High-speed maglev track irregularity data generation method and device based on time sequence generative adversarial network and storage medium

Abstract

The invention relates to a high-speed maglev track irregularity data generation method and device based on a time sequence generative adversarial network and a storage medium. The method comprises the following steps: acquiring track irregularity data in a running state of a maglev train, preprocessing the track irregularity data, and inputting the preprocessed track irregularity data into a pre-constructed time sequence generative adversarial network model for training and evaluation to obtain track irregularity generation data; wherein the time sequence generative adversarial network model comprises an embedded function, a recovery function, a generator and a discriminator, the embedded function and the recovery function form an embedded network, the generator and the discriminator form a generative adversarial network, the embedded network is used for providing a potential space, and the generative adversarial network runs in the potential space. Compared with the prior art, the method has the advantages that the irregularity characteristics of the high-speed maglev track are accurately reflected, and the precision of overall data is improved while time domain information is reserved.

Description

Technical Field

[0001] The present invention relates to the field of rail transit technology, and in particular to a method, device and storage medium for generating high-speed maglev track irregularity data based on a time series generative adversarial network. Background Art

[0002] At present, high-speed maglev has achieved a leap in speed through its contactless operation mode. The proposal of a high-speed maglev system with a speed of 600 kilometers per hour provides a new option for inter-city passenger transportation needs. High-speed maglev can achieve higher speeds at a lower cost, thereby realizing short-time commuting between long-distance cities. In addition, since wheel-rail contact is avoided, the mechanical noise during operation is effectively reduced, which is another reason why existing wheel-rail trains cannot match it.

[0003] In the process of high-speed maglev construction, its safety and stability are crucial considerations, and track unevenness is the main cause of the dynamic force between the vehicle and the track, thus becoming the main source of vibration of high-speed maglev trains. On the one hand, the collection of accurate track unevenness data depends on high-end sensors and precision measuring instruments. The installation and operation and maintenance costs of these devices are high, and the special operating environment of maglev makes the acquisition of large-scale, continuous data extremely complex and limited; on the other hand, the engineering field usually tends to use spectral modeling to process track unevenness data. Although it simplifies the analysis process to a certain extent, it may miss important information on spatiotemporal variation characteristics. Therefore, the existing track unevenness data cannot fully meet the needs of simulation experiments for massive data, and the lack of data directly affects the universality and accuracy of simulation experiments. Summary of the invention

[0004] The purpose of the present invention is to overcome the defects of the above-mentioned prior art in the high-speed maglev track unevenness data such as insufficient accuracy, insufficient data volume and missing information on temporal and spatial variation characteristics, and to provide a high-speed maglev track unevenness data generation method, device and storage medium based on a time series generative adversarial network, which can be applied to the fields of track detection, maintenance and optimization of maglev trains, and provide high-precision simulation data and technical support for track design and train operation.

[0005] The purpose of the present invention can be achieved by the following technical solutions:

[0006] According to a first aspect of the present invention, a method for generating high-speed maglev track irregularity data based on a time series generative adversarial network is provided, comprising the following steps: obtaining and preprocessing the track irregularity data under the running state of a maglev train; inputting the preprocessed track irregularity data into a pre-constructed time series generative adversarial network model for training and evaluation to obtain track irregularity generation data; wherein the time series generative adversarial network model comprises an embedding function, a recovery function, a generator and a discriminator, the embedding function and the recovery function form an embedding network, the generator and the discriminator form a generative adversarial network, the embedding network is used to provide a latent space, and the generative adversarial network operates in the latent space.

[0007] As a preferred technical solution, the training process of the time series generative adversarial network model specifically includes: based on the preprocessed track unevenness data, training the embedding network using a pre-constructed first loss function, and training the generative adversarial network using a pre-constructed second loss function.

[0008] As a preferred technical solution, the expression of the first loss function is:

[0009]

[0010] In the formula, θ e ,θ r denote the parameters of the embedding network and the recovery network respectively, represents the supervised loss of the generator in closed-loop mode, represents the reconstruction loss, and λ is the balance and and λ≥0.

[0011] As a preferred technical solution, the expression of the second loss function is:

[0012]

[0013] In the formula, θ g ,θ d They represent the parameters of the generating network and the discriminating network respectively, represents the supervised loss of the generator in closed-loop mode, represents the unsupervised loss of the generator in open-loop mode, and η is the balance and Hyperparameters of .

[0014] As a preferred technical solution, the embedding network and the generative adversarial network respectively adopt recursive neural network structures.

[0015] As an optimal technical solution, the method uses an inertial reference method to obtain track irregularity data under the running state of a maglev train. The specific process includes: obtaining an acceleration signal under the running state of a maglev train, and performing uniform resampling and preprocessing; obtaining an air gap signal; integrating the processed acceleration signal and superimposing it with the air gap signal to obtain a track irregularity signal.

[0016] As a preferred technical solution, the specific process of the preprocessing includes: drawing a box plot, and using the outlier limits in the box plot to calculate the outlier limits, the box plot is used to identify and determine potential outliers in the track irregularity data; based on the outlier limits, using the sliding window method to cross-analyze the track irregularity data to obtain the location of the outlier; replacing the outlier with the linear difference of the normal data adjacent to it.

[0017] As a preferred technical solution, the evaluation process of the time series generative adversarial network model specifically includes: after each iteration, after obtaining the track irregularity generation data, the classification error and the mean absolute error between the track irregularity generation data and the track irregularity data before preprocessing are calculated using a pre-built classification model and a regression prediction model, and the generation capability of the model is quantified using the classification error and the mean absolute error.

[0018] According to a second aspect of the present invention, there is provided a high-speed maglev track irregularity data generating device based on a time series generative adversarial network, comprising a memory, a processor, and a program stored in the memory, wherein the method described is implemented when the processor executes the program.

[0019] According to a third aspect of the present invention, there is provided a storage medium having a program stored thereon, wherein the program implements the method described above when executed.

[0020] Compared with the prior art, the present invention has the following beneficial effects:

[0021] 1. Compared with the traditional method of directly modeling the data spectrum, the present invention designs and constructs a time series generative adversarial network model, and uses the track irregularity data of the maglev train in operation to train and evaluate the model. It can directly process and generate data in the time domain, so that the generated data can more accurately reflect the irregularity characteristics of the maglev track at each time point, and improve the accuracy of the overall data while retaining the time domain information;

[0022] 2. Traditional generative adversarial networks (GANs) are usually unable to effectively handle the temporal dependencies and complex dynamic behaviors in data when learning and generating time series distributions. In contrast, the model proposed in this invention combines the embedding network with the generative adversarial network during training to achieve unified processing of feature encoding, data generation representation, and iterative learning across time series. The loss function includes the reconstruction loss of the embedding network, the unsupervised loss of the generative adversarial network, and the supervised loss based on the original data. The potential dynamics of the real data and the synthetic data are synchronized through the supervised loss, which can not only learn and reproduce the dynamic characteristics of track irregularity data changing over time, but also accurately capture the temporal and spatial dependencies of the data and enrich the track irregularity data.

[0023] 3. The present invention has independent time series generation capability and can capture complex patterns in track irregularity data distribution. It can completely generate new time series from data distribution without relying on existing time series, thus truly realizing the complete generation process from data distribution to time series. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 A schematic diagram of the steps of providing a method in an embodiment of the present invention;

[0025] Figure 2 A block diagram of component functions and objectives of a time series generation adversarial network model in an embodiment of the present invention;

[0026] Figure 3 This is an illustrative implementation of the time series generation adversarial network model in an embodiment of the present invention. DETAILED DESCRIPTION

[0027] Existing high-speed maglev track irregularity modeling and data generation methods are insufficient in data accuracy and comprehensiveness, and it is difficult to accurately simulate the complex dynamic characteristics of the track, resulting in difficulty in identifying potential structural problems in advance, and limited support for maintenance and emergency response. The high-speed maglev track irregularity data generation method based on time series generative adversarial network provided by the present invention, its modeling process and data generation process can effectively overcome the shortcomings of the prior art, accurately simulate track irregularity data, identify and optimize track structure in advance, predict track wear and damage trends, and quickly evaluate their impact, provide scientific emergency decision support, and comprehensively improve the efficiency of track design, operation maintenance and safety management.

[0028] In addition, the present invention provides a new methodology for the research and application of future maglev systems. When developing new maglev systems, the model can be used for virtual testing and verification, greatly reducing the time and cost of physical testing. By simulating and evaluating the performance of tracks and trains under different unevenness scenarios, the method can guide technological innovation and system optimization, and accelerate the maturity and promotion of new technologies. In future transportation systems, the method provided by the present invention can provide solid technical support for improving design, operation and maintenance efficiency, and enhancing system reliability and safety.

[0029] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

[0030] Example

[0031] like Figure 1 As shown, this embodiment provides a method for generating high-speed maglev track irregularity data based on a time series generative adversarial network, comprising the following steps:

[0032] Step S1, obtaining track irregularity data under the running state of the maglev train and performing preprocessing;

[0033] Step S2, inputting the preprocessed track irregularity data into a pre-built time series generative adversarial network model for training and evaluation to obtain track irregularity generation data; wherein the time series generative adversarial network model includes an embedding function, a recovery function, a generator and a discriminator, the embedding function and the recovery function form an embedding network, the generator and the discriminator form a generative adversarial network, the embedding network is used to provide a latent space, and the generative adversarial network operates in the latent space.

[0034] Step S1 collects track irregularity data by monitoring the motion state changes of the maglev train, and pre-processes the data to remove structural outliers. Specifically:

[0035] 1. Track irregularity data

[0036] In this embodiment, the track unevenness data of the maglev train in operation is obtained using the inertial reference method, and the track unevenness data is a time series with non-stationary and non-Gaussian properties. Specifically, first, an inertial sensor and an accelerometer are installed at the bottom of the maglev train to collect the acceleration signal of the maglev train in operation, and uniformly resample and preprocess it; secondly, the air gap signal is obtained; finally, the processed acceleration signal is integrated and superimposed with the air gap signal to obtain the track unevenness signal, which is used to evaluate the flatness and unevenness of the track.

[0037] 2. Preprocessing of track irregularity data

[0038] In the detection of maglev tracks, periodic outliers will be generated due to the gaps between track beams, and due to the characteristics of time series data, specific methods need to be adopted to eliminate these structural anomalies. This embodiment is based on the box plot principle and uses the sliding window method to remove structural anomalies. Specifically: First, draw a box plot to identify and determine potential outliers in track unevenness data, and use the box plot to calculate the outlier limits; secondly, based on the outlier limits, cross-apply the box plot through two sliding windows to achieve cross-analysis of the time series data. Only data points selected as outliers in both sliding windows are identified as true anomalies, and the positions of the outliers are recorded respectively; these outliers are replaced by the linear difference between the normal data on the left and right to ensure the integrity of the time series data and the accuracy of subsequent analysis.

[0039] In step S2, the construction of the time series generative adversarial network model aims to simulate the complex characteristics of high-speed maglev track unevenness data, and the training and evaluation of the model aims to ensure that the model meets the expected standards in capturing the spatiotemporal dynamic characteristics of track unevenness.

[0040] 1. Structure of the Time Series Generative Adversarial Network Model

[0041] The time series generative adversarial network provided in this embodiment combines the characteristics of supervised models (such as RNN) and unsupervised models (such as GAN) to capture time-related features and generate new sequences. Its architecture includes an embedding function, a recovery function, a generator, and a discriminator. First, the embedding network converts input features into low-dimensional latent features to capture the temporal dynamics of complex systems. Then, by jointly training the embedding network and the generator, the model captures the conditional distribution of data at each step in the latent space through step-by-step supervision loss and unsupervised adversarial loss. The discriminator operates in the latent space to ensure the consistency of the synthetic data with the real data in temporal dynamics. Among them, the synthetic data is also the output track unevenness generated data.

[0042] The block diagram of each component's function and purpose is shown in Figure 2 As shown, the functions and objectives of each component are described in detail below:

[0043] (1) Embedding function and recovery function:

[0044] Since the input features are represented as potential features, embedding functions and recovery functions are required as reversible mappings for mutual conversion. This ensures that the subsequent model can always maintain low-dimensional input when learning the potential temporal dynamic features of the data, thereby ensuring the stability of the algorithm.

[0045] make Representing feature space The corresponding latent vector space. Then the embedding function Convert static and temporal features into latent codes Here we use a recurrent network to implement e,

[0046]

[0047]

[0048] in, is the embedding network of static features, It is a recursive embedding network of temporal dynamic features.

[0049] In the reverse direction, the recovery function Return static and temporal encodings to their feature representations Here each step is implemented through the feedforward network,

[0050]

[0051]

[0052] in, and is a static and time-dynamic embedded restoration network.

[0053] It should be noted that both the embedding function and the recovery function, their architecture designs are customizable, and the only constraint is that they must follow the autoregressive principle.

[0054] (2) Generator and Discriminator

[0055] First, the sequence generator outputs into the embedding space. Let represents a vector space with a known distribution, and draws random vectors from it as the generated input to Then, the function is generated Get static and temporal random vector tuples to synthesize latent codes Implement g through a recurrent network,

[0056]

[0057]

[0058] in, is a generator network of static features, is a loop generator to generate temporal dynamic features. is a random vector, z t is a random process.

[0059] Then, the discriminator also operates in the embedding space. The discriminant function Receive static and time codes, return categories Indicates actual (h * ) or synthesis embedding; similarly, The symbol represents the actual y * or synthetic Classification of data. Here d is implemented by a bidirectional recurrent network with a feedforward output layer:

[0060]

[0061]

[0062] in, and denote the forward hidden state sequence and the backward hidden state sequence respectively, and is a recursive function, and is the classification function of the output. Again, the only constraint here is that the generator must follow the autoregressive principle.

[0063] 2. Training of Time Series Generative Adversarial Network Model

[0064] First, the embedding and recovery functions should be able to pass the original data s, x 1:T Potential representation of h S ,h 1:T To achieve accurate reconstruction, we can obtain The original data is the track unevenness data without preprocessing. Therefore, the first objective function is the reconstruction loss, which is expressed as:

[0065]

[0066] In the model, the generator receives two types of input during training: synthetic embeddings and the output of the previous time step to generate the next synthetic vector. In open-loop mode, the generator receives Generate the next resultant vector And the gradient is calculated based on the unsupervised loss. The unsupervised loss is expressed as:

[0067]

[0068] Given that simple feedback from the discriminator alone is not sufficient to accurately capture the continuous conditional distribution characteristics in the dataset, in order to achieve this goal more efficiently, this model adds an additional loss term as a guide during the training process. Through alternating training methods, the generator receives the actual data embedding sequence h calculated by the embedding network in a closed-loop mode.1:t-1 , to generate the next latent vector. Then, a loss term is constructed to capture the distribution p(H t |H S ,H 1:t-1 )and The difference between them is calculated by the maximum likelihood method to achieve supervision of the loss. The supervised loss is expressed as:

[0069]

[0070] Let θ e ,θ r ,θ g ,θ d They represent the parameters of the embedding network, the restoration network, the generation network, and the identification network respectively. The training of the embedding network and the restoration network is performed on the reconstruction loss and the supervised loss, that is, the first loss function is expressed as:

[0071]

[0072] Among them, λ ≥ 0 is a hyperparameter to balance the two losses.

[0073] Next, the generator and discriminator are trained adversarially as follows (i.e., the second loss function):

[0074]

[0075] Among them, η ≥ 0 is another hyperparameter to balance the two losses.

[0076] In summary, the illustrative implementation of the time series generative adversarial network model is as follows Figure 3 .

[0077] 3. Evaluation of Time Series Generative Adversarial Network Model

[0078] To evaluate the quality of generated data, three necessary conditions are observed: (1) diversity: the distribution of samples should cover the real data; (2) fidelity: samples should be indistinguishable from the real data; and (3) usefulness: samples should be as useful as the real data when used for the same prediction purpose.

[0079] Visualization: Data processing is performed on the original dataset (including track irregularity data before preprocessing) and the synthetic dataset (including track irregularity generated data) to obtain the projection of the high-dimensional data in two-dimensional space, which is convenient for visualization and provides a qualitative evaluation of the above condition (1).

[0080] Discrimination score: First, label each original sequence as “real” and each generated sequence as “generated”. Then, build a neural network classification model based on the above training data and train it to distinguish these two types of sequences. Finally, report the classification error on the reserved test set, which provides a quantitative evaluation of condition (2) above.

[0081] Prediction score: Build a time series regression prediction model and use the generated data as the training set. Then use the original data set as the test set, and the loss function is the mean absolute error. This solution provides a quantitative evaluation for condition (3) above.

[0082] The aforementioned model evaluation method evaluates the model performance in detail by comparing the model-generated data with the real original input data. Visualization tools can help provide intuitive feedback on the performance of the model. The classification model and regression prediction model compare the similarity between the generated data and the real data from the perspective of discrimination and prediction. Based on the evaluation results, the structure and parameters of the model are continuously optimized to ensure that it can accurately capture the spatiotemporal dynamic characteristics of track irregularities.

[0083] Based on the above detailed description of steps S1 to S2, the specific implementation process of the method provided in this embodiment is as follows:

[0084] like Figure 1 As shown, the real space maglev train suspension system transmits measurement and report data to the virtual space maglev train suspension system, and the virtual space maglev train suspension system transmits control parameters and reports to the real space maglev train suspension system.

[0085] (1) Data collection and preprocessing

[0086] Track irregularity data is collected using inertial sensors installed on maglev trains. The collected data includes information such as acceleration, velocity, and displacement during train operation. The collected data is preprocessed to remove outliers. The specific method is to use the box plot principle and cross-apply the box plot method through two sliding windows to record and replace outliers.

[0087] (2) Construction of time series generative adversarial network model

[0088] Build as Figure 2 The time series generative adversarial network model shown in the figure includes an embedding function, a recovery function, a generator, and a discriminator. The embedding network consists of an embedding function and a recovery function, and adopts a recurrent neural network (RNN) structure to retain temporal dynamic information using its hidden layer features to adapt to the temporal dependencies of time series data. The generator and the discriminator constitute a generative adversarial network, which also adopts an RNN structure to capture the original data distribution through unsupervised learning and generate realistic data similar to the original data.

[0089] (3) Model training and evaluation

[0090] The constructed model is trained using the preprocessed track irregularity dataset. The training process includes the following steps: mapping the data into the latent space through the embedding network; running the generative adversarial network in the latent space; training by combining supervised loss and unsupervised loss; and evaluating the performance of the trained model, using the classification model and regression prediction model to evaluate the classification error and mean absolute error of the generated data.

[0091] Furthermore, this embodiment also provides a high-speed maglev track unevenness data generation device based on a time series generative adversarial network, including a memory, a processor, and a program stored in the memory. When the processor executes the program, one or more steps of the aforementioned method are implemented, and the specific implementation process will not be repeated.

[0092] Furthermore, this embodiment also provides a storage medium on which a program is stored. When the program is executed, one or more steps of the aforementioned method are implemented, and the specific implementation process will not be repeated.

[0093] In some other embodiments, the following hardware and software devices may be involved to implement the data generation method provided by the present invention:

[0094] First, the data acquisition equipment includes a variety of sensors installed on the maglev train, such as acceleration sensors, laser rangefinders and displacement sensors, which can capture the dynamic changes of the train in operation in real time and process them into track unevenness data.

[0095] Secondly, the collected raw data needs to be processed by a preprocessing device to remove structural outliers. This step uses a sliding window to cross-apply a box plot to identify outliers, record the location of the outliers, and replace the outliers with linear differences. The preprocessing device includes a central processing unit, a programmable device or other computer processing unit, which stores a program that can implement the above-mentioned outlier identification process.

[0096] Finally, the time series generative adversarial network model is the core of the method provided by the present invention. In order to build and train the model, a high-performance computing platform is required to support the computing requirements of the deep learning model. The model consists of an embedding function, a recovery function, a generator, and a discriminator. The training of the model combines the first two parts of the autoencoder with the last two parts of the generative adversarial network to achieve unified processing of feature encoding, data generation representation, and iterative learning across time series.

[0097] The preferred specific embodiments of the present invention are described in detail above. It should be understood that a person skilled in the art can make many modifications and changes based on the concept of the present invention without creative work. Therefore, any technical solution that can be obtained by a person skilled in the art through logical analysis, reasoning or limited experiments based on the concept of the present invention on the basis of the prior art should be within the scope of protection determined by the claims.

Claims

1. A method for generating high-speed maglev track irregularity data based on a time series generative adversarial network, characterized in that: The following steps are involved: Obtain track irregularity data during the operation of the maglev train and perform preprocessing; The preprocessed track irregularity data is input into the pre-built time series generative adversarial network model for training and evaluation to obtain the track irregularity generation data; Among them, the time series generative adversarial network model includes an embedding function, a recovery function, a generator and a discriminator. The embedding function and the recovery function form an embedding network, the generator and the discriminator form a generative adversarial network, the embedding network is used to provide a latent space, and the generative adversarial network operates in the latent space.

2. The method for generating high-speed maglev track irregularity data based on a time series generative adversarial network according to claim 1 is characterized in that: The training process of the time series generative adversarial network model specifically includes: Based on the preprocessed track irregularity data, the embedding network is trained using a pre-constructed first loss function, and the generative adversarial network is trained using a pre-constructed second loss function.

3. The method for generating high-speed maglev track irregularity data based on time series generative adversarial network according to claim 2 is characterized in that: The expression of the first loss function is: In the formula, θ e ,θ r denote the parameters of the embedding network and the recovery network respectively, represents the supervised loss of the generator in closed-loop mode, represents the reconstruction loss, and λ is the balance and and λ≥0.

4. The method for generating high-speed maglev track irregularity data based on a time series generative adversarial network according to claim 2 is characterized in that: The expression of the second loss function is: In the formula, θ g ,θ d They represent the parameters of the generating network and the discriminating network respectively, represents the supervised loss of the generator in closed-loop mode, represents the unsupervised loss of the generator in open-loop mode, and η is the balance and Hyperparameters of .

5. The method for generating high-speed maglev track irregularity data based on time series generative adversarial network according to claim 1 is characterized in that: The embedding network and the generative adversarial network respectively adopt recursive neural network structures.

6. The method for generating high-speed maglev track irregularity data based on a time series generative adversarial network according to claim 1 is characterized in that: The method uses the inertial reference method to obtain track irregularity data under the running state of the maglev train, and the specific process includes: Obtain the acceleration signal of the maglev train in operation, and perform unified resampling and preprocessing; Acquire air gap signal; The processed acceleration signal is integrated and superimposed with the air gap signal to obtain a track irregularity signal.

7. The method for generating high-speed maglev track irregularity data based on a time series generative adversarial network according to claim 1 is characterized in that: The specific process of the pretreatment includes: plotting a box plot and calculating outlier limits using whiskers in the box plot, the box plot being used to identify and determine potential outliers in track irregularity data; Based on the outlier limit, a sliding window method is used to perform cross analysis on the track irregularity data to obtain the position of the outlier; The outliers are replaced by the linear difference of the normal data adjacent to them.

8. The method for generating high-speed maglev track irregularity data based on a time series generative adversarial network according to claim 1 is characterized in that: The evaluation process of the time series generative adversarial network model specifically includes: After obtaining the track irregularity generation data in each iteration, the classification error and the mean absolute error between the track irregularity generation data and the track irregularity data before preprocessing are calculated using the pre-built classification model and regression prediction model, and the generation capability of the model is quantified using the classification error and the mean absolute error.

9. A high-speed maglev track irregularity data generation device based on a time series generative adversarial network, comprising a memory, a processor, and a program stored in the memory, characterized in that: When the processor executes the program, the method according to any one of claims 1 to 8 is implemented.

10. A storage medium having a program stored thereon, characterized in that: When the program is executed, the method according to any one of claims 1 to 8 is implemented.

Images