Railway Track Wear Assessment Method and System

Abstract

The present invention provides a method and system for evaluating the wear amount of railway tracks, which relates to the technical field of evaluating the wear amount of railway tracks. The method includes obtaining train operation data and track historical wear data, and performing data cleaning and completion through a preprocessing model based on a variational autoencoder and a Gaussian process regression model. Then, wavelet transform and locally weighted regression processing are performed on the preprocessed data to extract local features and linear relationships, and a feature data set is constructed. Further, dimensionality reduction is carried out through an extreme learning machine combined with a principal component analysis model, and classification is performed using a Gaussian mixture model to generate a set of class features. The feature set is input into a prediction model and the prediction result is optimized by combining a convolutional neural network. Subsequently, reliability analysis is carried out, and the results with a reliability higher than the threshold are used as the final evaluation results of the track wear amount. The present invention improves the accuracy and reliability of wear prediction and is applicable to the intelligent maintenance and management of railway tracks.

Description

Technical Field

[0001] The present invention relates to the technical field of railway track wear amount assessment, and more particularly, to a method and system for assessing railway track wear amount. Background Art

[0002] With the continuous development of railway transportation, the increase in train running speed and load has had a significant impact on the wear degree of railway tracks. Track wear not only affects the driving safety of trains but also increases maintenance costs. Traditionally, the assessment of railway track wear amount relies on manual inspection and empirical calculation, usually requiring regular track inspections, measurements, or sampling to obtain wear data. Although these methods can provide basic data support for track maintenance, due to manual inspection errors, sampling incompleteness, and their time-consuming nature, these traditional methods are difficult to meet the requirements of accuracy and real-time performance in the increasingly complex railway operation environment.

[0003] Most existing technologies use mechanical sensors or simple statistical models to assess track wear. However, these methods often show certain limitations when faced with a large amount of multi-dimensional operation data. For example, a simple linear regression model has low accuracy in dealing with non-linear relationships and cannot effectively handle noise and missing values in track data. At the same time, traditional methods cannot achieve high accuracy and reliability in long-term prediction of track wear. In addition, existing technologies do not adequately consider non-linear relationships and the interaction of multiple influencing factors in complex systems, resulting in limitations in the adaptability and accuracy of the model.

[0004] Therefore, there is an urgent need for a method and system for assessing railway track wear amount to solve the above problems. Summary of the Invention

[0005] The purpose of the present invention is to provide a method and system for assessing railway track wear amount to improve the above problems. To achieve the above purpose, the technical solutions adopted by the present invention are as follows:

[0006] In a first aspect, the present application provides a method for assessing railway track wear amount, including:

[0007] Obtain first data, where the first data includes train operation data and track historical wear data;

[0008] Send the first data to a preprocessing model for preprocessing, where the first data is sequentially sent to a variational autoencoder and a Gaussian process regression model based on a preset for cleaning and complementing to obtain preprocessed first data;

[0009] Perform wavelet transform and locally weighted regression processing on the preprocessed first data to obtain local features and their local linear relationships of all the first data, and use them as a feature dataset;

[0010] Send the feature data set to a preset dimensionality reduction model for processing. Among them, the feature data set is sequentially subjected to feature relationship processing by an extreme learning machine and principal component analysis for dimensionality reduction, and is processed by a preset Gaussian mixture model to obtain at least one category feature set;

[0011] Send the feature sets of all categories to a preset prediction model for processing, and send the prediction results to a convolutional neural network model for optimization to obtain optimized prediction results;

[0012] Perform reliability analysis on the optimized prediction results, and use the prediction results with reliability greater than a preset threshold as the evaluation results of the rail wear amount.

[0013] In a second aspect, the present application also provides a railway track wear amount evaluation system, including:

[0014] An acquisition unit for acquiring first data, where the first data includes train operation data and track historical wear data;

[0015] A first processing unit for sending the first data to a preprocessing model for preprocessing. Among them, the first data is sequentially sent to a variational autoencoder and a Gaussian process regression model based on a preset for cleaning and complementing to obtain preprocessed first data;

[0016] A second processing unit for performing wavelet transform and locally weighted regression processing on the preprocessed first data to obtain the local features and their local linear relationships of all the first data, and using them as a feature data set;

[0017] A third processing unit for sending the feature data set to a preset dimensionality reduction model for processing. Among them, the feature data set is sequentially subjected to feature relationship processing by an extreme learning machine and principal component analysis for dimensionality reduction, and is processed by a preset Gaussian mixture model to obtain at least one category feature set;

[0018] A fourth processing unit for sending the feature sets of all categories to a preset prediction model for processing, and sending the prediction results to a convolutional neural network model for optimization to obtain optimized prediction results;

[0019] A fifth processing unit for performing reliability analysis on the optimized prediction results, and using the prediction results with reliability greater than a preset threshold as the evaluation results of the rail wear amount.

[0020] The beneficial effects of the present invention are:

[0021] It is understandable that the present invention provides a method and system for evaluating the wear amount of railway tracks, aiming to solve the deficiencies in data processing ability, accuracy, and real-time performance in the prior art. The present invention combines a variational autoencoder and a Gaussian process regression model for data cleaning and completion, uses wavelet transform and locally weighted regression processing to extract local features of track wear data, and combines extreme learning machine and principal component analysis for data dimensionality reduction. Furthermore, the present invention realizes accurate feature extraction through Gaussian mixture model clustering processing. In addition, the present invention optimizes the prediction results by integrating a convolutional neural network, and uses adaptive weighted learning and gradient boosting decision tree for error adjustment. Finally, high-confidence evaluation results are screened through reliability analysis. This method can effectively improve the accuracy, reliability, and real-time performance of track wear amount evaluation, especially showing significant advantages in multi-dimensional and diverse data processing and non-linear relationship modeling, and having higher prediction accuracy and stronger adaptability compared with traditional technologies.

[0022] Other features and advantages of the present invention will be described in the subsequent specification, and part of them will become obvious from the specification, or can be understood by implementing the embodiments of the present invention. The objectives and other advantages of the present invention can be achieved and obtained through the structures specifically pointed out in the written specification, claims, and drawings. Brief Description of the Drawings

[0023] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required for the embodiments. It should be understood that the following drawings only show some embodiments of the present invention, and therefore should not be regarded as limiting the scope. For those of ordinary skill in the art, other related drawings can be obtained based on these drawings without creative efforts.

[0024] Figure 1 It is a schematic flow chart of the method for evaluating the wear amount of railway tracks described in the embodiments of the present invention;

[0025] Figure 2 It is a schematic structural diagram of the system for evaluating the wear amount of railway tracks described in the embodiments of the present invention.

[0026] In the figure: 701, acquisition unit; 702, first processing unit; 703, second processing unit; 704, third processing unit; 705, fourth processing unit; 706, fifth processing unit. Detailed Embodiments

[0027] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and illustrated herein generally can be arranged and designed in a variety of different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the claimed invention, but merely represents selected embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts fall within the scope of protection of the present invention.

[0028] It should be noted that like reference numerals and letters denote like items in the following drawings. Therefore, once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings. At the same time, in the description of the present invention, the terms "first", "second", etc. are only used for differential description and cannot be construed as indicating or implying relative importance.

[0029] Embodiment 1:

[0030] This embodiment provides a method for evaluating the wear amount of a railway track.

[0031] See Figure 1 , which shows that this method includes step S1, step S2, step S3, step S4, step S5, and step S6.

[0032] Step S1: Obtain first data, where the first data includes train operation data and track historical wear data;

[0033] It can be understood that train operation data covers a variety of dynamic and static parameters, such as train speed, axle load, acceleration, operation time, line type, ambient temperature, humidity, and vibration frequency. These data are often collected in real time by a variety of sensors (such as acceleration sensors, laser displacement sensors, temperature and humidity sensors) installed on trains or tracks. In addition, to ensure the spatio-temporal synchronization of the data, a high-precision GPS positioning system and a time synchronization protocol (PTP) are also used for data timestamp calibration to ensure that all collected data is in the same time coordinate system. Track historical wear data usually comes from regular track inspection vehicles or manual inspection records, and the recorded parameters include track side wear depth, vertical wear, corrugation frequency, wear rate, and track material characteristics.

[0034] The first data obtained through the above process has a high degree of spatio-temporal consistency, data diversity, and integrity, providing a solid foundation for subsequent data preprocessing, feature extraction, and wear prediction. Such high-quality data collection can not only accurately reflect the load changes during train operation and their impact on track wear but also comprehensively capture the wear patterns of the track under different working conditions, improving the accuracy and stability of the entire railway track wear assessment method and providing more scientific and reliable data support for track maintenance decisions.

[0035] Step S2: Send the first data to a preprocessing model for preprocessing, where the first data is sequentially sent to a variational autoencoder and a Gaussian process regression model based on a preset for cleaning and completion to obtain the preprocessed first data.

[0036] It can be understood that in this step, by combining the variational autoencoder and the Gaussian process regression, the quality and integrity of the first data can be effectively improved. The variational autoencoder can capture the latent features of the data during the noise reduction process, thus reducing the impact of noise on subsequent analysis; while the Gaussian process regression uses the covariance relationship between data to ensure that the completed data is smooth and consistent. This combined method has significant advantages in dealing with high-dimensional, non-linear, and noisy data, providing a reliable basis for subsequent wavelet transform, feature extraction, and prediction modeling, and ultimately improving the accuracy and stability of the entire railway track wear assessment process. In this step, step S2 includes steps S21, S22, S23, and S24.

[0037] Step S21: Input the first data into a variational autoencoder for noise reduction processing. Among them, the variational autoencoder maps the input data into a latent space, the variational autoencoder learns the internal distribution of the data, distinguishes the noise from the key features in the data, and maps the data points in the latent space back to the original space to obtain the denoised data.

[0038] It can be understood that in this step, during the encoding stage, the first data (including train operation data and track historical wear data) is input into the encoder network, which is usually composed of multiple layers of neural networks. The encoder maps the high-dimensional, noisy data into a low-dimensional latent space. During this process, the encoder does not directly output a definite latent representation but outputs a set of parameters, namely the mean vector and variance vector of the latent variables, which define the probability distribution of the latent variables. This processing method enables the variational autoencoder to capture the uncertainty of the data, helping to distinguish noise and key features. For example, abnormal fluctuations in train speed or track wear data will be regarded as noise, while the distribution in the latent space can retain the important features related to real track wear.

[0039] During the decoding stage, the sampled data in the latent space is mapped back to the original data space through the decoder network. The decoder is also composed of multiple layers of neural networks and can reconstruct the low-dimensional latent representation into the denoised high-dimensional data. During the reconstruction process, the decoder tries to retain the key features of the original data while suppressing the presence of noise. This reconstruction mechanism ensures that the denoised data is closer to the real track wear and train operation conditions.

[0040] Step S22: Extract the latent variables of the first data based on the variational autoencoder, and map the latent variables to be represented in a low-dimensional space to obtain the latent variable data in the low-dimensional space. The latent variables are data of track wear, train load, running speed, and track type.

[0041] It can be understood that the latent variables extracted from the latent space of the variational autoencoder are denoted as , where represents the set of latent variables, is a vector containing features such as track wear, train load, running speed, and track type. These latent variables are already high-dimensional representations that have been denoised and feature-compressed and contain the non-linear coupling relationships between various types of data. However, to improve the efficiency of subsequent analysis and prediction, these latent variables need to be further mapped to a low-dimensional space. First, the latent variables are initially reduced in dimension through the principal component analysis method. The goal of the principal component analysis method is to find the directions with the largest variance in the data and project the data onto these directions.

[0042] Step S23: Learn the covariance structure between the latent variable data in the low-dimensional space based on a preset Gaussian process regression model, and predict the missing values based on the covariance structure.

[0043] It can be understood that Gaussian process regression is a non-parametric Bayesian regression method used to describe the function distribution. The Gaussian process regression model defines the similarity between data points through a covariance function (also known as a kernel function), thereby learning the internal structure of the data and predicting the missing values based on the observed data. Its basic model is expressed as:

[0044] ;

[0045] where is the objective function to be fitted, is the mean function, is the covariance function, and the covariance function is a commonly used Gaussian function, represents the Gaussian process.

[0046] In this step, the covariance matrix between latent variables is calculated through the covariance function, and then the kernel function parameters are optimized using the gradient descent method to ensure that the covariance structure can accurately describe the relationship between latent variables. After learning the covariance structure, the missing values can be predicted through the Gaussian process. Among them, the Gaussian process assumes that the joint distribution of the observed values and the predicted values follows a multivariate Gaussian distribution.

[0047] Step S24: Clean and complete the first data based on the denoised data and the missing values to obtain the preprocessed first data.

[0048] It can be understood that data cleaning and data completion are used to improve the data quality, making the subsequent feature extraction and analysis more accurate and effective. The whole process involves the reasonable combination of data processing algorithms to ensure the integrity and consistency of the data, while effectively removing noise and filling in the missing data.

[0049] Step S3: Perform wavelet transform and locally weighted regression processing on the preprocessed first data to obtain the local features and their local linear relationships of all the first data, and use them as the feature dataset;

[0050] It can be understood that this step reveals the subtle but important local patterns in the data through multi-scale analysis and local regression methods, which is especially suitable for the dynamic characteristic analysis of track wear and train operation data. These methods can effectively capture the features of the data at different time and frequency scales while retaining the details of the track wear evolution. In this step, step S3 includes step S31, step S32, and step S33.

[0051] Step S31: Select a preset wavelet transform function, preset the number of decomposition layers to three, and decompose the preprocessed first data in at least two frequency bands based on the preset wavelet transform function to obtain the signal information at at least two scales;

[0052] It can be understood that the preset wavelet transform function is as follows:

[0053] ;

[0054] Among them, is the result of the wavelet transform, which represents the input signal at the scale parameter and the translation position of the transform result, represents the input signal, represents the wavelet basis function, represents the time variable, represents the scale parameter, represents the translation position.

[0055] Step S32: Input the signal information at all scales into a preset locally weighted regression model to calculate the weights of adjacent data points, and use the weighted data points to train the locally weighted regression model to obtain a trained locally weighted regression model;

[0056] It can be understood that locally weighted regression is a non-parametric regression method using weighted least squares. In this model, the weights are calculated through distance metrics. Data closer to the target data point will be assigned higher weights, while data farther from the target point will be assigned lower weights. The advantage of locally weighted regression is that it can capture more accurate trends and relationships within the local region of the data. After calculating the weights of each data point, the regression model is trained based on the weighted data points. By minimizing the weighted error function, the model can find an optimal fitting curve for each data point. This means that the local linear relationship between signal data points will be preserved, and the model will adaptively adjust to emphasize those regions with significant local data changes and ignore regions far from the target. Its training process is that when using the weighted data points to train the model, locally weighted regression performs linear regression on each set of adjacent data points and their weights to obtain local fitting coefficients. This training method can flexibly adjust the model parameters according to local changes at different scales, thereby capturing the non-linear change characteristics of track wear data.

[0057] Step S33: Send the signal information at the at least two scales to the trained locally weighted regression model for prediction to obtain a prediction result, where the prediction result represents the local characteristics and non-linear relationship of the target point corresponding to the preprocessed first data.

[0058] It can be understood that in this step, the signal information at multiple scales is input into the trained locally weighted regression model. Locally weighted regression adjusts the regression function according to the local neighborhood of each data point through weighting, so it has significant advantages in capturing local changes in the data. For the track wear amount evaluation task, the locally weighted regression model will fit a linear model near each data point based on signal changes at different scales. Since higher weights are assigned to the adjacent points of each data point, the model can more accurately predict the local characteristics of track wear.

[0059] After prediction by the locally weighted regression model, the obtained prediction result can reflect the local characteristics and non-linear relationship of the track wear amount data. Specifically, the result predicted by the model not only gives an estimated value of the track wear amount but also can capture the complex non-linear relationship between track wear and factors such as train load, running speed, and track type.

[0060] Step S4: Send the feature dataset to a preset dimensionality reduction model for processing. Specifically, perform dimensionality reduction on the feature dataset by successively conducting feature relationship processing using an extreme learning machine and principal component analysis, and then process it through a preset Gaussian mixture model to obtain at least one category feature set;

[0061] It can be understood that in this step, through the mapping of the non - linear activation function, the complex non - linear relationship between train operation and track wear can be recognized. Especially when dealing with dynamic and high - dimensional track wear data, the extreme learning machine can extract effective features. The principal component analysis method maximizes the variance of the data while reducing dimensionality, reducing redundant information, which helps to improve the computational efficiency and model accuracy. By clustering through the Gaussian mixture model, the data can be automatically grouped, thus providing clear category features for the subsequent prediction model. This process helps to identify different wear patterns, and further improve the prediction accuracy of the track wear amount. Step S4 includes steps S41, S42, S43, S44, and S45.

[0062] Step S41: Send the feature dataset into a preset extreme learning machine model for processing. Specifically, the extreme learning machine model maps the feature dataset through the ReLU function of the hidden - layer neurons, captures the non - linear relationship of the feature dataset in the high - dimensional space, and continuously performs backpropagation and gradient descent until the non - linear relationships of all data in the high - dimensional space are captured, obtaining a feature dataset containing the non - linear relationships of all data;

[0063] It can be understood that the extreme learning machine maps the input feature dataset to a high - dimensional space through the ReLU function of the hidden - layer neurons and searches for the non - linear relationship between features in this space. During this process, the extreme learning machine continuously adjusts the weights between the hidden layer and the output layer. Although the training process of the extreme learning machine does not require backpropagation and gradient descent in traditional neural networks to update all weights, it still uses the least - squares method to optimize the weights of the output layer according to the difference between the input data and the expected output. In this way, the extreme learning machine can gradually capture the non - linear features of the data.

[0064] In this step, through the mapping and optimization process, the extreme learning machine can effectively learn all the non - linear relationships in the input dataset and provide an efficient and accurate feature dataset for subsequent modeling. Although the extreme learning machine does not perform conventional backpropagation during training, its method of directly determining the weights by randomly initializing the weights and using the least - squares method in the output layer ensures the full capture of non - linear relationships.

[0065] Step S42: Calculate the covariance matrix of the feature dataset containing the non-linear relationships of all data, where each covariance in the covariance matrix represents the non-linear relationship between each feature;

[0066] It can be understood that calculating the covariance matrix of the feature dataset enables the model to better understand and capture the complex relationships between different factors (such as train operation load, speed, track type, etc.). This is particularly important for improving the prediction accuracy of track wear. For example, certain train operation conditions (such as high load and low speed) may generate wear patterns different from those under normal operation conditions, and traditional linear methods may not be able to accurately capture this. Through the calculation of the non-linear covariance matrix, the model can better handle this complex multi-factor interaction and thus make more accurate predictions. The calculation formula of the covariance matrix is as follows:

[0067] ;

[0068] where, is the covariance matrix, is the th data point, is the mean vector of the features, is the number of data samples.

[0069] Step S43: Perform eigenvalue decomposition on the covariance matrix to obtain eigenvectors and eigenvalues. The eigenvalues represent the weights of each eigenvector in the data, and the eigenvectors represent the directions of each eigenvector;

[0070] It can be understood that through eigenvalue decomposition in this step, we can identify the principal components in the data and extract the most representative features, which is crucial for subsequent dimensionality reduction, feature selection, and improving the computational efficiency of the model. For the railway track wear assessment scenario, eigenvalue decomposition can help us find the factors most relevant to track wear, remove noise features, and thus make the subsequent prediction model more accurate and efficient.

[0071] Step S44: Select the eigenvectors corresponding to the top preset number of the largest eigenvalues from largest to smallest, and map the feature dataset containing the non-linear relationships of all data to the eigenvectors corresponding to the top preset number of the largest eigenvalues to obtain the dimensionality-reduced feature dataset;

[0072] Step S45: Input the dimensionality-reduced feature dataset into a preset Gaussian mixture model for clustering. Assume that each data point in the dimensionality-reduced dataset is randomly drawn from a Gaussian distribution, estimate the probability that each data point belongs to the Gaussian distribution based on the maximum expectation algorithm, and classify each data in the dimensionality-reduced dataset based on the probability to obtain at least one class feature set.

[0073] It can be understood that in this step, by selecting the most important feature vectors and mapping the data to a new feature space, we can significantly reduce the data dimension, thereby reducing the computational complexity and improving the computational efficiency. In the assessment of railway track wear, dimensionality reduction helps to remove irrelevant or redundant features, highlighting the features most relevant to the wear amount, thus enhancing the accuracy and robustness of the prediction model. Since a large amount of train operation data and track wear data are usually collected in the railway system, through this dimensionality reduction method, we can quickly and effectively identify several key factors most relevant to track wear, providing a more streamlined and efficient dataset for subsequent predictive modeling.

[0074] Step S5: Send the feature sets of all categories to a preset prediction model for processing, and send the prediction results to a convolutional neural network model for optimization to obtain optimized prediction results;

[0075] It can be understood that in this step, through the further optimization of the prediction results by the convolutional neural network, the accuracy of railway track wear prediction can be greatly improved. Especially when dealing with high-dimensional and complex data, the convolutional neural network can identify the information most useful for prediction through the extraction of local features and hierarchical learning. Step S5 includes Step S51, Step S52, and Step S53.

[0076] Step S51: Train the track historical wear data and feature data in the feature sets of all categories based on a random forest regression model. Among them, through multiple iterations of calculating the mean squared error until the error between the prediction result and the track historical wear data is less than a preset threshold, a trained random forest regression model is obtained;

[0077] It can be understood that in this step, first, multiple decision trees are trained using the feature dataset. Each tree is trained from a random subset of the original training set and a subset of features, generating a subset of training data through sampling with replacement of the data and selecting different subsets of features for training. Each tree splits nodes according to different features of the training data during training until a preset stopping condition is reached, and the stopping condition is the maximum depth of the preset tree.

[0078] The multiple iterations of calculating the mean squared error are to repeatedly calculate the mean squared error and continuously adjust the prediction of each tree until the prediction error is less than a preset threshold. The setting of this threshold is usually determined by domain experts according to actual needs, aiming to ensure that the model has sufficient prediction accuracy.

[0079] Step S52: Send the feature sets of all categories to the trained random forest regression model for prediction to obtain the first wear result of the track;

[0080] It can be understood that through repeated training and optimization, the random forest can effectively capture the complex non-linear relationship between the input features and the track wear. By integrating the prediction results of multiple decision trees, the model can reduce the overfitting risk of a single tree, thereby improving the prediction accuracy. The random forest has strong robustness to noise data and missing data. In practical applications, the track wear data may be affected by various factors (such as sensor errors, environmental changes, etc.), and the random forest can better handle these problems through the way of ensemble learning, improving the stability of the model.

[0081] Step S53: Input the first wear result of the track into the convolutional neural network for optimization. Among them, the convolutional layer of the convolutional neural network performs a sliding window process on the first wear result through a set of convolutional kernels, performs weighted summation on the local regions at different positions in the first wear result, and performs a max-pooling operation on the local features obtained by the weighted summation through the pooling layer. Finally, each local feature is extracted through the fully connected layer to obtain the final prediction result.

[0082] It can be understood that the convolutional layer is the core component of the convolutional neural network, and its main function is to extract local features in the data through convolutional operations. The convolutional layer performs a sliding window process on the input first wear result through a set of convolutional kernels (filters). The specific steps are as follows:

[0083] Sliding window operation: The convolutional kernel slides on the input data and performs convolutional calculations on each local region. The convolutional calculation obtains feature information through weighted summation of the local region.

[0084] Weighted summation of local regions: Each convolutional kernel performs a dot product operation with the local region in the first wear result and performs weighted summation to extract the features of this region. In this way, the convolutional layer can capture the local features in the data, such as the local change rules or trends of track wear.

[0085] Multiple convolutional kernels: Using convolutional kernels of different sizes and shapes can extract local features of different scales and capture different levels of information in the data.

[0086] The main task of the pooling layer in this step is to perform a dimensionality reduction operation on the local features extracted by the convolutional layer, and concentrate the information through max-pooling, thereby reducing the computational amount and preventing overfitting. The steps of the pooling layer are as follows:

[0087] Max-pooling: The pooling layer usually processes the convolutional feature map in a fixed window (such as 2x2 or 3x3). In each pooling window, the maximum value is selected as the representative feature of this region to obtain more concise feature information.

[0088] Dimensionality Reduction and Retaining Important Features: Max pooling can effectively retain the most significant features while eliminating noise and redundant information, enabling the model to focus more on key features, thereby enhancing the robustness and accuracy of predictions.

[0089] In this step, after extracting and reducing the dimensionality of the local features of the input data in the convolutional layer and the pooling layer, the fully connected layer is responsible for further processing and integrating these features to finally obtain the optimized prediction results. The specific process is as follows:

[0090] Feature Extraction: In the fully connected layer, each local feature after convolutional and pooling processing is transformed into a one-dimensional vector, and then these features are integrated through fully connected operations. These features can represent the global information of rail wear.

[0091] Output of the Prediction Result: The fully connected layer usually consists of multiple neurons, and each neuron is connected to all the outputs of the previous layer. After being processed by the ReLU activation function, the finally generated output is the prediction result of rail wear. For a regression task, this is usually a continuous value representing the amount of rail wear.

[0092] Step S6: Conduct reliability analysis on the optimized prediction results, and use the prediction results with reliability greater than the preset threshold as the evaluation results of the rail wear amount.

[0093] It can be understood that through reliability analysis, prediction results with high credibility can be effectively screened out, and those unreliable or with large errors can be excluded. The technical effect of this step is to improve the accuracy and practicality of the evaluation results of the rail wear amount, ensure that the final evaluation results can be used as a decision-making basis, and provide accurate data support for rail maintenance work. At the same time, reliability analysis can help identify and handle potential problems in the model, improving the overall stability and reliability of the system. In this step, Step S6 includes Step S61, Step S62, and Step S63.

[0094] Step S61: Calculate the maximum information coefficient between the optimized prediction results and the preset influencing factors, where the preset influencing factors include train speed, train load, and train temperature;

[0095] It can be understood that the maximum information coefficient is a statistical method used to measure the degree of non-linear association between two variables. Different from the traditional Pearson correlation coefficient, the maximum information coefficient can capture the complex relationships between variables, including linear and non-linear relationships. Therefore, the maximum information coefficient can perform well under various data types and relationship forms, especially suitable for data with complex or non-linear relationships.

[0096] The optimized prediction results and the preset influencing factors (train speed, load, and temperature) are respectively formed into two vectors. The elements of each vector represent the prediction results at different times or conditions and the corresponding values of the influencing factors. The maximum information coefficient algorithm is used to calculate the correlation between these pairs of variables. For example, calculate the maximum information coefficient value between the train speed and the prediction result, the maximum information coefficient value between the train load and the prediction result, and the maximum information coefficient value between the train temperature and the prediction result.

[0097] Among them, the calculation formula for the maximum information coefficient is as follows:

[0098] ;

[0099] Among them, represents the maximum information coefficient of two variables, represents the information content of two variables, represents the joint entropy of two variables.

[0100] Step S62: Construct a Bayesian network model based on the optimized prediction results and the preset influencing factors. Among them, each node in the Bayesian network model represents a random variable, and the edge represents a conditional dependence relationship;

[0101] It can be understood that the application of the Bayesian network can effectively integrate the historical data of train operation and the current train operation status (such as speed, load, temperature, etc.), and infer the future wear condition of the track based on the historical data. By constructing a Bayesian network model with clear conditional dependence relationships, the comprehensive influence of different factors (train speed, load, etc.) on track wear can be captured, providing more accurate and reliable prediction results for track maintenance.

[0102] In the Bayesian network, the edge represents a conditional dependence relationship, and the track wear prediction result (optimized prediction result) is a node. The train speed, train load, train temperature, etc. are used as influencing factors and are other nodes. That is, the state of a node depends on the state of its parent node. By modeling the conditional dependence relationships between variables, the Bayesian network can effectively capture the interactions between influencing factors. For example, the train load may directly affect track wear, while the train speed may have an indirect impact on both the train load and track wear. In this way, the Bayesian network can establish the dependence structure between these variables.

[0103] Step S63: Calculate the reliability of the prediction results based on the Bayesian network model, and use the prediction results corresponding to the reliability greater than the preset threshold as the evaluation results of the railway track wear amount.

[0104] It is understandable that the conditional probability table provided by the Bayesian network is the basis for calculating the reliability. Through the inference process, the posterior probability of a certain prediction result can be obtained, and this probability reflects the reliability of the result. For example, when information such as the speed, load, and temperature of the train is known, the Bayesian network can calculate the probability distribution of track wear based on this information. The maximum probability value of this distribution corresponds to the reliability of the prediction result.

[0105] Embodiment 2:

[0106] As Figure 2 shown, this embodiment provides a railway track wear amount evaluation system. Refer to Figure 2 The system includes an acquisition unit 701, a first processing unit 702, a second processing unit 703, a third processing unit 704, a fourth processing unit 705, and a fifth processing unit 706.

[0107] The acquisition unit 701 is used to acquire first data, and the first data includes train operation data and track historical wear data;

[0108] The first processing unit 702 is used to send the first data to a preprocessing model for preprocessing, where the first data is sequentially sent to a variational autoencoder and a Gaussian process regression model based on a preset for cleaning and completion to obtain preprocessed first data;

[0109] The second processing unit 703 is used to perform wavelet transform and locally weighted regression processing on the preprocessed first data to obtain the local features of all the first data and their local linear relationships, and use them as a feature data set;

[0110] The third processing unit 704 is used to send the feature data set to a preset dimensionality reduction model for processing, where the feature data set is sequentially processed by a feature relationship processing of an extreme learning machine and a principal component analysis for dimensionality reduction, and processed by a preset Gaussian mixture model to obtain at least one category feature set;

[0111] The fourth processing unit 705 is used to send the feature sets of all categories to a preset prediction model for processing, and send the prediction result to a convolutional neural network model for optimization to obtain an optimized prediction result;

[0112] The fifth processing unit 706 is used to perform reliability analysis on the optimized prediction result, and use the prediction result with a reliability greater than a preset threshold as the evaluation result of the track wear amount.

[0113] It should be noted that regarding the system in the embodiment, the specific manners in which each module performs operations have been described in detail in the embodiment related to the method, and will not be elaborated here in detail.

[0114] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

[0115] The above is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art can easily think of changes or replacements within the technical scope disclosed by the present invention, and all of them should be covered by the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims

1. A method for evaluating railway track wear, characterized in that: include: Acquiring first data, wherein the first data includes train operation data and track historical wear data; Sending the first data to a preprocessing model for preprocessing, wherein the first data is sequentially sent to a preset variational autoencoder and a Gaussian process regression model for cleaning and completion to obtain the preprocessed first data; Performing wavelet transform and local weighted regression processing on the preprocessed first data to obtain local features and local linear relationships of all the first data, and using them as feature data sets; Sending the feature data set to a preset dimensionality reduction model for processing, wherein the feature data set is sequentially subjected to feature relationship processing of an extreme learning machine and principal component analysis for dimensionality reduction, and is processed by a preset Gaussian mixture model to obtain at least one category feature set; Send the feature sets of all categories to the preset prediction model for processing, and send the prediction results to the convolutional neural network model for optimization to obtain the optimized prediction results; The optimized prediction results are subjected to reliability analysis, and the prediction results with a reliability greater than a preset threshold are used as the evaluation results of the track wear amount.

2. The railway track wear assessment method according to claim 1, characterized in that: The first data is sent to a preprocessing model for preprocessing, wherein the first data is sequentially sent to a preset variational autoencoder and a Gaussian process regression model for cleaning and completion, including: Inputting the first data into a variational autoencoder for denoising, wherein the variational autoencoder maps the input data into a latent space, the variational autoencoder learns the intrinsic distribution of the data, distinguishes noise from key features in the data, and maps the data points in the latent space back to the original space to obtain denoised data; Extracting latent variables of the first data based on the variational autoencoder, and mapping the latent variables to a low-dimensional space for representation, to obtain latent variable data in the low-dimensional space, wherein the latent variables are data of track wear and train load, running speed, and track type; Learning the covariance structure between the latent variable data in the low-dimensional space based on a preset Gaussian process regression model, and predicting missing values ​​based on the covariance structure; The first data is cleaned and completed based on the denoised data and missing values ​​to obtain preprocessed first data.

3. The railway track wear assessment method according to claim 1, characterized in that: The pre-processed first data is subjected to wavelet transformation and local weighted regression processing to obtain local features and local linear relationships of all the first data, and the local features and local linear relationships thereof are used as feature data sets, including: Selecting a preset wavelet transform function, and presetting the number of decomposition layers to three layers, and decomposing the preprocessed first data on at least two frequency bands based on the preset wavelet transform function to obtain signal information at at least two scales; Inputting signal information at all scales into a preset local weighted regression model to calculate weights of adjacent data points, and using the weighted data points to train the local weighted regression model to obtain a trained local weighted regression model; The signal information at the at least two scales is sent to the trained local weighted regression model for prediction to obtain a prediction result, wherein the prediction result represents the local features and nonlinear relationship of the target point corresponding to the preprocessed first data.

4. The railway track wear assessment method according to claim 1, characterized in that: The feature data set is sent to a preset dimensionality reduction model for processing, wherein the feature data set is sequentially subjected to feature relationship processing of an extreme learning machine and principal component analysis for dimensionality reduction, and is processed by a preset Gaussian mixture model, including: The feature data set is sent to a preset extreme learning machine model for processing, wherein the extreme learning machine model maps the feature data set through the ReLU function of the hidden layer neurons, and captures the nonlinear relationship of the feature data set in the high-dimensional space, and continuously performs back propagation and gradient descent until the nonlinear relationship of all data in the feature data set in the high-dimensional space is captured, thereby obtaining a feature data set containing the nonlinear relationship of all data; Calculate a covariance matrix of a feature data set containing nonlinear relationships of all data, wherein each covariance in the covariance matrix represents a nonlinear relationship between each feature; Perform eigenvalue decomposition on the covariance matrix to obtain eigenvectors and eigenvalues. The eigenvalue represents the weight of each eigenvector in the data, and the eigenvector represents the direction of each eigenvector. Select the eigenvectors corresponding to the preset number of largest eigenvalues ​​from large to small, and map the feature data set containing the nonlinear relationship of all data to the eigenvectors corresponding to the preset number of largest eigenvalues ​​to obtain the feature data set after dimensionality reduction; The feature data set after dimensionality reduction is input into a preset Gaussian mixture model for clustering processing, wherein it is assumed that each data point in the data set after dimensionality reduction is randomly extracted from a Gaussian distribution, and the probability that each data point belongs to the Gaussian distribution is estimated based on the maximum expected value algorithm, and each data in the data set after dimensionality reduction is classified based on the probability to obtain at least one category feature set.

5. The railway track wear assessment method according to claim 1, characterized in that: The feature sets of all categories are sent to the preset prediction model for processing, and the prediction results are sent to the convolutional neural network model for optimization to obtain the optimized prediction results, including: Based on the random forest regression model, the track historical wear data and feature data in all categories of feature sets are trained, wherein the mean square error is iteratively calculated multiple times until the error between the prediction result and the track historical wear data is less than a preset threshold, thereby obtaining a trained random forest regression model; The feature sets of all categories are sent to the trained random forest regression model for prediction to obtain the first wear result of the track; The first wear result of the track is input into a convolutional neural network for optimization, wherein the convolution layer of the convolutional neural network performs sliding window processing on the first wear result through a set of convolution kernels, performs weighted summation on local areas at different positions in the first wear result, and performs maximum pooling operation on local features of the first wear result obtained by the weighted sum through a pooling layer. Finally, each local feature is extracted through a fully connected layer to obtain the final prediction result.

6. A railway track wear assessment system, characterized in that: include: An acquisition unit, configured to acquire first data, wherein the first data includes train operation data and track historical wear data; A first processing unit, configured to send the first data to a preprocessing model for preprocessing, wherein the first data is sequentially sent to a preset variational autoencoder and a Gaussian process regression model for cleaning and completion to obtain the preprocessed first data; A second processing unit is used to perform wavelet transform and local weighted regression processing on the preprocessed first data to obtain local features and local linear relationships of all the first data, and use them as a feature data set; A third processing unit is used to send the feature data set to a preset dimensionality reduction model for processing, wherein the feature data set is sequentially subjected to feature relationship processing of an extreme learning machine and dimensionality reduction by principal component analysis, and is processed by a preset Gaussian mixture model to obtain at least one category feature set; A fourth processing unit is used to send feature sets of all categories to a preset prediction model for processing, and send the prediction results to a convolutional neural network model for optimization to obtain optimized prediction results; The fifth processing unit is used to perform reliability analysis on the optimized prediction results, and use the prediction results with reliability greater than a preset threshold as the evaluation results of the track wear amount.

7. The railway track wear assessment system according to claim 6, characterized in that: The first processing unit comprises: A first processing subunit is used to input the first data into a variational autoencoder for noise reduction processing, wherein the variational autoencoder maps the input data into a latent space, the variational autoencoder learns the intrinsic distribution of the data, distinguishes noise from key features in the data, and maps the data points in the latent space back to the original space to obtain denoised data; A second processing subunit is used to extract latent variables of the first data based on the variational autoencoder, and map the latent variables to a low-dimensional space for representation to obtain latent variable data in the low-dimensional space, wherein the latent variables are data of track wear and train load, running speed, and track type; A first prediction subunit, configured to learn a covariance structure between latent variable data in the low-dimensional space based on a preset Gaussian process regression model, and predict missing values ​​based on the covariance structure; The third processing subunit is used to clean and complete the first data based on the denoised data and missing values ​​to obtain preprocessed first data.

8. The railway track wear assessment system according to claim 6, characterized in that: The second processing unit comprises: A first decomposition subunit is used to select a preset wavelet transform function, preset the number of decomposition layers to three layers, and decompose the preprocessed first data on at least two frequency bands based on the preset wavelet transform function to obtain signal information at at least two scales; A first calculation subunit is used to input the signal information at all scales into a preset local weighted regression model to perform weight calculation of adjacent data points, and use the weighted data points to train the local weighted regression model to obtain a trained local weighted regression model; The second prediction subunit is used to send the signal information at the at least two scales to the trained local weighted regression model for prediction to obtain a prediction result, wherein the prediction result represents the local features and nonlinear relationship of the target point corresponding to the preprocessed first data.

9. The railway track wear assessment system according to claim 6, characterized in that: The third processing unit comprises: A fourth processing subunit is used to send the feature data set to a preset extreme learning machine model for processing, wherein the extreme learning machine model maps the feature data set through the ReLU function of the hidden layer neurons, and captures the nonlinear relationship of the feature data set in the high-dimensional space, and continuously performs back propagation and gradient descent until the nonlinear relationship of all data in the feature data set in the high-dimensional space is captured, thereby obtaining a feature data set containing the nonlinear relationship of all data; A second calculation subunit is used to calculate a covariance matrix of a feature data set containing nonlinear relationships of all data, wherein each covariance in the covariance matrix represents a nonlinear relationship between each feature; The second decomposition subunit is used to perform eigenvalue decomposition on the covariance matrix to obtain eigenvectors and eigenvalues, where the eigenvalue represents the weight of each eigenvector in the data, and the eigenvector represents the direction of each eigenvector; A fifth processing subunit is used to select eigenvectors corresponding to a preset number of maximum eigenvalues ​​from large to small, and map a feature data set containing nonlinear relationships of all data to eigenvectors corresponding to a preset number of maximum eigenvalues ​​to obtain a feature data set after dimensionality reduction; The sixth processing subunit is used to input the feature data set after dimensionality reduction into a preset Gaussian mixture model for clustering processing, wherein it is assumed that each data point in the data set after dimensionality reduction is randomly selected from a Gaussian distribution, and the probability that each data point belongs to the Gaussian distribution is estimated based on the maximum expectation value algorithm, and each data in the data set after dimensionality reduction is classified based on the probability to obtain at least one category feature set.

10. The railway track wear assessment system according to claim 6, characterized in that: The fourth processing unit comprises: a seventh processing subunit, configured to train the track historical wear data and feature data in all categories of feature sets based on a random forest regression model, wherein the mean square error is iterated multiple times until the error between the prediction result and the track historical wear data is less than a preset threshold, thereby obtaining a trained random forest regression model; The third prediction subunit is used to send the feature sets of all categories to the trained random forest regression model for prediction, so as to obtain the first wear result of the track; The fourth prediction subunit is used to input the first wear result of the track into a convolutional neural network for optimization, wherein the convolution layer of the convolutional neural network performs sliding window processing on the first wear result through a set of convolution kernels, performs weighted summation on local areas at different positions in the first wear result, and performs maximum pooling operation on local features of the first wear result obtained by the weighted sum through a pooling layer, and finally, extracts each local feature through a fully connected layer to obtain a final prediction result.

Images