A method, system, equipment and medium for predicting track irregularities (TQI) on high-speed railways.

Abstract

This invention discloses a method, system, equipment, and medium for predicting track irregularities (TQI) on high-speed railways, relating to the field of high-speed railway track prediction technology. The steps include: extracting multiple track irregularity feature variables from track inspection measurement data; adaptively optimizing the hyperparameters of the three basic models (SVM, ELM, and BP) using the GWO algorithm to generate an improved Stacking model; processing the track irregularity feature variables using the improved Stacking model to output a preliminary predicted TQI value; and performing multiple linear regression to obtain the final TQI prediction result for the target high-speed railway track section. This invention uses TQI as a track smoothness evaluation index, employs machine learning and ensemble thinking, integrates multiple sub-models, uses a multiple linear regression algorithm to extract the optimal fusion weights, and establishes an improved Stacking model, significantly improving the accuracy of TQI prediction and providing technical support for precise track maintenance.

Description

Technical Field

[0001] This invention relates to the field of high-speed rail track prediction technology, and in particular to a method, system, equipment and medium for predicting track irregularities (TQI) on high-speed rail tracks. Background Technology

[0002] As a crucial infrastructure, high-speed railways rely heavily on the Track Quality Index (TQI), a key indicator for measuring the overall performance and load-bearing capacity of railway structures. This directly impacts the operational safety of high-speed trains and passenger comfort. Accurate diagnosis, prediction, and control of track irregularities are indispensable components of intelligent railway operation and maintenance, playing a vital role in guiding track maintenance work, preventing accidents, and ensuring safe train operation.

[0003] In the traditional research field of track irregularity prediction technology, numerous scholars have proposed various methods. He Yongchun et al. focused on the prediction and application of track elevation irregularities. Xu Yude et al. proposed a method in 2003 using a feature matrix to describe the track state and predict track elevation irregularities. In 2006, Chen Xianmai et al., considering the influence of multiple factors, proposed a composite factor method integrating digital statistics, signal processing, and detection data. With the rapid development of neural network technology, deep learning algorithms have been gradually applied to track irregularity state prediction, using artificial intelligence technology to solve engineering problems. Peng Liyu et al. conducted research on track irregularity prediction based on backpropagation neural networks (BP). Subsequently, Yu Yao et al. used support vector machines (SVM) models to find the optimal decision boundary to improve prediction accuracy. Jia Chaolong et al. conducted a systematic study and prediction of track irregularities based on grey models and autoregressive (AR) models. Dewey et al. used an improved particle swarm optimization algorithm (IPSO) and a long short-term memory network (LSTM) to build a track quality index prediction model to predict the development trend of track irregularities in high-speed railway ballastless tracks. Wang Yingjie et al. used a joint GM(1,1) grey model and a correlation vector machine (RVM) algorithm to propose an interval track irregularity evolution model.

[0004] In summary, most existing technologies rely on single models constructed using the aforementioned algorithms for predicting track irregularities. These technologies have long computation and training times, or the model hyperparameters depend on manual tuning. The training process is prone to getting stuck in local optima, especially when the network structure is complex and there are many parameters. This results in the inability to obtain the best prediction results and makes it difficult to meet the needs of track condition prediction and precise maintenance. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of the prior art by providing a method, system, device, and medium for predicting track irregularities (TQI) on high-speed railway tracks, thereby solving the problems in the prior art.

[0006] This invention specifically provides the following technical solution: a method for predicting track irregularities (TQI) on high-speed railways, comprising:

[0007] Collect actual track inspection data for high-speed railways;

[0008] Collect track inspection data for the target track section of the high-speed railway;

[0009] Extract multiple track irregularity feature variables from the track inspection measured data;

[0010] The different track irregularity feature variables are input into the improved Stacking model, wherein each sub-model of the improved Stacking model uses the optimal model parameters determined by the optimization algorithm and outputs a preliminary track quality index (TQI) prediction value respectively.

[0011] Multiple preliminary track quality index (TQI) predictions were processed by multiple linear regression to obtain multiple regression coefficients. Based on the integration concept, the predicted preliminary TQI values ​​and their corresponding regression coefficients were weighted and summed to obtain the final TQI prediction result for the high-speed rail target track section.

[0012] Preferably, the improved Stacking model is as follows:

[0013] An improved Stacking model is constructed by integrating GWO-SVM, GWO-ELM, and GWO-BP sub-models based on the Stacking model.

[0014] Preferably, each sub-model uses the optimal model parameters determined by the optimization algorithm and outputs a preliminary predicted orbital quality index (TQI) value, specifically:

[0015] The optimal model parameter values ​​of SVM, ELM and BP models are obtained using the Grey Wolf Optimization Algorithm (GWO), and the wolf population size and number of evolutionary iterations are selected to construct GWO-SVM, GWO-ELM and GWO-BP models.

[0016] Among them, the GWO-SVM, GWO-ELM and GWO-BP models use key feature variables of track irregularity as input variables and the corresponding measured TQI values ​​as output variables. The relationship between the input variables and the output variables is obtained through a linear regression function.

[0017] Multiple track irregularity characteristic variables from track inspection data are input into GWO-SVM, GWO-ELM, and GWO-BP models, and preliminary track quality index (TQI) prediction values ​​are output through linear regression functions.

[0018] Preferably, the step of using the Grey Wolf Optimization Algorithm (GWO) to obtain the optimal model parameter values ​​for SVM, ELM, and BP models specifically involves:

[0019] The gray wolf optimization algorithm is used to calculate the fitness of each gray wolf and update the position of the gray wolf group by simulating the three stages of gray wolf hunting behavior: search, encirclement and attack. When a set threshold is met, the optimal parameters in the corresponding processing algorithm are output. When the gray wolf optimization algorithm stops iterating, the gray wolf with the best fitness is selected and the corresponding position vector is the final output optimal model parameter value.

[0020] Preferably, multiple track irregularity feature variables are extracted from the track inspection measured data, specifically as follows:

[0021] Multiple track irregularity indices are calculated from the track inspection measured data to generate multiple basic features with different numbers. Then, the random forest method is used to select multiple basic features as track irregularity feature variables according to the importance of feature information. The track irregularity indices include track gauge, level, triangular crater, left elevation difference, right elevation difference, left track orientation, and right track orientation. The multiple basic features include the maximum value, minimum value, average value, standard deviation, skewness, and kurtosis of the track irregularity indices.

[0022] Preferably, after obtaining the final TQI prediction result for the high-speed rail target track segment, the method further includes:

[0023] The final TQI prediction results are compared with the original data to obtain the residual map within the prediction unit;

[0024] The sample dataset is predicted using a residual plot to obtain a comparison chart of predicted TQI and measured values. The horizontal axis of the comparison chart represents the model's fitted value, which is the TQI value predicted by the model. The vertical axis represents the residual, which is the difference between the actual observed TQI value and the fitted value of the predicted model.

[0025] Preferably, after collecting the track inspection measurement data of the target track section of the high-speed railway, the method further includes:

[0026] The track inspection measurement data were denoised using complete empirical mode decomposition based on adaptive noise.

[0027] Outliers in the track inspection data were checked and removed. The blank data after removing outliers were then processed using a linear interpolation method to obtain the processed track inspection data.

[0028] This invention provides a high-speed rail track irregularity (TQI) prediction system, comprising:

[0029] The data acquisition module is used to collect track inspection measurement data of the target track section of the high-speed railway;

[0030] The feature extraction module is used to extract multiple track irregularity feature variables from the track inspection measured data;

[0031] The initial prediction module is used to input the different orbital irregularity feature variables into the improved Stacking model, wherein each sub-model in the improved Stacking model uses the optimal model parameters determined by the optimization algorithm and outputs the preliminary orbital quality index (TQI) prediction value respectively.

[0032] The final prediction module is used to perform multiple linear regression processing on multiple preliminary track quality index (TQI) prediction values ​​to obtain multiple regression coefficients. Based on the integration concept, the module performs weighted summation on each of the preliminary TQI prediction values ​​and their corresponding regression coefficients to obtain the final TQI prediction result for the high-speed rail target track section.

[0033] The present invention provides a computer device, including a memory and a processor. The memory stores a program, and when the program is executed by the processor, the processor performs the steps of the above-mentioned method for predicting track irregularities (TQI) of high-speed railways.

[0034] The present invention provides a storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the above-mentioned method for predicting track irregularities (TQI) of high-speed railways.

[0035] Compared with the prior art, the present invention has the following significant advantages:

[0036] This invention extracts track irregularity features from collected high-speed railway track inspection data, utilizes machine learning and ensemble thinking, selects Track Quality Index (TQI) as the track smoothness evaluation index, and employs ensemble thinking to incorporate multiple sub-models. The optimal model parameters are used to obtain preliminary TQI prediction values. A multiple linear regression algorithm is then used to obtain multiple regression coefficients, thereby obtaining the final TQI prediction results for the target high-speed railway section. By first optimizing individual models and then using an ensemble strategy to fuse the results, the influence of numerous network structures and parameters is reduced, further improving the accuracy of model predictions. Attached Figure Description

[0037] Figure 1 The GWO-SVM flowchart provided by this invention;

[0038] Figure 2 The GWO fitness image provided by this invention;

[0039] Figure 3 This is a comparison chart of GWO-SVM data provided by the present invention; wherein Figure 3 (a) shows the comparison between the validation set and the GWO-SVM prediction data. Figure 3(b) shows a comparison between the test set and the GWO-SVM prediction data;

[0040] Figure 4 This is a comparison chart of GWO-ELM data provided by the present invention; wherein Figure 4 (a) is a comparison chart of the validation set and the GWO-ELM prediction data. Figure 4 (b) is a comparison chart of the test set and the GWO-ELM prediction data;

[0041] Figure 5 This is a comparison chart of GWO-BP model data provided by the present invention; wherein Figure 5 (a) is a comparison chart of the validation set and the GWO-BP prediction data. Figure 5 (b) is a comparison chart of the test set and the GWO-BP prediction data;

[0042] Figure 6 This is a TQI prediction diagram for high-speed railway track irregularities based on an improved Stacking model, as presented in this invention.

[0043] Figure 7 This is a TQI residual-fit value diagram provided by the present invention;

[0044] Figure 8 This is a comparison chart of TQI predicted values ​​and measured values ​​for different models provided by this invention; wherein... Figure 8 (a) is a comparison of SVM. Figure 8 (b) is a comparison of GWO-SVM. Figure 8 (c) is a comparison of Stacking integration;

[0045] Figure 9 The flowchart for predicting TQI using the machine learning stacking model provided by this invention;

[0046] Figure 10 The horizontal irregularity raw signal and preprocessing provided for this invention Figure 1 ;in Figure 10 (a) is the original signal diagram with horizontal irregularities. Figure 10 (b) is a spectrum diagram;

[0047] Figure 11 The horizontal irregularity raw signal and preprocessing provided for this invention Figure 2 ;in Figure 11 (a) is the CEEMDAN decomposition plot of the non-rough data. Figure 11 (b) is the corresponding spectrum diagram of the decomposition;

[0048] Figure 12 This invention provides a comparison diagram of track irregularities before and after processing. Figure 12 (a) represents the original signal with horizontal irregularities. Figure 12 (b) is a diagram of horizontal irregularities after processing based on CEEMDAN.

[0049] Figure 13 This is a comparison chart of the TQI prediction accuracy of the training set of the sub-models (SVM / GWO-SVM / STACKING) provided in this invention; wherein... Figure 13 (a) is a comparison chart of MAE. Figure 13 (b) is a comparison chart of RMSE;

[0050] Figure 14 This is a comparison chart of the TQI predicted by the Stacking / SVM model provided by this invention and the measured value at 20km; wherein, Figure 14 (a) is a comparison chart of the predicted TQI and the measured values ​​of the two models (SVM and GWO-SVM). Figure 14 (b) is a comparison chart of the Stacking model predictions and the measured values;

[0051] Figure 15 The flowchart of the Stacking model based on multiple linear regression provided by this invention. Detailed Implementation

[0052] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0053] This invention is based on the concept of machine deep learning, uses MATLAB programming, and takes into account the advantages of integration. It uses the Grey Wolf Algorithm (GWO) to optimize the model parameters of three algorithms: Support Vector Machine (SVM), Extreme Learning Machine (ELM), and Backpropagation Neural Network (BP). It uses the multiple linear regression method to construct a Stacking model and conducts in-depth research on TQI prediction of high-speed railway track irregularities using big data.

[0054] The advantages and disadvantages of existing single-model SVM, ELM, and BP are as follows:

[0055] SVM (Support Vector Machine): Advantages: Performs well with small sample sizes and high-dimensional data; effectively handles linear and nonlinear problems; maps data to a high-dimensional space using kernel functions to find the optimal classification hyperplane or regression function. For classification problems, it can find the global optimum; for regression problems, the accuracy of regression can be controlled by adjusting parameters. Disadvantages: High computational complexity and long training time, especially noticeable on large-scale datasets. Sensitive to parameter and kernel function selection; different parameters and kernel functions can lead to significant differences in model performance, and there is a lack of universal parameter selection methods, requiring extensive experimental parameter tuning.

[0056] ELM (Extreme Learning Machine): Advantages: Fast training speed; its hidden layer weights can be randomly generated without iterative adjustment, greatly reducing training time and demonstrating significant efficiency advantages when handling large-scale data. Good generalization ability; it has a certain adaptability to different types of data and can effectively avoid overfitting. Disadvantages: The selection of the number of hidden layer nodes lacks theoretical guidance and needs to be determined based on experience or experiments. High requirements for data preprocessing; different preprocessing methods may have a significant impact on model performance.

[0057] Backpropagation (BP) Neural Network: Advantages: It has a strong ability to map nonlinear functions, approximating arbitrarily complex nonlinear functions, and is widely used in pattern recognition, function approximation, data classification, and other fields. It can flexibly adapt to different task requirements by adjusting the network structure and training parameters. Disadvantages: The training process is prone to getting stuck in local optima, especially when the network structure is complex and has many parameters. Training time is long, requiring multiple iterations to update weights, resulting in high computational cost. It is sensitive to initial weights and learning rate; different settings can lead to significant differences in training results.

[0058] The advantages of stacking models include: Improved prediction performance: By combining multiple different single models (such as SVM, ELM, and BP), stacking models fully utilize the advantages of each model in different aspects and compensate for the shortcomings of a single model. For example, SVM performs well when handling small-sample, high-dimensional data, ELM has a fast training speed, and BP has strong non-linear fitting ability. After ensemble integration, more accurate prediction results can be obtained, reducing prediction errors. Enhanced generalization ability: Different single models may perform differently on different datasets or data distributions. Stacking models integrate the prediction results of multiple models, making the model more robust when facing new data, reducing overfitting or underfitting problems caused by the limitations of the models themselves, and improving the model's generalization ability. High flexibility: Different single models can be freely selected as base models according to specific task requirements and data characteristics, and then fused through appropriate meta-models. Furthermore, the performance of the ensemble model can be optimized by adjusting the type and number of base models and the selection of meta-models to adapt to different application scenarios. Strong robustness: When there is noise or outliers in the data, a single model may be greatly affected, leading to large deviations in the prediction results. The Stacking model combines the results of multiple models, which weakens the impact of noise or outliers on individual models, thus making the overall model more robust.

[0059] Figure 6 The key to the flowchart design is: (1) Improved Stacking concept: Traditional Stacking uses ordinary base models, but here we use GWO (Grey Wolf Optimization) to enhance the base models (SVM / ELM / BP), and then use multiple linear regression to perform "meta-model" fusion, which is an advanced solution of "optimizing base models + ensemble learning". (2) Layered division of labor: The base model is responsible for "local accurate prediction", and the meta-model is responsible for "global fusion decision". The prediction capability is improved through multi-layer collaboration, which is often used for regression prediction tasks (such as fault diagnosis). In summary, this is a prediction system that "optimizes a single model first and then uses an ensemble strategy to fuse the results". The core is to use the Grey Wolf algorithm to enhance the base model and then integrate it through multiple linear regression to finally output more reliable prediction results.

[0060] like Figure 6To address the issues of long cycles and high costs associated with existing TQI detection methods that rely on track inspection vehicles, and the difficulty in optimizing hyperparameters and insufficient accuracy of traditional single prediction models, this paper proposes a technical solution based on Grey Wolf Optimization (GWO) and improved Stacking integration to achieve efficient and accurate TQI prediction. The steps include: ① track inspection data preprocessing and feature engineering construction; ② adaptive optimization of hyperparameters of the three base models (SVM, ELM, and BP) using the GWO algorithm to generate improved base models and corresponding meta-features; ③ optimization of meta-features through residual correction technology, combined with a multiple linear regression algorithm to construct a Stacking meta-model, ultimately completing TQI reconstruction and prediction. This achieves efficient TQI prediction through a path of "adaptive parameter optimization - automatic meta-feature correction - hierarchical model integration." This embodiment also includes a method for predicting TQI for high-speed rail track irregularities, comprising:

[0061] Step S1: Collect track inspection data of the target track section of the high-speed railway.

[0062] The track inspection measured data were denoised using complete empirical mode decomposition based on adaptive noise. Outliers in the track inspection measured data were checked and removed. The blank data after removing outliers were processed using a linear interpolation method to obtain the processed track inspection measured data.

[0063] Step S2: Extract multiple track irregularity feature variables from the track inspection measured data.

[0064] The acquisition of feature variables is specifically as follows: calculate multiple indicators of the track inspection measured data TQI, generate multiple basic features, and use the random forest method to select the basic features in the track inspection measured data; the indicators include track gauge, level, triangular pit, left elevation, right elevation, left track orientation and right track orientation, and the multiple basic features include the maximum value, minimum value, average value, standard deviation, skewness and kurtosis of the indicators.

[0065] Step S3: Input different track irregularity feature variables into the improved Stacking model, wherein each sub-model in the improved Stacking model uses the optimal model parameters determined by the optimization algorithm and outputs the preliminary track quality index (TQI) prediction value respectively.

[0066] The track inspection measured data is input into the improved Stacking model to obtain the prediction results of multiple sub-models in the improved Stacking model. The prediction results of multiple sub-models are applied to the prediction results of multiple sub-models to obtain multiple regression coefficients. The final TQI prediction result is output as the sum of the products of multiple regression coefficients and the corresponding prediction results. The improved Stacking model is constructed by integrating GWO-SVM, GWO-ELM and GWO-BP sub-models based on the Stacking model.

[0067] Model accuracy evaluation metrics: The mean absolute error (MAE) and root mean square error (RMSE) are used to measure the accuracy and precision of the model training regression.

[0068] Among them, the predicted value y ∗ ={ , , , ... The actual value is y = {y1, y2, y3, y4, ..., y}. n}, and obtain the sample MAE and RMSE, see formula (1) ~ formula (2).

[0069] ;

[0070] Formula (1) The smaller the MAE value, the higher the accuracy of the model. When MAE is 0, the prediction accuracy is 100%. The larger the error, the larger the RMSE value. Conversely, the smaller the RMSE value, the higher the prediction accuracy and the more perfect the model. When RMSE is 0, the prediction accuracy is 100%.

[0071] The gray wolf optimization algorithm is adopted, and the fitness of each gray wolf is calculated using gray wolves as model parameters for each processing algorithm. The position of the gray wolf group is updated by simulating the three stages of gray wolf hunting behavior: search, encirclement and attack. When a set threshold is met, the optimal parameters in the corresponding processing algorithm are output. When the gray wolf optimization algorithm stops iterating, the gray wolf with the best fitness is selected, and the corresponding position vector is the final output optimal model parameter value.

[0072] The Gray Wolf Optimization Algorithm (GWO) is an intelligent optimization algorithm that simulates the leadership and hunting hierarchy of wolf packs in nature. Wolf packs have a strict hierarchy, using a pyramid to represent the relationships between wolves. From top to bottom, the pyramid consists of α, β, δ, and ω. α is at the top of the pyramid, playing a leadership role in the pack, while β assists α in decision-making and problem-solving. α leads the pack to complete the following tasks:

[0073] Encircling prey: During the hunt, gray wolves first search for their prey, then surround it. This behavior of wolf packs encircling prey is described as follows:

[0074] ;

[0075] Among them, X M ( t () represents the location of the prey at this time. t The number of iterations is represented by A and B, which are coefficients, as shown in formulas (2) and (3). r1 and r2 are random simulated values ​​in the interval [0,1].

[0076] ;

[0077] Attacking prey: After the wolf pack spots its prey, it... α Leading the wolf pack in an attack, specifically:

[0078] ;

[0079] DIS α X represents the position of alpha wolf relative to the other wolves. α This represents the current position of wolf α, and X represents the position of the individual gray wolf. Other wolf packs are influenced by the wolf packs at the top of the pyramid, resulting in the positions of individual wolves within the pack, as follows:

[0080] ;

[0081] Take the average value for each individual wolf in the pack.

[0082] ;

[0083] After surrounding and attacking the gray wolves and continuously updating their location information, record the current optimal fitness value among the gray wolves. The specific steps are as follows:

[0084] 1) Initialize the position and parameters of the gray wolf pack. 2) Calculate the fitness of each gray wolf. 3) Determine the three best gray wolves in the pack, representing the optimal solution α, the second-best solution β, and the next-nearest solution δ. 4) Update the position of the gray wolf pack; this is done by simulating the three stages of gray wolf hunting behavior, where tracking and approaching prey are simulated by simulating the interaction between the gray wolves and α, β, and δ, and attacking prey is simulated by simulating the gray wolves surrounding and attacking prey. 5) If the predetermined number of iterations is reached or other stopping criteria are met, stop the algorithm and output the optimal solution; otherwise, return to step 2. 6) Use the optimized parameters for prediction.

[0085] The Grey Wolf Optimization Algorithm (GWO) was used to obtain the optimal model parameter values ​​for SVM, ELM, and BP models, resulting in fitness images. The wolf population size and number of evolutionary iterations were selected to construct GWO-SVM, GWO-ELM, and GWO-BP models. All three models used the key feature variable of track irregularity as input variables and the corresponding measured TQI values ​​as output variables. A linear regression function was constructed to determine the relationship between the input and output variables. Comparison of the predicted TQI values ​​from the three sub-models with the measured values ​​is shown below. Figure 13 As shown, multiple track irregularity characteristic variables from the track inspection data are input into the GWO-SVM, GWO-ELM, and GWO-BP models, and preliminary track quality index (TQI) prediction values ​​are output using linear regression functions.

[0086] In the Grey Wolf optimization algorithm, fitness is an evaluation metric used to measure the quality of a "grey wolf position" (i.e., a set of model parameters). It is a numerical value (such as error value, accuracy value, etc.) and is only used as a selection criterion, not directly as the parameter result to be optimized. Position is the location vector of each grey wolf, directly corresponding to a set of "processing algorithm model parameters" (such as hyperparameters, coefficients, etc.) to be optimized. Each dimension of the position represents a specific parameter value. When the algorithm stops meeting a threshold (such as the number of iterations, fitness accuracy), the location vector corresponding to the grey wolf with the best fitness is the optimal model parameter required by the processing algorithm. When the algorithm stops iterating, it selects the grey wolf with the best fitness (such as the smallest error and highest accuracy), and its corresponding location vector is the final optimal parameter to be output.

[0087] Step S4: Perform multiple linear regression on the multiple preliminary track quality index (TQI) predictions to obtain multiple regression coefficients. Based on the integration concept, perform weighted summation on each preliminary TQI prediction and its corresponding regression coefficient to obtain the final TQI prediction result for the high-speed rail target track section.

[0088] GWO-SVM sub-model: The SVM model dataset is T={( x 1, y 1),( x 2, y 2),( x 3, y 3)......( x n , y n The feature variables are used as input variables for training. x n , y n This represents the corresponding measured TQI value. g( x Given a nonlinear mapping function, we seek the optimal... a and b Establishing a linear regression function in a high-dimensional feature space f ( x ),as follows:

[0089] ;

[0090] The large dataset used for machine learning training comes from track inspection data of a high-speed railway section from K720+000m to K1108+200m. Each sample unit is set to 200m, resulting in 1941 unit segments after preprocessing. The dataset is divided into a training set and a prediction set. The training set contains 1581 unit segments, evenly divided into three subsets T1, T2, and T3; the prediction set contains 180 unit segments. The GWO-SVM model uses the Grey Wolf Optimization Algorithm to optimize the parameter selection of the Support Vector Machine Regression (SVM), thereby improving the accuracy of regression predictions. For details, see [link to details]. Figure 1 During training, the T1 of the 527 units in the training subset was divided into a training set:validation set:test set ratio of 8:1:1. The Grey Wolf optimization algorithm was used to find the optimal model parameter values ​​for the SVM. Figure 2 Based on the fitness images, the wolf population size was ultimately selected as 10, the number of evolutionary iterations was 20, and the GWO-SVM algorithm model was constructed. The prediction set P1 of GWO-SVM was output and compared with the original data to verify the accuracy of the GWO-SVM model algorithm.

[0091] As shown in Table 1, comparing the MAE and RMSE metrics, GWO-SVM's errors are significantly lower than those of traditional SVM, with MAE and RMSE reduced by more than 50% across all datasets. This demonstrates that the GWO-SVM model optimized by the Grey Wolf algorithm completely outperforms traditional SVM in prediction accuracy. Regarding generalization stability: GWO-SVM exhibits smaller error fluctuations (e.g., MAE 3.524%~3.992%) across the training, validation, and test sets, while traditional SVM shows larger and generally higher error fluctuations. This indicates that GWO-SVM has superior generalization ability and is more suitable for real-world applications, making it a suitable choice as the first sub-model in the first layer of a stacking model.

[0092] Table 1 Comparison and Validation of GWO-SVM Model

[0093]

[0094] Figure 3 The graph shows a comparison between the TQI prediction data output by the GWO-SVM model and the original data. The TQI prediction set P1, consisting of approximately 53 units, matches the original validation set, indicating the feasibility of the GWO-SVM model.

[0095] GWO-ELM Sub-model: The GWO-ELM model uses the Grey Wolf Optimization algorithm to optimize the hidden layer network parameters of the Extreme Learning Machine (ELM), including the number of hidden layer neurons and activation functions. Its process also includes model training and dataset prediction. The training process is similar to the GWO-SVM model, selecting... Figure 2The wolf population was 10, with 20 generations. After optimization using the Gray Wolf Optimization Algorithm (GWO), the corresponding validation and test set data were obtained. Table 2 shows the MAE and RMSE data for the GWO-ELM model on the validation and test sets. The table uses MAE and RMSE as the core evaluation indicators to quantitatively compare the predictive performance of GWO-ELM with that of traditional ELM. The results show that: 1) Optimization effectiveness: In both the validation and test sets, the MAE and RMSE of GWO-ELM are lower than those of traditional ELM, with the highest error reduction reaching 52.19%, proving the effectiveness of the scheme of optimizing ELM parameters using the Gray Wolf Optimization Algorithm (GWO). 2) Scene adaptability: The error reduction in the test set (MAE 35.63%, RMSE 52.19%) is significantly higher than that in the validation set (MAE 14.52%, RMSE 39.96%), indicating that GWO-ELM has a more prominent advantage in real-world unknown data scenarios and stronger generalization ability. 3) Key Indicator Focus: The reduction in RMSE (39.96%-52.19%) was generally higher than that in MAE (14.52%-35.63%), indicating that GWO optimization had a more significant effect on "improving the model's resistance to extreme biases and enhancing stability" than "simply reducing the mean bias," demonstrating that the GWO-ELM model has higher prediction accuracy. The GWO-ELM model was selected as the second sub-model of the improved Stacking model, and multiple linear regression using the GWO-ELM algorithm was used for prediction, resulting in the prediction set P2.

[0096] Table 2 Comparison and Validation of GWO-ELM Model

[0097]

[0098] Depend on Figure 4 As can be seen, the TQI prediction results for the approximately 10km orbit are better than the original data, and the prediction results of the GWO-ELM model after optimization are good, meeting the accuracy and stability requirements.

[0099] GWO-BP Sub-model Establishment and Validation: GWO-BP combines the Grey Wolf Algorithm (GWO) with the Backpropagation Neural Network (BP) to seek the optimal solution for the weights and biases of the BP neural network, fully utilizing the global search capability of the Grey Wolf Algorithm and the nonlinear modeling capability of the BP neural network. For example... Figure 15As shown, during the training process, the sample dataset T3 was programmed using MATLAB, and the weights and biases of the BP neural network model were optimized using the Grey Wolf Optimization Algorithm (GWO) to improve the accuracy and stability of regression prediction. Similarly, a population size of 10 and a generation count of 20 were selected. After GWO optimization, the output results for the validation and test sets were obtained. Table 3 shows a comparison of MAE and RMSE data for the validation and test sets of GWO-BP. The results show that: 1) The optimization is effective: regardless of the validation or test set, the errors of GWO-BP in MAE and RMSE are lower than those of traditional BP, and the error reduction exceeds 28%, proving that the scheme of "using the Grey Wolf Optimization Algorithm (GWO) to optimize BP network parameters" is effective and can significantly improve the model's prediction performance. 2) Higher practical application value: The error reduction on the test set (MAE 48.78%, RMSE 43.26%) is significantly higher than that on the validation set (MAE 28.92%, RMSE 31.76%), indicating that GWO-BP is more adaptable to "unknown data scenarios" and can effectively avoid the "overfitting" problem that is prone to occur in traditional BP, making it more suitable for practical prediction tasks. 3) Dual improvement in stability and accuracy: Under both datasets, the error reduction in RMSE is no less than 31%. Combined with the significant optimization of MAE, it shows that GWO-BP not only improves prediction accuracy but also enhances prediction stability, achieving a dual optimization of "accuracy + stability". Therefore, the GWO-BP model was selected as the third sub-model of the Stacking model, and the prediction set P3 was obtained by using the GWO-BP regression model for prediction. Figure 5 As shown, the prediction results of the GWO-BP model meet the error accuracy requirements compared with the original data.

[0100] Table 3 Comparison and Validation of GWO-BP Model

[0101]

[0102] Figure 2 Explanation: Horizontal axis: Number of evolution steps (iteration steps), representing the algorithm optimization process. Vertical axis: Fitness, measuring the quality of the solution (higher / lower values ​​depend on the objective; here, it is high initially and decreases later, possibly indicating a minimization problem). Figure 2 Fitness trend: The fitness fluctuates and decreases in the early stages, then stabilizes after 7-8 iterations, indicating that the algorithm gradually seeks the optimal solution during iterations and converges to a better solution in the later stages. Ultimately, 20 iterations were selected. In short... Figure 2 It reflects the process by which the algorithm approaches the optimal solution from an initial random solution through iterative (evolutionary) optimization, and can be used to evaluate the algorithm's convergence and optimization efficiency.

[0103] An Improved Stacking Model: Ensemble learning is a crucial method in machine learning for enhancing prediction accuracy. The improved Stacking model combines the prediction sets (P1, P2, P3) from multiple learners (GWO-SVM, GWO-ELM, GWO-BP) as input to the first layer of the model. This input is then passed to the second layer for multiple linear regression fitting, and finally synthesized to obtain an improved ensemble learner result, outputting a prediction set P. This algorithm fully leverages the strengths of each sub-model, employing multiple linear regression for prediction, thus improving generalization ability and enhancing the accuracy and precision of the prediction results.

[0104] Formula (10) uses the prediction sets P1, P2, and P3 as inputs to perform multiple linear regression in the second layer, obtaining three fitted parameters: 0.4966, 0.4491, and 0.0486. The ordinary coefficient of determination R0 in this formula is... 2 The coefficient of determination is 0.9898, and the adjusted coefficient of determination R0 is... 2 The value is 0.9896, indicating a very high degree of model fit, and the sub-model predictions can explain the overall variability of the predictions well. The fitted output dataset P, i.e., the final TQI prediction result, is compared with the original data to obtain the residual plot within the prediction unit. Figure 7 ), and make predictions on a large sample prediction dataset to obtain a comparison chart of predicted TQI and measured TQI (). Figure 8 According to formulas (1) and (2), the mean relative error and root mean square error of the improved Stacking model are 0.0052 and 0.0722, respectively. It can be seen that the MAE and RMSE predicted by any single model constructed using the six algorithms are smaller, indicating that the improved ensemble model has higher prediction accuracy and possesses the advantages of ensemble integration. Figure 8 It can be seen that the Stacking model best matches the measured data and has the highest prediction accuracy, followed by GWO-SVM, and the SVM model is the worst.

[0105] ;

[0106] Figure 9This is a flowchart of a high-speed railway track irregularity prediction model based on an improved Stacking model. Extensive collection of measured track irregularity data from track inspections was conducted, including noise reduction, outlier removal, and trend term removal. Selected sample data were partitioned, and several algorithm models (such as SVM, ELM, and BP) were constructed for training on the datasets. The Grey Wolf algorithm was used for optimization and validation. An improved Stacking-based track irregularity prediction model was built using multiple linear regression, employing the TQI index to predict track irregularity. The mean absolute error and root mean square error of the predicted set were compared with those of the measured set to evaluate the model's prediction accuracy and stability. Based on machine learning and ensemble thinking, the safety of high-speed train operation was predicted and controlled.

[0107] Track irregularity data preprocessing: An outlier criterion of 3‰ track irregularity variation rate was used to check and remove errors from the dynamic track inspection data. Sample denoising was performed using Complete Empirical Mode Decomposition (CEEMDAN) based on adaptive noise. A 200m interval of sample data was selected. Taking horizontal irregularity data as an example, after CEEMDAN denoising, a comparison chart of the sample data before and after denoising was obtained, as shown below. Figure 10 and Figure 11 .

[0108] Figure 10 and Figure 11 A comparison chart showing the process before and after removing trend terms and outliers from the track irregularities. Figure 10 (a) The original signal of horizontal irregularity shows the fluctuation curve of the original monitoring data in the horizontal direction of the track (such as rail height, left and right deviation, etc.), which includes real irregularity characteristics + noise (such as measurement error, environmental interference), and the signal may be messy and have many spikes. Figure 10 (b) CEEMDAN decomposition demonstrates that the original signal is decomposed into several IMFs (Intrinsic Mode Functions) and trend terms. Different IMFs represent the components of the signal at different frequencies (time scales), which can separate noise and true features, thus preparing for denoising. Figure 11 (a) is the waveform after horizontal irregularity denoising. The horizontal irregularity features (such as periodicity and trend) are clearer after the noise is suppressed. Figure 10 and Figure 11 The core objective is to address the fact that track unevenness is a key indicator affecting train operation safety and comfort, and noise can interfere with fault diagnosis (such as identifying loose fasteners and track deformation). By decomposing and separating noise using CEEMDAN, the reconstructed signal is cleaner, which is beneficial for building TQI prediction models and dynamic analysis.

[0109] CEEMDAN Advantages: Compared to traditional EMD (Empirical Mode Decomposition), CEEMDAN solves the mode aliasing problem by introducing adaptive noise, resulting in more accurate decomposition and more reliable noise reduction.

[0110] In short, this set of images of horizontal irregularities demonstrates the complete process of "original noisy signal → CEEMDAN decomposition and noise reduction → clean signal output", reflecting the application value of CEEMDAN in track irregularity signal processing and providing a high-quality data foundation for track condition health monitoring (such as TQI) and defect diagnosis. The sample selected a total of 20km of processing data in the K720+000m~K740+000m interval. The programming considers that the rate of change of irregularity data in adjacent 1m intervals is no greater than 3‰, as shown in formulas (11) and (12), in order to check and remove outliers and other errors in the measured track irregularity data. After removing outliers, the blank data is then processed by linear interpolation. The track irregularity comparison chart is shown below. Figure 12 .

[0111] Estimated value: ;

[0112] Comparison of data within a 1m range: ;

[0113] TQI Feature Selection: The large dataset used for machine learning training originated from data from a high-speed railway section from K720+000m to K1130+000m. After preprocessing, 1941 unit segments were obtained. For each unit segment, the maximum, minimum, average, standard deviation, skewness, and kurtosis of seven TQI indicators (gauge, level, triangular crater, left elevation / recession, right elevation / recession, left track orientation, and right track orientation) were calculated, numbered 1-42, for a total of 42 basic features (see Table 4). Random forest was used for important feature selection, prioritizing the construction of a supervised model for feature selection. Considering the weight of track irregularity prediction indicators (42 codes) in the actual TQI values, several important track irregularity TQI features were output using MATLAB programming.

[0114] Table 4. Selection of Track Irregularity Data Features

[0115]

[0116] Based on the importance of the 42 encoded output track irregularity prediction indicators in Table 4, the top 10 most important indicators are 11, 16, 19, and 20. Indicators 22-28 correspond to the maximum right elevation / lower bound, level, and right elevation / lower bound, the average left track orientation, and the standard deviations of left track orientation, right elevation / lower bound, right track orientation, triangular irregularities, left elevation / lower bound, and level. This indicates that the standard deviation of each indicator has the greatest impact on TQI, followed by the peak and average values ​​of elevation / lower bound, and finally the average value of level. Seven key features, including the standard deviations of TQI (22-28), along with other features such as odometer readings, are adaptively selected as the input layer of the prediction model to establish an improved Stacking model.

[0117] To verify the applicability and accuracy of the improved Stacking model in predicting TQI, track irregularity data from a 2km section within the K940+000m-K942+000m high-speed railway was randomly selected from the sample data. This data was then divided into 10 interval units using the method described above. A single model was used, and the TQI values ​​predicted by the Stacking model were compared with the measured TQI values. Figure 13 , Figure 14 As shown, the overall trend of the model prediction results is consistent with the measured data. The sub-models optimized by the Grey Wolf optimization algorithm are close to the measured data. Among them, the SVM prediction results are the worst, while the improved Stacking model has the best prediction results.

[0118] Of the three models mentioned above, the SVM model shows the worst prediction accuracy based on the MAE and RMSE evaluation metrics. The Stacking model has the highest prediction accuracy and precision, while the GWO-SVM model's accuracy is close to that of Stacking. The Grey Wolf optimization model performs well. Overall, as the number of interval units increases and the length of the track length grows, the MAE and RMSE of the intervals also increase, as shown in the MAE and RMSE of the validation and test sets in Tables 1-3. If the intervals are divided into groups of 10 units, the model's prediction accuracy will fluctuate with the track smoothness of the intervals, especially in curves and turnout sections.

[0119] To illustrate the results more intuitively, mean squared error (MSE) and root mean square error (RMSE) were used as accuracy evaluation metrics. Since MSE and RMSE have similar evaluation criteria, the improved Stacking model exhibited the lowest MAE and the best stability. Secondly, the three sub-models optimized by the Grey Wolf algorithm—GWO-SVM, GWO-ELM, and GWO-BP—showed similar MAE values ​​across different intervals, with relatively small MSE amplitudes, indicating a significant improvement effect of the Grey Wolf algorithm on individual models. The worst performing model was SVM, with the highest MAE, indicating the worst prediction accuracy. Two histogram peaks also showed unusual spikes, indicating instability in the interval unit prediction results. The improved Stacking model, compared to other algorithms, improves the prediction accuracy and stability of the TQI model and is overall superior to other single algorithms.

[0120] like Figure 15 As shown, this invention provides a flowchart of a Stacking model based on multiple linear regression. The target variable, the predata of track irregularity processing, is used as the TQI (Trajectory Quality Index). The dataset is divided into training, testing, and validation sets based on continuous mileage segments. In the base model training phase (first layer), the optimization objective is to minimize the TQI prediction error. The optimization parameters are: SVM kernel parameters / ELM hidden nodes / BP fitness. The optimization method is: GWO (Gross WORD) optimization of initial weights / learning rate. Meta-feature generation: the base model's TQI prediction results on the training set (3 columns of meta-features). Meta-dataset construction: meta-feature moments adapted to the track data. The array, with each row corresponding to an integrated output of a mileage segment; Meta-model training phase (layer 2): Select a robust meta-model (at this time, obtain the actual TQI value by performing multiple linear regression on the meta-model, as well as by correlation, comparison, selection, and 5-fold cross-validation), and perform the final prediction and evaluation of the Stacking model (test set and validation set): Compare with the measured TQI and output the TQI prediction evaluation results: 1. Base model prediction on the test set: meta-features; 2. Meta-model fusion: final TQI prediction; 3. Evaluation indicators: RMSE and MAE; 4. Visual output of track status evaluation (mileage-TQI-warning interval).

[0121] This invention utilizes machine learning and ensemble thinking, selecting Track Quality Index (TQI) as the evaluation index for track smoothness. By comparing the accuracy values ​​of the mean error (MAE) and root mean square error (RMSE) between predicted and measured TQI values, an improved Stacking model is proposed to predict track irregularities in high-speed railways. The results are as follows:

[0122] (1) Using the actual track inspection data of a high-speed railway line in northern my country with a total length of about 410km in a certain year, including data of roadbed, bridge, tunnel, curve and turnout sections, six algorithm models (SVM, ELM, BP, GWO-SVM, GWO-ELM and GWO-BP) were trained, established, optimized and verified, and the machine learning training and prediction accuracy of the algorithm models were compared. It should be noted that: the measured and predicted TQI values ​​of this invention are sometimes higher than 5, for two reasons (1) the curve section and turnout section are weak links, and the TQI is larger here; (2) it shows that the track quality status of this section and time period is indeed not good. The Stacking model constructed by this invention can provide a method for subsequent track quality status monitoring and evaluation. (2) Incorporating the Grey Wolf algorithm can improve the prediction accuracy of the sub-models (GWO-SVM, GWO-ELM and GWO-BP), especially the support vector machine SVM algorithm model. (3) By adopting the ensemble approach and incorporating three seed models, and using the multiple linear regression algorithm, three regression coefficients (0.4966, 0.4491, and 0.0486) were obtained to establish an improved Stacking model, further enhancing the accuracy of Stacking model predictions. Based on the error accuracy evaluation index, the improved Stacking model achieved a TQI prediction accuracy of up to 99.52%, and can be given priority for promotion.

[0123] This invention proposes a high-speed rail track irregularity (TQI) prediction system, specifically comprising: a data acquisition module for collecting track inspection measurement data of the target high-speed rail track section; a feature extraction module for extracting multiple track irregularity feature variables from the track inspection measurement data; an initial prediction module for inputting different track irregularity feature variables into an improved Stacking model, wherein each sub-model in the improved Stacking model uses the optimal model parameters determined by an optimization algorithm and outputs preliminary TQI prediction values ​​respectively; and a final prediction module for processing multiple TQI prediction results using a multiple linear regression algorithm to obtain multiple regression coefficients, and based on the ensemble concept, performing a weighted summation of each preliminary TQI prediction value and its corresponding regression coefficient to obtain the final TQI prediction result for the target high-speed rail track section.

[0124] The present invention provides a computer device, including a memory and a processor. The memory stores a program, and when the program is executed by the processor, the processor performs the steps of a high-speed rail track irregularity (TQI) prediction method.

[0125] According to the disclosed embodiments, the computer device can communicate with one or more external devices (e.g., keyboard, pointing device, Bluetooth communication, etc.) or with any device that enables the computing device to communicate with one or more other computing devices (e.g., router, demodulator, etc.).

[0126] The present invention provides a storage medium storing a computer program thereon, wherein the computer program, when executed by a processor, implements the steps of a method for predicting track irregularities (TQI) of high-speed railways.

[0127] According to the disclosed embodiments, the storage medium can be a non-volatile computer-readable storage medium, such as, but not limited to: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this invention, the storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0128] The above description, in conjunction with specific preferred embodiments, provides a more detailed explanation of the present invention. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such deductions or substitutions should be considered to fall within the scope of protection of the present invention.

Claims

1. A method for predicting track irregularities (TQI) in high-speed rail, characterized in that, include: Collect track inspection data for the target track section of the high-speed railway; Extract multiple track irregularity feature variables from the track inspection measured data; The different track irregularity feature variables are input into the improved Stacking model. Specifically, the improved Stacking model is constructed by integrating GWO-SVM, GWO-ELM and GWO-BP sub-models based on the Stacking model. Each sub-model in the improved Stacking model uses the optimal model parameters determined by the optimization algorithm and outputs the preliminary track quality index (TQI) prediction value respectively. Multiple preliminary track quality index (TQI) predictions were processed by multiple linear regression to obtain multiple regression coefficients. Based on the integration concept, the preliminary TQI predictions and their corresponding regression coefficients were weighted and summed to obtain the final TQI prediction result for the high-speed rail target track section. Each sub-model uses the optimal model parameters determined by the optimization algorithm and outputs a preliminary predicted orbital quality index (TQI) value, specifically: The optimal model parameter values ​​of SVM, ELM and BP models are obtained using the Grey Wolf Optimization Algorithm (GWO), and the wolf population size and number of evolutionary iterations are selected to construct GWO-SVM, GWO-ELM and GWO-BP models. Among them, the GWO-SVM, GWO-ELM and GWO-BP models use key feature variables of track irregularity as input variables and the corresponding measured TQI values ​​as output variables. The relationship between the input variables and the output variables is obtained through a linear regression function. Multiple track irregularity characteristic variables from track inspection data are input into GWO-SVM, GWO-ELM, and GWO-BP models, and preliminary track quality index (TQI) prediction values ​​are output through linear regression functions. The Gray Wolf Optimization Algorithm (GWO) is used to obtain the optimal model parameter values ​​for SVM, ELM, and BP models, specifically as follows: The gray wolf optimization algorithm is adopted, with gray wolves as the model parameters of each processing algorithm. The fitness of each gray wolf is calculated, and the position of the gray wolf group is updated by simulating the three stages of gray wolf hunting behavior: search, encirclement and attack. When a set threshold is met, the optimal parameters in the corresponding processing algorithm are output. When the gray wolf optimization algorithm stops iterating, the gray wolf with the best fitness is selected, and the corresponding position vector is the final output optimal model parameter value. Extracting multiple track irregularity feature variables from the track inspection measured data, specifically: Multiple track irregularity indices are calculated from the track inspection measured data to generate multiple basic features with different numbers. Then, the random forest method is used to select multiple basic features as track irregularity feature variables according to the importance of feature information. The track irregularity indices include track gauge, level, triangular crater, left elevation, right elevation, left track orientation, and right track orientation. The multiple basic features include the maximum value, minimum value, average value, standard deviation, skewness, and kurtosis of the track irregularity indices. After obtaining the final TQI prediction result for the target track section of the high-speed railway, the method further includes: The track inspection measurement data were denoised using complete empirical mode decomposition based on adaptive noise. Outliers in the track inspection data were checked and removed. The blank data after removing outliers were then processed using a linear interpolation method to obtain the processed track inspection data.

2. The method for predicting TQI (Traffic Quality Index) of high-speed rail tracks as described in claim 1, characterized in that, After collecting the track inspection measurement data of the target track section of the high-speed railway, the following is also included: The track inspection measurement data were denoised using complete empirical mode decomposition based on adaptive noise. Outliers in the track inspection data were checked and removed. The blank data after removing outliers were then processed using a linear interpolation method to obtain the processed track inspection data.

3. A high-speed rail track irregularity (TQI) prediction system, characterized in that, include: The data acquisition module is used to collect track inspection measurement data of the target track section of the high-speed railway; The feature extraction module is used to extract multiple track irregularity feature variables from the track inspection measured data; The initial prediction module is used to input different orbital irregularity feature variables into the improved Stacking model. Specifically, the improved Stacking model is constructed by fusing GWO-SVM, GWO-ELM and GWO-BP sub-models based on the Stacking model. Each sub-model in the improved Stacking model uses the optimal model parameters determined by the optimization algorithm and outputs the preliminary orbital quality index (TQI) prediction value respectively. The final prediction module is used to perform multiple linear regression on multiple preliminary track quality index (TQI) prediction values ​​to obtain multiple regression coefficients. Based on the integration concept, the module performs a weighted summation of each preliminary TQI prediction value and its corresponding regression coefficient to obtain the final TQI prediction result for the high-speed rail target track section. Each sub-model uses the optimal model parameters determined by the optimization algorithm and outputs a preliminary predicted orbital quality index (TQI) value, specifically: The optimal model parameter values ​​of SVM, ELM and BP models are obtained using the Grey Wolf Optimization Algorithm (GWO), and the wolf population size and number of evolutionary iterations are selected to construct GWO-SVM, GWO-ELM and GWO-BP models. Among them, the GWO-SVM, GWO-ELM and GWO-BP models use key feature variables of track irregularity as input variables and the corresponding measured TQI values ​​as output variables. The relationship between the input variables and the output variables is obtained through a linear regression function. Multiple track irregularity characteristic variables from track inspection data are input into GWO-SVM, GWO-ELM, and GWO-BP models, and preliminary track quality index (TQI) prediction values ​​are output through linear regression functions. The Gray Wolf Optimization Algorithm (GWO) is used to obtain the optimal model parameter values ​​for SVM, ELM, and BP models, specifically as follows: The gray wolf optimization algorithm is adopted, with gray wolves as the model parameters of each processing algorithm. The fitness of each gray wolf is calculated, and the position of the gray wolf group is updated by simulating the three stages of gray wolf hunting behavior: search, encirclement and attack. When a set threshold is met, the optimal parameters in the corresponding processing algorithm are output. When the gray wolf optimization algorithm stops iterating, the gray wolf with the best fitness is selected, and the corresponding position vector is the final output optimal model parameter value. Extracting multiple track irregularity feature variables from the track inspection measured data, specifically: Multiple track irregularity indices are calculated from the track inspection measured data to generate multiple basic features with different numbers. Then, the random forest method is used to select multiple basic features as track irregularity feature variables according to the importance of feature information. The track irregularity indices include track gauge, level, triangular crater, left elevation, right elevation, left track orientation, and right track orientation. The multiple basic features include the maximum value, minimum value, average value, standard deviation, skewness, and kurtosis of the track irregularity indices. After obtaining the final TQI prediction result for the target track section of the high-speed railway, the method further includes: The track inspection measurement data were denoised using complete empirical mode decomposition based on adaptive noise. Outliers in the track inspection data were checked and removed. The blank data after removing outliers were then processed using a linear interpolation method to obtain the processed track inspection data.

4. A computer device, characterized in that, The system includes a memory and a processor, wherein the memory stores a program, and when the program is executed by the processor, the processor performs the steps of the high-speed rail track irregularity (TQI) prediction method as described in any one of claims 1 to 2.

5. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the high-speed rail track irregularity (TQI) prediction method according to any one of claims 1 to 2.

Images