Wheel-rail component process optimization method based on double-loop inverse design

Abstract

The application provides a wheel-rail component process optimization method based on double-cycle reverse design, and belongs to the technical field of wheel-rail material design and optimization. The method comprises the following steps: S1, constructing a data set one; S2, constructing a data set two; S3, data preprocessing and feature selection; S4, constructing a second-cycle forward model; S5, constructing a second-cycle reverse model; S6, constructing a first-cycle forward model; and S7, constructing a first-cycle reverse model. The application aims to break the dependence on the complete data chain of "component-process-performance" and a single powerful algorithm in the prior art, and through the construction of a double-cycle optimization framework, the accurate prediction of performance under specific service conditions is realized on the basis of sparse and heterogeneous data, and the optimal component and process parameter combination is directly and quickly solved in reverse.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of wheel rail material design and optimization, and particularly relates to a wheel rail composition process optimization method based on double-cycle reverse design. BACKGROUND

[0002] Currently, the prediction of the final service performance (such as wear resistance, fatigue resistance, fracture toughness, etc.) of wheel rail mainly relies on empirical models based on physical metallurgy and a large number of laboratory tests. These traditional models have limited precision and generalization ability when dealing with complex systems with multiple components, nonlinearities, and multiple process links coupling, and it is difficult to accurately quantify the comprehensive influence of all factors on the final performance. Accordingly, the design of wheel rail composition and process for specific service performance requirements (such as designing high wear-resistant wheel rail for heavy-load lines) is essentially a complex "inverse problem". Currently, the industry generally adopts the "trial and error method", that is, based on expert experience to propose a preliminary scheme, and through multiple laboratory trials, performance tests and scheme adjustments, the cycle is long, the cost is high, and it is difficult to find a global optimal solution.

[0003] With the wheel rail preparation and maintenance entering the stage of "intelligentization + greenization" driven by two wheels, machine learning technology shows broad application prospects in it, especially in improving production efficiency and predicting maintenance capability. Currently, existing researches mainly learn the mapping relationship between "composition-process" and "service performance" from massive historical production data, experimental data and service data, focusing on the forward prediction of service performance. Or use simple machine learning algorithms to predict individual mechanical performance indicators, and convert the reverse design into an optimization problem under the constraint of the forward model (such as using genetic algorithm, particle swarm optimization algorithm, etc. to find the optimal solution). However, these methods have not formed a full-chain closed loop covering from composition process to mechanical performance to service performance.

[0004] Existing technologies are as follows:

[0005] Scheme one: This paper (Performance prediction and reverse design of weathering steel based on machine learning and multi-objective optimization research_He Xin) proposes a performance prediction and reverse design model of machine learning + multi-objective optimization to solve the problems of complex, long cycle and high cost in traditional weathering steel research and development, introduces physical metallurgy (PM) parameters modeling, constructs SVR, RF, GPR models, and GPR performs best (95.19% of sample error <10% for tensile strength test set, 97.67% for GPR), GPR (80.4% of sample error <8%) and GA-SVM (accuracy 80.93%) for corrosion prediction, and then optimizes the composition process of Q700NH with SPEA2, and the obtained steel performance is excellent. However, this method uses only hundreds of samples, the elongation prediction error is large, and the generalization of other steel types is not verified.

[0006] ​Scheme two: This paper (Research on Steel Rail Wear Prediction Method Based on Machine Learning) aims to solve the problem of inaccurate prediction of rail wear and affect the safety of operation and maintenance. Based on the vertical / side wear data collected by the rail inspection vehicle, the speed, track gauge, axle load and other characteristics are extracted, the single class support vector machine is used to separate the out-of-limit data, and the combination algorithm of RBF fuzzy neural network and wavelet analysis-particle swarm optimization least squares support vector machine is compared. The results show that the combination algorithm has stronger generalization and higher prediction accuracy, which can provide basis for rail grinding / replacement.

[0007] Disadvantages are:

[0008] 1. Traditional wheel rail research and development mainly relies on experience and trial and error method and a large number of experiments, which has problems of long cycle, high cost and low efficiency. With the addition of machine learning method, the performance of forward prediction has been improved to a certain extent, but its development is limited by the biggest problem of insufficient data dimension and quality: the "complete sample" of composition-process-mechanical property-service performance chain is scarce. The production data only include composition content and process parameters, and the subsequent organization-performance correlation model is distorted; the laboratory data mainly include room temperature, quasi-static, small size sample, and lack of intrinsic properties under-60℃ to 200℃, to Strain rate and multi-axial fatigue working condition, which causes error amplification when extrapolating the service environment; the field service data often only have macro damage records, and lack of key parameters such as fine organization, residual stress and loading history, which leads to extremely sparse training set and large label noise, and is difficult to support high-order machine learning model.

[0009] 2. The existing technology usually regards "composition design", "process optimization" and "service performance prediction" as independent links. For example, scheme one optimizes composition and process, but its forward prediction model is usually based on mechanical properties, not real long-term service performance; scheme two predicts service performance, but does not link back to the initial composition and process design, and lacks a unified model that connects "composition-process-mechanical property-service performance-life" chain. SUMMARY

[0010] The present application provides a wheel rail composition process optimization method based on double cycle reverse design, which aims to break the dependence of existing technology on "composition-process-performance" complete data chain and single powerful algorithm, and realizes accurate prediction of performance under specific service conditions based on sparse and heterogeneous data, and directly and quickly solves the optimal composition and process parameter combination.

[0011] To achieve the above purpose, the present application adopts the following technical scheme:

[0012] The wheel rail composition process optimization method based on double cycle reverse design comprises:

[0013] S1, constructing dataset one: collecting the chemical composition, processing technology and mechanical property characteristic original data of wheel rail as initial samples, all initial samples constitute initial dataset one;

[0014] S2, constructing dataset two: collecting the mechanical property, service condition and service performance characteristic original data of wheel rail as initial samples, all initial samples constitute initial dataset two;

[0015] S3, data preprocessing and feature selection: preprocessing initial dataset one and initial dataset two to obtain dataset, performing Min-max standardization on the dataset, analyzing feature correlation and performing feature elimination by combining Pearson correlation coefficient method, and then reducing dimension of the retained features by principal component analysis to construct modeling dataset;

[0016] S4, constructing second cycle forward model: dividing the corresponding part of the modeling dataset two into training set two and test set two, constructing machine learning forward prediction model two with mechanical property and service condition as input and service performance as output, training the model with training set two, evaluating accuracy with test set two, optimizing hyperparameters by combining particle swarm optimization algorithm and cross validating the model;

[0017] S5, constructing second cycle reverse model: constructing a multi-objective optimization framework based on evolutionary algorithm, setting the constraint range of mechanical property and service condition and initializing the population, taking the trained machine learning forward prediction model two as fitness evaluation tool, and iteratively optimizing the mechanical property optimal solution;

[0018] S6, constructing first cycle forward model: dividing the corresponding part of the modeling dataset one into training set one and test set one, constructing machine learning forward prediction model one with chemical composition and processing technology as input and mechanical property as output, training the model with training set one, evaluating accuracy with test set one, optimizing hyperparameters by combining grid search, and cross validating to improve model stability and generalization ability;

[0019] S7, constructing first cycle reverse model: constructing a multi-objective optimization framework based on evolutionary algorithm, taking the mechanical property optimal solution as optimization target, setting the constraint range of chemical composition and processing technology parameters and initializing the population, taking the trained machine learning forward prediction model one as fitness evaluation tool, and iteratively optimizing the optimal chemical composition and processing technology parameters.

[0020] In the specification, the wheel rail composition process optimization method based on double cycle reverse design further comprises S8, closed loop iterative optimization: feeding the performance results corresponding to the optimal chemical composition and processing technology parameters to machine learning forward prediction model two and machine learning forward prediction model one, updating dataset one and dataset two, and re-executing steps S4 to S7 to realize the composition-process-mechanical property-service performance closed loop system.

[0021] In the specification, in step S1, the chemical composition includes the content of elements C, Si, Mn, P, S, Cr, V, Mo; the processing technology includes the finishing exit thickness, the finishing roughing temperature, the finishing finish rolling temperature, the start cooling temperature, the final cooling temperature, the quenching temperature, the tempering time; the mechanical properties include the tensile strength, the yield strength, the elongation, the hardness.

[0022] In the specification, in step S2, the mechanical properties include the tensile strength, the yield strength, the elongation, the hardness; the service conditions include the creep rate, the axle load, the running speed; the service performance includes the wear rate, the fatigue life.

[0023] In the specification, in step S3, the Min-max standardization is to scale the sample values of the data set to a preset numerical range; the principal component analysis is to extract the principal components of the cumulative variance contribution rate of the remaining features that meet the preset requirements as the modeling features.

[0024] In the specification, in step S4, the sample quantity ratio of the training set two to the test set two is 4:1; the machine learning forward prediction model two adopts the extremely gradient boosting tree.

[0025] In the specification, in steps S5 and S7, the evolutionary algorithm adopts the genetic algorithm or the particle swarm algorithm; the target of the iterative optimization is to minimize the deviation between the service performance output by the machine learning forward prediction model two and the preset target service performance.

[0026] In the specification, in the closed-loop iterative optimization of step S8, the conditions for triggering the iterative update include at least one of the following: ① the deviation between the actual test value of the mechanical properties corresponding to the optimal chemical composition and processing technology parameters and the output value of the machine learning forward prediction model one exceeds 5%; ② the deviation between the actual test value of the service performance corresponding to the optimal mechanical properties and the output value of the machine learning forward prediction model two exceeds 8%; ③ the accumulated newly added full-chain sample data reaches 50 groups or more; the iteration termination condition is that after continuous 3 iterations, the test set evaluation accuracy of the two forward prediction models is improved to 92% or more, and the performance deviation is stable within 3%; wherein the full-chain sample data refers to complete sample data that simultaneously covers the chemical composition parameters, the processing technology parameters, the mechanical property test results, and the service performance test results under the corresponding service conditions of the wheel-rail.

[0027] In the specification, in step S4, the machine learning forward prediction model two adopts the extremely gradient boosting tree as the basic model, and adopts the particle swarm optimization algorithm combined with 10-fold cross-validation to optimize the model hyperparameters, including: learning rate, tree maximum depth, subsample ratio, base learner number, split minimum loss reduction, and L2 regularization weight.

[0028] In the specification, in step S6, the machine learning forward prediction model uses a random forest as a base model, and model hyperparameter optimization is performed by combining a grid search, and the hyperparameters include: the number of decision trees, the maximum depth, the minimum number of samples for splitting a node, and the maximum number of features.

[0029] To sum up, the present application has at least the following beneficial effects:

[0030] The present application realizes reverse design from service performance to component process by constructing two data sets and a double cycle model, solves the problems of incomplete wheel-rail data samples and dependence on experience in the design process, improves the prediction accuracy and design efficiency by combining machine learning with optimization algorithms, and supports multi-objective and multi-constraint optimization requirements. BRIEF DESCRIPTION OF DRAWINGS

[0031] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiment description. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor.

[0032] Figure 1 A schematic diagram of the technical route of the wheel-rail component process optimization method based on double cycle reverse design involved in the present application;

[0033] Figure 2 A flowchart of the wheel-rail component process optimization method based on double cycle reverse design involved in the present application. DETAILED DESCRIPTION

[0034] In the following, only some exemplary embodiments are simply described. As those skilled in the art can recognize, the described embodiments can be modified in various different ways without departing from the spirit or scope of the embodiments of the present application. Therefore, the drawings and the description are considered to be exemplary in nature rather than limiting.

[0035] The following disclosure provides many different embodiments or examples for implementing the various structures of the embodiments of the present application. In order to simplify the disclosure of the embodiments of the present application, the components and settings of specific examples are described in the following. Of course, they are only examples, and the purpose is not to limit the embodiments of the present application. In addition, the embodiments of the present application can repeatedly refer to numerals and / or reference letters in different examples, and such repetition is for the purpose of simplification and clarity, which itself does not indicate the relationship between the various embodiments and / or settings discussed.

[0036] The embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0037] As Figure 1 and Figure 2 shown, the embodiment provides a wheel-rail composition process optimization method based on double-cycle reverse design, comprising:

[0038] S1, constructing dataset one: collecting the chemical composition, processing technology and mechanical property characteristic original data of wheel-rail as initial samples, and all initial samples constitute initial dataset one;

[0039] S2, constructing dataset two: collecting the mechanical property, service condition and service performance characteristic original data of wheel-rail as initial samples, and all initial samples constitute initial dataset two;

[0040] S3, data preprocessing and feature selection: preprocessing initial dataset one and initial dataset two to obtain dataset, performing Min-max standardization on the dataset, analyzing feature correlation and performing feature elimination by combining Pearson correlation coefficient method, and then reducing dimension through principal component analysis to construct modeling dataset;

[0041] S4, constructing second-cycle forward model: dividing the corresponding part of the modeling dataset two into training set two and test set two, constructing machine learning forward prediction model two with mechanical property and service condition as input and service performance as output, training the model with training set two, evaluating accuracy with test set two, optimizing hyperparameters by combining particle swarm optimization algorithm and cross validating the model;

[0042] S5, constructing second-cycle reverse model: constructing a multi-objective optimization framework based on evolutionary algorithm, setting the constraint range of mechanical property and service condition and initializing the population, taking the trained machine learning forward prediction model two as the fitness evaluation tool, and iteratively optimizing the output mechanical property optimal solution;

[0043] S6, constructing first-cycle forward model: dividing the corresponding part of the modeling dataset one into training set one and test set one, constructing machine learning forward prediction model one with chemical composition and processing technology as input and mechanical property as output, training the model with training set one, evaluating accuracy with test set one, optimizing hyperparameters by combining grid search, and cross validating to improve model stability and generalization ability;

[0044] S7, constructing first-cycle reverse model: constructing a multi-objective optimization framework based on evolutionary algorithm, taking the mechanical property optimal solution as the optimization target, setting the constraint range of chemical composition and processing technology parameters and initializing the population, taking the trained machine learning forward prediction model one as the fitness evaluation tool, and iteratively optimizing the output optimal chemical composition and processing technology parameters.

[0045] In some embodiments, the wheel-rail composition-process optimization method based on the dual-cycle inverse design further comprises S8, closed-loop iterative optimization: feeding the optimal chemical composition and the performance results corresponding to the processing parameters to the machine learning forward prediction model two and the machine learning forward prediction model one, updating the dataset one and the dataset two, re-executing steps S4 to S7, and realizing a composition-process-mechanical property-service performance closed-loop system.

[0046] In some embodiments, in step S1, the chemical composition includes the contents of elements C, Si, Mn, P, S, Cr, V, and Mo; the processing process includes the finishing rolling outlet thickness, the finishing rolling open rolling temperature, the finishing rolling final rolling temperature, the start cooling temperature, the final cooling temperature, the quenching temperature, and the tempering time; and the mechanical properties include the tensile strength, the yield strength, the elongation, and the hardness.

[0047] In some embodiments, in step S2, the mechanical properties include the tensile strength, the yield strength, the elongation, and the hardness; the service conditions include the creep rate, the axle load, and the running speed; and the service performance includes the wear rate and the fatigue life.

[0048] In some embodiments, in step S3, the Min-max standardization is to scale the sample values of the dataset to a preset numerical range; and the principal component analysis is to extract the principal components of the remaining features whose cumulative variance contribution rate meets the preset requirement as the modeling features.

[0049] In some embodiments, in step S4, the sample number ratio of the training set two to the test set two is 4:1; and the machine learning forward prediction model two adopts the extremely gradient boosting tree.

[0050] In some embodiments, in steps S5 and S7, the evolutionary algorithm adopts the genetic algorithm or the Bayesian optimization algorithm; and the target of the iterative optimization is to minimize the deviation between the service performance output by the machine learning forward prediction model two and the preset target service performance.

[0051] In some embodiments, in the closed-loop iterative optimization of step S8, the conditions triggering the iterative update include at least one of the following: ① the deviation between the actual test value of the mechanical properties corresponding to the optimal chemical composition and the processing parameters and the output value of the machine learning forward prediction model one exceeds 5%; ② the deviation between the actual test value of the service performance corresponding to the optimal mechanical properties and the output value of the machine learning forward prediction model two exceeds 8%; and ③ the accumulated newly added full-chain sample data reaches 50 groups or more; and the iteration termination condition is that after three consecutive iterations, the test set evaluation accuracy of the two forward prediction models is improved to 92% or more, and the performance deviation is stable within 3%; wherein the full-chain sample data refers to complete sample data covering the chemical composition parameters, the processing parameters, the mechanical property test results, and the service performance test results under the corresponding service conditions of the wheel-rail.

[0052] In some embodiments, in step S4, the machine learning forward prediction model two adopts an extreme gradient boosting tree as a base model, adopts a particle swarm optimization algorithm combined with 10-fold cross-validation to perform model hyperparameter optimization, and the hyperparameters include: learning rate, maximum tree depth, subsample ratio, base learner number, minimum split loss reduction, and L2 regularization weight.

[0053] In some embodiments, in step S6, the machine learning forward prediction model one adopts a random forest as a base model, and combines a grid search to perform model hyperparameter optimization, and the hyperparameters include: number of decision trees, maximum depth, minimum number of split samples, and maximum number of features.

[0054] The technical concept of the present application is as follows:

[0055] The present application aims to provide a wheel-rail composition process optimization method based on double-cycle reverse design, so as to solve the problem that it is difficult to accurately predict the wheel-rail service performance and directly reverse optimize the composition and process under the condition of incomplete data samples and single algorithm. The present application is especially suitable for rapid and accurate design of wheel-rail materials under specific working conditions such as heavy load and high speed.

[0056] The present application comprises the following steps:

[0057] S1, constructing a data set one: collecting the feature original data of the chemical composition, processing technology and mechanical properties of the wheel-rail, taking it as an initial sample, and collecting all the initial sample sets as an initial data set one;

[0058] S2, constructing a data set two: collecting the feature original data of the mechanical properties, service conditions (such as single working conditions of heavy load, high speed and general speed) and service performance of the wheel-rail, taking it as an initial sample, and collecting all the initial sample sets as an initial data set two;

[0059] S3, data preprocessing and feature selection: for the initial sample with more missing feature original data, directly deleting; for the initial sample with less missing feature original data, using the mean value of the missing feature in the initial data set to replace the missing feature original data, obtaining the sample after preprocessing, and the whole sample set is a data set; then the data set is processed by Min-max standardization; the Pearson correlation coefficient method is used to analyze the feature correlation and perform feature elimination, and then the principal component analysis method is used to further reduce the dimension of the remaining features, and a high-quality modeling data set is constructed;

[0060] S4, based on the second data set, a second cycle forward model is constructed: the second data set in step S3 is divided into two parts of training set and test set, a mechanical property, service condition as input, a service performance as output is constructed, a mechanical property, service condition→service performance machine learning forward prediction model is constructed, the model is trained by the training set, and the accuracy of the forward prediction model is evaluated by the test set; the hyperparameters are optimized, and the cross-validation method is used to optimize the model;

[0061] S5, based on the second data set, a second cycle inverse model is constructed: a multi-objective optimization framework is constructed based on evolutionary algorithm, and the target performance is optimized; the mechanical property constraint range is set, the population is initialized, the trained forward model is used as the fitness evaluation tool, the optimization is iterated, and the mechanical property optimal solution is output;

[0062] S6, based on the first data set, a first cycle forward model is constructed: the first data set in step S3 is divided into two parts of training set and test set, alloy composition, processing technology as input, mechanical property as output is constructed, alloy composition, processing technology→mechanical property machine learning forward prediction model is constructed, the model is trained by the training set, and the accuracy of the forward prediction model is evaluated by the test set; the model hyperparameters are optimized by grid search (GS), and the cross-validation method is used to optimize the model, and the stability and generalization ability of the model are improved;

[0063] S7, based on the first data set, a first cycle inverse model is constructed: a multi-objective optimization framework is constructed based on evolutionary algorithm, and the target performance is optimized; the composition and process parameter constraint range is set, the population is initialized, the trained forward model is used as the fitness evaluation tool, the optimization is iterated, and the optimal chemical composition and processing technology parameter are output;

[0064] S8, through iterative optimization, a closed-loop system of composition-process-mechanical property-service performance is realized;

[0065] S9, alloy design scheme output and verification: the scheme with optimal comprehensive performance is selected from the Pareto solution set, and experimental verification is carried out.

[0066] Further, in step S1, the chemical composition includes C, Si, Mn, P, S, Cr, V, MO element content; the processing technology parameters include finish rolling outlet thickness, finish rolling opening temperature, finish rolling final rolling temperature, starting cooling temperature, final cooling temperature, quenching temperature, tempering time; the mechanical properties include tensile strength, yield strength, elongation, hardness;

[0067] Further, in step S2, the mechanical properties include tensile strength, yield strength, elongation, hardness; the service conditions include creep rate, axle load, running speed; the service performance includes wear rate, fatigue life;

[0068] Further, in step S4, the sample number ratio in the model training set and the model test set is 4:1; the second cycle forward model adopts an extremely gradient boosting tree to serve performance forward prediction;

[0069] Further, in step S5, the second cycle inverse model adopts a genetic algorithm or a particle swarm algorithm to inversely solve a mechanical performance optimal solution, and an optimization objective is to minimize an objective function, and specifically satisfies:

[0070] ;

[0071] ;

[0072] Wherein, TP is a target service performance vector, MP is a mechanical performance vector, SC is a service condition, and F is a second cycle forward prediction model. is a physical constraint of MP and SC; and m is a total number of constraint conditions of MP and SC. represents an L2 norm; , are upper and lower limits of component parameters, respectively. , are upper and lower limits of process parameters, respectively; and s.t. indicates that a constraint condition is satisfied.

[0073] Further, in step S6, the sample number ratio in the model training set and the model test set is 4:1; the first cycle forward model adopts a random forest to establish a forward mapping of components-process to mechanical performance.

[0074] Further, in step S7, the first cycle inverse model adopts Bayesian optimization to perform inverse design and solve the following optimization problem:

[0075] ;

[0076] ;

[0077] Wherein, MP* is a second cycle output mechanical performance optimal solution, CP, PP is a component and process parameter vector, and G is a first cycle forward prediction model. is a physical constraint of CP and PP; and k is a total number of constraint conditions of CP and PP. represents an L2 norm; , are upper and lower limits of component parameters, respectively. , are upper and lower limits of process parameters, respectively; and s.t. indicates that a constraint condition is satisfied.

[0078] Further, in step S6, the inverse design results are experimentally verified through two-cycle iterative feedback, and the data set one and the data set two are updated, so as to gradually improve the model precision.

[0079] Further, the method further comprises a user interface module, which supports user input of target service performance and service conditions, and automatically outputs recommended rail chemical composition and process parameters.

[0080] Taking the development of a high-wear-resistant rail suitable for a "30-ton axle load and 100 million tons of annual traffic" heavy haul line as an example, the specific implementation steps are as follows:

[0081] I. Data collection:

[0082] 1) Constructing the rail data set one: collecting historical production data and experimental data to form an original database, wherein the database content includes the chemical composition (C, Si, Mn, Cr, V, N, etc.) of the rail, the processing process parameters (rolling temperature, cooling rate, heat treatment process, etc.) and the mechanical properties (tensile strength, elongation, hardness, impact toughness, etc.).

[0083] 2) Constructing the rail data set two: collecting the performance data of the rail in actual service to form an original database, wherein the database content includes the mechanical properties (same as the data set one, as a bridge connecting the two data sets), the service conditions (heavy load working conditions, i.e. creep rate of 0.5% to 3%, axle load of 24 tons to 36 tons, running speed of 80 to 120 km / h) and the service performance (wear rate and fatigue life).

[0084] 3) Data preprocessing and feature selection: for the initial sample with more missing feature original data, directly delete; for the initial sample with less missing feature original data, use the mean value of the missing feature in the initial data set to replace the missing feature original data, and obtain the sample after preprocessing, and the whole sample set is the data set; then the data set is normalized by using the normalization function mapminmax to realize the dimensionless processing, and the sample value is scaled to [-1, 1], and the calculation method is as follows:

[0085] ;

[0086] In the formula: y is the value to be calculated; and are -1 and 1, respectively; and are the minimum value and the maximum value in the current row, respectively;

[0087] The Pearson correlation coefficient method is used to determine the correlation between each feature and the output target, and the correlation coefficient The characteristics of the features; then the remaining features are further reduced by principal component analysis method, a high-quality modeling data set is constructed, and the data characteristics obtained are: C, Si, Mn, Cr, V, N, final rolling temperature, quenching temperature, tempering temperature, hardness, tensile strength, impact toughness, elongation, wear rate and fatigue life;

[0088] Pearson correlation coefficient The calculation formula is as follows:

[0089] ;

[0090] Wherein and represent the value of any two features in the ith rail (i=1, 2,..., 18); and represent the average value of the two features in the data set;

[0091] The PCA feature reduction formula is as follows:

[0092] ;

[0093] Where: Y is the new data matrix after dimensionality reduction; X is the original matrix, the dimension is (n samples, p features); is the mean vector of each feature; is the inverse of the original data covariance matrix; is the principal component projection matrix; T is the transpose;

[0094] II. Establishment of the second circulation model:

[0095] For the prediction of service performance (TP), since the data set two may contain complex nonlinear relationships, the gradient boosting tree (XGBoost) is preferred as the basic model. It establishes an independent prediction model for each service performance index (such as wear rate, fatigue life). Particle swarm optimization algorithm (PSO) is used to optimize the hyperparameters (such as learning rate, maximum depth, subsampling ratio, etc.) of each XGBoost model. The average mean square error (MSE) of 10-fold cross-validation is used as the fitness function of PSO to evaluate the influence of different parameter combinations in the parameter space on the performance of the model.

[0096] The specific implementation method is as follows:

[0097] Encode the hyperparameters to be optimized into a 6-dimensional particle vector:

[0098] ;

[0099] : learning rate; : maximum tree depth; Subsampling ratio; Number of base learners; Minimum loss from splitting; L2 regularization weights;

[0100] The value ranges of each component are shown in Table 1;

[0101] Table 1. Component Value Range

[0102]

[0103] For particles The training set is randomly divided into 10 folds, and the k-th fold is used as the validation set. The remaining 90% is used as the training set to build the XGBoost model. The verification error is as follows:

[0104] ;

[0105] in For particles The mean squared error on the k-th fold validation set; k is the fold index of the cross-validation (an integer from 1 to 10). For the k-th fold verification set The number of samples in the sample; For the j-th sample in the validation set, where These are input features. These are real labels; For particle-based The hyperparameters of the model are the predicted values ​​of the XGBoost model trained on the training set corresponding to the k-th fold.

[0106] Particle fitness formula (10-fold average error):

[0107] ;

[0108] in For particles The fitness value is smaller, indicating better generalization performance.

[0109] Let the population size be =30, at iteration t, the velocity and position of particle i are updated according to standard PSO:

[0110] ;

[0111] ;

[0112] in Let be the velocity vector of the i-th particle at the t-th iteration; is the optimal position of the ith particle from the beginning of the iteration to the tth iteration; is the global optimal position of the entire population from the beginning of the iteration to the tth iteration; ω = 0.9 is the inertia weight; = 1.8 is the acceleration constant; , ~ U(0, 1); is the position vector of the ith particle at the tth iteration; if a component exceeds the boundary of Table 1 after updating, it is reflected by mirroring the boundary value.

[0113] The search is stopped when the iteration reaches the maximum number = 100 or the fitness does not improve for 20 consecutive times, and the global optimal hyperparameters are output:

[0114] ;

[0115] wherein is the optimal parameter combination; is the parameter combination that minimizes the fitness ;

[0116] The optimal values of the hyperparameters of XGBoost optimized by PSO are learning rate 0.07, maximum tree depth 8, subsample ratio 0.85, base learner number 650, minimum split loss reduction 0.03, and L2 regularization weight 1.2, respectively, and XGBoost is retrained on the complete training set to obtain the final prediction model corresponding to the service performance indicators. The above process is independently performed for indicators such as wear rate and fatigue life, achieving multi-indicator parallel modeling;

[0117] The optimal solution of the mechanical performance that meets the target service performance is solved in reverse by the multi-objective optimization algorithm NSGA-II (population 100, generation number 200, crossover probability 0.9, and mutation probability 0.1). The objective function of the second cycle reverse design can be further represented as:

[0118]

[0119] ;

[0120] wherein: R(MP) is the mechanical performance regularization term, TP is the target service performance vector; MP is the mechanical performance vector; SC is the service condition; F is the second cycle forward prediction model; is the physical constraint of MP and SC; m is the total number of MP and SC constraint conditions; represents the L2 norm; ,​ upper and lower limits of ingredient parameters, respectively; , upper and lower limits of process parameters, respectively; s.t. denotes satisfying constraints;

[0121] Setting target service performance TP_target: wear rate <0.008 / MGT, fatigue life > 15 million tons passed total weight Running the second cycle reverse design, and solving the optimal mechanical performance combination MP* satisfying TP_target requirement by NSGA-II as shown in Table 2;

[0122] Table 2. Optimal mechanical performance combination

[0123]

[0124] Taking the optimal mechanical performance combination MP* as input, performing the first cycle forward prediction, and obtaining the predicted service performance: wear rate = 0.0072 / MGT (satisfying <0.008), fatigue life = 16.8 million tons passed total weight (satisfying > 150).

[0125] III. Establishment of the first cycle model:

[0126] For the mapping from ingredients-process to mechanical performance, due to the high nonlinearity of the relationship and possible uncertainties, a random forest (RF) is preferred as the base model. It establishes an independent prediction model for each mechanical performance index (such as tensile strength, hardness). The advantage of RF is that it not only gives the predicted value, but also gives the uncertainty estimate of the prediction. Grid search (GS) combined with cross-validation is used to optimize the hyperparameters of the RF model. The specific implementation is as follows: four independent RF solvers are established for the four performances; hyperparameters are set: number of decision trees 500, maximum depth 25, minimum number of node splitting samples 5, maximum number of features 3; five-fold cross-validation is used.

[0127] Taking the optimal mechanical performance solution of the second cycle as the target, the first cycle reverse optimization problem is extended to:

[0128] ;

[0129] ;

[0130] Where: R(CP,PP) is the ingredient-process regularization term, which constrains the complexity and cost of parameters; MP* is the optimal solution of mechanical performance output by the second cycle, CP,PP is the ingredient and process parameter vector, G is the first cycle forward prediction model; is the physical constraint of CP and PP; k is the total number of constraints of CP and PP; ​denotes the L2 norm; , denote the upper and lower bounds of the composition parameters, respectively; , denote the upper and lower bounds of the process parameters, respectively; s.t. denotes satisfying the constraints;

[0131] Constraint C: 0.7% ~ 0.9%, Si: 0.4% ~ 0.8%, Mn: 0.8% ~ 1.2%, Cr: 0.2% ~ 0.5%, V: 0.03% ~ 0.1%, Ni: 0.05% ~ 0.2%, Cu: 0.05% ~ 0.2%, finish rolling temperature 850 ~ 950°C, quenching temperature 200 ~ 300°C, tempering temperature 250-350°C; the Pareto optimal solution set is obtained by Bayesian optimization (acquisition function EI, kernel function Matern5 / 2, iteration budget 60 times, 4 kernel asynchronous evaluation) inverse solution: composition (C: 0.79%, Si: 0.63%, Mn: 1.02%, Cr: 0.33%, V: 0.06, Ni: 0.12, Cu: 0.10) and process (finish rolling temperature 875°C, quenching temperature 260°C, tempering temperature 260°C).

[0132] With the optimal composition process combination as input, forward prediction is carried out in the second cycle to obtain the predicted mechanical properties: hardness 403HB, tensile strength 1285MPa, impact toughness 23J, elongation 13%, and the deviation of the target 402HB / 1280MPa / 22J / 14% is only +1HB, +5MPa, +1J, +1%, respectively.

[0133] Four, closed-loop system iterative optimization:

[0134] Experimental verification: the optimal composition CP * and process parameters PP * obtained by the first cycle inverse solution are used for small-scale laboratory smelting, rolling and heat treatment test, and then the mechanical properties of the trial steel rail are tested to obtain the actual mechanical properties MP_actual. If conditions permit, further processing into test pieces for simulated service performance testing can be carried out to obtain the actual service performance TP_actual.

[0135] Data feedback: add the new data pair (CP * , PP * , MP_actual) generated by this trial to data set one. If service performance data is obtained, add (MP_actual, SC, TP_actual) to data set two.

[0136] Model update: periodically (e.g. every 50 new data sets accumulated) or when new data reaches a certain size, retrain the models of the first and second loops with the augmented data set one and data set two.

[0137] Iteration: repeat steps 3 to 7, so that the two prediction models evolve with the accumulation of data, the prediction accuracy and the reliability of the inverse design are continuously improved, forming a self-improving intelligent design closed loop.

[0138] The above-described embodiments are intended to illustrate the present application, not to limit the present application, so the change of example values or the replacement of equivalent elements should still belong to the scope of the present application.

[0139] From the above detailed description, it is clear to those skilled in the art that the present application can achieve the above-mentioned purposes, and has met the requirements of the Patent Law.

[0140] Although the preferred embodiments of the present application have been described, those skilled in the art can make further changes and modifications to these embodiments once they know the basic inventive concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications falling within the scope of the present application. The above description is only the preferred embodiments of the present application and is not intended to limit the present application. It should be noted that any modification, equivalent replacement and improvement made within the spirit and principle of the present application should be included in the protection scope of the present application.

[0141] It should be noted that the above description of the process is only for example and illustration, and does not limit the scope of the present application. Those skilled in the art can make various modifications and changes to the process under the guidance of the present application. However, these modifications and changes are still within the scope of the present application.

[0142] The above has described the basic concept, and it is obvious that the above-mentioned invention disclosure is only as an example and does not constitute a limitation to the present application for those skilled in the art after reading this application. Although it is not explicitly stated here, those skilled in the art can make various modifications, improvements and modifications to the present application. Such modifications, improvements and modifications are suggested in the present application, so such modifications, improvements and modifications still belong to the spirit and scope of the exemplary embodiments of the present application.

[0143] Also, certain terminology can also be used in the description for the purpose of reference only, and thus are not necessarily intended to be limiting. For example, the terms "one embodiment", "an embodiment" and / or "some embodiments", means a certain feature, structure, or characteristic is included in at least one embodiment of the disclosure. Therefore, these terms are

[0144] Also, those skilled in the art will appreciate that the various aspects of the present disclosure can be illustrated and described by means of certain embodiments or examples that have a variety of uses and / or that are implemented in a variety of suitable contexts. One of the aspects of the present disclosure is to provide for such embodiments and examples. Therefore, it is to be understood that the aspects of the present disclosure can be implemented in various ways, including software, hardware, firmware, special-purpose computers, or a combination thereof. In one embodiment, various aspects of the present disclosure can be implemented by one or more computer programs executing on one or more computers or computer- readable media. In an embodiment, a computer program is a set of instructions that can be used, directly or indirectly, in order to cause one or more computers or computer- readable media to perform a certain activity or bring about a certain result. A computer program can be written in any form of programming language, including compiled or interpreted languages, and it can be deployed in any form, including as a stand-alone program or as a module, component, subroutine, or other unit suitable for use in a computing environment. A computer program may, but need not, correspond to a file in a file system. Any component that can perform the functions described herein can be used as or to implement a computer program.

[0145] Computer program code for carrying out operations for aspects of the present disclosure can be written in any combination of one or more programming languages, including an object-oriented programming language such as Java, Scala, Smalltalk, Eiffel, JADE, Emerald, C++, C#, VB.NET, Python, or the like, conventional procedural programming languages, such as the C programming language, Visual Basic, Fortran 2103, Perl, COBOL 2102, PHP, ABAP, dynamic programming languages such as Python, Ruby, and Groovy, or another programming language. The program code can execute entirely on the user's computer, as a stand-alone software package, partly on the user's computer and partly on a remote computer, or entirely on the remote computer or server. In the latter scenario, the remote computer can be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or the connection can be made to an external computer (for example, through the Internet using an Internet Service Provider). In some embodiments, electronic program code can be downloaded from an on-demand computing service over the network (for example, on-demand software), or it can be downloaded across the network from the on-demand computing service (for example, software as a service (SaaS)).

[0146] Furthermore, the order of processing elements or sequences, or the use or appearance of certain terminology, throughout the above description should not be construed as limiting the application. Other steps, components, or configurations can be determined and implemented in a manner most beneficial to a particular application. For example, although the implementation of the various components described above can be embodied in hardware devices, it can also be implemented as a pure software solution, for example, as an installation on an existing server or mobile device.

[0147] Similarly, it is to be noticed that the term "comprising", used in the description, should not be interpreted as being restricted to the means listed thereafter; it does not exclude other elements or steps. It is thus to be interpreted as specifying the presence of the stated features, integers, steps or components as referred to, but does not preclude the presence or addition of one or more other features, integers, steps or components, or groups thereof. Furthermore, the description of the application is not intended to limit the application to the form disclosed herein. Various modifications and changes can be made without departing from the spirit and scope of the application as set forth in the following claims.

Claims

1. A method for optimizing wheel-rail composition processes based on dual-cycle reverse design, characterized in that, include: S1. Constructing Dataset 1: Collect raw data on the chemical composition, processing technology, and mechanical performance characteristics of wheel rails as initial samples. All initial samples constitute Initial Dataset 1. S2. Construct dataset two: Collect raw data on the mechanical properties, service conditions and service performance characteristics of wheel and rail as initial samples. All initial samples constitute initial dataset two. S3. Data Preprocessing and Feature Selection: Preprocess the initial dataset 1 and initial dataset 2 to obtain the dataset. Perform Min-max standardization on the dataset, analyze the features using the Pearson correlation coefficient method and remove features, and then reduce the dimensionality of the retained features through principal component analysis to construct the modeling dataset. S4. Construct the second loop forward model: Divide the part of the modeling dataset corresponding to dataset two into training set two and test set two. Construct machine learning forward prediction model two with mechanical performance and service conditions as input and service performance as output. Train the model with training set two and evaluate the accuracy with test set two. Combine particle swarm optimization algorithm to optimize hyperparameters and cross-validate the model. S5. Constructing the second loop inverse model: Based on the evolutionary algorithm, construct a multi-objective optimization framework, set the constraint range of mechanical performance and service conditions and initialize the population, use the trained machine learning positive prediction model 2 as the fitness evaluation tool, and iteratively optimize to output the optimal solution of mechanical performance. S6. Construct the first loop forward model: Divide the part of the modeling dataset corresponding to dataset 1 into training set 1 and test set 1. Construct machine learning forward prediction model 1 with chemical composition and processing technology as input and mechanical properties as output. Train the model with training set 1 and evaluate the accuracy with test set 1. Combine grid search to optimize hyperparameters and cross-validation to improve model stability and generalization ability. S7. Construct the first loop inverse model: Based on the evolutionary algorithm, construct a multi-objective optimization framework, take the optimal solution of mechanical performance as the optimization objective, set the constraint range of chemical composition and processing parameters and initialize the population, use the trained machine learning positive prediction model as the fitness evaluation tool, and iteratively optimize to output the optimal chemical composition and processing parameters. S8. Closed-loop iterative optimization: Feed the performance results corresponding to the optimal chemical composition and processing parameters back to the machine learning positive prediction model 2 and machine learning positive prediction model 1, update dataset 1 and dataset 2, and re-execute steps S4 to S7 to realize the closed-loop system of composition-process-mechanical properties-service performance.

2. The wheel-rail composition process optimization method based on dual-cycle reverse design according to claim 1, characterized in that, In step S1, the chemical composition includes the content of elements C, Si, Mn, P, S, Cr, V, and Mo; the processing technology includes the finishing mill exit thickness, finishing mill start temperature, finishing mill finish temperature, start cooling temperature, finish cooling temperature, quenching temperature, and tempering time; the mechanical properties include tensile strength, yield strength, elongation, and hardness.

3. The wheel-rail composition process optimization method based on dual-cycle reverse design according to claim 1, characterized in that, In step S2, the mechanical properties include tensile strength, yield strength, elongation, and hardness; the service conditions include creep rate, axle load, and operating speed; and the service performance includes wear rate and fatigue life.

4. The wheel-rail composition process optimization method based on dual-cycle reverse design according to claim 1, characterized in that, In step S3, the Min-max standardization is to scale the sample values ​​of the dataset to a preset numerical range; the principal component analysis is to extract the principal components whose cumulative variance contribution rate meets the preset requirements from the remaining features as modeling features.

5. The wheel-rail composition process optimization method based on dual-cycle reverse design according to claim 1, characterized in that, In step S4, the ratio of the number of samples in training set 2 to test set 2 is 4:1; the machine learning positive prediction model 2 adopts the extreme gradient boosting tree.

6. The wheel-rail composition process optimization method based on dual-cycle reverse design according to claim 1, characterized in that, In steps S5 and S7, the evolutionary algorithm adopts either a genetic algorithm or a particle swarm optimization algorithm; the goal of iterative optimization is to minimize the deviation between the service performance output by the machine learning positive prediction model 2 and the preset target service performance.

7. The wheel-rail composition process optimization method based on dual-cycle reverse design according to claim 1, characterized in that, In the closed-loop iterative optimization of step S8, the conditions for triggering the iterative update include at least one of the following: ① The actual test value of the mechanical properties corresponding to the optimal chemical composition and processing parameters deviates from the output value of the machine learning positive prediction model one by more than 5%; ② The actual test value of the service performance corresponding to the optimal mechanical properties deviates from the output value of the machine learning positive prediction model two by more than 8%; ③ The accumulated new full-chain sample data reaches 50 sets or more; The iteration termination condition is that after three consecutive iterations, the test set evaluation accuracy of the two positive prediction models is improved to more than 92%, and the performance deviation is stabilized within 3%.

8. The wheel-rail composition process optimization method based on dual-cycle reverse design according to claim 1, characterized in that, In step S4, the machine learning positive prediction model 2 uses an extreme gradient boosting tree as the base model and employs a particle swarm optimization algorithm combined with 10-fold cross-validation to optimize the model's hyperparameters. The hyperparameters include: learning rate, maximum tree depth, subsampling ratio, number of base learners, minimum loss reduction for splits, and L2 regularization weights.

9. The wheel-rail composition process optimization method based on dual-cycle reverse design according to claim 1, characterized in that, In step S6, the machine learning positive prediction model 1 uses random forest as the base model and combines grid search to optimize the model hyperparameters. The hyperparameters include: number of decision trees, maximum depth, minimum number of split samples per node, and maximum number of features.

Images