Wrapper roller service life prediction method based on single factor test

Abstract

According to the wrapper roller service life prediction method based on the single-factor test, the data set and the initial feature pool are constructed by using the existing single-factor measurement data and the wrapper roller service life; a prediction model is constructed by independently measuring single measurement values such as coating thickness, hardness, porosity, bonding strength, friction coefficient, abrasion loss, corrosion potential, current density, elastic modulus, yield strength, tensile strength and heat conductivity; in the model, obtaining an optimal feature subset to evaluate the performance of the optimal feature subset in the aspect of deriving complexity performance, and further constructing a nonlinear expression among the service life of the wrapper roller, the coating thickness, the hardness, the porosity, the bonding strength, the friction coefficient, the abrasion loss, the corrosion potential, the current density, the elastic modulus, the yield strength, the tensile strength, the heat conductivity and the like; therefore, the service life of the wrapper roller is predicted; the method can quickly predict the service life of the wrapper roller, and can effectively save the experiment cost.

Description

Technical Field

[0001] The invention relates to a method for predicting the life of a winding roller based on single factor testing. Background Art

[0002] In the continuous rolling process of the metallurgical industry, winder rollers, as critical and vulnerable components, are subjected to the multiple effects of high temperatures, alternating stresses, and abrasive wear over long periods of time. The abrasive failure of their surface materials directly impacts equipment life and production continuity. Existing research is largely limited to laboratory testing of single performance indicators (such as ASTM G65 wear loss or electrochemical corrosion rate), lacking long-term performance prediction methods under the coupled effects of complex dynamic loads, temperature gradients, and medium pH fluctuations. Furthermore, online verification of winder roller life is expensive and time-consuming, making the development of new materials for winder rollers extremely difficult. Summary of the Invention

[0003] A method for predicting the life of a winder roller based on single-factor testing is proposed. The method utilizes existing single-factor measurement data and the life of the winder roller to construct a data set and an initial feature pool. A prediction model is constructed by separately measuring the single measurement values ​​of characteristic parameters of material properties, environmental parameters, and stress conditions. In the model, the optimal feature subset is obtained to evaluate its performance in deriving complex properties, and then a nonlinear expression is constructed between the life of the winder roller and the parameter values ​​of the material properties, environmental parameters, and stress conditions, thereby realizing the prediction of the life of the winder roller.

[0004] Preferably, the method specifically includes the following steps:

[0005] S1: Based on historical information data, a dataset of single-factor measurement data and the life of the winder roller is established; the single measurement values ​​of characteristic parameters of material properties, environmental parameters, and stress conditions are used as initial characteristic input variables; and the life of the winder roller is used as the target variable;

[0006] S2: Preprocess the initial data. The specific steps of preprocessing are: (1) draw a box plot and remove data with large deviations; (2) normalize the data; (3) use Box-Cox transformation and logarithmic transformation on the feature variables and target variables respectively to reduce the skewness of the data and make the data closer to normal distribution;

[0007] S3: Using the cross-validation method, the dataset was randomly divided into training and validation sets in proportion, and repeated 5 times;

[0008] S4: Select multiple algorithms;

[0009] S5: Define model performance evaluation indicators;

[0010] S6: Input the training set obtained in S3 into the algorithm;

[0011] S7: Calculate the model performance evaluation index defined in S5, and take the comprehensive optimal model as the final model;

[0012] S8: Determine the target measurement items, set the limit range of each single factor measurement value, and construct the measurement system space with a certain step size;

[0013] S9: Taking the best feature combination adapted to the model obtained in S7 as the standard, calculate each feature input variable in the measurement system space obtained in S8 and determine the data set;

[0014] S10: Input the data set obtained in S9 into the model determined in S7 to predict the specific life span in the measurement system space;

[0015] S11: Combining the prediction results of S9 and S10, construct a nonlinear expression between the lifespan and the characteristic parameters of material properties, environmental parameters and stress conditions, and conduct experimental verification;

[0016] S12: Determine whether the error between the predicted life and the actual life is less than 10%; if so, the design is successful; if not, add the data to the initial data set of S1 and repeat S2 to S12.

[0017] Preferably, in step S1, the following requirements are made for the feature collection specification:

[0018] Feature parameter dimension list:

[0019] Material properties: coating properties including coating thickness h c , porosity φ and bonding strength σb; mechanical properties include hardness Hv, elastic modulus E, yield strength σy and tensile strength σuts; friction and wear include friction coefficient μ and wear loss W; thermophysical properties include thermal conductivity κ;

[0020] Environmental parameters: Corrosion parameters include corrosion potential E corr and corrosion current density I corr ; Working conditions include temperature T;

[0021] Stress conditions: Dynamic loads include cyclic load amplitude Δσ and frequency f.

[0022] Preferably, in step S1, different measuring equipment is selected for different characteristic parameters: 1) for coating characteristics, a coating thickness gauge is selected, using a magnetic induction method and a probe diameter of 2 mm; 2) for porosity, a mercury intrusion instrument is selected, using a pressure range of 0.1-400 MPa; 3) for bonding strength, a tensile testing machine is selected, using a loading rate of 0.5 mm / min; 4) for hardness, a microhardness tester is selected, using a load of 500 gf and a holding time of 15 s; 5) for corrosion parameters, an electrochemical workstation is selected, using a scanning rate of 0.167 mV / s; 6) for thermal conductivity, a laser flash method is selected, using a pulse energy of 15 J , temperature range 25-300℃; different measurement standards are selected for different characteristics: 1) For dynamic parameters, such as wear volume W: record once every 1000 cycles, and collect friction coefficient μ simultaneously; 2) For static parameters, such as coating thickness / hardness, etc.: take the average of 3 measurements, and sample spatially distributed; 3) For corrosion monitoring, such as potential / current density: record once every 30 seconds, and start collecting after the open circuit potential stabilizes; for the elimination of outliers in the measurement, the dynamic 3σ criterion + physical range verification method is adopted, and hardness values ​​exceeding HV100-800 are rejected. Data interpolation is performed for missing values, and the missing values ​​are processed using the KNN interpolation method.

[0023] Preferably, in step S2, a multi-stage data cleaning and transformation method is implemented:

[0024] 1) Outlier detection: Box plot threshold setting: Q3+3.0IQR / Q1-3.0IQR;

[0025] 2) Grubbs test, to assist in judgment and eliminate 7 groups of abnormal data;

[0026] 3) Data standardization: Min-Max normalization is used: Where x represents the original data, x max Represents the maximum value of this group of data, x min Represents the minimum value of this group of data, x norm Represents normalized data.

[0027] 4) Logarithmic transformation of target variable: y trans =log 10 (y+10);

[0028] 5) Distribution optimization: Box-Cox transformation parameter: λ = 0.32;

[0029] 6) Skewness improvement: The skewness of porosity data was reduced from 1.85 to 0.13.

[0030] Preferably, in step S3, the data division method is:

[0031] The training set: validation set ratio is 8:2. Stratified sampling ensures balanced material batches. The random seeds for the five repeated experiments are set as follows: 2023, 2024, 2025, 2026, and 2027.

[0032] Time series verification: Implement training on the first 80% of the time nodes and verification on the last 20% of the field data;

[0033] Stability control: Model performance index fluctuation range: RMSE ± 2.1%.

[0034] Preferably, in step S4, the model includes ridge regression, lasso regression, elasticnet regression, support vector machine regression, k-nearest neighbor regression, gradient boosting tree, maximum gradient boosting tree, random forest, linear support vector machine, polynomial support vector machine and Gaussian radial basis function support vector machine; the model parameters and optimization method are as follows:

[0035] The key parameters for ridge regression are: regularization coefficient α = 1.2, convergence tolerance tol = 1e-5, optimization logic: α is determined by observing the coefficient stability through ridge trace analysis. When α>1.0, the variance of each feature coefficient drops to the stable zone; 5-fold cross-validation is used to evaluate the RMSE change of α in the interval [0.5, 2.0]. The optimal α corresponds to the lowest point of validation error, and RMSE = 2.37 when α is 1.2.

[0036] The key parameters for XGBoost regression are: learning_rate = 0.08, max_depth = 7. The Bayesian hyperparameter optimization process is: use the Hyperopt library to perform 100 iterations to collect the expected improvement of function selection, and the kernel function is Matérn 5 / 2, which is suitable for continuous parameter space; make Bayesian optimization make the validation set R 2 From 0.91 to 0.96;

[0037] The parameters of the Gaussian kernel SVR were set as follows: C = 1.8, γ = 0.12, ε = 0.05, and a grid search strategy was used. The parameter range used was C: logarithmic uniform sampling with a step size of 0.2, and γ: a heuristic setting based on feature spacing. The cross-validation early stopping mechanism improved performance, and the MAE of the test set was reduced by 37% after the grid search.

[0038] The parameter configuration for random forest regression is: n_estimators = 200, max_features = 0.6, and the OOB error monitoring method is adopted: Dynamic adjustment process: Initially set n_estimators = 50, calculate the OOB error every time 10 trees are grown, and stop growing when the OOB error changes by < 0.5% for 3 consecutive times. Finally, select n_estimators = 200 to leave a safety margin.

[0039] Preferably, in step S5, the model performance evaluation indicators include mean square error, root mean square error, cross validation score and determination coefficient; and a composite evaluation indicator is defined:

[0040] 1) Core indicators:

[0041] Where n represents the total number of test samples, y i represents the actual life span of the i-th sample, The predicted life span of the i-th sample is, represents the actual life expectancy average;

[0042] 2) Stability index: 5-fold cross-validation standard deviation: σ < 2.5%;

[0043] 3) Engineering applicability: Prediction deviation rate: <10% of data points exceed the ±15% error band.

[0044] Preferably, in step S8, specifically, the target measurement value range is between the maximum and minimum measurement values, wherein the coating thickness h c The range of the material is 1-5 mm, the porosity φ range is 5-15%, the bonding strength σb range is 300-600 MPa; the hardness Hv range is 600-800 HV, the elastic modulus E range is 200-400 GPa, the yield strength σy range is 500-1200 MPa, the tensile strength σuts range is 800-1200 MPa; the friction coefficient μ range is 0.1-0.2, and the wear loss W range is <1×10 -5 ; Thermal conductivity κ range is 2-400W / (m·K); Corrosion potential E corr The range is -0.3-0.1V and the corrosion current density I corr Range: <1μA / cm 2 ; The operating conditions include temperature T ranging from 500-800℃; cyclic load amplitude Δσ ranging from 100-300MPa, and frequency f ranging from 10-50Hz.

[0045] Preferably, in step S11, specifically, the life of the winding roller is verified according to each single factor measurement value, and the following expression is obtained:

[0046]

[0047] where h c represents the thickness of the material matrix, φ represents the porosity of the coating, σ b represents the coating bonding strength, Hv represents the material hardness, σy represents the yield strength, E represents the elastic modulus, σuts represents the tensile strength, μ represents the friction coefficient, W represents the wear rate, I corr represents the corrosion current density, E corrrepresents corrosion potential, T represents operating temperature, Δσ represents cyclic stress amplitude, f represents load frequency, and κ represents thermal conductivity; Life represents predicted life; coefficient range: A=850±50; B=320±20; C=1.8±0.3; D=0.15±0.02.

[0048] The present invention provides a method for predicting the life of a winding roller based on single-factor testing. The method constructs a data set and an initial feature pool by utilizing existing single-factor measurement data and the life of the winding roller, and constructs a prediction model by separately measuring single measurement values ​​such as coating thickness, hardness, porosity, bonding strength, friction coefficient, wear amount, corrosion potential, current density, elastic modulus, yield strength, tensile strength, and thermal conductivity. In the model, the optimal feature subset is obtained to evaluate its performance in deriving complex properties, and then a nonlinear expression between the life of the winding roller and the coating thickness, hardness, porosity, bonding strength, friction coefficient, wear amount, corrosion potential, current density, elastic modulus, yield strength, tensile strength, and thermal conductivity is constructed, thereby realizing the prediction of the life of the winding roller. The method can quickly predict the life of the winding roller and can effectively save experimental costs. The present method The main features are reflected in: 1) Multi-scale feature fusion: integrating experimental measurement values ​​(thickness, hardness, porosity, bonding strength, friction coefficient, wear volume, corrosion potential, current density, elastic modulus, yield strength, tensile strength, thermal conductivity, etc.) to construct a feature system covering atomic-mesoscopic-macroscopic scales; 2) Dynamic coupling modeling: by strengthening parameters such as the energy term A and electronegativity difference Δχ in the model, the material degradation mechanism under the synergistic effect of friction-corrosion-impact is quantified; 3) Life prediction optimization: taking life as the target variable, combined with the time convolutional network (TCN), captures the nonlinear influence of microstructure evolution on macroscopic performance, breaking through the long-term prediction bottleneck of traditional regression models; it can provide methodological support for the development of materials with high working condition adaptability, significantly reduce experimental costs and shorten the R&D cycle, and has important engineering application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 The figure is a flow chart of the method for predicting the life of the auxiliary winding roller based on single factor testing of the present invention.

[0050] Figure 2 This is a model performance diagram of the winding roller life prediction method based on single factor test of the present invention. DETAILED DESCRIPTION

[0051] like Figures 1 to 2As shown in the figure, a method for predicting the life of a winder roller based on single-factor testing utilizes existing single-factor measurement data and the life of the winder roller to construct a data set and an initial feature pool, and constructs a prediction model by separately measuring single measurement values ​​such as coating thickness, hardness, porosity, bonding strength, friction coefficient, wear volume, corrosion potential, current density, elastic modulus, yield strength, tensile strength, and thermal conductivity. In the model, the optimal feature subset is obtained to evaluate its performance in deriving complex properties, and then a nonlinear expression is constructed between the life of the winder roller and the coating thickness, hardness, porosity, bonding strength, friction coefficient, wear volume, corrosion potential, current density, elastic modulus, yield strength, tensile strength, and thermal conductivity, thereby realizing the prediction of the life of the winder roller. The method specifically includes the following steps:

[0052] S1: Establish a data set of single-factor measurement data and winder roller life based on historical information data;

[0053] Specifically, single measurement values ​​such as coating thickness, hardness, porosity, bonding strength, friction coefficient, wear volume, corrosion potential, current density, elastic modulus, yield strength, tensile strength, and thermal conductivity were input as initial feature input variables; ultimately, 100 pieces of data for predicting the life of the winder roller were collected.

[0054] Specifically, the following requirements are made for feature collection specifications:

[0055] Feature dimension list:

[0056] Material properties: coating properties including coating thickness h c , porosity φ and bonding strength σb; mechanical properties include hardness Hv, elastic modulus E, yield strength σy and tensile strength σuts; friction and wear include friction coefficient μ and wear loss W; thermophysical properties include thermal conductivity κ;

[0057] Environmental parameters: Corrosion parameters include corrosion potential E corr and corrosion current density I corr ; Working conditions include temperature T;

[0058] Stress conditions: Dynamic loads include cyclic load amplitude Δσ and frequency f;

[0059] Different measurement equipment was selected based on the above characteristics: 1) For coating properties, a coating thickness gauge (ISO 2178) was selected, using the magnetic induction method with a probe diameter of 2 mm; 2) For porosity, a mercury intrusion porosimeter (ASTM D4404) was selected, using a pressure range of 0.1-400 MPa; 3) For bond strength, a tensile testing machine (ASTM C633) was selected, using a loading rate of 0.5 mm / min; 4) For hardness, a microhardness tester (ISO 6507) was selected, using a load of 500 gf and a hold time of 15 s; 5) For corrosion parameters, an electrochemical workstation (Gamry Reference 3000) was selected, using a scan rate of 0.167 mV / s; 6) For thermal conductivity, a laser flash method (ASTM E1461) was selected, using a pulse energy of 15 J and a temperature range of 25-300°C.

[0060] The measurement criteria for the above feature selection are as follows: 1) For dynamic parameters, such as wear volume W: record once every 1000 cycles, and simultaneously collect the friction coefficient μ; 2) For static parameters, such as coating thickness / hardness, etc.: take the average of 3 measurements, and spatially distribute sampling (5 points on the surface); 3) For corrosion monitoring, such as potential / current density: record once every 30 seconds, and start collecting after the open circuit potential stabilizes.

[0061] For the elimination of outliers in the measurement, the dynamic 3σ criterion + physical range calibration method is adopted. Hardness values ​​exceeding HV100-800 are rejected, and data interpolation is performed for missing values. The missing values ​​are processed using the KNN interpolation method (k=5).

[0062] S2: Preprocess the initial data;

[0063] Specifically, the preprocessing steps are as follows: (1) draw a box plot and remove data with large deviations; (2) normalize the data; (3) use Box-Cox transformation and logarithmic transformation on the feature variables and target variables respectively to reduce the skewness of the data and make the data closer to normal distribution.

[0064] The method for implementing multi-stage data cleaning and transformation is:

[0065] 1) Outlier detection: Box plot threshold setting: Q3+3.0IQR / Q1-3.0IQR;

[0066] 2) Grubbs test (α = 0.01) was used to assist in the judgment and eliminate 7 groups of abnormal data;

[0067] 3) Data standardization: Min-Max normalization is used: Where x represents the original data, x max Represents the maximum value of this group of data, x min Represents the minimum value of this group of data, x normRepresents normalized data.

[0068] 4) Logarithmic transformation of target variable: y trans =log 10 (y+10);

[0069] 5) Distribution optimization: Box-Cox transformation parameter: λ = 0.32 (determined by maximum likelihood estimation);

[0070] 6) Skewness improvement: The skewness of the porosity data was reduced from 1.85 to 0.13 (Shapiro-Wilk test p>0.05).

[0071] S3: Using the cross-validation method, the data set is randomly divided into training set and validation set in proportion, and repeated 5 times; the data division method is:

[0072] The training set: validation set ratio was 8:2. Stratified sampling was used to ensure balanced material batches. The random seeds for the five repeated experiments were set to 2023, 2024, 2025, 2026, and 2027.

[0073] Time series verification: Implement training on the first 80% of the time nodes and verification on the last 20% of the field data;

[0074] Stability control: Model performance index fluctuation range: RMSE ± 2.1% (95% confidence interval).

[0075] S4: Select multiple algorithms;

[0076] Specifically, the models include ridge regression, lasso regression, elasticnet regression, support vector machine regression, k-nearest neighbor regression, gradient boosted tree, maximum gradient boosted tree, random forest, linear support vector machine, polynomial support vector machine and Gaussian radial basis function support vector machine.

[0077] The parameters and optimization methods are as follows:

[0078] The key parameters for ridge regression are: regularization coefficient α = 1.2, convergence tolerance tol = 1e-5, and optimization logic: α is determined by observing coefficient stability through ridge trace analysis. When α > 1.0, the variance of each feature coefficient drops to a stable region. 5-fold cross-validation is used to evaluate the RMSE change of α in the interval [0.5, 2.0]. The optimal α corresponds to the lowest point of validation error (RMSE = 2.37 when 1.2).

[0079] The key parameters for XGBoost regression are: learning_rate = 0.08, max_depth = 7. The Bayesian hyperparameter optimization process is: use the Hyperopt library to perform 100 iterations of the acquisition function to select the expected improvement (EI) kernel function as Matérn 5 / 2 (suitable for continuous parameter space); make Bayesian optimization make the validation set R 2 From 0.91 to 0.96;

[0080] The parameters of Gaussian kernel SVR are set as follows: C = 1.8, γ = 0.12, ε = 0.05, and a grid search strategy is used. The parameter range is C: logarithmic uniform sampling (10~10 2 , step size 0.2), γ: heuristic setting based on feature spacing; cross-validation early stopping mechanism improves performance, and the MAE of the test set is reduced by 37% after grid search (from 3.42 to 2.15);

[0081] Parameter configuration for random forest regression: n_estimators = 200, max_features = 0.6, using the OOB error monitoring method: Dynamic adjustment process: Initially set n_estimators = 50, calculate the OOB error every 10 trees grown, and stop growing when the OOB error changes by < 0.5% for three consecutive times (actually stop at 195 trees). Finally, select n_estimators = 200 to leave a safety margin.

[0082] S5: Define model performance evaluation indicators;

[0083] Specifically, the model performance evaluation indicators include mean square error, root mean square error, cross-validation score and determination coefficient.

[0084] Define composite evaluation metrics:

[0085] 1) Core indicators:

[0086] Where n represents the total number of test samples, y i represents the actual life span of the i-th sample, The predicted life span of the i-th sample is, represents the actual life expectancy average;

[0087] 2) Stability index: 5-fold cross-validation standard deviation: σ < 2.5%;

[0088] 3) Engineering applicability: Prediction deviation rate: <10% of data points exceed the ±15% error band.

[0089] S6: Input the training set obtained in S3 into the algorithm;

[0090] S7: Calculate the model performance evaluation index defined in S5, and take the best comprehensive model as the final model; the final result shows that the gradient boosting tree is the best model, R 2 Reaching 0.98. Through the correlation analysis between input parameters and error sources, the model has achieved a prediction error of ≤8.5% within the 95% confidence interval;

[0091] S8: Determine the target measurement items, set the limit range of each single factor measurement value, and construct the measurement system space with a certain step size;

[0092] Specifically, the target measurement value range is between the maximum and minimum measurement values; wherein the coating thickness h c The range of the material is 1-5 mm, the porosity φ range is 5-15%, the bonding strength σb range is 300-600 MPa; the hardness Hv range is 600-800 HV, the elastic modulus E range is 200-400 GPa, the yield strength σy range is 500-1200 MPa, the tensile strength σuts range is 800-1200 MPa; the friction coefficient μ range is 0.1-0.2, and the wear loss W range is <1×10 -5 ; Thermal conductivity κ range is 2-400W / (m·K); Corrosion potential E corr The range is: 0.3-0.1V and the corrosion current density Icorr range is <1μA / cm 2 ; The operating conditions include temperature T ranging from 500-800℃; cyclic load amplitude Δσ ranging from 100-300MPa, and frequency f ranging from 10-50Hz.

[0093] S9: Taking the best feature combination adapted to the model obtained in S7 as the standard, calculate each feature input variable in the measurement system space obtained in S8 and determine the data set;

[0094] S10: Input the data set obtained in S9 into the model determined in S7 to predict the specific life span in the measurement system space;

[0095] S11: Combining the prediction results of S9 and S10, construct nonlinear expressions between life and coating thickness, hardness, porosity, bonding strength, friction coefficient, wear volume, corrosion potential, current density, elastic modulus, yield strength, tensile strength, thermal conductivity, etc., and conduct experimental verification;

[0096] Specifically, the life of the winding roller is verified according to the measured values ​​of each single factor. The following expression is obtained:

[0097]

[0098] where h c represents the thickness of the material matrix, φ represents the porosity of the coating, σb represents the coating bonding strength, Hv represents the material hardness, σy represents the yield strength, E represents the elastic modulus, σuts represents the tensile strength, μ represents the friction coefficient, W represents the wear rate, I corr represents the corrosion current density, E corr represents corrosion potential, T represents operating temperature, Δσ represents cyclic stress amplitude, f represents load frequency, κ represents thermal conductivity; Life represents predicted life.

[0099] Specifically, the coefficient ranges obtained by fitting 100 sets of test data are: A = 850 ± 50 (coating dominant term amplification coefficient); B = 320 ± 20 (mechanical property benchmark value); C = 1.8 ± 0.3 (wear-corrosion coupling coefficient); D = 0.15 ± 0.02 (corrosion sensitivity parameter).

[0100] S12: Determine whether the error between the predicted life and the actual life is less than 10%; if so, the design is successful; if not, add the data to the initial data set of S1 and repeat S2 to S12.

[0101] The present invention provides a method for predicting the life of a winding roller based on single-factor testing. The method constructs a data set and an initial feature pool by utilizing existing single-factor measurement data and the life of the winding roller, and constructs a prediction model by separately measuring single measurement values ​​such as coating thickness, hardness, porosity, bonding strength, friction coefficient, wear amount, corrosion potential, current density, elastic modulus, yield strength, tensile strength, and thermal conductivity. In the model, the optimal feature subset is obtained to evaluate its performance in deriving complex properties, and then a nonlinear expression between the life of the winding roller and the coating thickness, hardness, porosity, bonding strength, friction coefficient, wear amount, corrosion potential, current density, elastic modulus, yield strength, tensile strength, and thermal conductivity is constructed, thereby realizing the prediction of the life of the winding roller. The method can quickly predict the life of the winding roller and can effectively save experimental costs. The present method The main features are reflected in: 1) Multi-scale feature fusion: integrating experimental measurement values ​​(thickness, hardness, porosity, bonding strength, friction coefficient, wear volume, corrosion potential, current density, elastic modulus, yield strength, tensile strength, thermal conductivity, etc.) to construct a feature system covering atomic-mesoscopic-macroscopic scales; 2) Dynamic coupling modeling: by strengthening parameters such as the energy term A and electronegativity difference Δχ in the model, the material degradation mechanism under the synergistic effect of friction-corrosion-impact is quantified; 3) Life prediction optimization: taking life as the target variable, combined with the time convolutional network (TCN), captures the nonlinear influence of microstructure evolution on macroscopic performance, breaking through the long-term prediction bottleneck of traditional regression models; it can provide methodological support for the development of materials with high working condition adaptability, significantly reduce experimental costs and shorten the R&D cycle, and has important engineering application value.

[0102] The verification results are as follows:

[0103] The first group of coatings has a thickness of hc = 500 μm, a porosity of φ = 4.2%, a bonding strength σb = 85 MPa, a hardness of Hv = 520 HV, and a corrosion current density of I corr =3.8μA / cm 2 , temperature T = 65 ° C, cyclic load Δσ = 120 MPa. The predicted lifespan was 7.13 (months), the measured lifespan was 6.71 (months), with an error of 6.1%;

[0104] The second group of coating thickness hc = 300 μm, porosity φ = 5.8%, bonding strength σb = 72 MPa, hardness Hv = 380 HV, corrosion current density I corr =5.1μA / cm 2 , temperature T = 85 ° C, cyclic load Δσ = 150 MPa. The predicted lifespan is 5.61 (months), the measured lifespan is 5.23 (months), and the error is 7.3%;

[0105] The third group of coating thickness hc = 800 μm, porosity φ = 3.5%, bonding strength σb = 95 MPa, hardness Hv = 550 HV, corrosion current density I corr =2.3μA / cm 2 , temperature T = 45 ° C, cyclic load Δσ = 90 MPa. The predicted lifespan is 9.33 (months), while the measured lifespan is 8.81 (months), with an error of 5.6%.

[0106] Specifically, various models have been implemented. Figure 2 The performance evaluation shown here shows the prediction error, except for the K-nearest neighbor regression R 2 is 0.24, and R 2 All are above 0.96, reaching the expected prediction life error; the highest accuracy is the gradient boosting tree R 2 Reaching 0.98; through the correlation analysis between input parameters and error sources, the model has achieved a prediction error of ≤8.5% within a 95% confidence interval, meeting the requirements of engineering applications.

[0107] Finally, it should be noted that the above embodiments only illustrate the technical solutions of the present invention and do not limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments or replace some of the technical features therein with equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A method for predicting the life of a winding roller based on single factor testing, characterized in that: The data set and initial feature pool are constructed using the existing single-factor measurement data and the life of the winder roller. The prediction model is constructed by separately measuring the single measurement values ​​of the characteristic parameters of material properties, environmental parameters and stress conditions. In the model, the optimal feature subset is obtained to evaluate its performance in deriving complex properties, and then a nonlinear expression is constructed between the life of the winder roller and the parameter values ​​of material properties, environmental parameters and stress conditions, thereby realizing the prediction of the life of the winder roller.

2. The method for predicting the life of a winding roller based on single factor testing according to claim 1 is characterized in that: The specific steps include: S1: Establish a data set of single-factor measurement data and winder roller life based on historical information data; use single measurement values ​​of characteristic parameters of material properties, environmental parameters and stress conditions as initial characteristic input variables; The life of the winding roller is used as the target variable; S2: Preprocess the initial data. The specific steps of preprocessing are: (1) draw a box plot and remove data with large deviations; (2) normalize the data; (3) use Box-Cox transformation and logarithmic transformation on the feature variables and target variables respectively to reduce the skewness of the data and make the data closer to normal distribution; S3: Using the cross-validation method, the dataset was randomly divided into training and validation sets in proportion, and repeated 5 times; S4: Select multiple algorithms; S5: Define model performance evaluation metrics; S6: Input the training set obtained in S3 into the algorithm; S7: Calculate the model performance evaluation index defined in S5, and take the comprehensive optimal model as the final model; S8: Determine the target measurement items, set the limit range of each single factor measurement value, and construct the measurement system space with a certain step size; S9: Taking the best feature combination adapted to the model obtained in S7 as the standard, calculate each feature input variable in the measurement system space obtained in S8 and determine the data set; S10: Input the data set obtained in S9 into the model determined in S7 to predict the specific life span in the measurement system space; S11: Combining the prediction results of S9 and S10, construct nonlinear expressions between life and characteristic parameters of material properties, environmental parameters and stress conditions, and conduct experimental verification; S12: Determine whether the error between the predicted life and the actual life is less than 10%; if so, the design is successful; if not, add the data to the initial data set of S1 and repeat S2 to S12.

3. The method for predicting the life of a winding roller based on single factor testing according to claim 2 is characterized in that: In step S1, the following requirements are made for feature collection specifications: Feature parameter dimension list: Material properties: coating properties including coating thickness h c , porosity φ and bonding strength σb; mechanical properties include hardness Hv, elastic modulus E, yield strength σy and tensile strength σuts; friction and wear include friction coefficient μ and wear loss W; thermophysical properties include thermal conductivity κ; Environmental parameters: Corrosion parameters include corrosion potential E corr and corrosion current density I corr ; Working conditions include temperature T; Stress conditions: Dynamic loads include cyclic load amplitude Δσ and frequency f.

4. The method for predicting the life of a winding roller based on single factor testing according to claim 3 is characterized in that: In step S1, different measuring equipment is selected for different characteristic parameters: 1) for coating characteristics, a coating thickness gauge is selected, using the magnetic induction method and a probe diameter of 2 mm; 2) for porosity, a mercury intrusion instrument is selected, using a pressure range of 0.1-400 MPa; 3) for bonding strength, a tensile testing machine is selected, using a loading rate of 0.5 mm / min; 4) for hardness, a microhardness tester is selected, using a load of 500 gf and a holding time of 15 s; 5) for corrosion parameters, an electrochemical workstation is selected, using a scanning rate of 0.167 mV / s; 6) for thermal conductivity, a laser flash method is selected, using a pulse energy of 15 J and a temperature of 0. The temperature range is 25-300°C; different measurement standards are selected for different characteristics: 1) For dynamic parameters, such as wear volume W: record once every 1000 cycles, and simultaneously collect the friction coefficient μ; 2) For static parameters, such as coating thickness / hardness, etc.: take the average of three measurements, and sample spatially distributed; 3) For corrosion monitoring, such as potential / current density: record once every 30 seconds, and start collecting after the open circuit potential stabilizes; for the removal of outliers in the measurement, the dynamic 3σ criterion + physical range verification method is adopted. Hardness values ​​exceeding HV100-800 are rejected, and data interpolation is performed for missing values, and the missing values ​​are processed using the KNN interpolation method.

5. The method for predicting the life of a winding roller based on single factor testing according to claim 2, characterized in that: In step S2, a multi-stage data cleaning and transformation method is implemented: 1) Outlier detection: Box plot threshold setting: Q3+3.0IQR / Q1-3.0IQR; 2) Grubbs test, to assist in judgment and eliminate 7 groups of abnormal data; 3) Data standardization: Min-Max normalization is used: Where x represents the original data, x max Represents the maximum value of this group of data, x min Represents the minimum value of this group of data, x norm Represents normalized data. 4) Logarithmic transformation of target variable: y trans =log 10 (y+10); 5) Distribution optimization: Box-Cox transformation parameter: λ = 0.32; 6) Skewness improvement: The skewness of porosity data was reduced from 1.85 to 0.

13.

6. The method for predicting the life of a winding roller based on single factor testing according to claim 2, characterized in that: In step S3, the data is divided into: The training set: validation set ratio is 8:

2. Stratified sampling ensures balanced material batches. The random seeds for the five repeated experiments are set as follows: 2023, 2024, 2025, 2026, and 2027. Time series verification: Implement training on the first 80% of the time nodes and verification on the last 20% of the field data; Stability control: Model performance index fluctuation range: RMSE ± 2.1%.

7. The method for predicting the life of a winding roller based on single factor testing according to claim 2, characterized in that: In step S4, the models include ridge regression, lasso regression, elasticnet regression, support vector machine regression, k-nearest neighbor regression, gradient boosted tree, maximum gradient boosted tree, random forest, linear support vector machine, polynomial support vector machine, and Gaussian radial basis function support vector machine; the model parameters and optimization methods are as follows: The key parameters for ridge regression are: regularization coefficient α = 1.2, convergence tolerance tol = 1e-5, optimization logic: α is determined by observing the coefficient stability through ridge trace analysis. When α>1.0, the variance of each feature coefficient drops to the stable zone; 5-fold cross-validation is used to evaluate the RMSE change of α in the interval [0.5, 2.0]. The optimal α corresponds to the lowest point of validation error, and RMSE = 2.37 when α is 1.

2. The key parameters for XGBoost regression are: learning_rate = 0.08, max_depth = 7. The Bayesian hyperparameter optimization process is: use the Hyperopt library to perform 100 iterations to collect the expected improvement of function selection, and the kernel function is Matérn 5 / 2, which is suitable for continuous parameter space; make Bayesian optimization make the validation set R 2 From 0.91 to 0.96; The parameters of the Gaussian kernel SVR were set as follows: C = 1.8, γ = 0.12, ε = 0.05, and a grid search strategy was used. The parameter range used was C: log-uniform sampling with a step size of 0.2, and γ: a heuristic setting based on feature spacing. The cross-validation early stopping mechanism improved performance, and the MAE of the test set was reduced by 37% after the grid search. The parameter configuration for random forest regression is: n_estimators = 200, max_features = 0.6, and the OOB error monitoring method is adopted: Dynamic adjustment process: Initially set n_estimators = 50, calculate the OOB error every time 10 trees are grown, and stop growing when the OOB error changes by < 0.5% for 3 consecutive times. Finally, select n_estimators = 200 to leave a safety margin.

8. The method for predicting the life of a winding roller based on single factor testing according to claim 2, characterized in that: In step S5, the model performance evaluation indicators include mean square error, root mean square error, cross-validation score and determination coefficient; the composite evaluation indicators are defined as follows: 1) Core indicators: Where n represents the total number of test samples, y i represents the actual life span of the i-th sample, The predicted life span of the i-th sample is, represents the actual life expectancy average; 2) Stability index: 5-fold cross-validation standard deviation: σ < 2.5%; 3) Engineering applicability: Prediction deviation rate: <10% of data points exceed the ±15% error band.

9. The method for predicting the life of a winding roller based on single factor testing according to claim 2, characterized in that: In step S8, specifically, the target measurement value range is between the maximum and minimum measurement values, where the coating thickness h c The range of the material is 1-5 mm, the porosity φ range is 5-15%, the bonding strength σb range is 300-600 MPa; the hardness Hv range is 600-800 HV, the elastic modulus E range is 200-400 GPa, the yield strength σy range is 500-1200 MPa, the tensile strength σuts range is 800-1200 MPa; the friction coefficient μ range is 0.1-0.2, and the wear loss W range is <1×10 -5 ; Thermal conductivity κ range is 2-400W / (m·K); Corrosion potential E corr The range is -0.3-0.1V and the corrosion current density I_corr range is <1μA / cm 2 The operating conditions include a temperature T range of 500-800°C, a cyclic load amplitude Δσ range of 100-300 MPa, and a frequency f range of 10-50 Hz.

10. The method for predicting the life of a winding roller based on single factor testing according to claim 2, characterized in that: In step S11, specifically, the life of the winding roller is verified according to the measured values ​​of each single factor, and the following expression is obtained: where h c represents the thickness of the material matrix, φ represents the porosity of the coating, σ b represents the coating bonding strength, Hv represents the material hardness, σy represents the yield strength, E represents the elastic modulus, σuts represents the tensile strength, μ represents the friction coefficient, W represents the wear rate, I corr represents the corrosion current density, E corr represents corrosion potential, T represents operating temperature, Δσ represents cyclic stress amplitude, f represents load frequency, and κ represents thermal conductivity; Life represents predicted life; coefficient range: A=850±50; B=320±20; C=1.8±0.3; D=0.15±0.02.

Images