blocking logic's permission signal combination, for example, limiting f under yellow light conditions. c =550Hz and f l =18Hz. This constraint is enforced during matrix updates via the projection operator, mathematically equivalent to mapping illegal signal parameters to the most recently permitted value, ensuring that the virtual model output always conforms to the operating rules.

[0101] The linear relationship of track circuit attenuation constraints is derived based on transmission line theory. Ideally, the attenuation coefficient should be proportional to the section length, but in reality, impedance discontinuities at rail joints can lead to nonlinear deviations. The linear relationship between the attenuation coefficient α and the section length L is defined as follows:

[0102] α=k·L+∈

[0103] In the formula, k is the material property coefficient, and ∈ represents the environmental noise term. This constraint is added to the optimization function as a regularization term, allowing for a moderate deviation while ensuring the physical rationality of the model, thus avoiding model rigidity caused by overfitting.

[0104] The sliding window mechanism of the matrix dynamic update algorithm considers a balance between computational efficiency and information integrity. While a fixed window length can reduce computational complexity, it cannot adapt to the time-varying characteristics of different devices. Therefore, an adaptive window adjustment strategy is preferred, where the window length is adjusted every time a new sampling time t is acquired. n Based on the data, perform the following operations:

[0105] 1. Remove the oldest time slice t from the matrix. n-T Corresponding column vector

[0106] 10 will t n Moment α (i) (t n ), Insert a new column by device type

[0107] 3. Correcting historical data bias using constrained least squares method:

[0108]

[0109] Where Φ(·) is the physical constraint violation metric function, and λ is the tradeoff coefficient. This process can eliminate the accumulated error of sensor noise while ensuring the spatiotemporal continuity of the matrix.

[0110] The residual threshold for virtual-to-real mapping verification is set based on the criticality level of the equipment. For equipment such as relays related to driving safety, a strict threshold (e.g., 0.05 pu) is set; for non-safety-related parameters such as frequency shifting equipment, a more lenient threshold (e.g., 0.1 pu) is used. Define the state reconstruction residual:

[0111]

[0112] In the formula s represents the actual measured value from the equipment. i (t) represents the corresponding element in the matrix. When the residual exceeds the threshold, the constraint weight adaptive adjustment mechanism is triggered. Preferably, the regularization coefficient λ is dynamically adjusted using a PID control algorithm to form a closed-loop correction circuit.

[0113] This implementation method achieves a reliable two-way mapping between the digital twin model and the physical system through multi-level constraint coupling and dynamic optimization mechanisms. The structured design of the spatiotemporal state matrix is ​​compatible with the dimensionality expansion requirements of subsequent tensor decomposition algorithms, and its physical constraint embedding method provides initial conditions that conform to actual laws for fault evolution simulation.

[0114] Dynamic prediction module

[0115] This invention's dynamic prediction module achieves multi-dimensional coupled prediction of equipment state based on tensor decomposition theory. This module transforms the spatiotemporal state matrix output from digital twin modeling into analytical feature components, and estimates future states through extrapolation of latent feature vectors. The specific implementation process is as follows:

[0116] The three-dimensional tensor extension design stems from the need for multi-parameter correlation analysis. Because device state parameters (carrier frequency, attenuation, voltage) have physically coupled characteristics, traditional two-dimensional matrices cannot characterize their interactions. Therefore, the spatiotemporal state matrix is ​​extended... Expanded into a three-dimensional tensor The third dimension, D=3, corresponds to the parameter category index. This structure can explicitly preserve nonlinear relationships such as the modulation effect of the carrier frequency on attenuation, providing complete input for subsequent decomposition.

[0117] The choice of CP tensor decomposition is based on the requirements of feature decoupling and dimensionality reduction. The tensor decomposition form is defined as follows:

[0118]

[0119] In the formula:

[0120] The latent feature vectors representing devices, whose element values ​​reflect the strength of the association between devices;

[0121] Characterizes latent feature vectors over time, describing common temporal patterns of state changes;

[0122] Characterizing the hidden eigenvectors of parameters reveals the coupling weights between multiple physical quantities.

[0123] The decomposition rank R is preferably dynamically adjusted based on the system topology complexity, for example, by increasing R when adding new devices to maintain decomposition accuracy.

[0124] The latent feature vector initialization strategy employs randomized orthogonal projection. To avoid local optima, the tensor is first initialized... Perform High-Order Singular Value Decomposition (HOSVD) to obtain the initial projection direction:

[0125]

[0126] Among them U (1) U (1) U (3) For the factor matrix, For the core tensor. Let U (1) The first R columns are used as {a r The initial value is determined similarly for {b}. r},{c r This strategy can accelerate the convergence speed of the ALS (Alternating Least Squares) algorithm.

[0127] The dynamic prediction model is constructed based on latent feature extrapolation and noise injection mechanisms. To achieve multi-step prediction, this implementation modifies the temporal latent feature vector b. r Perform trend extrapolation:

[0128] b r (t+1)=b r (t)+Φ(b r (t-τ:t))

[0129] Where Φ(·) is the ARIMA prediction operator, and τ is the backtracking time window length. Substituting the extrapolation result into the CP decomposition form yields the prediction tensor:

[0130]

[0131] The covariance matrix of the noise term ∈(t) can simulate the effects of unmodeled dynamics by estimating the historical residuals.

[0132] A decomposition rank adaptive adjustment mechanism is used to handle system topology changes. When a new device is detected, this implementation expands the device dimension to N+1 and initializes the corresponding row vector:

[0133]

[0134] in c is the set of physical adjacent devices for the newly added device. adj This is the adjacency coupling coefficient vector. This strategy utilizes prior knowledge of the feature distribution of existing devices to reduce model oscillations.

[0135] The prediction results are validated using statistical process control methods. The standardized residual Mahalanobis distance is defined as follows:

[0136]

[0137] In the formula For the residual vector, This represents the historical residual mean. When D... M Model recalibration is triggered when (t) > 3σ, where σ is the standard deviation of the residuals. This verification mechanism can effectively identify equipment malfunctions or model mismatch issues.

[0138] The above implementation achieves high-precision prediction of multi-parameter coupled states through the synergistic effect of tensor decomposition and dynamic extrapolation. The three-dimensional tensor structure is inherited from the spatiotemporal matrix output of digital twin modeling, the CP decomposition method decouples complex coupling relationships into interpretable latent feature components, and the dynamic adjustment mechanism ensures the model's adaptability to changes in system topology. These technical components form a closed-loop optimization chain, providing reliable predictive input for subsequent coordinated control.

[0139] Multi-agent cooperative control module

[0140] This invention's multi-agent cooperative control module achieves autonomous and coordinated control of high-speed railway block sections through distributed optimization. The module models each block section as an independent agent, achieving a globally optimal control strategy through information exchange and collaborative computation. The specific implementation process is as follows:

[0141] A multi-agent architecture is designed to address the scalability issues of centralized control. Because high-speed rail lines contain numerous physically dispersed block sections, traditional centralized controllers face communication latency and computational bottlenecks. Therefore, this implementation defines each block section as an independent agent, with its control input u... i Includes frequency shift parameter adjustment Δf c (i) Low-frequency modulation amount Δf l (i), output y i Includes track circuit current I in (i), I out (i) and the status displayed by the signal controller. This distributed architecture enables parallel processing of control tasks, reducing system response time.

[0142] The objective function is constructed by integrating local performance optimization and global consistency constraints. To achieve coordination in control of adjacent sections, the optimization objective is defined as follows:

[0143]

[0144] In the formula, N represents the total number of agents. Let represent the set of segments adjacent to agent i. The first term ensures that each agent tracks the reference output, the second term suppresses abrupt changes in control input, and the third term forces a smooth transition between adjacent control strategies. The parameters λ and μ are preferably determined through sensitivity analysis to balance the requirements of control accuracy and stability.

[0145] The reference output dynamic adjustment mechanism is inherited from the four-signal blocking logic rule. Define the reference output. Generation rules:

[0146]

[0147] in This represents the number of consecutive free segments in front of agent i. The mapping function ψ(·) converts the number of free segments into a signal indicating the status. For example, when Output green light signal at time. A yellow light signal is output at certain times. This mechanism ensures that the control strategy complies with railway operation safety regulations and avoids signal display conflicts.

[0148] The distributed solution algorithm employs the Alternating Direction Multiplier Method (ADMM) for parallel computation. First, the global objective function is decomposed into local subproblems:

[0149]

[0150] Where z ij This is the consensus variable for the control variables of neighboring agents. This is achieved by introducing the Lagrange multiplier η. ij Construct the augmented Lagrange function:

[0151]

[0152] The iterative update consists of three steps: solving for local control variables, updating consensus variables, and correcting multipliers, until the residual convergence condition is met. This algorithm can approximate the global optimal solution without centralizing all the data.

[0153] Control constraint processing ensures execution feasibility through projection operators. Physical limiting constraints are defined for the frequency shift parameter adjustment:

[0154]

[0155] During the ADMM iteration process, a projection operation is performed after each local optimization:

[0156]

[0157] in Project the solution onto the feasible set. This operation prevents the algorithm from generating control commands that exceed the physical limits of the device, ensuring the feasibility of the control strategy.

[0158] Communication topology construction is based on the physical connectivity of block segments. An adjacency matrix is ​​defined. Where element a ij =1 if and only if there is a direct track circuit connection between segments i and j. This topology restricts information exchange to occur only between physically adjacent segments, reducing communication load. Preferably, a periodic broadcast mechanism is used to synchronize control quantities, with the broadcast interval consistent with the heartbeat period of the block system.

[0159] The stability of the control strategy was verified using Lyapunov function analysis. The energy function was constructed as follows:

[0160]

[0161] Prove that during the ADMM iteration process, V(u) (k+1) )≤V(u (k) This ensures the convergence of the control sequence. This analysis provides a theoretical stability guarantee for the engineering application of distributed algorithms.

[0162] The above implementation method, through objective function decomposition and parallel optimization mechanisms, enables coordinated control of large-scale block sections. The multi-agent architecture inherits from the system's physical distribution characteristics, the ADMM algorithm fully utilizes local computing resources, and the dynamic reference adjustment mechanism ensures that the control strategy conforms to operational rules. All technical components form a closed-loop optimization link, providing adaptive control capabilities for the high-speed rail block system.

[0163] Fault chain evolution module

[0164] This invention's fault chain evolution module achieves causal constraint simulation of complex fault scenarios through a directed graph model and dynamic equations. This module constructs a fault propagation network based on the physical relationships between devices and dynamically optimizes the propagation path weights using a reinforcement learning strategy. The specific implementation process is as follows:

[0165] The fault propagation graph is constructed based on the causal dependencies of the equipment's physical topology. Since fault propagation in real-world systems is directional (e.g., a relay fault leads to track circuit failure, but the reverse is not true), this implementation defines a directed graph G = (V, E), where the node set V contains all equipment entities, and the edge set E is determined by the physical connections between the devices. Edge weight initialization follows the inverse distance principle:

[0166]

[0167] In the formula, d(u,v) represents the physical distance between devices u and v, and γ is the normalization coefficient. This initialization strategy reflects the actual pattern that the probability of fault propagation is higher between physically adjacent devices, providing a reasonable initial value for subsequent dynamic adjustments.

[0168] The dynamic fault propagation equation is designed to quantify the time-series cumulative effect of fault probability. The discretized differential equation is defined as follows:

[0169]

[0170] Where x v (t)∈[0,1] represents the failure probability of device v at time t, k v The self-recovery coefficient is determined by the equipment reliability parameters. This equation models the decay of the failure probability (by -k). v x v Item) and adjacent device fault injection (by ∑w uv x u The competition mechanism of the items can simulate complex chain reactions of failures.

[0171] Reinforcement learning constructs a state space that integrates the real-time state and topology information of the system. State vectors are defined as follows:

[0172] S(t)={x v (t),w uv (t),E}

[0173] Where x v (t) represents the current failure probability of each device, w uv (t) represents the edge weight matrix, and E represents the set of edges in the propagation graph. This state representation encapsulates the complete dynamic characteristics of the system, providing an information basis for agent decision-making. The action space is defined as the edge weight adjustment amount:

[0174] A(t)={Δw uv (t)∈[-0.1w uv (t), 0.1w uv (t)]}

[0175] Action amplitude constraints prevent excessive single-step adjustments from causing model instability, while maintaining non-negativity of weights to conform to probability propagation semantics.

[0176] The reward function design balances prediction accuracy with causal logic constraints. A multi-objective reward is defined:

[0177] R(t)=β1·Accuracy(t)-β2·CausalityViolation(t)

[0178] Accuracy(t) is calculated by comparing the set of predicted faulty devices. Calculated with the actual fault set F(t):

[0179]

[0180] CausalityViolation(t) counts the number of edge weight adjustments for violating causality, such as for reverse edges. Illegal modifications. The coefficients β1 and β2 are preferably determined through Pareto front analysis to achieve the optimal trade-off between accuracy and logical rationality.

[0181] The policy gradient algorithm is applied to achieve online adaptive adjustment of edge weights. The update rules for the policy network parameters θ are defined as follows:

[0182]

[0183] Q π (s t ,a t The state-action value function is estimated and updated through temporal difference learning. The action selection probability π... θ (a t |s t The algorithm is generated by a Gaussian strategy, and its mean and variance are output by a neural network. This algorithm can explore the optimal propagation path without relying on pre-stored fault samples.

[0184] Enforcing causal constraints is achieved through topology graph editing operations. Constraint validation is performed after weight adjustment.

[0185]

[0186] Where E DAG Given a directed acyclic graph edge set, loop edges are detected and removed through periodic topological sorting. This mechanism ensures that fault propagation paths always conform to the causal logic of the actual system, avoiding inconsistencies in simulation results.

[0187] Complex fault scenarios are generated based on the Monte Carlo sampling method. The following steps are performed at each simulation step:

[0188] 1. Perform Bernoulli sampling on the equipment fault state based on the current xv(t);

[0189] 2. If device VV fails, it triggers a failure probability update for its downstream device UU:

[0190] x u (t+1)=min(1,x u (t)+w vu (t)·x v (t))

[0191] 3. Record the sequence of fault propagation paths to form a complex fault chain;

[0192] This process can generate complex scenarios that include multiple devices and multiple failure modes, providing test cases for emergency strategy verification.

[0193] The above implementation method achieves adaptive simulation of fault propagation paths through the collaborative optimization of dynamic equations and reinforcement learning. The directed graph model inherits from the physical connection relationships of the equipment, the dynamic equations quantify the cumulative effect of faults, and the reinforcement learning strategy explores the propagation law under causal constraints.

[0194] Simulation verification module

[0195] The simulation verification module of this invention maintains the consistency between the digital twin model and the physical system through a two-way closed-loop verification mechanism. This module constructs a collaborative link between forward control verification and reverse model correction, and the specific implementation process is as follows:

[0196] The forward verification unit is designed to ensure the physical executability of virtual control commands. Since the virtual control quantities generated by the digital twin model may exceed the physical limits of the actual device, this implementation defines a control deviation measurement function:

[0197]

[0198] In the formula u v (t) is the virtual control command vector, u p (t) is the actual device response vector, and ∈ is a small constant to prevent division by zero. When δ(t) > δ th When this occurs, the control command regeneration process is triggered. This mechanism can identify control mismatch problems caused by model simplification or parameter drift.

[0199] The reverse verification unit constructs a virtual model with online correction capabilities based on residual analysis. Define the state residual vector:

[0200] r(t) = x p (t)-x v (t)

[0201] Where x p (t) represents the measured state of the physical equipment, x v (t) represents the output state of the virtual model. Abnormal residuals are detected using Mahalanobis distance.

[0202]

[0203] In the formula, Σ is the historical residual covariance matrix. When D M When (t)>3σ, initiate the calibration of digital twin model parameters, preferably using the recursive least squares method to update the spatiotemporal state matrix elements.

[0204] The control command regeneration process integrates model prediction and equipment constraints. When the positive verification exceeds the limit, a constrained quadratic programming problem is constructed:

[0205]

[0206] in y is the set of allowable control variables for the device. d The desired output is obtained by solving the problem using the effective set method, which generates the corrected safety control instructions.

[0207] The model parameter calibration mechanism enables dynamic adaptation of the virtual system. For residual anomalies detected by reverse verification, parameter update rules are defined:

[0208] θ(t+1)=θ(t)+K(t)r(t)

[0209] The Kalman gain matrix K(t) is calculated recursively using the following formula:

[0210] K(t)=P(t)H T (HP(t)H T +R) -1

[0211] In the formula, H is the observation matrix, R is the measurement noise covariance, and P(t) is the parameter error covariance matrix. This algorithm can stably update the model parameters under noisy conditions.

[0212] Timing alignment addresses the issue of asynchronous timing between virtual and real data. A timestamp compensation function is defined as follows:

[0213] x′ v (t)=x v (t-Δτ)

[0214] The time delay estimate Δτ is calculated using the cross-correlation function:

[0215]

[0216] This process eliminates false discrepancies caused by communication delays, improving the reliability of verification results.

[0217] The rule base supports multi-level exception handling strategies. Define the mapping relationship between rule triggering conditions and corresponding actions:

[0218] IFδ(t)∈[δ k ,δ k+1 THEEN Action k

[0219] The action set includes hierarchical responses such as instruction fine-tuning, local model updates, and global reconstruction. The rule interval δ kThe preferred approach is to classify equipment according to its safety level to achieve refined control over anomaly handling.

[0220] The above implementation method achieves dynamic consistency maintenance between the digital twin system and physical devices through bidirectional data stream collaborative verification. Forward verification inherits the output instructions from the multi-agent collaborative control module, while reverse correction and residual detection from the fault evolution module share computing resources. The timing alignment mechanism eliminates time base errors in cross-module data interaction.

[0221] Simulation example:

[0222] Test environment configuration

[0223] A simulation platform was built based on the block system of a high-speed railway line, including 12 block sections, 48 ​​sets of frequency shifting devices, and 156 relay nodes. The test data came from fault logs and sensor historical records in actual operation. The simulation duration was set to 72 hours, and the hardware platform used an Intel Xeon 6248R processor and an NVIDIA A100 GPU cluster.

[0224] 1. Performance verification of the dynamic prediction module

[0225] We employ CP tensor decomposition (rank=8) combined with an ARIMA (2,1,2) model to compare the prediction accuracy of traditional LSTM (128 hidden layers) and pure ARIMA methods:

[0226]

[0227] CP tensor decomposition decouples the device state tensor (device × time × parameters) into low-dimensional features, and the ARIMA operator extrapolates the temporal latent features, solving the problem of insufficient processing of multi-dimensional coupled data by traditional methods. The multi-step prediction error growth rate is reduced by 42%, verifying the effectiveness of tensor decomposition for long-term prediction.

[0228] 2. Multi-agent cooperative control efficiency test

[0229] Comparing the performance of distributed ADMM (Adjacency Communication Topology), centralized QP optimization, and conventional distributed gradient descent:

[0230]

[0231]

[0232] ADMM decomposes the global optimization into local subproblems (with the objective function including adjacent segment consistency constraints), making the communication overhead linearly dependent only on the number of adjacent nodes. Tests show that, in the 12-segment scenario, ADMM's convergence time decreases by 58% with the rate of increase in the number of segments compared to gradient descent.

[0233] 3. Fault chain evolution coverage analysis

[0234] Based on a directed fault propagation graph (nodes = 156, edges = 384) and a reinforcement learning dynamic adjustment strategy, the coverage of a predefined script and a static weight propagation graph are compared:

[0235]

[0236] The reinforcement learning strategy network adjusts the propagation weights w based on the real-time device status. uv This ensures that fault path generation conforms to the actual physical connections between devices (such as track circuit-relay electrical coupling constraints). During testing, 32 novel composite faults not previously recorded were detected, demonstrating the effectiveness of the dynamic evolution mechanism.

[0237] This embodiment verifies the technical advantages of the system in terms of state prediction accuracy (MSE reduced by 35%), control real-time performance (convergence time shortened by 62%), and fault scenario coverage (path number increased by 144%). The synergistic effect of each module meets the high reliability simulation requirements of the high-speed rail block system.

[0238] In summary, this invention acquires multi-dimensional state parameters of frequency shifting equipment, track circuits, and relays in real time through a data acquisition module, constructs a spatiotemporal state matrix to establish a digital twin model, and embeds physical constraint rules. It utilizes tensor decomposition to achieve dynamic prediction of equipment states and combines a multi-agent distributed optimization algorithm to generate collaborative control strategies. Based on a directed fault propagation graph and reinforcement learning mechanism, it simulates the chain evolution path of complex faults. Finally, a bidirectional closed-loop verification module performs consistency verification of virtual and real data and dynamic model correction, forming a full-link simulation system from physical equipment perception and virtual model deduction to closed-loop feedback correction. This system enables real-time state mapping, fault evolution simulation, and adaptive control verification of high-speed rail block systems.

[0239] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A high-speed rail frequency shift automatic block virtual-real fusion intelligent simulation system based on digital twinning, characterized in that, The system comprises: a data acquisition module for real-time acquisition of electrical parameters and logical states of physical devices in a high-speed rail frequency shift automatic block system, and output of the acquired data to a digital twin modeling module; a digital twin modeling module for constructing a virtual-real mapping space-time state matrix and embedding block system physical constraints, the output of which is connected to a dynamic prediction module and a multi-agent collaborative control module; a dynamic prediction module for dynamic prediction of device states based on a tensor decomposition method, and input of the prediction results to the multi-agent collaborative control module as prior information for collaborative strategy optimization; a multi-agent collaborative control module for achieving collaborative control of multiple block sections through a distributed optimization algorithm, and output of the generated control strategy to a fault chain evolution module and a simulation verification module; a fault chain evolution module for generating a complex fault scenario based on a fault propagation graph and a dynamic evolution equation, and feedback of the fault path data output by the module to the simulation verification module for logical reasonableness checking; a simulation verification module for performing virtual-real data consistency verification and closed-loop control correction, and input of the model correction parameters output by the module to the digital twin modeling module for closed-loop updating, and feedback of the checked control instructions to the multi-agent collaborative control module for strategy iterative optimization.

2. The high-speed rail frequency shift automatic block virtual-real fusion intelligent simulation system based on digital twinning according to claim 1, characterized in that, The data acquisition module comprises: a frequency shift device parameter acquisition unit for measuring carrier frequency, low frequency and output current; a track circuit parameter acquisition unit for calculating transmission loss and measuring inlet current and outlet current; a relay state acquisition unit for monitoring contact state and coil voltage.

3. The high-speed rail frequency-shift automatic block based on digital twin virtual-real fusion intelligent simulation system according to claim 1, characterized in that, In the digital twin modeling module: the spatio-temporal state matrix is defined as: where N is the total number of devices, T is the length of the predefined time window, and Δt is the sampling interval, α (i) , denote the carrier frequency, transmission loss, and coil voltage of device i, respectively, j = 0, 1,..., T - 1, and t is the current time stamp. the physical constraints comprise: Frequency shift signal constraint: carrier frequency f c in combination with low frequency f l must satisfy a preset blocking logic rule; a track circuit loss constraint: transmission loss α and section length L satisfy α=k·L, where k is a material characteristic coefficient.

4. The high-speed rail frequency shift automatic block based on digital twin virtual-real fusion intelligent simulation system according to claim 3, characterized in that, The steps for dynamic prediction of device states by the dynamic prediction module comprise: extending the spatio-temporal state matrix into a three-dimensional tensor where D is the dimension of the state parameters, including carrier frequency, transmission loss and coil voltage; CP decomposition of the three-dimensional tensor: wherein, denotes the rth device latent feature vector, reflecting the correlation between devices; denotes the rth time latent feature vector, reflecting the state time series change; denotes the rth state latent feature vector, reflecting the multi-parameter coupling relationship; R is the decomposition rank, which is dynamically adjusted according to the physical topology of the closed system; prediction of device states at the next time based on the decomposition results: where ∈(t) is a Gaussian noise term.

5. The high-speed rail frequency shift automatic block based on digital twin virtual-real fusion intelligent simulation system according to claim 1, characterized in that, The steps for achieving collaborative control of multiple block sections by the multi-agent collaborative control module through a distributed optimization algorithm comprise: defining each block section as an independent agent, with control input including frequency shift parameter adjustment and output including track circuit current and signal display state; designing a distributed optimization objective function, which contains four display block logical constraints, i.e., reference output is dynamically set according to the number of idle block sections in front; solving the distributed optimization problem by using an alternating direction multiplier method.

6. The high-speed rail frequency shift automatic block based on digital twin virtual-real fusion intelligent simulation system according to claim 5, characterized in that, The objective function is: where N1is the total number of agents, u i represents the control input of agent i, y i represents the output of agent i, represents the reference output of agent i, N i represents the set of adjacent segments of agent i, j∈N i is the index of adjacent segments, λ and μ are weight coefficients.

7. The high-speed rail frequency-shift automatic block based on digital twin virtual-real fusion intelligent simulation system according to claim 1, characterized in that, The steps for generating a complex fault scenario based on a fault propagation graph and a dynamic evolution equation by the fault chain evolution module comprise: constructing a directed fault propagation graph G=(V,E), where node v∈V represents a device, and edge (u,v)∈E represents a path for fault propagation from device u to device v; Edge weight w uv According to the initialization of the physical connection relationship between devices, the following is satisfied d(u,v) is the physical distance between devices u and v; defining a dynamic fault propagation equation to describe the time sequence evolution of node fault probability by using a differential equation; dynamically adjusting edge weights based on a reinforcement learning strategy, while constraining the causality of fault propagation.

8. The high-speed rail frequency shift automatic block based on digital twin virtual-real fusion intelligent simulation system according to claim 7, characterized in that, The dynamic fault propagation equation is: where x v (t) denotes the failure probability of device v at time t, x u (t) denotes the failure probability of device u at time t, k v denotes the self-recovery coefficient of device v, and Δt is a predefined time step.

9. The high-speed rail frequency-shift automatic block based on digital twin virtual-real fusion intelligent simulation system according to claim 8, characterized in that, The steps for dynamically adjusting edge weights based on a reinforcement learning strategy, while constraining the causality of fault propagation, comprise: Definition of reinforcement learning state space S(t) = {x v (t), w uv (t), E}, including current device failure probability, edge weight and propagation graph structure; Define the action space A(t) = {Δw uv (t)}, which represents the adjustment amount of the edge weight; The reward function R(t) is designed as R(t) = β1·Accuracy(t) - β2·CausalityViolation(t), wherein: Accuracy(t) is a matching degree of a fault prediction result and an actual fault; CausalityViolation(t) is a number of times of edge weight adjustment violating causality; β1 and β2 are respectively a fault prediction accuracy weight and a causality constraint weight; The edge weight is updated through a policy gradient algorithm, and the constraint condition is: w uv (t+1) = w uv (t) + Δw uv (t) And (u,v) e E DAG where E DAG is a set of directed acyclic graph edges, forbidden backpropagation edges weight adjustments.

10. The high-speed rail frequency shift automatic block based on digital twin virtual-real fusion intelligent simulation system according to claim 1, characterized in that, The simulation verification module comprises: A forward verification unit configured to verify control consistency of the virtual operation on the physical device; A reverse verification unit configured to correct the virtual model according to a state mutation of the physical device.

Images