A track irregularity prediction method based on knowledge linkage

Abstract

The application discloses a track irregularity prediction method based on knowledge linkage, and comprises the following steps: S1, determining whether it is the first time to predict, if yes, executing S2, and if no, executing S4; S2, cleaning historical data of track irregularity indexes, arranging each type of data corresponding to multiple types of track irregularity indexes in time sequence to obtain multiple sets of data; S3, training multiple LSTM prediction models by using the multiple sets of data obtained in S2; S4, collecting track irregularity data in real time; S5, cleaning the collected track irregularity data to obtain multiple sets of normalized data; S6, arranging each type of data in the multiple sets of normalized data obtained in S5 in time sequence and segmenting, and performing fuzzy clustering operation, and selecting main indexes from the track irregularity indexes as indexes after dimension reduction; and S7, inputting the normalized data of the indexes after dimension reduction obtained in S6 into the prediction model trained in S3 to predict the change trend of the track irregularity indexes.

Description

Technical Field

[0001] This invention relates to the field of railway surveying and tamping, and specifically to a method for predicting track irregularities based on knowledge linkage. Background Technology

[0002] Track irregularity is a core parameter in the condition assessment of high-speed railway infrastructure, and its dynamic changes directly affect train operation safety and passenger comfort.

[0003] Traditional detection methods mainly rely on manual inspections or offline testing equipment, which suffer from technical bottlenecks such as long data acquisition cycles, poor real-time performance, and insufficient analysis of multi-indicator correlations. As train speeds increase, the dynamic coupling effect of seven core indicators, including track gauge, track alignment, and elevation, intensifies. However, existing technologies generally employ single sensors for independent monitoring, making it difficult to acquire comprehensive spatiotemporal data, primarily in the following two aspects:

[0004] 1. At the data processing level, traditional methods have significant drawbacks: traditional methods generally crudely select a single indicator or all seven indicators for analysis. Single random selection lacks scientific rigor and is difficult to effectively represent the essential changes in smoothness. When all seven indicators are analyzed, the coupling relationship between data is difficult to decouple, and the same feature as the representative data will also lead to data redundancy. This not only increases the analysis time but also affects the data analysis.

[0005] 2. At the level of predictive model construction, methods based on statistical regression or shallow neural networks are difficult to capture nonlinear time series dynamic characteristics and lack in-depth analysis of the time series extrapolation patterns of multi-source data, making it difficult to achieve high-precision dynamic prediction of time series data. Summary of the Invention

[0006] To address the problems existing in the background technology, this invention provides an efficient knowledge-linked method for predicting track irregularities.

[0007] Therefore, the present invention adopts the following technical solution:

[0008] A knowledge-based method for predicting track irregularities includes the following steps:

[0009] S1, determine if this is the first prediction; if yes, execute S2; otherwise, execute S4.

[0010] S2, clean the historical data of the track irregularity index. After cleaning, arrange the data of each class of data corresponding to multiple track irregularity indices in chronological order to obtain a multi-class dataset.

[0011] S3 uses the multi-class dataset obtained in S2 to train multiple LSTM prediction models;

[0012] S4, real-time acquisition of track irregularity data;

[0013] S5 uses the same cleaning method as in S2 to clean gross errors and redundant data from the real-time track irregularity data, and unifies the data units to obtain multiple types of standardized data.

[0014] S6 arranges each category of the multi-class standardized data according to the time series, then segments the data according to the track length, and performs fuzzy clustering operation on multiple indicators in each segment using multi-source clustering analysis method. Indicators with similar influencing factors are grouped into one category, and then a representative indicator is selected from each category as the dimensionality-reduced indicator.

[0015] S7, Track irregularity prediction: Input the normalized data of the dimensionality-reduced index obtained in S6 into the prediction model trained in S3 to predict the changing trend of the track irregularity index.

[0016] The fuzzy clustering operation described in S6 uses the fuzzy K-means method, specifically:

[0017] First, based on the actual physical meaning of the track irregularity index, it is classified into N major categories. The objective function of the fuzzy K-means index is:

[0018]

[0019] in, For indicator data, The number of items in the indicator, here =1~7; Centered on the category; The number of major categories. ; Membership matrix Total number of indicators For hyperparameters; for Indicator data points to category center The distance;

[0020] The membership matrix and class centers are solved by iterative updates:

[0021] When the category centers are fixed, the membership matrix is:

[0022] ,

[0023] Where k is a variable, ;

[0024] when When fixed, the category center is:

[0025] ,

[0026] The category center is obtained through alternating iterative updates. After determining N category centers, the normalized data of multiple categories are classified into N categories according to the category centers, thereby achieving category dimensionality reduction.

[0027] The specific steps for S7 are as follows:

[0028] (1) Input time step data: Let each time step be... Normalized data of the dimensionality-reduced indicators Compared with the previous time period's trajectory irregularity prediction data Input the LSTM prediction model;

[0029] (2) Forget gate calculation: The forget gate determines the state to be retained from the previous state. The weight, The calculation formula is as follows:

[0030]

[0031] in, It is the weight of a fully connected network; It is a bias; It is the sigmoid function;

[0032] (3) Input gate calculation, which is divided into two parts:

[0033] First, determine the status of deep orbital irregularities data. t Weights to be retained , The calculation formula is as follows:

[0034]

[0035] in, These are the weights of a fully connected network, used to extract track irregularity data at the current moment. Data status of deep orbital irregularities t The relationship between them It is a bias;

[0036] Then, the deep orbital irregularity data state is calculated. t The calculation formula is as follows:

[0037] t

[0038] in, These are the weights of a fully connected network, used to map the state of deep orbital irregularity data. t ; It is the bias; tanh is the tanh activation function;

[0039] (4) Update state: extract the previous time-sequence state. The state of deep orbital irregularity data obtained in step (3) t Merge and update the state at time t. The update formula is as follows:

[0040] t

[0041] (5) Output gate calculation: State weights when the output gate controls the output. This is used to characterize the information loss of track irregularity data during the prediction and transmission process, and the calculation formula is as follows:

[0042]

[0043] in, It is the weight of a fully connected network; It is a bias;

[0044] (6) Hidden state calculation: through the aforementioned information loss and the state at time t Obtain the predicted value of the track irregularity data at the current moment. The calculation formula is as follows:

[0045] .

[0046] The track irregularity prediction method requires first defining the scope of track irregularity detection. Currently, within the railway system, track irregularities mainly include seven track irregularity indicators: track gauge, left track alignment, right track alignment, left elevation difference, right elevation difference, levelness, and triangular pits. The track quality index (TQI), which comprehensively considers these seven indicators, is used to express irregularities.

[0047] The data cleaning methods in S2 include: using bilinear interpolation to handle missing values; using one-SVM to detect and handle outliers; using duplicate values ​​at the same time point to detect and handle duplicate data; and using a standardized unit system to standardize the data.

[0048] The specific steps of S3 include: For each type of dataset, firstly, select the first M data points in chronological order as input, and select the (M+1)th data point as output to construct multiple sets of data pairs. Use the dataset composed of these data pairs as the training dataset. Then, input the data pairs in the training dataset into the LSTM model input interface according to the time series to train the LSTM model. Repeated training is performed by adjusting the hyperparameters, including the learning rate, optimization method, and optimization method parameters, to obtain a prediction model with higher accuracy.

[0049] In S4, a track inspection trolley is used to collect track irregularity data in real time, and the collected data is transmitted to the track maintenance section via the network to achieve the purpose of receiving inspection data in real time.

[0050] In one embodiment of the present invention, the number of predictions in S7 is set to 8.

[0051] In one embodiment of the present invention, in S6, the system is divided into segments of 200 meters each; during the fuzzy clustering operation, N=3; and the hyperparameter m=2.

[0052] Preferably, in S3, M=4.

[0053] This invention addresses the intelligent operation and maintenance needs of rail transit by providing a knowledge-based method for predicting track irregularities. This method achieves track irregularity prediction by integrating cloud-edge collaborative data acquisition (using an integrated mobile acquisition device that integrates seven index acquisition sensors), multi-source clustering analysis, and LSTM prediction.

[0054] Compared with the prior art, the present invention has the following beneficial effects:

[0055] 1. The track irregularity prediction method of the present invention reduces the dimensionality of the detection indicators through fuzzy clustering operation, retains the core features, and then selects the most representative indicators for the railway section for detection. While reducing interference indicators, it also achieves dimensionality reduction of massive data and improves the prediction efficiency.

[0056] 2. This invention uses an LSTM neural network to construct a time series prediction model, fully explores the temporal patterns and spatial correlations in historical data, effectively utilizes historical information, breaks through the technical bottleneck of traditional technology in dynamic prediction, and provides an efficient solution for preventive maintenance of rail infrastructure. Attached Figure Description

[0057] Figure 1 This is a flowchart of the track irregularity prediction method of the present invention. Detailed Implementation

[0058] The technical solution of the present invention will be clearly and completely described below with reference to the accompanying drawings and embodiments. Obviously, the following embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0059] First, it is necessary to define the scope of track irregularity detection. Currently, within the railway system, track irregularity mainly includes seven track irregularity indicators: track gauge, left track alignment, right track alignment, left elevation difference, right elevation difference, levelness, and triangular pits. The Track Quality Index (TQI) is mainly used to express irregularity by comprehensively considering these seven indicators.

[0060] The following uses seven track irregularity indicators as examples to illustrate the knowledge-linked track irregularity prediction method of this invention. Figure 1 As shown, the method includes the following steps:

[0061] S1, determine if this is the first prediction; if yes, execute S2; otherwise, execute S4.

[0062] S2, Historical data cleaning and dataset construction for track irregularity indicators:

[0063] The data cleaning methods were as follows: bilinear interpolation was used to handle missing values; one-SVM was used to detect and handle outliers; duplicate data was handled by detecting duplicate values ​​at the same time point; and data standardization was achieved using a standardized unit system. After data cleaning, the data in each of the seven categories corresponding to the seven track irregularity indicators were arranged in chronological order, resulting in seven datasets.

[0064] S3, LSTM Model Training: Seven LSTM prediction models are trained using the seven-class dataset obtained in S2. The specific steps include:

[0065] For each dataset, the first M data points are selected sequentially as inputs, and the (M+1)th data point is selected as output, thus constructing multiple data pairs. The dataset consisting of these data pairs is used as the training dataset. Then, the data pairs in the training dataset are grouped according to time series and input into the input interface of the LSTM model to train the LSTM model. The hyperparameters, including the learning rate, optimization method, and optimization method parameters, are repeatedly trained to obtain a prediction model with high accuracy.

[0066] S4, Data Acquisition: Real-time acquisition of track irregularity data, and transmission of the acquired data to the track maintenance section via network, to achieve the purpose of receiving detection data in real time;

[0067] S5, Data Cleaning: The real-time track irregularity data is cleaned using the data cleaning method in S2 to remove gross errors and redundant data, and the data units are standardized to obtain 7 types of standardized data.

[0068] S6. Arrange each of the seven types of standardized data obtained according to the time series, and then segment the data according to the track length (e.g., segment by 200 meters). Perform fuzzy clustering operation on the seven indicators in each segment of data using multi-source clustering analysis. Group the indicators with similar influencing factors into one category, and then select a representative indicator from each category as the representative class as the dimensionality-reduced indicator.

[0069] In embodiments of the present invention, the fuzzy clustering operation employs the fuzzy K-means method. Specifically:

[0070] First, based on the actual physical meaning of the track irregularity index, it is classified into N major categories. The objective function of the fuzzy K-means index is:

[0071]

[0072] in, For indicator data, The number of items in the indicator, here =1~7; Centered on the category; The number of major categories. ; Membership matrix Total number of indicators For hyperparameters; for Indicator data points to category center The distance.

[0073] When solving for the membership matrix and class centers, iterative updates are required. The Lagrange method can be used to obtain the following:

[0074] When the category centers are fixed, the membership matrix is:

[0075] ,

[0076] Where k is a variable, .

[0077] when When fixed, the category center is:

[0078] ,

[0079] The category center is obtained through alternating iterative updates. After determining N category centers, the 7 categories of normalized data are classified into N categories according to the category centers, thus achieving category dimensionality reduction.

[0080] S7. Input the normalized data of the dimensionality-reduced index obtained in S6 into the prediction model trained in S3 to predict the changing trend of the track irregularity index. The specific steps are as follows:

[0081] (1) Input time step data: Let each time step be... Normalized data of the dimensionality-reduced indicators Compared with the previous time period's trajectory irregularity prediction data Input the LSTM prediction model.

[0082] (2) Forget gate calculation: The forget gate determines the state to be retained from the previous state. The weight, The calculation formula is as follows:

[0083]

[0084] in, It is the weight of a fully connected network; It is a bias (obtained through training); It is the sigmoid function (outputs the forgetting weight, i.e. the proportion of forgotten information, used to filter out useless information in the track irregularity data);

[0085] (3) Input gate calculation, which is divided into two parts:

[0086] First, determine the status of deep orbital irregularities data. t Weights to be retained , The calculation formula is as follows:

[0087]

[0088] in, These are the weights of a fully connected network, used to extract track irregularity data at the current moment. Data status of deep orbital irregularities t The relationship between them It is a bias;

[0089] Then, the deep orbital irregularity data state is calculated. t The calculation formula is as follows:

[0090] t

[0091] in, These are the weights of a fully connected network, used to map the state of deep orbital irregularity data. t ; It is the bias; tanh is the tanh activation function.

[0092] (4) Update state: extract the previous time-sequence state. The state of deep orbital irregularity data obtained in step (3) t Merge and update the state at time t. The update formula is as follows:

[0093] t

[0094] (5) Output gate calculation: State weights when the output gate controls the output. This is used to characterize the information loss of track irregularity data during the prediction and transmission process, and the calculation formula is as follows:

[0095]

[0096] in, It is the weight of a fully connected network; It is a bias.

[0097] (6) Hidden state calculation: through the aforementioned information loss and the state at time t Obtain the predicted value of the track irregularity data at the current moment. The calculation formula is as follows:

[0098]

[0099] Example

[0100] First, a track inspection trolley is used to acquire track irregularity data in real time, and the data is cleaned in real time to obtain 7 types of standardized data;

[0101] Then, the seven types of standardized data are segmented in 200-meter segments; fuzzy clustering is performed on the seven indicators in each segment. Through fuzzy clustering, the seven track irregularity indicators are classified into three major categories (N=3) based on their actual physical meaning. The specific operation is as follows:

[0102] First, three initial category centers are randomly selected, and the membership degree of each indicator's data point to the initial category center is calculated. Then, the membership degree is fixed, new category centers are obtained, and the process is repeated iteratively to obtain three final category centers.

[0103] Subsequently, the seven track irregularity indicators were clustered into three categories: the first category was track gauge; the second category was left track alignment and right track alignment; and the third category was left elevation, right elevation, horizontal alignment, and triangular pits.

[0104] From the three major categories, track gauge, left track alignment, and right elevation were selected as representative categories, thus reducing the original seven track irregularity indicators to three representative categories.

[0105] In this embodiment, during the fuzzy clustering operation, the hyperparameters =2.

[0106] Based on the data of the three representative indicators, the trained prediction models are used to predict the trends of the three types of data respectively. The track gauge, left track direction, and right elevation and depression data from the first four tests are input into the track gauge prediction model, the left track direction prediction model, and the right elevation and depression prediction model respectively, and the number of predictions is set.

[0107] In this embodiment, the prediction count is set to 8 times. Through continuous prediction, 8 predicted values ​​of 3 representative index data are obtained. Then, based on the predicted values ​​of the 3 representative index data, the TQI is calculated to achieve track irregularity evaluation.

Claims

1. A method for predicting track irregularities based on knowledge linkage, characterized in that, Includes the following steps: S1, determine if this is the first prediction; if yes, execute S2; otherwise, execute S4. S2, clean the historical data of the track irregularity index. After cleaning, arrange the data of each class of data corresponding to multiple track irregularity indices in chronological order to obtain a multi-class dataset. S3 uses the multi-class dataset obtained in S2 to train multiple LSTM prediction models; S4, real-time acquisition of track irregularity data; S5 uses the same cleaning method as in S2 to clean gross errors and redundant data from the real-time track irregularity data, and unifies the data units to obtain multiple types of standardized data. S6 arranges each category of the multi-class standardized data according to the time series, then segments the data according to the track length, and performs fuzzy clustering operation on multiple indicators in each segment using multi-source clustering analysis method. Indicators with similar influencing factors are grouped into one category, and then a representative indicator is selected from each category as the dimensionality-reduced indicator. S7, Track irregularity prediction: Input the normalized data of the dimensionality-reduced index obtained in S6 into the prediction model trained in S3 to predict the trend of track irregularity index changes.

2. The method for predicting track irregularities according to claim 1, characterized in that, The fuzzy clustering operation described in S6 uses the fuzzy K-means method, specifically: First, based on the actual physical meaning of the track irregularity index, it is classified into N major categories. The objective function of the fuzzy K-means index is: ， in, For indicator data, The number of items in the indicator, here =1~7; Centered on the category; The number of major categories. ; Membership matrix This represents the total number of indicators; For hyperparameters; for Indicator data points to category center The distance; The membership matrix and class centers are solved by iterative updates: When the category centers are fixed, the membership matrix is: ， Where k is a variable, ; when When fixed, the category center is: ， The category center is obtained through alternating iterative updates. After determining N category centers, the normalized data of multiple categories are classified into N categories according to the category centers, thereby achieving category dimensionality reduction.

3. The method for predicting track irregularities according to claim 1, characterized in that... The specific steps for S7 are as follows: (1) Input time step data: Let each time step be... Normalized data of the dimensionality-reduced indicators Compared with the previous time period's trajectory irregularity prediction data Input the LSTM prediction model; (2) Forget gate calculation: The forget gate determines the state to be retained from the previous state. The weight, The calculation formula is as follows: ， in, It is the weight of a fully connected network; It is a bias; It is the sigmoid function; (3) Input gate calculation, which is divided into two parts: First, determine the status of deep orbital irregularities data. t Weights to be retained , The calculation formula is as follows: ， in, These are the weights of a fully connected network, used to extract track irregularity data at the current moment. Data status of deep orbital irregularities t The relationship between them It is a bias; Then, the deep orbital irregularity data state is calculated. t The calculation formula is as follows: t ， in, These are the weights of a fully connected network, used to map the state of deep orbital irregularity data. t ; It is the bias; tanh is the tanh activation function; (4) Update state: extract the previous time-sequence state. The state of deep orbital irregularity data obtained in step (3) t Merge and update the state at time t. The update formula is as follows: t ， (5) Output gate calculation: State weights when the output gate controls the output. This is used to characterize the information loss of track irregularity data during the prediction and transmission process, and the calculation formula is as follows: ， in, It is the weight of a fully connected network; It is a bias; (6) Hidden state calculation: through the aforementioned information loss and the state at time t Obtain the predicted value of the track irregularity data at the current moment. The calculation formula is as follows: 。 4. The method for predicting track irregularities according to claim 1, characterized in that: The track irregularity indicators mentioned in S2 include track gauge, left track alignment, right track alignment, left elevation difference, right elevation difference, levelness, and triangular pits.

5. The method for predicting track irregularities according to claim 1, characterized in that, Historical data cleaning in S2 includes: using bilinear interpolation to handle missing values; using the one-SVM method to detect and handle outliers; using duplicate values ​​at the same time point to detect and handle duplicate data; and using a standardized unit system to standardize the data.

6. The method for predicting track irregularities according to claim 1, characterized in that, The specific steps of S3 include: For each dataset, the first M data points are selected sequentially as inputs, and the (M+1)th data point is selected as the output to construct multiple data pairs. These data pairs form the training dataset. Then, the data pairs in the training dataset are grouped according to time series and input into the LSTM model input interface to train the LSTM model. The hyperparameters, including the learning rate, optimization method, and optimization method parameters, are repeatedly trained to obtain a prediction model with high accuracy.

7. The method for predicting track irregularities according to claim 1, characterized in that, In S4, a track inspection trolley is used to collect track irregularity data in real time, and the collected data is transmitted to the track maintenance section via the network to achieve the purpose of receiving inspection data in real time.

8. The method for predicting track irregularities according to claim 1, characterized in that, In S7, the number of predictions is set to 8.

9. The method for predicting track irregularities according to claim 2, characterized in that, In S6, the system is segmented into 200-meter segments; during fuzzy clustering, N=3; hyperparameters... =2.

10. The method for predicting track irregularities according to claim 6, characterized in that: In S3, M=4.

Images