Method and system for automatically generating LKJ driving simulation real-time data

Abstract

The invention discloses a method and system for automatically generating LKJ driving simulation real-time data, and the method comprises the following steps: S1, collecting multi-source data, and constructing an original data set; s2, initializing a multi-manifold adaptive generative adversarial network, and generating initial latent variable distribution; s3, optimizing the initial latent variable distribution by using a variational reasoning mechanism, and generating initial LKJ driving simulation data; s4, generating real-time driving simulation data in combination with a virtual instantaneous sensing module; s5, adjusting the real-time driving simulation data through a global-local adaptive correction mechanism, and generating driving simulation optimization data; s6, inputting the driving simulation optimization data into the virtual track system, and updating a generator and a correction mechanism according to a verification result; and S7, carrying out dimension reduction processing on the verified driving simulation optimization data, and outputting final driving simulation data. According to the method, the LKJ driving simulation data is accurately generated by combining a multi-manifold generative adversarial network, variational reasoning and other methods.

Description

Technical Field

[0001] The present invention relates to the technical field of rail transit, and particularly to a method and a system for automatically generating real-time LKJ train operation simulation data. Background Art

[0002] With the development of rail transit and the increasing complexity of train operation scheduling, the application of train operation control systems (such as the LKJ system) in railway transportation has become increasingly widespread. The LKJ system provides reliable technical support for train operation scheduling, fault analysis, and safety monitoring by recording dynamic data during train operation, including information such as train speed, position, acceleration, signal status, and track environment. However, in practical applications, due to the high dynamicity of the track environment and train status, as well as a large amount of complex multi-dimensional data, the existing LKJ system faces many technical challenges in data simulation and generation.

[0003] Currently, the generation of LKJ train operation simulation data mostly relies on traditional mathematical modeling or rule-driven simulation techniques, which usually generate data through preset track environments and train operation models. However, these traditional methods have the following defects. First, due to the difficulty of comprehensively modeling the dynamic characteristics of the track environment and train status, the data generated by traditional methods often lack a true reflection of complex environments and dynamic changes, resulting in limited reliability and applicability of the data. Second, rule-driven simulation techniques are inefficient in the face of high-dimensional complex data, especially in scenarios with high real-time requirements, and it is difficult to meet the actual needs of the rail transit system. In addition, traditional methods have insufficient response capabilities to abnormal environments (such as signal anomalies, track damage, extreme weather, etc.), and the generated data cannot truly reflect these abnormal scenarios, further limiting the application value of LKJ train operation simulation data.

[0004] With the rapid development of artificial intelligence technology, deep learning methods such as generative adversarial networks (GANs) have gradually been applied to the field of data generation, providing new ideas for solving the technical problems in LKJ train operation simulation data generation. GANs can generate high-quality data through the adversarial training of generators and discriminators, and are suitable for processing multi-dimensional dynamic characteristics. However, the application of existing GAN technologies in the field of rail transit still faces significant technical bottlenecks. First, traditional GANs mainly model based on a single latent variable space and cannot fully capture the multiple dynamic characteristics of track conditions, train status, and environmental changes, resulting in the lack of relevance and consistency between multi-dimensional characteristics in the generated data. Second, the training process of GANs usually relies on static data distributions, and the dynamic and real-time requirements of the track environment make it impossible to directly apply traditional GANs to the LKJ train operation simulation data generation scenario. In addition, due to the large number of iterative optimizations required for the adversarial training of generators and discriminators, existing GANs have huge computational overhead when processing high-dimensional complex data and are difficult to meet the real-time requirements of the rail transit system.

[0005] In the generation of rail transit simulation data, existing research has attempted to introduce variational inference and manifold learning techniques to optimize the structure and efficiency of data generation models. Variational inference optimizes the latent variables in the generation model by constructing prior and posterior distributions, improving the accuracy and flexibility of data generation; manifold learning simplifies the data modeling process by identifying the low-dimensional nested structure of high-dimensional data. However, when these techniques are applied alone, there are still certain limitations. Variational inference usually ignores the dynamic associations between different data characteristics and is difficult to achieve efficient modeling in multi-modal data scenarios; manifold learning relies on static data distributions and lacks the ability to adapt to real-time dynamic changes. In addition, existing technologies have insufficient real-time perception and dynamic adjustment capabilities of the environment during the data generation process, and the generated data cannot accurately reflect the dynamic characteristics of the rail environment and the actual state of train operation.

[0006] Therefore, how to provide a method and system for automatically generating real-time LKJ train operation simulation data is an urgent problem to be solved by those skilled in the art. Summary of the Invention

[0007] An object of the present invention is to propose a method and system for automatically generating real-time LKJ train operation simulation data. The present invention combines a multi-manifold adaptive generative adversarial network, variational inference, a virtual instantaneous perception module, and a global-local adaptive correction mechanism to model rail conditions, train states, and environmental changes, and accurately generate LKJ train operation simulation data. By dynamically optimizing the latent variable distribution, real-time perceiving the dynamic changes of the rail environment, and verifying the authenticity and adaptability of the generated data, high-precision, high dynamic adaptability, and real-time response capabilities are achieved, providing strong support for the operation optimization, safety monitoring, and dispatching management of rail transit, and having the remarkable advantages of strong real-time performance, high adaptability, and excellent accuracy.

[0008] A method for automatically generating real-time LKJ train operation simulation data according to an embodiment of the present invention includes the following steps:

[0009] S1. Collect multi-source data of the rail environment and train operation, and construct an original data set;

[0010] S2. Based on the original data set, initialize a multi-manifold adaptive generative adversarial network, and construct a rail condition manifold, a train state manifold, and an environmental change manifold in the multi-manifold adaptive generative adversarial network to generate an initial latent variable distribution;

[0011] S3. Use the variational inference mechanism to optimize the initial latent variable distribution, input the optimized latent variable distribution into the generator of the multi-manifold adaptive generative adversarial network, and generate initial LKJ train operation simulation data through the adversarial training between the generator and the discriminator;

[0012] S4. Combine with the virtual instantaneous perception module, and adjust the generation process of the initial LKJ train operation simulation data based on the real-time perception data of the track environment to generate real-time train operation simulation data;

[0013] S5. Through the global-local adaptive correction mechanism, perform global consistency optimization and local environment adjustment on the real-time train operation simulation data to generate optimized train operation simulation data;

[0014] S6. Input the optimized train operation simulation data into the virtual track system constructed by digital twin technology to verify authenticity and adaptability, and update the generator and correction mechanism according to the verification results;

[0015] S7. Perform dimensionality reduction processing on the verified optimized train operation simulation data, map the high-dimensional characteristics of track conditions, train behaviors, and environmental changes to a low-dimensional space, and output the final train operation simulation data.

[0016] Optionally, the S2 specifically includes:

[0017] S21. Classify the features of the original data set, and divide the track condition data, train state data, and environmental change data into independent data sets respectively;

[0018] S22. In the multi-manifold adaptive generative adversarial network, construct a track condition manifold, a train state manifold, and an environmental change manifold respectively. Each manifold corresponds to an independent feature subspace, which is used to represent the internal characteristics and dynamic change relationships of the track condition data, train state data, and environmental change data respectively;

[0019] S23. Use the network embedding layer to encode the data features of each manifold, map the high-dimensional features to the low-dimensional manifold space, and generate the initial feature distributions of the track condition manifold, the train state manifold, and the environmental change manifold by maintaining the structural relationship of the original data in the low-dimensional space;

[0020] S24. Construct a generator in the multi-manifold adaptive generative adversarial network. The generator receives the initial feature distributions of the track condition manifold, the train state manifold, and the environmental change manifold as inputs, and generates latent variables based on the preset network weights and activation functions;

[0021] S25. Construct a discriminator in the multi-manifold adaptive generative adversarial network. The discriminator optimizes the generator parameters through adversarial training;

[0022] S26. After the adversarial training reaches the convergence state, the generator generates the initial latent variable distributions of the track condition manifold, the train state manifold, and the environmental change manifold.

[0023] Optionally, S3 specifically includes:

[0024] S31. Optimize the initial latent variable distribution using a variational inference mechanism, and define the prior distribution and posterior distribution of the initial latent variable;

[0025] S32. Extract the initial mean and variance of the initial latent variable through the encoder network to generate variational posterior distribution parameters;

[0026] S33. Optimize the initial latent variable distribution, and define an extended variational inference optimization objective function:

[0027]

[0028] where L represents the loss function, N represents the total dimension of the initial latent variable, μ i represents the mean of the optimized latent variable, represents the variance of the optimized latent variable, μ prior,i represents the mean of the prior distribution, represents the variance of the prior distribution;

[0029] S34. Adopt an adaptive iterative optimization strategy to dynamically adjust the initial mean and variance of the initial latent variable distribution to generate an optimized latent variable distribution;

[0030] S35. Input the optimized latent variable distribution into the generator of the multi-manifold adaptive generative adversarial network, and generate initial LKJ driving simulation data through the adversarial training between the generator and the discriminator.

[0031] Optionally, S4 specifically includes:

[0032] S41. Construct a virtual instantaneous perception module, which includes a real-time perception unit, a feature extraction unit, a dynamic weight allocation unit, and a feedback adjustment unit;

[0033] S42. Collect dynamic data of the track environment through the real-time perception unit, including signal status, slope change, weather conditions, and track anomalies, and generate an original perception matrix X raw ={x ij}, where x ij represents the value of the environmental feature j at the i-th time step;

[0034] S43. Use the feature extraction unit to perform denoising and dimensionality reduction processing on the original perception matrix to generate a standardized matrix X

[0035] ={x' norm}, where x' ij} ijRepresents the value of environmental feature j at time step i after standardization. Combine with the dynamic weight allocation unit to calculate the comprehensive environmental feature vector:

[0036]

[0037] Among them, X combined Represents the comprehensive environmental feature vector, M represents the total number of features, α j Represents the importance adjustment coefficient of environmental feature j, w j Represents the dynamic weight of environmental feature j, exp represents the exponential function, β j Represents the smoothing factor of environmental feature j, μ j Represents the mean value of environmental feature j;

[0038] S44. Input the comprehensive environmental feature vector X combined And the optimized latent variable distribution Z opt Into the generator:

[0039]

[0040] Among them, Y real-time Represents the real-time driving simulation data, G represents the generator, θ represents the generator parameters, γ represents the step size parameter, Represents the gradient operation;

[0041] S45. Use the feedback adjustment unit to perform dynamic verification on the generated real-time driving simulation data:

[0042]

[0043] Among them, Δθ represents the amplitude of parameter optimization, η represents the learning rate, Q represents the total number of samples, y a Represents the eigenvalue of the generated real-time driving simulation data, x a Represents the eigenvalue of the perception data, Represents the gradient of the generator parameters, L gen Represents the generation loss function, p and q represent the adjustment coefficients, σ a Represents the standard deviation;

[0044] S46. After the feedback adjustment reaches convergence, output the dynamically verified real-time driving simulation data.

[0045] Optionally, the S5 specifically includes:

[0046] S51. Construct a global-local adaptive correction mechanism, which includes a global optimization module and a local adjustment module. The global optimization module is used to optimize the overall characteristics of real-time driving simulation data in a consistent manner, and the local adjustment module is used to further adjust the real-time driving simulation data according to the detailed characteristics of the track environment.

[0047] S52. Use the global optimization module to conduct an overall analysis of the real-time driving simulation data, extract the global features of the real-time driving simulation data, and compare the extracted global features with the preset benchmark features of the track environment to identify the overall characteristic deviation.

[0048] S53. According to the identified overall characteristic deviation, adjust the global parameters of the generator to make the real-time driving simulation data consistent with the benchmark features of the track environment in terms of overall characteristics, and at the same time ensure that the global dynamic changes of the generated data meet the actual requirements of the track environment.

[0049] S54. Through the local adjustment module, conduct a further analysis of the local characteristics of the real-time driving simulation data, extract local features based on the dynamic changes of the track environment, including signal status, slope change, and weather conditions, and compare the local features with the real-time perception data to identify local deviations.

[0050] S55. According to the identified local deviations, optimize and adjust the local parameters of the real-time driving simulation data, and ensure that the local adjustment does not damage the consistency of the overall characteristics.

[0051] S56. Combine the correction results of the global optimization module and the local adjustment module to generate optimized driving simulation data.

[0052] Optionally, the specific content of S6 includes:

[0053] S61. Construct a virtual track system based on digital twin technology. The virtual track system includes a track geometry model, a signal status model, and a train dynamics model, and establish the global dynamic characteristic equation of the virtual track system:

[0054] M virtual =A·T track +B·S signal +C·D train ;

[0055] Among them, M virtual represents the virtual track system model, T track represents the track geometric characteristics, S signal represents the signal status, D train represents the train dynamics characteristics, and A, B, and C represent weight matrices;

[0056] S62. Map the driving simulation optimization data to the virtual track system, generate dynamic operation data through the virtual track system, and the dynamic operation data is jointly driven by a track geometry model, a signal status model, and a train dynamics model:

[0057]

[0058] Among them, Y sim represents the dynamic operation data, ⊙ represents the element-wise multiplication operation, represents the tensor outer product operation;

[0059] S63. Conduct multi-level comparison between the generated dynamic operation data Y sim and the real track data, and construct a dynamic characteristic deviation model:

[0060]

[0061] Among them, ΔY represents the deviation value, L represents the number of characteristic layers, γ l represents the layer weight, represents the dynamic operation data of the l-th layer, represents the real track data of the l-th layer, and κ represents the adjustment parameter;

[0062] S64. Use the dynamic characteristic deviation model to optimize and update the generator parameters and the correction mechanism parameters:

[0063]

[0064] Among them, θ represents the generator parameters, φ represents the correction mechanism parameters, and ξ1 and ξ2 represent the regularization parameters;

[0065] S65. After the deviation value ΔY converges to the set threshold, output the driving simulation optimization data verified by the virtual track system.

[0066] A system for automatically generating real-time LKJ driving simulation data according to an embodiment of the present invention includes the following modules:

[0067] A data acquisition module, which acquires multi-source data of the track environment and train operation, and constructs an original data set;

[0068] A multi-manifold generation module, which initializes a multi-manifold adaptive generative adversarial network based on the original data set, constructs a track condition manifold, a train state manifold, and an environment change manifold, and generates an initial latent variable distribution;

[0069] A variational inference optimization module, which optimizes the initial latent variable distribution to generate an optimized latent variable distribution;

[0070] The virtual instantaneous perception module adjusts the initial LKJ train operation simulation data based on the real-time perception data of the track environment to generate real-time train operation simulation data;

[0071] The global-local correction module generates optimized train operation simulation data through global consistency optimization and local environment adjustment;

[0072] The virtual track verification module inputs the optimized train operation simulation data into the virtual track system to verify authenticity and adaptability, and updates the generator and correction mechanism;

[0073] The dimensionality reduction output module performs dimensionality reduction processing on the verified optimized train operation simulation data and outputs the final train operation simulation data.

[0074] The beneficial effects of the present invention are as follows:

[0075] First of all, by introducing the multi-manifold adaptive generative adversarial network (V-MAGAN), the present invention models the track conditions, train states, and environmental change characteristics as independent manifold spaces respectively, and collaboratively optimizes these manifolds during the generation process. This method breaks through the limitation of the traditional generative adversarial network in modeling a single latent variable space, effectively improving the ability to capture the complex dynamic characteristics of the track environment, so that the generated train operation simulation data can truly reflect the multi-dimensional characteristics and dynamic correlations of the rail transit system. Compared with the existing methods, the data generated by the present invention is significantly improved in terms of track environment adaptability and train state accuracy.

[0076] Secondly, by combining the variational inference mechanism, the present invention realizes dynamic adjustment and flexible modeling during the optimization process of the latent variable distribution. Variational inference optimizes the prior distribution and posterior distribution of the latent variable, improving the adaptability of the generated data to complex track environments and abnormal scenarios. At the same time, the efficient inference ability of variational inference significantly reduces the computational complexity of the generative model, providing reliable technical support for rail transit scenarios with high real-time requirements. This optimization process ensures the accuracy and flexibility of the generated data, especially showing excellent performance when dealing with high-dimensional, multi-modal dynamic data.

[0077] In addition, the virtual instantaneous perception module designed by the present invention realizes the dynamic response ability of the data generation process by real-time perceiving the dynamic changes of the track environment and feeding back to the generator for adjustment. The real-time perception module can capture key characteristics such as signal status, slope change, and weather conditions, and dynamically optimize the generator parameters to ensure that the generated data can timely reflect environmental changes. Compared with traditional methods, this module shows higher sensitivity and adaptability when dealing with sudden scenarios and complex environments.

[0078] Through the global-local adaptive correction mechanism, the present invention achieves a balance between global consistency and local accuracy during the generation process. The global optimization module adjusts the overall characteristics of the generated data to ensure the global dynamic consistency of the track system; the local adjustment module refines the generated data based on the detailed characteristics of the track environment, enabling the data to accurately reflect the local changes in the track environment. This collaborative optimization mechanism between the global and local levels ensures the comprehensive performance of the generated data at both the macroscopic and microscopic levels, and the generated optimized train operation simulation data has been significantly improved in terms of authenticity and application value.

[0079] Finally, based on the virtual track system of the digital twin technology, the present invention verifies the authenticity and adaptability of the generated data. The virtual track system combines the track geometric model, signal state model, and train dynamics model to conduct multi-level comparison and verification of the generated data, ensuring the reliability of the generated data in practical applications. Through the verification feedback mechanism, the parameters of the generator and correction mechanism are further optimized, enabling the generated data to achieve a high degree of accuracy and consistency in reflecting the characteristics of the track environment and the dynamics of train operation. Brief Description of the Drawings

[0080] The drawings are used to provide a further understanding of the present invention, and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention, and do not constitute a limitation to the present invention. In the drawings:

[0081] Figure 1 is the overall flowchart of a method for automatically generating real-time LKJ train operation simulation data proposed by the present invention;

[0082] Figure 2 is the structural schematic diagram of a system for automatically generating real-time LKJ train operation simulation data proposed by the present invention. Detailed Embodiment

[0083] Now, the present invention will be further described in detail with reference to the drawings. These drawings are all simplified schematic diagrams, only illustrating the basic structure of the present invention in a schematic manner, so they only show the components related to the present invention.

[0084] Refer to Figure 1 , a method for automatically generating real-time LKJ train operation simulation data, includes the following steps:

[0085] S1. Collect multi-source data of the track environment and train operation, and construct an original data set;

[0086] S2. Based on the original data set, initialize a multi-manifold adaptive generative adversarial network, and construct a track condition manifold, a train state manifold, and an environmental change manifold in the multi-manifold adaptive generative adversarial network to generate an initial latent variable distribution;

[0087] S3. Optimize the initial latent variable distribution using a variational inference mechanism, and input the optimized latent variable distribution into the generator of the multi-manifold adaptive generative adversarial network. Generate initial LKJ train operation simulation data through the adversarial training between the generator and the discriminator;

[0088] S4. Combine a virtual instantaneous perception module, and adjust the generation process of the initial LKJ train operation simulation data based on the real-time perception data of the track environment to generate real-time train operation simulation data;

[0089] S5. Perform global consistency optimization and local environment adjustment on the real-time train operation simulation data through a global-local adaptive correction mechanism to generate optimized train operation simulation data;

[0090] S6. Input the optimized train operation simulation data into the virtual track system constructed by digital twin technology to verify authenticity and adaptability, and update the generator and the correction mechanism according to the verification results;

[0091] S7. Perform dimensionality reduction on the verified optimized train operation simulation data, map the high-dimensional characteristics of track conditions, train behavior, and environmental changes to a low-dimensional space, and output the final train operation simulation data.

[0092] In this embodiment, the specific steps of S2 include:

[0093] S21. Classify the features of the original data set, and divide the track condition data, train state data, and environmental change data into independent data sets respectively;

[0094] S22. In the multi-manifold adaptive generative adversarial network, construct a track condition manifold, a train state manifold, and an environmental change manifold respectively. Each manifold corresponds to an independent feature subspace, which is used to represent the internal characteristics and dynamic change relationships of the track condition data, train state data, and environmental change data respectively;

[0095] S23. Use a network embedding layer to encode the data features of each manifold, map the high-dimensional features to a low-dimensional manifold space, and generate the initial feature distribution of the track condition manifold, the initial feature distribution of the train state manifold, and the initial feature distribution of the environmental change manifold by maintaining the structural relationship of the original data in the low-dimensional space;

[0096] S24. Construct the generator in the multi-manifold adaptive generative adversarial network. The generator receives the initial feature distributions of the track condition manifold, the train state manifold, and the environmental change manifold as inputs, and generates latent variables based on preset network weights and activation functions;

[0097] S25. Construct the discriminator in the multi-manifold adaptive generative adversarial network, and the discriminator optimizes the generator parameters through adversarial training;

[0098] S26. After the adversarial training reaches the convergence state, the generator generates the initial latent variable distributions of the track condition manifold, the train state manifold, and the environmental change manifold.

[0099] In this embodiment, the S3 specifically includes:

[0100] S31. Use the variational inference mechanism to optimize the initial latent variable distribution, and define the prior distribution and posterior distribution of the initial latent variables;

[0101] S32. Extract the initial mean and variance of the initial latent variables through the encoder network to generate the variational posterior distribution parameters;

[0102] S33. Optimize the initial latent variable distribution, and define the extended variational inference optimization objective function:

[0103]

[0104] where L represents the loss function, N represents the total dimension of the initial latent variables, μ i represents the optimized latent variable mean, represents the optimized latent variable variance, μ prior,i represents the mean of the prior distribution, represents the variance of the prior distribution;

[0105] S34. Adopt an adaptive iterative optimization strategy to dynamically adjust the initial mean and variance of the initial latent variable distribution to generate the optimized latent variable distribution;

[0106] S35. Input the optimized latent variable distribution into the generator of the multi-manifold adaptive generative adversarial network, and generate the initial LKJ train operation simulation data through the adversarial training between the generator and the discriminator.

[0107] In this embodiment, the S4 specifically includes:

[0108] S41. Construct a virtual instantaneous perception module, which includes a real-time perception unit, a feature extraction unit, a dynamic weight allocation unit, and a feedback adjustment unit;

[0109] S42. Collect the dynamic data of the track environment through the real-time perception unit, including signal status, slope change, weather conditions, and track anomalies, to generate the original perception matrix X raw ={x ij}, where x ij represents the value of the environmental feature j at the time step i;

[0110] S43. Use the feature extraction unit to perform denoising and dimensionality reduction processing on the original perception matrix to generate the standardized matrix Xnorm = {x' ij}, where x' ij represents the value of the environmental feature j at the time step i after standardization. Combining with the dynamic weight allocation unit, a comprehensive environmental feature vector is calculated:

[0111]

[0112] where X combined represents the comprehensive environmental feature vector, M represents the total number of features, α j represents the importance adjustment coefficient of the environmental feature j, w j represents the dynamic weight of the environmental feature j, exp represents the exponential function, β j represents the smoothing factor of the environmental feature j, μ j represents the mean value of the environmental feature j;

[0113] S44. Input the comprehensive environmental feature vector X combined and the optimized latent variable distribution Z opt into the generator:

[0114]

[0115] where Y real-time represents the real-time driving simulation data, G represents the generator, θ represents the generator parameters, γ represents the step size parameter, represents the gradient operation;

[0116] S45. Use the feedback adjustment unit to perform dynamic verification on the generated real-time driving simulation data:

[0117]

[0118] where Δθ represents the amplitude of parameter optimization, η represents the learning rate, Q represents the total number of samples, y a represents the eigenvalue of the generated real-time driving simulation data, x a represents the eigenvalue of the perception data, represents the gradient of the generator parameters, L gen represents the generation loss function, p and q represent the adjustment coefficients, σ a represents the standard deviation;

[0119] S46. After the feedback adjustment reaches convergence, output the dynamically verified real-time driving simulation data.

[0120] In this embodiment, the S5 specifically includes:

[0121] S51. Construct a global-local adaptive correction mechanism. The global-local adaptive correction mechanism includes a global optimization module and a local adjustment module. The global optimization module is used to optimize the overall characteristics of real-time driving simulation data for consistency, and the local adjustment module is used to further adjust the real-time driving simulation data according to the detailed characteristics of the track environment.

[0122] S52. Use the global optimization module to perform an overall analysis of the real-time driving simulation data, extract the global features of the real-time driving simulation data, and compare the extracted global features with the preset track environment reference features to identify overall characteristic deviations.

[0123] S53. According to the identified overall characteristic deviations, adjust the global parameters of the generator so that the real-time driving simulation data is consistent with the track environment reference features in terms of overall characteristics, and at the same time ensure that the global dynamic changes of the generated data meet the actual requirements of the track environment.

[0124] S54. Through the local adjustment module, further analyze the local characteristics of the real-time driving simulation data, extract local features based on the dynamic changes of the track environment, including signal status, slope changes, and weather conditions, and compare the local features with the real-time perception data to identify local deviations.

[0125] S55. According to the identified local deviations, optimize and adjust the local parameters of the real-time driving simulation data, and ensure that the local adjustment does not damage the consistency of the overall characteristics.

[0126] S56. Combine the correction results of the global optimization module and the local adjustment module to generate optimized driving simulation data.

[0127] In this embodiment, the specific steps of S6 are as follows:

[0128] S61. Construct a virtual track system based on digital twin technology. The virtual track system includes a track geometry model, a signal status model, and a train dynamics model, and establish the global dynamic characteristic equation of the virtual track system:

[0129] M virtual =A·T track +B·S signal +C·D train ;

[0130] Where, M virtual represents the virtual track system model, T track represents the track geometric characteristics, S signal represents the signal status, D train represents the train dynamics characteristics, and A, B, and C represent weight matrices.

[0131] S62. Map the driving simulation optimization data to the virtual track system, and generate dynamic operation data through the virtual track system. The dynamic operation data is jointly driven by a track geometry model, a signal status model, and a train dynamics model:

[0132]

[0133] Among them, Y sim represents the dynamic operation data, ⊙ represents the element-wise multiplication operation, represents the tensor outer product operation;

[0134] S63. Perform multi-level comparison between the generated dynamic operation data Y sim and the real track data to construct a dynamic characteristic deviation model:

[0135]

[0136] Among them, ΔY represents the deviation value, L represents the number of characteristic layers, γ l represents the layer weight, represents the dynamic operation data of the l-th layer, represents the real track data of the l-th layer, and κ represents the adjustment parameter;

[0137] S64. Use the dynamic characteristic deviation model to optimize and update the generator parameters and the correction mechanism parameters:

[0138]

[0139] Among them, θ represents the generator parameters, φ represents the correction mechanism parameters, and ξ1 and ξ2 represent the regularization parameters;

[0140] S65. After the deviation value ΔY converges to the set threshold, output the driving simulation optimization data verified by the virtual track system.

[0141] Reference Figure 2 , a system for automatically generating real-time LKJ driving simulation data, includes the following modules:

[0142] A data acquisition module, which acquires multi-source data of the track environment and train operation, and constructs an original data set;

[0143] A multi-manifold generation module, which initializes a multi-manifold adaptive generative adversarial network based on the original data set, constructs a track condition manifold, a train state manifold, and an environment change manifold, and generates an initial latent variable distribution;

[0144] A variational inference optimization module, which optimizes the initial latent variable distribution to generate an optimized latent variable distribution;

[0145] The virtual instantaneous perception module adjusts the initial LKJ train operation simulation data based on the real-time perception data of the track environment to generate real-time train operation simulation data;

[0146] The global-local correction module generates optimized train operation simulation data through global consistency optimization and local environment adjustment;

[0147] The virtual track verification module inputs the optimized train operation simulation data into the virtual track system to verify authenticity and adaptability, and updates the generator and correction mechanism;

[0148] The dimensionality reduction output module performs dimensionality reduction processing on the verified optimized train operation simulation data and outputs the final train operation simulation data.

[0149] Embodiment 1:

[0150] To verify the feasibility of the present invention in implementation, the present invention is applied to the operation simulation and optimization scenario of a certain railway system. The railway system consists of 3 main lines, with a complex operation environment. The lines pass through multiple curves, slopes and intersections, and at the same time need to adapt to the impact of different weather conditions on operation efficiency and safety. In the process of train operation data simulation of the existing system, there are problems such as insufficient response to complex environmental changes, low accuracy of generated data and poor real-time performance, which cannot meet the requirements of refined scheduling and operation optimization.

[0151] The system of the present invention is deployed in the railway dispatching center to generate real-time train operation simulation data and assist in optimizing dispatching decisions. The data acquisition module collects multi-source data on track conditions, train status and environmental changes, including track geometric characteristics (such as curve radius, slope), train operation characteristics (such as speed, acceleration, stop time) and environmental data (such as weather conditions, signal status) in the past 12 months. These data constitute the original data set.

[0152] In the application, the original data is first processed through a multi-manifold adaptive generative adversarial network to respectively construct a track condition manifold, a train status manifold and an environmental change manifold, and generate an initial latent variable distribution. Subsequently, a variational inference mechanism is used to optimize the latent variable distribution to improve the adaptability of the generated data to complex environments and dynamic changes. In the real-time application scenario, the virtual instantaneous perception module real-time collects dynamic change data of the track environment, such as signal light change, slope adjustment and sudden weather impact, and generates real-time train operation simulation data by dynamically adjusting the generator parameters.

[0153] In a practical application on a certain day, the system conducted simulation and optimization tests on real-time operation data for a line with multiple ramps and curves. In the generated simulation data, for the train speed control on curves, the system provided dynamic adjustment suggestions, controlling the passing speed of curves between 30 - 40 km / h. Meanwhile, the signal light response time was optimized, reducing the average train stop time by 5 seconds. In the ramp scenario, the system optimized the acceleration control according to the track slope change, avoiding passenger discomfort and increased energy consumption caused by excessive acceleration.

[0154] To verify the beneficial effects of the present invention, during the test process, the comparison results of the generated data and the real data under different lines and working conditions were collected, and the accuracy, real-time performance, and response ability to sudden environments of the data were mainly analyzed. Through comparison, it was found that the train operation simulation data generated by the system was significantly superior to the existing methods in terms of track environment adaptability and train operation accuracy.

[0155] The following table records the specific experimental data of 3 main lines.

[0156] Table 1 Comparison Table of Simulation Data and Real Data under Different Lines and Working Conditions

[0157]

[0158] From the data in Table 1 above, it can be seen that the real-time LKJ train operation simulation data generated by the present invention under different lines and environmental characteristics has extremely high accuracy and response ability. In the simple working conditions of a straight track and sunny environment (Line A), the accuracy of the simulation data reaches 98.5%, with a very small difference from the 99.1% of the real data. The signal response delay is only 0.5 seconds, the average speed error is 0.3 km / h, and the response time to sudden environments is 1.2 seconds, fully demonstrating the high efficiency and stability of the system in conventional scenarios.

[0159] In a more complex curve and rainy environment (Line B), the accuracy of the simulation data still reaches 96.2%, with only a small gap from the 97.8% of the real data. The signal response delay is 0.8 seconds, and the average speed error is 0.6 km / h. This shows that the present invention can still generate high-precision data and respond to environmental changes in a timely manner when dealing with complex environments such as curves and rainy days.

[0160] For severe working conditions such as ramps and extreme weather (Line C), the accuracy of the simulation data is 94.7%, slightly inferior to the 96.3% of the real data. Nevertheless, the performance of the present invention in terms of signal response delay (1.1 seconds) and average speed error (0.8 km / h) is still remarkable, and the response time to sudden environments is 2.5 seconds, indicating that the system still has strong adaptability and data generation ability under extreme conditions.

[0161] Generally speaking, the data generated by the present invention under different lines and working conditions all show high accuracy, low response latency, and strong adaptability to sudden environments. These results verify the practicability and reliability of the present invention in the field of rail transit. In particular, it demonstrates significant advantages in complex dynamic scenarios, providing solid technical support for the operation optimization and safety guarantee of rail transit systems.

[0162] As described above, only the preferred specific embodiments of the present invention are provided, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, making equivalent substitutions or changes, shall be covered by the protection scope of the present invention.

Claims

1. A method for automatically generating real-time LKJ train operation simulation data, characterized in that, It includes the following steps: S1. Collect multi-source data of the track environment and train operation, and construct an original data set; S2. Based on the original data set, initialize a multi-manifold adaptive generative adversarial network, and construct a track condition manifold, a train state manifold and an environmental change manifold in the multi-manifold adaptive generative adversarial network to generate an initial latent variable distribution; S3. Use the variational inference mechanism to optimize the initial latent variable distribution, input the optimized latent variable distribution into the generator of the multi-manifold adaptive generative adversarial network, and generate initial LKJ train operation simulation data through the adversarial training between the generator and the discriminator; S4. Combine the virtual instantaneous perception module, and adjust the generation process of the initial LKJ train operation simulation data based on the real-time perception data of the track environment to generate real-time train operation simulation data; S5. Perform global consistency optimization and local environment adjustment on the real-time train operation simulation data through a global-local adaptive correction mechanism to generate optimized train operation simulation data; S6. Input the optimized train operation simulation data into the virtual track system constructed by digital twin technology to verify authenticity and adaptability, and update the generator and the correction mechanism according to the verification results; S7. Perform dimensionality reduction processing on the verified optimized train operation simulation data, map the high-dimensional characteristics of the track state, train behavior and environmental change to a low-dimensional space, and output the final train operation simulation data.

2. The method for automatically generating real-time LKJ train operation simulation data according to claim 1, wherein, The specific content of S2 includes: S21. Classify the features of the original data set, and divide the track condition data, train state data and environmental change data into independent data sets respectively; S22. In the multi-manifold adaptive generative adversarial network, construct a track condition manifold, a train state manifold and an environmental change manifold respectively. Each manifold corresponds to an independent feature subspace, which is used to represent the internal characteristics and dynamic change relationships of the track condition data, train state data and environmental change data respectively; S23. Use the network embedding layer to encode the data features of each manifold, map the high-dimensional features to the low-dimensional manifold space, and generate the initial feature distributions of the track condition manifold, the train state manifold and the environmental change manifold by maintaining the structural relationship of the original data in the low-dimensional space; S24. Construct the generator in the multi-manifold adaptive generative adversarial network. The generator receives the initial feature distributions of the track condition manifold, the train state manifold and the environmental change manifold as inputs, and generates latent variables based on the preset network weights and activation functions; S25. Construct the discriminator in the multi-manifold adaptive generative adversarial network, and the discriminator optimizes the generator parameters through adversarial training; S26. After the adversarial training reaches the convergence state, the generator generates the initial latent variable distributions of the track condition manifold, the train state manifold and the environmental change manifold.

3. A method for automatically generating real-time LKJ train operation simulation data according to claim 1, characterized in that, The specific content of S3 includes: S31. Use the variational inference mechanism to optimize the initial latent variable distribution, and define the prior distribution and posterior distribution of the initial latent variables; S32. Extract the initial mean and variance of the initial latent variables through the encoder network to generate variational posterior distribution parameters; S33. Optimize the initial latent variable distribution and define the extended variational inference optimization objective function: Among them, L represents the loss function, N represents the total dimension of the initial latent variables, μ i represents the mean of the optimized latent variables, represents the variance of the optimized latent variables, μ prior,i represents the mean of the prior distribution, represents the variance of the prior distribution; S34, using an adaptive iterative optimization strategy to dynamically adjust the initial mean and variance of the initial latent variable distribution to generate an optimized latent variable distribution; S35. Input the optimized latent variable distribution into the generator of the multi-manifold adaptive generative adversarial network, and generate the initial LKJ driving simulation data through adversarial training between the generator and the discriminator.

4. A method for automatically generating real-time LKJ train operation simulation data according to claim 1, characterized in that, The S4 specifically includes: S41, constructing a virtual instantaneous perception module, wherein the virtual instantaneous perception module includes a real-time perception unit, a feature extraction unit, a dynamic weight allocation unit, and a feedback adjustment unit; S42. Collect dynamic data of the track environment through the real-time perception unit, including signal status, slope change, weather conditions, and track anomalies, and generate an original perception matrix X raw ={x ij}, where x ij represents the value of environmental feature j at time step i; S43. Use the feature extraction unit to perform denoising and dimensionality reduction processing on the original perception matrix to generate a standardized matrix X norm ={x′ ij}, where x′ ij represents the value of the environmental feature j at time step i after standardization. Combine with the dynamic weight allocation unit to calculate the comprehensive environmental feature vector: Among them, X combined represents the comprehensive environmental feature vector, M represents the total number of features, and α j represents the importance adjustment coefficient of environmental feature j, w j represents the dynamic weight of environmental feature j, exp represents the exponential function, and β j represents the smoothing factor of environmental feature j, and μ j represents the mean value of environmental feature j; S44. Input the comprehensive environmental feature vector X combined and the optimized latent variable distribution Z opt into the generator: Among them, Y real-time represents real-time driving simulation data, G represents a generator, θ represents generator parameters, and γ represents step size parameters, represents gradient operation; S45, using the feedback adjustment unit to dynamically check the generated real-time driving simulation data: Among them, Δθ represents the amplitude of parameter optimization, η represents the learning rate, Q represents the total number of samples, and y a represents the eigenvalue of the generated real-time driving simulation data, and x a represents the eigenvalue of the perception data, represents the generator parameter gradient, and L gen represents the generation loss function, p and q represent the adjustment coefficients, and σ a represents the standard deviation; S46. After the feedback adjustment reaches convergence, the real-time driving simulation data after dynamic verification is output.

5. A method and system for automatically generating real-time LKJ train operation simulation data according to claim 1, characterized in that The S5 specifically includes: S51, constructing a global-local adaptive correction mechanism, wherein the global-local adaptive correction mechanism includes a global optimization module and a local adjustment module, wherein the global optimization module is used to optimize the consistency of the overall characteristics of the real-time driving simulation data, and the local adjustment module is used to further adjust the real-time driving simulation data according to the detailed characteristics of the track environment; S52, using a global optimization module to perform an overall analysis on the real-time driving simulation data, extracting global features of the real-time driving simulation data, and comparing the extracted global features with preset track environment benchmark features to identify overall feature deviations; S53, adjusting the global parameters of the generator according to the identified overall characteristic deviation, so that the overall characteristics of the real-time driving simulation data are consistent with the track environment benchmark characteristics, and at the same time ensuring that the global dynamic changes of the generated data meet the actual needs of the track environment; S54, further analyzing the local characteristics of the real-time driving simulation data through the local adjustment module, extracting local features based on the dynamic changes of the track environment, including signal status, slope changes, and weather conditions, and comparing the local features with the real-time perception data to identify local deviations; S55, optimizing and adjusting local parameters of the real-time driving simulation data according to the identified local deviation, and ensuring that the local adjustment does not destroy the consistency of the overall characteristics; S56. Combining the correction results of the global optimization module and the local adjustment module, generating driving simulation optimization data.

6. A method for automatically generating real-time LKJ train operation simulation data according to claim 1, characterized in that, The S6 specifically includes: S61. Construct a virtual track system based on digital twin technology, wherein the virtual track system includes a track geometry model, a signal state model, and a train dynamics model, and establish a global dynamic characteristic equation of the virtual track system: M virtual = A·T track + B·S signal + C·D train ; Among them, M virtual represents the virtual track system model, T track represents the track geometric characteristics, S signal represents the signal state, D train represents the train dynamic characteristics, and A, B, and C represent weight matrices; S62, mapping the driving simulation optimization data to the virtual track system, and generating dynamic operation data through the virtual track system, the dynamic operation data being jointly driven by the track geometry model, the signal state model and the train dynamics model: Among them, Y sim represents dynamic operation data, ⊙ represents element-wise multiplication operation, represents the outer product operation of tensors; S63. Compare the generated dynamic operation data Y sim with the real track data at multiple levels to construct a dynamic characteristic deviation model: where ΔY represents the deviation value, L represents the number of characteristic layers, and γ l represents the layer weight, represents the dynamic operation data of the l-th layer, represents the true orbit data of the l-th layer, and κ represents the adjustment parameter; S64. Optimize and update the generator parameters and correction mechanism parameters using the dynamic characteristic deviation model: Among them, θ represents the generator parameter, φ represents the correction mechanism parameter, ξ1 and ξ2 represent the regularization parameters; S65. After the deviation value ΔY converges to the set threshold, output the optimized train operation simulation data verified by the virtual track system.

7. A system for automatically generating real-time LKJ train operation simulation data, which executes the method for automatically generating real-time LKJ train operation simulation data according to any one of claims 1 to 6, characterized in that, It includes the following modules: Data acquisition module, which acquires multi-source data of the track environment and train operation and constructs an original data set; Multiple manifold generation module, which initializes a multiple manifold adaptive generative adversarial network based on the original data set, constructs a track condition manifold, a train state manifold and an environmental change manifold, and generates an initial latent variable distribution; Variational inference optimization module, which optimizes the initial latent variable distribution to generate an optimized latent variable distribution; Virtual instantaneous perception module, which adjusts the initial LKJ train operation simulation data based on the real-time perception data of the track environment to generate real-time train operation simulation data; Global-local correction module, which generates optimized train operation simulation data through global consistency optimization and local environment adjustment; Virtual track verification module, which inputs the optimized train operation simulation data into the virtual track system to verify the authenticity and adaptability, and updates the generator and the correction mechanism; Dimensionality reduction output module, which performs dimensionality reduction processing on the verified optimized train operation simulation data and outputs the final train operation simulation data.

Images