Steel rail polishing decision-making method and system based on deep neural network

Abstract

The invention provides a steel rail polishing decision-making method and system based on a deep neural network, and belongs to the technical field of railway track maintenance. The steel rail polishing decision-making method comprises the steps of collecting polishing historical data of a typical line, and establishing a basic database; processing the polishing historical data in the basic database, and dividing the polishing historical data into a training set, a verification set and a test set; constructing a steel rail polishing decision model based on the multilayer neural network, wherein the steel rail polishing decision model at least comprises an input layer, a hidden layer and an output layer; carrying out iterative training on the steel rail polishing decision model through the training set, and regularly evaluating the performance of the model by using the verification set in the iterative training process; finally evaluating the steel rail polishing decision model by using the test set; and the finally evaluated steel rail polishing decision model is used for carrying out polishing decision on a to-be-polished line. By means of the model structure, accurate mapping from steel rail state data to grinding parameters is achieved, and the intelligent level of grinding decision making is greatly improved.

Description

Technical Field

[0002] This invention relates to the field of railway track maintenance technology, and in particular to a rail grinding decision-making method and system based on deep neural networks. Background Technology

[0004] Rail grinding, as an important railway line maintenance technology, can effectively repair or reduce fatigue damage to rails, extend their service life, improve wheel-rail contact, and enhance the safety and stability of vehicle operation. It has been widely used worldwide.

[0005] Currently, there are two main methods for developing rail grinding modes: one is to rely on experienced technicians to select suitable grinding modes from the existing mode library of the grinding vehicle. Grinding vehicles usually have multiple pre-set grinding modes, each corresponding to a specific combination of grinding wheel angle and motor power, which can specifically address single or combined defects such as rail fatigue cracks, crushed edges, corrugation, side grinding, and spalling; the other is not to rely on the onboard mode library, but rather for technicians to subjectively set parameters such as grinding angle and power based on the rail profile data detected on-site and their personal experience.

[0006] However, both methods heavily rely on the experience of technicians. Different grinding modes have different grinding angle distributions. Deviations in the selection of grinding modes or mismatches with the key issues on site can lead to problems such as poor rail profiles, incomplete damage removal, and insufficient smoothness optimization, directly affecting maintenance and long-term service performance. Re-arranging grinding modes based on experience requires grinding tests before grinding, adjusting parameters based on test results, and checking rail profile deviations after each grinding pass to determine if further grinding is needed. It's impossible to determine the number of grinding passes in advance. This "trial and error" approach not only results in low grinding efficiency, increased wheel wear and labor costs, but also makes it difficult to guarantee the stability of grinding quality, easily leading to under-grinding or over-grinding of the rails.

[0007] Through years of practice, the railway engineering system has accumulated a wealth of historical grinding data. This data contains valuable expert experience and has been validated in practice, making it generally capable of handling various track conditions and rail service statuses. However, regrettably, this experience remains at the data level; current methods have failed to effectively extract its value and fully leverage its guiding role in grinding. If an intelligent data mining and analysis system could be established to extract the correlation between rail service status data and grinding plans from this historical grinding data, and transform it into quantifiable and reusable decision-making models, the potential value of these data assets could be fully realized.

[0008] Neural networks, as a powerful nonlinear modeling tool, can uncover the inherent correlations between complex data through deep learning. Their unique distributed representation capabilities and adaptive learning mechanisms can effectively integrate features from multi-source heterogeneous data to establish a nonlinear mapping relationship between rail service status parameters and grinding process parameters. By constructing deep neural network models, historical grinding data can be fully utilized to achieve intelligent grinding decisions for different service conditions.

[0009] In view of this, based on years of experience in production design in this and related fields, the inventor has designed a rail grinding decision method and system based on deep neural networks through repeated experiments, in order to solve the problems existing in the prior art. Summary of the Invention

[0011] The purpose of this invention is to provide a rail grinding decision-making method and system based on deep neural networks, which can organize rail grinding construction in a more scientific and rational manner based on the site conditions.

[0012] To achieve the above objectives, this invention proposes a rail grinding decision-making method based on deep neural networks, characterized in that the rail grinding decision-making method includes:

[0013] Collect historical data on the polishing of typical lines and establish a basic database;

[0014] The polishing history data in the basic database is processed, and the polishing history data is divided into a training set, a validation set, and a test set;

[0015] A rail grinding decision model is constructed based on a multi-layer neural network. The rail grinding decision model includes at least an input layer, a hidden layer, and an output layer.

[0016] The rail grinding decision model is iteratively trained using the training set, and the model performance is evaluated periodically using the validation set during the iterative training process.

[0017] The rail grinding decision model was finally evaluated using the test set.

[0018] The final evaluated rail grinding decision model is used to make grinding decisions for the track to be ground.

[0019] The present invention also proposes a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the above-described calculation method.

[0020] The present invention also proposes a computer device, including a memory and a processor, wherein a computer program capable of running on the processor is stored in the memory, wherein the processor executes the computer program to implement the steps of the above-described calculation method.

[0021] Compared with the prior art, the present invention has the following features and advantages:

[0022] This invention proposes a rail grinding decision-making method and system based on deep neural networks, which is an intelligent decision-making model for rail grinding based on deep neural networks. The model's input layer is connected to the rail grinding index, and the output layer provides specific grinding parameters such as grinding speed, number of grinding passes, grinding head distribution angle, and motor power. The optimal number of network layers and neurons is determined through hyperparameter search, and the hidden layers are activated using the Leaky ReLU function. Dropout layers and L2 regularization are also inserted to effectively prevent overfitting. This model structure achieves a precise mapping from rail state data to grinding parameters, significantly improving the intelligence level of grinding decision-making. Attached Figure Description

[0024] The accompanying drawings described herein are for illustrative purposes only and are not intended to limit the scope of the invention in any way. Furthermore, the shapes and proportions of the components in the drawings are merely illustrative to aid in understanding the invention and do not specifically limit the shapes and proportions of the components. Those skilled in the art, guided by the teachings of this invention, can select various possible shapes and proportions to implement the invention according to specific circumstances.

[0025] Figure 1 This is a flowchart of the rail grinding decision-making method of the present invention;

[0026] Figure 2 This is a schematic diagram of the neural network structure of the present invention. Detailed Implementation

[0028] The details of the present invention can be more clearly understood by referring to the accompanying drawings and the description of specific embodiments. However, the specific embodiments of the present invention described herein are for illustrative purposes only and should not be construed as limiting the invention in any way. Under the teachings of this invention, those skilled in the art can conceive of any possible modifications based on the invention, and these should all be considered to fall within the scope of the invention.

[0029] like Figure 1 As shown, this invention proposes a rail grinding decision-making method based on deep neural networks, wherein the rail grinding decision-making method includes:

[0030] Collect historical data on the polishing of typical lines and establish a basic database;

[0031] The historical polishing data in the basic database is processed and divided into training set, validation set and test set;

[0032] A rail grinding decision model is constructed based on a multi-layer neural network. The rail grinding decision model includes at least an input layer, a hidden layer, and an output layer.

[0033] The rail grinding decision model is iteratively trained using the training set, and the model performance is evaluated periodically using the validation set during the iterative training process.

[0034] The rail grinding decision model was finally evaluated using a test set.

[0035] The final evaluated rail grinding decision model is used to make grinding decisions for the rails to be ground.

[0036] This invention proposes a rail grinding decision-making method based on deep neural networks. By collecting historical data and constructing a decision-making model based on deep neural networks, it establishes an automated intelligent process from data to decision. Utilizing a fully trained deep neural network model, it can automatically learn and establish a complex mapping relationship between rail condition and grinding decisions from a large amount of historical data. This enables intelligent decision-making for rail grinding, effectively improving the automation level and objectivity of the decision-making process and avoiding grinding quality problems caused by human selection bias. Furthermore, by learning and modeling historical data, the grinding decision-making process shifts from experience-driven to data-driven, improving the scientific rigor, consistency, and efficiency of the decision.

[0037] In one optional embodiment of the present invention, the grinding history data includes rail profile data, rail surface damage data, rail surface smoothness data, and grinding parameters.

[0038] The data includes: rail profile data collected by a laser measuring instrument to describe the geometric shape of the rail head; rail surface damage data acquired by an ultrasonic flaw detector to record cracks, fish scale patterns, and other damage on the rail surface; rail surface smoothness data collected by an inertial measurement unit to reflect the degree of corrugation and unevenness on the rail surface; and grinding parameters including grinding speed, number of grinding passes, grinding head distribution angle, and motor power. These parameters are derived from historical grinding operation records and are integrated into the basic database through a data import module either manually entered or automatically parsed from common file formats (such as CSV and TXT), providing complete data support for the training and decision-making of the deep neural network model. By incorporating rail profile data, rail surface damage data, rail surface smoothness data, and grinding parameters into historical grinding data, this method can provide comprehensive and multi-dimensional input information for deep neural network models. This enables the models to more accurately learn the complex mapping relationship between rail service status and optimal grinding parameters, thereby improving the accuracy and reliability of grinding decisions, avoiding decision-making biases caused by missing or incomplete data, and ultimately realizing the intelligent transformation of rail grinding from experience-driven to data-driven.

[0039] In one alternative embodiment of this implementation, the grinding parameters include grinding speed, number of grinding passes, grinding head distribution angle, and motor power.

[0040] The grinding speed, expressed in kilometers per hour, represents the equipment's travel speed during grinding operations. The number of grinding passes is a positive integer, representing the number of cycles for grinding the same section. The grinding head distribution angle, expressed in degrees, describes the spatial orientation of each grinding head relative to the longitudinal axis of the rail. The motor power is expressed as a percentage, reflecting the magnitude of the motor's output power during grinding. These parameters serve as the output targets of the deep neural network model. During the model training phase, they, along with rail condition data, constitute the training samples. In the decision-making phase, the fully trained model automatically generates the corresponding parameter combinations based on the input rail condition data. By incorporating grinding speed, number of grinding passes, grinding head distribution angle, and motor power into the grinding parameter system, the deep neural network model can output a complete and executable grinding operation plan. This enables intelligent decision-making throughout the entire process, from rail condition assessment to the generation of specific process parameters, effectively ensuring the accuracy and operability of grinding operations and avoiding the subjectivity and inconsistency of traditional parameter selection based on human experience.

[0041] In an optional embodiment of the present invention, the polishing history data is processed by: removing outliers using the Z-score method, removing noise data using a low-pass filter, and then performing Min-Max normalization processing.

[0042] Removing outliers from historical rail grinding data using the Z-score method effectively eliminates the interference of extreme data on subsequent model training. Using a low-pass filter to remove noisy data preserves effective feature information related to the service status of rails in historical grinding data, preventing noise from obscuring key data features. After Min-Max normalization, historical grinding data of different dimensions and numerical ranges can be unified to the same scale, eliminating the impact of differences in data magnitude on model training. This improves the quality of historical grinding data, significantly reduces the disturbance of sample noise and dimensional differences on model training, and thus improves the convergence speed and prediction accuracy of the rail grinding decision model.

[0043] In an optional embodiment of the present invention, the processing of grinding history data further includes: calculating the rail grinding index based on the rail grinding index calculation formula using the grinding history data.

[0044] This calculation process uses a pre-established formula for calculating the rail grinding index. Processed rail profile data, rail surface damage data, and rail surface smoothness data are used as input parameters. The formula calculates quantitative evaluation results including the primary index (rail grinding index) and secondary indices (profile quality index, rail surface damage index, and rail surface smoothness index). These indices collectively constitute a comprehensive evaluation system for the rail's service condition. By processing historical grinding data and calculating the rail grinding index based on the formula, a quantitative evaluation of the rail's service condition is achieved. This provides standardized feature input for the deep neural network model, enabling the model to establish a mapping relationship between rail condition and grinding parameters based on a unified evaluation standard.

[0045] In one alternative embodiment of this implementation, the rail polishing index includes a primary index GI and a secondary index, wherein the secondary index includes the rail profile quality index PQI, the rail surface damage index SDI, and the rail surface smoothness index RCI.

[0046] Among them, GI serves as the primary indicator for comprehensively evaluating the overall service condition of rails, PQI is specifically used to quantitatively assess the quality of rail profile, SDI is specifically used to quantitatively assess the severity of rail surface damage, and RCI is specifically used to quantitatively assess the smoothness of rail surface. These indicators together constitute a multi-level rail condition evaluation system, which is calculated in parallel by the data processing module based on the corresponding calculation formulas. By constructing a rail grinding index system that includes the primary indicator GI and the secondary indicators PQI, SDI, and RCI, a multi-dimensional and refined quantitative assessment of the rail service condition is achieved. This provides comprehensive and structured feature inputs for the deep neural network model, enabling the model to more accurately grasp the degree of influence of different factors on grinding decisions.

[0047] In one alternative embodiment of this implementation, a typical track is divided into multiple sections based on the rail grinding index, and the grinding parameters corresponding to each section are added.

[0048] Specifically, the railway line is divided into multiple sections with similar service characteristics based on the calculated rail grinding index (GI). Within the same section, the difference in the GI does not exceed 0.4. After division, corresponding grinding parameters are added to each section, including grinding speed, number of grinding passes, grinding head distribution angle, and motor power. These parameters are derived from historical grinding operation records and are correlated with the rail condition characteristics of each section, forming a complete training sample dataset. By dividing the railway line into multiple sections based on the rail grinding index and adding corresponding grinding parameters to each section, refined management and targeted processing of rail condition are achieved. This enables the deep neural network model to learn the correspondence between different condition sections and optimal grinding parameters.

[0049] In one optional embodiment of the present invention, such as Figure 2 As shown, the input layer has four nodes x1, x2, x3, and x4, corresponding to the primary index GI, rail profile quality index PQI, rail surface damage index SDI, and rail surface smoothness index RCI, respectively. These four nodes constitute the data entry point of the neural network, each independently receiving the numerical input of its corresponding index. The preprocessed rail polishing index data is then passed as a feature vector to the subsequent hidden layers for further processing. The input layer establishes a one-to-one correspondence with GI, PQI, SDI, and RCI through four nodes, enabling it to comprehensively and accurately capture key information about the rail's service status. It receives GI, reflecting the overall rail condition, through x1, and PQI, SDI, and RCI, reflecting the subdivided dimensions of rail profile quality, rail surface damage, and rail surface smoothness, respectively, through x2, x3, and x4. This avoids the model's inability to accurately identify rail condition characteristics due to missing input data dimensions or chaotic correspondences. This explicit node-index correspondence ensures that the data input to the neural network can completely reproduce the actual rail service status, reducing information bias during data transmission. This lays a reliable foundation for the hidden layer to effectively extract features and the output layer to accurately calculate grinding parameters, thereby ensuring that the grinding scheme output by the deep neural network grinding decision model is highly adapted to the actual rail condition.

[0050] In one optional embodiment of the present invention, such as Figure 2As shown, the output layer has four nodes: y1, y2, y3, and y4, corresponding to the grinding speed v, the number of grinding passes n, the grinding head distribution angle θ, and the motor power P, respectively. These four nodes constitute the output of the neural network. Node y1 outputs the grinding speed value, node y2 outputs the number of grinding passes value, node y3 outputs the grinding head distribution angle value, and node y4 outputs the motor power value. Each node independently generates the corresponding grinding parameters, and the final grinding scheme is presented through the decision output module. By establishing a one-to-one correspondence between the output layer and four nodes—grinding speed v, number of grinding passes n, grinding head distribution angle θ, and motor power P—the system can directly output the specific operable parameters required for grinding operations. This eliminates the need for technicians to rely on experience to select or reprogram grinding modes, avoiding issues such as grinding mode selection deviations or mismatches with on-site requirements. The preset constraints of each parameter ensure that the output grinding parameters are within the reasonable range for actual construction, preventing problems such as poor profiles or incomplete damage removal due to parameter anomalies. At the same time, the corresponding settings of the number of grinding passes n with the grinding head distribution angle θ and motor power P can meet the precise operation requirements of different grinding stages, ensuring grinding quality and long-term service performance.

[0051] In one alternative embodiment of this implementation, constraints are added to the grinding speed v, the number of grinding passes n, the grinding head distribution angle θ, and the motor power P.

[0052] The grinding speed *v* is constrained within the range of 5 km / h to 25 km / h, and numerical mapping is achieved through a piecewise linear function. The number of grinding passes *n* is constrained to an integer from 1 to 6, and quantization is achieved through a rounding function. The grinding head distribution angle *θ* is constrained within the range of -10° to 50°. The motor power *P* is constrained within the range of 40% to 90%. These constraints take effect after processing in the neural network output layer, ensuring that all output grinding parameters are within a reasonable range that is operable in actual construction and achievable by the equipment. This avoids adverse consequences caused by parameters exceeding equipment capabilities (such as excessive grinding speed leading to equipment overload or grinding head angle exceeding adjustment limits) or not meeting rail repair requirements (such as insufficient motor power leading to incomplete damage removal or excessive grinding passes causing over-cutting). At the same time, the constraint settings ensure that the output grinding parameters always conform to the on-site construction conditions and the rail service condition repair requirements, guaranteeing stable and efficient grinding operations and effectively improving grinding quality and the long-term service performance of the rail.

[0053] In an alternative embodiment of the invention, the hidden layer determines the optimal number of layers and neurons through hyperparameter search.

[0054] The hyperparameter search employs a grid search method, systematically testing different combinations of the number of hidden layers and neurons per layer within a predefined parameter space. The optimal network structure configuration is determined by evaluating the performance metrics of each combination on the validation set. The final optimal number of layers and neurons is used to construct the hidden layer structure of the deep neural network model. Determining the optimal number of hidden layers and neurons through hyperparameter search avoids problems such as insufficient feature extraction and inability to fit the complex nonlinear relationship between the rail grinding index and grinding parameters due to too few hidden layers. It also avoids overfitting and reduced training efficiency due to too many layers or neurons. Determining the optimal number of layers and neurons enables the hidden layers to more efficiently extract and abstract features from the GI, PQI, SDI, and RCI data transmitted from the input layer, accurately capturing the correlation patterns between data, and laying a solid foundation for the output layer to accurately calculate grinding parameters.

[0055] In an alternative embodiment of the invention, the hidden layer outputs a nonlinear activation function.

[0056] Specifically, after weighted summation and the addition of a bias term, each neuron in the hidden layer immediately outputs through the nonlinear activation function Leaky ReLU, with the function form f(x) = max(0.01x, x). The activated signal is then passed to the next layer, and this process is repeated layer by layer until the output layer, forming a mapping of the nonlinear relationship between the rail grinding index and grinding parameters. The hidden layer's output through a nonlinear activation function introduces nonlinear factors into the neural network, overcoming the limitation of linear models in fitting the complex nonlinear relationship between rail grinding indices (GI, PQI, SDI, RCI) and grinding parameters (grinding speed, number of grinding passes, etc.). The Leaky ReLU function effectively avoids the gradient vanishing problem of traditional ReLU functions when the input is negative, ensuring that the hidden layer's processing of the input signal retains effective feature information while achieving accurate abstraction of complex data relationships. This allows the neural network to more accurately learn the correlation between rail service status and grinding schemes, providing crucial support for the output layer to output reasonable and accurate grinding parameters.

[0057] In an alternative embodiment of the invention, a Dropout layer is inserted between hidden layers and L2 regularization is applied.

[0058] Specifically, a Dropout layer is inserted between every two hidden layers with a dropout rate of 0.3. Simultaneously, L2 regularization is applied during weight updates, with a regularization coefficient λ = 1e-4. During training, 30% of neuron connections are randomly dropped; during testing, all connections are restored and multiplied by the retention probability. Weight decay terms are added to the loss function and participate in backpropagation, iterating layer by layer until convergence. Inserting Dropout layers between hidden layers effectively breaks the co-dependencies between neurons by randomly discarding some neuron outputs, preventing the model from overlearning noisy data in the training set and thus preventing overfitting. Applying L2 regularization further controls model complexity by limiting the weight size, preventing the model from becoming overly sensitive to small fluctuations in the training data due to excessively large weights. The synergistic effect of these two methods allows the deep neural network to fully learn the effective correlation between the rail grinding index and grinding parameters during training, while maintaining its adaptability to unseen rail service status data, improving the model's generalization performance and ensuring stable decision accuracy on both the validation and test sets.

[0059] In an optional embodiment of the present invention, the performance of the rail grinding decision model is periodically evaluated using a validation set during the iterative training process. The hyperparameters of the rail grinding decision model are adjusted based on the performance indicators on the validation set. When the performance of the rail grinding decision model on the validation set no longer improves, training is stopped, and the final rail grinding decision model is obtained.

[0060] Specifically, the operational procedure for iteratively training the rail grinding decision model is as follows: First, the historical grinding data is divided into a training set, a validation set, and a test set. During training, after a certain number of iterations (e.g., every 50 iterations), the performance of the current model is evaluated using the validation set. The performance metrics used for evaluation include root mean square error (RMSE) (≤0.15) and coefficient of determination (≥0.93). Based on the evaluation results on the validation set, the model's hyperparameters are adjusted. These hyperparameters specifically include the number of neural network layers, the number of hidden layer neurons, and the learning rate. For example, when the RMSE on the validation set is greater than 0.15, the number of hidden layers can be reduced or the learning rate adjusted to optimize the model. The process of iterative training, validation set evaluation, and hyperparameter adjustment is repeated until, after multiple consecutive evaluations (e.g., 3 consecutive evaluations), the model's performance on the validation set (RMSE no longer decreases, and the coefficient of determination no longer increases) no longer improves. At this point, training is stopped, and the final rail grinding decision model is obtained. Regularly evaluating model performance using a validation set during iterative training can promptly identify overfitting or underfitting issues during training, preventing the model from becoming overtrained and only adapting to the training set data, thus losing its ability to adapt to new data. Adjusting hyperparameters based on validation set performance metrics allows for targeted optimization of the model structure and training strategy, ensuring the model iterates towards better performance. Stopping training when validation set performance no longer improves avoids wasting time and resources due to ineffective training and ensures that the final rail grinding decision model has stable and excellent performance, accurately fitting the relationship between the rail grinding index and grinding parameters.

[0061] The present invention also proposes a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the above-described calculation method.

[0062] The present invention also proposes a computer device, including a memory and a processor, wherein a computer program capable of running on the processor is stored in the memory, wherein the processor executes the computer program to implement the steps of the above-described calculation method.

[0063] Example:

[0064] A smart decision-making system for rail grinding was used to make grinding decisions for a certain railway line.

[0065] A high-speed railway line (mileage K1234+500~K1236+200) has experienced corrugation, causing abnormal vibrations in the carriages, requiring preventative grinding.

[0066] (1) Data collection and import

[0067] A rail inspection vehicle integrating a laser measuring instrument, ultrasonic flaw detector, and inertial measurement unit is used to inspect the rails and obtain data on rail profile, rail surface damage, and smoothness.

[0068] The system automatically parses and imports data from CSV files via its data interface, and then associates the data with mileage coordinates.

[0069] (2) Data processing and segmentation

[0070] 1) Data Processing: Remove outliers caused by vibration from the vehicle data (acceleration > 1.5g is marked as noise). Use a low-pass filter (cutoff frequency 25Hz) to eliminate mechanical vibration noise from the test vehicle while retaining the wave-pattern characteristic frequency (5~15Hz).

[0071] 2) Based on the index calculation results, the segments are divided as shown in Table 1:

[0072] Table 1:

[0073]

[0074] 3) Determining the timing of polishing, as shown in Table 2:

[0075] Table 2:

[0076]

[0077] (3) Intelligent grinding parameter output, as shown in Table 3:

[0078] Table 3:

[0079]

[0080] The detailed explanations of the above embodiments are intended only to explain the present invention so as to facilitate a better understanding of the present invention. However, these descriptions should not be construed as limiting the present invention for any reason. In particular, the various features described in different embodiments can be arbitrarily combined with each other to form other embodiments. Unless there is an explicit description to the contrary, these features should be understood to be applicable to any embodiment, and not limited to the described embodiments.

Claims

1. A rail grinding decision-making method based on deep neural networks, characterized in that, The rail grinding decision-making method includes: Collect historical data on the polishing of typical lines and establish a basic database; The polishing history data in the basic database is processed, and the polishing history data is divided into a training set, a validation set, and a test set; A rail grinding decision model is constructed based on a multi-layer neural network. The rail grinding decision model includes at least an input layer, a hidden layer, and an output layer. The rail grinding decision model is iteratively trained using the training set, and the model performance is evaluated periodically using the validation set during the iterative training process. The rail grinding decision model was finally evaluated using the test set. The final evaluated rail grinding decision model is used to make grinding decisions for the track to be ground.

2. The rail grinding decision method based on deep neural networks as described in claim 1, characterized in that, The grinding history data includes rail profile data, rail surface damage data, rail surface smoothness data, and grinding parameters.

3. The rail grinding decision method based on deep neural networks as described in claim 2, characterized in that, The grinding parameters include grinding speed, number of grinding passes, grinding head distribution angle, and motor power.

4. The rail grinding decision method based on deep neural networks as described in claim 1 or 2, characterized in that, Processing the polishing history data includes: The Z-score method is used to remove outliers, and a low-pass filter is used to remove noisy data. Then, Min-Max normalization is performed.

5. The rail grinding decision method based on deep neural networks as described in claim 1, characterized in that, Processing the polishing history data also includes: The rail grinding index is calculated based on the formula for calculating the rail grinding index using the historical grinding data.

6. The rail grinding decision method based on deep neural networks as described in claim 5, characterized in that, The rail polishing index includes a primary index GI and a secondary index. The secondary index includes the rail profile quality index PQI, the rail surface damage index SDI, and the rail surface smoothness index RCI.

7. The rail grinding decision method based on deep neural networks as described in claim 5, characterized in that, The typical line is divided into multiple sections based on the rail grinding index, and the grinding parameters corresponding to each section are added.

8. The rail grinding decision method based on deep neural networks as described in claim 1, characterized in that, The input layer has four nodes x1, x2, x3, and x4, which correspond to the primary index GI, rail profile quality index PQI, rail surface damage index SDI, and rail surface smoothness index RCI, respectively.

9. The rail grinding decision method based on deep neural networks as described in claim 1, characterized in that, The output layer has four nodes y1, y2, y3, and y4, which correspond to the grinding speed v, the number of grinding passes n, the grinding head distribution angle θ, and the motor power P, respectively.

10. The rail grinding decision method based on deep neural networks as described in claim 9, characterized in that, Constraints were added for the grinding speed v, the number of grinding passes n, the grinding head distribution angle θ, and the motor power P.

11. The rail grinding decision method based on deep neural networks as described in claim 1, characterized in that, The hidden layer determines the optimal number of layers and neurons through hyperparameter search.

12. The rail grinding decision method based on deep neural networks as described in claim 1, characterized in that, The hidden layer outputs through a non-linear activation function.

13. The rail grinding decision method based on deep neural networks as described in claim 1, characterized in that, Dropout layers are inserted between the hidden layers and L2 regularization is applied.

14. The rail grinding decision method based on deep neural networks as described in claim 1, characterized in that, During iterative training, the performance of the rail grinding decision model is periodically evaluated using the validation set. The hyperparameters of the rail grinding decision model are adjusted based on the performance indicators on the validation set. When the performance of the rail grinding decision model on the validation set no longer improves, training is stopped, and the final rail grinding decision model is obtained.

15. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the calculation method according to any one of claims 1 to 14.

16. A computer device comprising a memory and a processor, wherein a computer program capable of running on the processor is stored in the memory, characterized in that, When the processor executes the computer program, it implements the steps of the computation method according to any one of claims 1 to 14.

Images