Milling surface integrity prediction method

Abstract

The application discloses a kind of milling surface integrity prediction methods, comprising: S1.milling test data set is collected;S2.milling test data set is preprocessed, and the test data set after processing is obtained;S3.the test data set after processing is input into milling surface integrity prediction model and model training is carried out, and the trained prediction model is obtained;S4.milling data to be measured is input into the trained prediction model, and the prediction result of milling surface integrity is output.The application can significantly improve the precision of milling surface integrity prediction, and help to improve the efficiency and quality of milling process.

Description

Technical Field

[0001] This invention relates to the field of milling, and more specifically to a method for predicting the integrity of milled surfaces. Background Technology

[0002] Surface integrity is a technical indicator used to describe, evaluate, and control the impact of machining processes on surface performance and the various changes that may occur within the machined surface layer. With the development of the manufacturing industry and the increasing demands for surface integrity of parts, the need to predict and control milling surface integrity is becoming increasingly urgent. However, due to the complexity and variability of the milling process, it is difficult to accurately predict milling surface integrity using only experience and experimental methods; therefore, computer simulation technology is needed for prediction and analysis.

[0003] Currently, methods or products for predicting the integrity of milled surfaces based on computer simulation technology often use existing model frameworks directly, resulting in overly simplistic construction processes. This makes it difficult to effectively identify the features of the milled surface, leading to suboptimal prediction results and an inability to effectively predict the integrity of the milled surface. Therefore, a method for predicting the integrity of milled surfaces is needed to address these issues. Summary of the Invention

[0004] In view of this, the purpose of this invention is to overcome the deficiencies in the prior art and provide a method for predicting the integrity of milled surfaces, which can significantly improve the accuracy of predicting the integrity of milled surfaces and help improve the efficiency and quality of the milling process.

[0005] The milled surface integrity prediction method of the present invention includes the following steps:

[0006] S1. Collect milling machining test dataset;

[0007] S2. Preprocess the milling test dataset to obtain the processed test dataset;

[0008] S3. Input the processed test dataset into the milling surface integrity prediction model to train the model and obtain the trained prediction model;

[0009] S4. Input the milling data to be tested into the trained prediction model and output the prediction results of the milling surface integrity.

[0010] Furthermore, the milling test dataset includes feature variable data and workpiece surface data after machining;

[0011] The feature variable data includes machining parameters, tool posture parameters, and tool geometry parameters;

[0012] The surface data of the processed workpiece includes surface roughness and surface residual stress; the surface residual stress includes surface residual stress in the feed direction and surface residual stress in the transverse direction.

[0013] Furthermore, the machining parameters include spindle speed, feed rate, and depth of cut;

[0014] The tool attitude parameters include the forward tilt angle and the tilt angle;

[0015] The tool geometry parameters include tool diameter, number of cutting edges, and helix angle.

[0016] Furthermore, the preprocessing includes:

[0017] The milling test dataset is dimensionless to obtain the dimensionless test dataset.

[0018] The dimensionless experimental dataset is divided into a training set and a test set.

[0019] Furthermore, the processed test dataset is input into the milled surface integrity prediction model for model training, resulting in a trained prediction model, which specifically includes:

[0020] Using the processed experimental dataset, k regression models were selected, and each of the k selected regression models was optimized to obtain k optimized regression models.

[0021] Different combinations of k optimized regression models are obtained to form base learners of different combinations. The same meta-learner is used to perform stacking model fusion on the base learners of different combinations. The optimal base learner combination is selected from the base learner combinations of several model fusions.

[0022] The optimal combination of base learners is used as the first layer learner. Based on the first layer learner, different regression models are selected as meta learners for model fusion comparison. The regression model with the greatest improvement in prediction accuracy in model fusion is selected as the meta learner for model fusion.

[0023] Furthermore, k regression models are selected, specifically including:

[0024] The training set is used to train m regression models respectively. Each regression model is trained 3 times with surface roughness, surface residual stress in the feed direction and surface residual stress in the transverse direction as label values; where m is greater than k.

[0025] After training, the trained model is used to make predictions on the test set.

[0026] Each regression model is scored using evaluation metrics, and the regression models whose average scores meet the set conditions are selected.

[0027] Furthermore, the selected regression model is optimized, specifically including:

[0028] The model iterates through all combinations of hyperparameters in the regression model and performs cross-validation on each set of hyperparameters to obtain the performance index of each set of hyperparameters. Finally, the set of hyperparameters with the best performance is selected as the final model parameters.

[0029] Furthermore, the milling surface integrity prediction model includes a surface roughness prediction model, a feed direction surface residual stress prediction model, and a transverse surface residual stress prediction model.

[0030] Furthermore, the optimal combination of base learners in the surface roughness prediction model includes Bagging regression model, Random Forest regression model, ExtraTree regression model, Gradient Boosting regression model, XGBoost regression model, and CatBoost regression model.

[0031] The optimal combination of base learners in the feed direction surface residual stress prediction model includes Bagging regression model, random forest regression model, ExtraTree regression model, XGBoost regression model and CatBoost regression model.

[0032] The optimal combination of base learners in the transverse surface residual stress prediction model includes the ExtraTree regression model, the XGBoost regression model, and the CatBoost regression model.

[0033] Furthermore, in the surface roughness prediction model, the gradient boosting regression model is used as the meta-learner during model fusion;

[0034] In the surface residual stress prediction model in the feed direction, the random forest regression model is used as the meta-learner when model fusion is performed.

[0035] In the prediction model of residual stress on transverse surfaces, the gradient boosting regression model is used as the meta-learner during model fusion.

[0036] The beneficial effects of this invention are as follows: The milling surface integrity prediction method disclosed in this invention preprocesses the collected milling test dataset, making the resulting dataset a strong support and guarantee for subsequent model training and testing; by constructing a milling surface integrity prediction model based on machine learning algorithms and combined with the influencing factors of the milling process, the prediction method of this invention is beneficial to optimizing milling process parameters, helping to improve the efficiency and quality of the milling process, and providing strong technical support for the development of the manufacturing industry. Attached Figure Description

[0037] The present invention will be further described below with reference to the accompanying drawings and embodiments:

[0038] Figure 1 This is a schematic diagram of the prediction method of the present invention;

[0039] Figure 2 This is a flowchart illustrating the construction process of the Stacking integration model of the present invention.

[0040] Figure 3 A flowchart illustrating the construction of the surface integrity prediction model for this invention. Detailed Implementation

[0041] The present invention will be further described below with reference to the accompanying drawings, as shown in the figures:

[0042] The milled surface integrity prediction method of the present invention includes the following steps:

[0043] S1. Collect milling machining test dataset;

[0044] S2. The milling test dataset is preprocessed to obtain the processed test dataset; the resulting dataset provides strong support and guarantee for subsequent model training and testing.

[0045] S3. Input the processed test dataset into the milling surface integrity prediction model to train the model and obtain the trained prediction model;

[0046] S4. Input the milling data to be tested into the trained prediction model and output the prediction results of the milling surface integrity.

[0047] This invention constructs a surface integrity prediction method based on Stacking ensemble learning, providing certain theoretical and practical references for surface integrity prediction, and offering some guidance and inspiration for research in related fields. It can be used for high-precision prediction of surface integrity in five-axis milling of aluminum alloys.

[0048] In this embodiment, constructing an accurate surface integrity prediction model requires the support of a large amount of experimental data, but milling the parts is a time-consuming task. Therefore, it is crucial to obtain a dataset that can meet the needs of model training and testing within a limited time and under limited experimental conditions.

[0049] In step S1, several factors that significantly affect the surface integrity of five-axis milling are analyzed. The sensitive parameters of the selected influencing factors mainly include the following aspects:

[0050] (1) Machining parameters: Five-axis milling requires the selection of appropriate cutting parameters such as cutting speed, feed rate, and depth of cut. Different cutting parameters have different effects on surface integrity. If the cutting speed is too high or the feed rate is too large, it will lead to an increase in the heat-affected zone on the surface, resulting in defects such as cracks or burns on the surface. If the depth of cut is too large, it will lead to defects such as tearing and extrusion on the machined surface.

[0051] (2) Milling cutters: Five-axis milling requires the selection of appropriate milling cutters. Different cutter materials, shapes and sizes have different effects on the integrity of the machined surface. Excessive hardness of the cutter material will lead to increased surface roughness, unsuitable cutter shape will lead to defects such as inclusions on the surface, and excessively large cutter size will lead to bulges on the machined surface.

[0052] (3) Tool Posture: In five-axis milling, when using a ball end mill, the speed of each point on the milling edge is proportional to the perpendicular distance from that point to the axis. However, the speed of the end point of the milling edge is almost zero, meaning that this point and its surrounding area hardly participate in milling, but the milling force is large, resulting in severe tool wear, which affects the surface quality of the machined workpiece. Therefore, to avoid the end area of ​​the milling edge participating in milling, it is necessary to adjust and control the tool posture. By controlling the tool posture, five-axis CNC machining can effectively reduce the milling force. The concept of tool posture originates from the fact that five-axis machining has two additional rotational degrees of freedom, resulting in higher machining efficiency, better surface quality, and easier avoidance of interference between the tool and the machined surface.

[0053] A combination of full factorial and orthogonal experimental design was used to determine different combinations of process parameters for processing analysis.

[0054] First, the influence of two types of parameters on the surface integrity of the workpiece after milling is studied: tool attitude parameters, including rake angle L and tilt angle T, and machining parameters, including depth of cut ap, feed rate f, and spindle speed n. With five input parameters, if each parameter has five levels, the full factorial design would require 55 trials, which is very time-consuming and expensive. A 5-factor, 4-level orthogonal design would only have 16 trials, insufficient for training and testing. To reasonably increase the number of trials and collect sufficient experimental data for constructing the surface integrity prediction model, a combination of full factorial and orthogonal experimental design is adopted.

[0055] Next, the influence of the geometric parameters of the milling cutter on the surface integrity of the workpiece after milling was studied. Different milling cutters were designed for milling, and the results were analyzed through combinations of different process parameters and different milling cutters.

[0056] After completing the milling analysis, it is also necessary to measure the surface roughness and residual stress of the machined workpiece to form a milling test dataset.

[0057] Therefore, the milling test dataset includes feature variable data and workpiece surface data after machining;

[0058] The feature variable data includes machining parameters, tool posture parameters, and tool geometry parameters;

[0059] The machining parameters include spindle speed, feed rate, and depth of cut;

[0060] The tool attitude parameters include the forward tilt angle and the tilt angle;

[0061] The tool geometry parameters include tool diameter, number of cutting edges, and helix angle.

[0062] The surface data of the processed workpiece includes surface roughness and surface residual stress; the surface residual stress includes surface residual stress in the feed direction and surface residual stress in the transverse direction.

[0063] Milling analysis was performed on 7075 aluminum alloy. After machining, the surface roughness Ra of the workpiece was measured using an FTS Intra roughness profiler, and the residual stress Rsf in the feed direction and Rst in the transverse direction on the workpiece surface were measured using a μ-X360s residual stress analyzer. During residual stress measurement, the scanning angle between the analyzer and the workpiece surface was set to 25°. A positive residual stress value indicates that the residual stress on the workpiece surface is under tensile stress, while a negative residual stress value indicates that the residual stress on the workpiece surface is under compressive stress.

[0064] To minimize measurement error, each set of experimental analyses requires selecting five evenly distributed points in the processed groove and marking them with a marker. Then, the surface roughness and residual stress in two directions of these five marked points are measured, and the average value of the five measurement results is taken as the final measurement value.

[0065] The four sets of experiments yielded a total of 177 sets of experimental data. Each set of data consisted of eight characteristic variable parameters before machining (including: spindle speed n, feed rate f, depth of cut ap, rake angle L, tilt angle T, tool diameter, number of cutting edges, and helix angle) and three target value parameters after machining (including: surface roughness Ra, surface residual stress Rsf in the feed direction, and surface residual stress Rst in the transverse direction).

[0066] In this embodiment, in step S2, after organizing the experimental data, preprocessing is required for subsequent model training. Since the feature variables selected in this invention have different numerical ranges and units, this may cause weight bias and affect the performance of the prediction model. Therefore, it is necessary to dimensionlessly transform the experimental data. Dimensionless transformation refers to transforming the data so that it has no units or scale, and no fixed range or distribution. In machine learning and data analysis, dimensionless transformation is often an important preprocessing step to ensure that the data has the same scale and the same weights for different features. There are various methods for dimensionless transformation; this invention uses data normalization, and the formula for data normalization is as follows:

[0067]

[0068] Where x represents the original feature variable data, x * These are the normalized data, where max(x) represents the maximum value of the feature variable and min(x) represents the minimum value of the feature variable.

[0069] Furthermore, to ensure the prediction accuracy of the surface integrity prediction model and avoid overfitting, the experimental dataset needs to be divided into a training set and a test set. The dataset in this invention is a compilation of data from four sets of experiments; therefore, the division of the dataset will have a significant impact on the model's training and final performance, making a reasonable dataset division method crucial.

[0070] To ensure that the data from the four experimental sets are randomly but uniformly distributed across the training and test sets, this invention uses the 'train_test_split' function from the Scikit-learn library to partition the dataset. The main function of this function is to randomly split the dataset and return the resulting training and test sets. The 'test_size' parameter in 'train_test_split' adjusts the size of the test set. This invention divides the dataset into training and test sets in a 7:3 ratio, therefore 'test_size' is set to 0.3. Adjusting the 'random_state' parameter in 'train_test_split' controls the random partitioning process; setting the same random seed ensures consistent partitioning results. After multiple tunings of the 'random_state' parameter, it was found that when 'random_state' is set to 273, the four experimental sets can be evenly distributed between the training and test sets, which helps improve the model's training performance and robustness. Ultimately, the 177 sets of experimental data were divided into 123 training sets and 54 test sets. The preprocessed data will be used for subsequent model training and model testing analysis.

[0071] In this embodiment, in step S3, an accurate milling surface integrity prediction model is a prerequisite for optimizing milling process parameters. In order to improve the prediction accuracy of the model as much as possible, this invention uses the Stacking model fusion method to construct an integrated model for predicting surface integrity based on the machine learning regression model.

[0072] The basic process for building a Stacking integration model is as follows:

[0073] (1) Prepare the dataset: The dataset is divided into a training set and a test set. The training set is used only to train the base model, and the test set is used to evaluate the predictive performance of the base model and the ensemble model.

[0074] (2) Constructing the base model: Select multiple different base models and train them using the training set. This invention first trains common machine learning regression models using the training set; then uses the test set to evaluate the predictive performance of the obtained models, and selects the model with higher prediction accuracy as the base model for surface integrity prediction; finally, the selected base model is optimized to improve its prediction accuracy.

[0075] (3) Forming a new training set: Use the training set to make predictions for each base model and use the prediction results as new features. The dataset composed of the new features will be used for training the meta-model.

[0076] (4) Constructing an ensemble model: Use the training set obtained in the previous step to train the meta-model. The resulting meta-model is the final ensemble prediction model.

[0077] Stacking ensemble learning can use different combinations of base learners and different meta learners to build ensemble models. This invention ablates different combinations of base learners and different meta learners to determine the base learners and meta learners used to build the ensemble model.

[0078] (5) Model evaluation: Use evaluation metrics on the test set to evaluate the predictive performance of the ensemble model.

[0079] The surface integrity prediction model consists of a surface roughness prediction model, a feed direction residual stress prediction model, and a transverse residual stress prediction model. These three models are independent of each other, but their model construction processes are completely identical. The overall construction process of the prediction model is as follows: Figure 2 As shown.

[0080] Building a base prediction model is one of the key steps in stacking model fusion, so it is necessary to select a suitable algorithm to ensure that the base model also has sufficient prediction accuracy.

[0081] Surface integrity prediction is essentially the prediction of one or more indicators that constitute surface integrity, such as surface roughness prediction and surface residual stress prediction. The prediction of a single indicator can be regarded as a regression problem, that is, finding a function or model that can describe the relationship between independent variables and dependent variables.

[0082] Currently, many models are available for solving regression problems, but only a small fraction of them have been validated for surface integrity prediction. Therefore, this invention uses a dataset to analyze the surface integrity prediction of commonly used regression models in machine learning, selecting a total of 18 models.

[0083] The 18 models can be roughly divided into the following categories:

[0084] (1) Models based on linear regression: including linear regression, ridge regression, Lasso regression, elastic network regression, and adaptive regression.

[0085] (2) Models based on decision tree and Bagging algorithms: including decision tree regression, ExtraTree regression, random forest regression, and Bagging regression.

[0086] (3) Support vector machine-based models: including support vector regression and linear support vector regression.

[0087] (4) Neural network-based models: including multilayer perceptron regression.

[0088] (5) Models based on the nearest neighbor algorithm: including KNN regression.

[0089] (6) Models based on gradient boosting algorithm: including gradient boosting regression, HistGradientBoosting regression, AdaBoost regression, XGBoost regression, and CatBoost regression.

[0090] The analysis includes the following steps:

[0091] (1) Eighteen models were trained using the training set. Each model was trained three times with surface roughness Ra, feed direction surface residual stress Rsf, and transverse surface residual stress Rst as label values. This was used for preliminary analysis of the models. All models used default parameters to ensure the consistency of the analysis.

[0092] (2) After training, the trained model was used to predict the test set. A total of 18×3 models were used to obtain 54 sets of prediction results.

[0093] (3) Each model was scored using evaluation metrics. Considering the large range of surface roughness and residual stress values, RMSE or MAE as evaluation metrics could not intuitively reflect the model's predictive performance at different target values. Therefore, only R² was selected as the model evaluation metric. R² values ​​range from 0 to 1, with values ​​closer to 1 indicating better model performance.

[0094] For the dataset of this invention, linear regression models performed poorly in prediction. When trained with default parameters, the R² scores of linear regression models did not exceed 0.5, and their prediction performance on Rsf was even worse. This indicates that surface integrity prediction models involve non-linear relationships, making linear regression models unsuitable. Furthermore, the nearest neighbor algorithm, support vector machine, and multi-measure perceptron models also performed poorly on this dataset. In contrast, decision tree-based models (including Bagging and Boosting models) showed better prediction performance on this dataset. This is because decision tree models are highly flexible, can adapt to non-linear relationships in the data, and can automatically perform feature selection.

[0095] Therefore, this invention selects seven regression models as the chosen regression models: Decision Tree Regression, Bagging Regression, Random Forest Regression, ExtraTree Regression, Gradient Boosting Regression, XGBoost Regression, and CatBoost Regression. All of these models achieve prediction scores of 0.75 or higher in Ra, Rsf, and Rst. Specifically, Decision Tree (DT), Bagging (BAG), Random Forest (RF), ExtraTrees (ET), Gradient Boosting Decision Tree (GB), XGBoost (XGB), and CatBoost (CAT) are selected.

[0096] In this embodiment, in step S3, the seven selected regression models have not yet achieved accurate prediction of surface integrity, and parameter tuning and optimization of these models are still required to improve their prediction accuracy. Commonly used parameter tuning methods include: grid search, random search, and Bayesian optimization.

[0097] Grid search is a commonly used hyperparameter tuning method. Its basic idea is to traverse all hyperparameter combinations, perform cross-validation on each set of hyperparameters, obtain the performance index of each set of hyperparameters, and finally select the set of hyperparameters with the best performance as the final model parameters.

[0098] Since the feature dimension of the milling machining dataset is not high and the data sample size is not large, this invention uses grid search for parameter tuning to obtain the optimal parameters of the model.

[0099] The following are the parameter tuning optimizations and model evaluations for seven regression models:

[0100] (1) Surface integrity prediction based on decision tree

[0101] With default parameters, the decision tree regression model achieves a good prediction result for surface roughness Ra, with an R² score exceeding 0.9. This indicates that the decision tree regression model has excellent performance in surface roughness prediction. However, its performance in predicting feed direction residual stress Rsf and transverse residual stress Rst is less satisfactory. This may be due to the more complex generation mechanism of residual stress on milled surfaces. Therefore, further parameter tuning of the decision tree regression model is needed to improve prediction accuracy.

[0102] When tuning hyperparameters, the impact of hyperparameters needs to be considered. Hyperparameters are parameters that need to be manually set before training the model. In decision tree regression models, the main hyperparameters include decision tree depth, minimum sample split, and minimum number of samples per leaf node.

[0103] As shown in Table 4-2, during the mesh search process, the prediction models for surface roughness and residual stress in both directions were tuned. The optimal hyperparameter combination was finally obtained, as shown in Table 4-3.

[0104] Table 4-2

[0105]

[0106] Table 4-3

[0107]

[0108] The decision tree regression model is trained on the training set using the above optimal hyperparameter combination. The performance of the trained model is evaluated on both the training and test sets using MAE, RMSE, and R2 as evaluation metrics. CS-DT represents the decision tree regression model optimized using grid search.

[0109] (2) Bagging-based surface integrity prediction

[0110] Bagging is an ensemble learning method. Bagging regression can select different base regression models for ensemble learning. This invention uses decision tree regression as the base regression model for ensemble learning. To improve the predictive performance of the Bagging regression model, a grid search method is used to determine the optimal parameters of the model. The main hyperparameters of the Bagging regression model are shown in Table 4-5, and the optimal hyperparameters after grid search optimization are shown in Table 4-6.

[0111] Table 4-5

[0112]

[0113] Table 4-6

[0114]

[0115] The Bagging regression model is trained using the training set with the obtained optimal hyperparameters. CS-BAG represents the Bagging regression model optimized by network search. The trained model is then used to predict feature vectors in both the training and test sets, and the prediction results are scored using MAE, RMSE, and R2 as evaluation metrics.

[0116] (3) Surface Integrity Prediction Model Based on Random Forest

[0117] Random forest regression performs well in handling high-dimensional and noisy data and has good generalization ability. The main hyperparameters of random forest regression are shown in Table 4-8. The optimal parameter combination obtained by using grid search to tune the random forest regression model is shown in Table 4-9.

[0118] Table 4-8

[0119]

[0120] Table 4-9

[0121]

[0122]

[0123] The prediction model is trained on the training set using the above optimal hyperparameters to obtain the optimized random forest regression model, i.e., the CS-RF model. The prediction performance of the CS-RF model is scored using MAE, RMSE, and R2 as evaluation metrics.

[0124] (4) Surface integrity prediction based on Extra Trees

[0125] The hyperparameters of the ExtraTree regression model are basically the same as those of the random forest regression model, except for the 'bootstrap' parameter. This is because, unlike the ordinary random forest model, each decision tree in the ExtraTree regression model is trained on the full dataset of the original dataset, without the need for bootstrap sampling to generate different training sets. The hyperparameters of the ExtraTree regression model are optimized using grid search to obtain the CS-ET model. The optimized hyperparameter combinations are shown in Table 4-11.

[0126] Table 4-11

[0127]

[0128] The model was trained using the hyperparameters described above, and the trained model was scored using MAE, RMSE, and R2 as evaluation metrics. The CS-ET model optimized using grid search showed a significant improvement in prediction performance compared to the default parameters.

[0129] (5) Surface integrity prediction based on gradient boosting tree

[0130] Boosting is an ensemble learning method different from Bagging. Besides the decision tree-related hyperparameters, the gradient boosting regression model based on Boosting also has some unique hyperparameters that affect its performance. Table 4-13 lists the hyperparameters that need to be adjusted when tuning the gradient boosting regression model. Since this model has many hyperparameters, manual tuning and grid search are used to find the optimal parameters for the gradient boosting regression model. The optimal parameter combinations after tuning are shown in Table 4-14.

[0131] Table 4-13

[0132]

[0133] Table 4-14

[0134]

[0135] After parameter tuning, three CS-GB models were obtained by training on the training set. Their prediction scores for surface roughness and residual stress on the training and test sets show that their prediction performance is improved compared to before parameter tuning and optimization.

[0136] (6) Surface Integrity Prediction Based on XGBoost

[0137] XGBoost is a machine learning algorithm based on gradient boosting trees. The main hyperparameters affecting the prediction performance of the XGBoost regression model are shown in Table 4-16. The XGBoost regression model is optimized using a grid search algorithm, and the optimal hyperparameter combination is shown in Table 4-17.

[0138] Table 4-16

[0139]

[0140]

[0141] Table 4-17

[0142]

[0143] Similarly, the optimized XGBoost regression model was trained on the training set, and three evaluation metrics were used to score the prediction performance of the obtained CS-XGB model on the training and test sets. The CS-XGB model showed good performance in predicting surface integrity.

[0144] (7) Surface Integrity Prediction Based on CatBoost

[0145] CatBoost introduces techniques such as automatic processing of categorical features and parallel computation based on symmetric trees on top of gradient boosting trees. Therefore, CatBoost does not require manual adjustment of the 'learning_rate' parameter, but uses an adaptive learning rate. The main hyperparameters affecting the prediction performance of the CatBoost regression model are shown in Table 4-19. These hyperparameters were optimized using grid search, and the results are shown in Table 4-20.

[0146] Table 4-19

[0147]

[0148] Table 4-20

[0149]

[0150] The optimized CS-CAT model was scored on its predictions of surface roughness and residual stress in two directions. It can be seen that the mesh search has a very significant effect on optimizing the CatBoost regression model, especially in improving the prediction accuracy of Ra and Rsf.

[0151] After optimization, the seven regression models have shown good results in predicting different indicators of surface integrity. However, the reliability of the optimization results largely depends on the prediction accuracy of the prediction models. Therefore, further improvement of the prediction models is needed. This invention uses the Stacking model fusion method to perform model fusion analysis on the seven given regression models in different combinations, obtaining an ensemble learning model with better prediction performance, thereby improving the prediction accuracy of the milling surface integrity prediction model. Different base prediction models often have different advantages and disadvantages, and the best-performing base model does not necessarily have the best prediction performance on all feature vectors. The Stacking model fusion method has the characteristic of fusing different types of models, which not only improves the model's generalization ability and performance but also helps reduce the risk of model overfitting.

[0152] This invention selects different combinations of regression models for Stacking model fusion analysis to obtain the optimal Stacking ensemble model for surface integrity prediction.

[0153] In Stacking ensemble learning, the base learner (also known as the primary learner) refers to multiple independent models used to generate the input to the meta learner. These models can be homogeneous or heterogeneous, and can come from different domains. The goal is to achieve better performance in the combined model.

[0154] Generally, a base learner should meet the following conditions:

[0155] (1) The base learner itself should have high accuracy in order to provide high-quality prediction results.

[0156] The accuracy of base learners is one of the most important factors in ensemble learning. Only when the predictions of base learners are highly accurate can the ensemble model achieve higher prediction accuracy. Therefore, when selecting base learners, it is necessary to choose models with high accuracy and evaluate their performance through methods such as cross-validation.

[0157] (2) There should be certain differences between different base learners in order to reduce the variance in ensemble learning.

[0158] The predictions of the base learners should be as close as possible to the true values ​​to provide high-quality information. At the same time, the predictions of the base learners should also exhibit diversity so that the meta-learner can learn more from them. If the predictions of all base learners are very similar, the meta-learner cannot gain more information, resulting in no improvement in the performance of the ensemble model.

[0159] When selecting base learners, the characteristics and requirements of the dataset must be considered to choose the most suitable model. Furthermore, the number and types of base learners also affect the performance of the ensemble model; therefore, analysis is needed to determine the optimal combination.

[0160] The design concept is as follows: By arranging and combining the 7 optimized prediction models in different ways, and using the same meta-learner to perform stacking model fusion on the base learners of different combinations, the optimal combination of base learners can be selected.

[0161] Gradient boosting regression models were used uniformly as meta-learners to control for variables. The specific steps were as follows: First, model fusion, model training, and prediction performance scoring were performed on the seven models. Then, the number of base learners was reduced, and the same analysis was performed. Models with relatively poor prediction performance among the base learners were prioritized for reduction. Eight sets of analyses were performed for each surface integrity evaluation index. The analysis results are shown in Tables 4-22, 4-23, and 4-24.

[0162] Table 4-22

[0163]

[0164] Table 4-22 shows the R² evaluation results of different base learner combinations for surface roughness prediction on the test set. As can be seen from the table, when selecting the optimized Bagging regression, Random Forest regression, ExtraTree regression, Gradient Boosting regression, XGBoost regression, and CatBoost regression (a total of 6 models) as base learners, the Stacking ensemble model achieves the best prediction performance for surface roughness.

[0165] Table 4-23

[0166]

[0167]

[0168] Table 4-23 shows the R² scores of Stacking ensemble models with different base learner combinations for predicting residual stress in the feed direction. The results in the table show that the Stacking ensemble model using optimized Bagging regression, Random Forest regression, ExtraTree regression, XGBoost regression, and CatBoost regression as base learners achieves the best prediction performance when predicting residual stress in the feed direction.

[0169] Table 4-24

[0170]

[0171] The R2 scores of Stacking ensemble models with different base learner combinations for predicting transverse residual stress are shown in Table 4-24. The results in the table show that the ensemble model obtained by selecting the optimized ExtraTree regression, XGBoost regression and CatBoost regression as base learners to fuse the Stacking model has the best prediction effect when predicting transverse residual stress.

[0172] When using Stacking for model fusion, the choice of meta-learners (also known as secondary learners) is crucial, as they will be used to combine the predictions from the base models. Typically, a meta-learner should be chosen that has high accuracy and strong generalization ability. In Stacking model fusion of regression prediction models, linear regression, random forest regression, or gradient boosting regression are commonly chosen as meta-learners.

[0173] To determine a suitable meta-learner, this invention uses the aforementioned optimal combination of base learners as the first-layer learner. Based on the base learners, different regression models are selected as meta-learners for analysis. After comparison, the regression model with the best model fusion effect, i.e., the greatest improvement in prediction accuracy, is selected as the meta-learner for model fusion. The analysis results are shown in Tables 4-25, 4-26, and 4-27.

[0174] Table 4-25

[0175]

[0176] Table 4-26

[0177]

[0178] Table 4-27

[0179]

[0180] Based on the analysis results, gradient boosting regression was ultimately selected as the meta-learner for the Stacking ensemble model when predicting surface roughness and transverse residual stress, while random forest regression was selected as the meta-learner for the Stacking ensemble model when predicting residual stress in the feed direction.

[0181] In summary, this invention first identified seven regression models for predicting surface integrity. All seven models underwent hyperparameter optimization using grid search to achieve optimal prediction results. Subsequently, the Stacking ensemble learning method was used to further improve the prediction accuracy of the surface integrity prediction models. The final surface integrity prediction model is as follows: Figure 3As shown, the model comprises three sub-models: a surface roughness prediction model, a feed direction surface residual stress prediction model, and a transverse surface residual stress prediction model. These models are used to predict different indices of surface integrity, and the prediction results of the three sub-models together constitute the final predicted value of surface integrity.

[0182] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for predicting the integrity of milled surfaces, characterized in that: Includes the following steps: S1. Collect milling machining test dataset; S2. Preprocess the milling test dataset to obtain the processed test dataset; S3. Input the processed test dataset into the milling surface integrity prediction model for model training to obtain the trained prediction model, which specifically includes: Using the processed experimental dataset, select A regression model, and on Each of the selected regression models was optimized to obtain... An optimized regression model was selected; The regression models include: Using the training set respectively Several regression models were trained, and each model was trained three times, using surface roughness, residual surface stress in the feed direction, and residual surface stress in the transverse direction as label values; among them... Greater than ; After training, the trained model is used to make predictions on the test set. Each regression model is scored using evaluation metrics, and regression models whose scores meet the set conditions are selected. right The optimized regression model is combined in different ways to obtain base learners with different combinations. The same meta-learner is used to perform stacking model fusion on the base learners with different combinations. The optimal base learner combination is selected from the base learner combinations after several model fusions. The optimal combination of base learners is used as the first layer learner. Based on the first layer learner, different regression models are selected as meta learners for model fusion comparison. The regression model with the greatest improvement in prediction accuracy in model fusion is selected as the meta learner for model fusion. S4. Input the milling data to be tested into the trained prediction model and output the prediction results of the milling surface integrity.

2. The method for predicting the integrity of milled surfaces according to claim 1, characterized in that: The milling test dataset includes feature variable data and workpiece surface data after machining; The feature variable data includes machining parameters, tool posture parameters, and tool geometry parameters; The surface data of the processed workpiece includes surface roughness and surface residual stress; the surface residual stress includes surface residual stress in the feed direction and surface residual stress in the transverse direction.

3. The milling surface integrity prediction method according to claim 2, characterized in that: The machining parameters include spindle speed, feed rate, and depth of cut; The tool attitude parameters include the forward tilt angle and the tilt angle; The tool geometry parameters include tool diameter, number of cutting edges, and helix angle.

4. The milling surface integrity prediction method according to claim 1, characterized in that: The preprocessing includes: The milling test dataset is dimensionless to obtain the dimensionless test dataset. The dimensionless experimental dataset is divided into a training set and a test set.

5. The method for predicting the integrity of milled surfaces according to claim 1, characterized in that: The selected regression model is optimized, specifically including: The model iterates through all combinations of hyperparameters in the regression model and performs cross-validation on each set of hyperparameters to obtain the performance index of each set of hyperparameters. Finally, the set of hyperparameters with the best performance is selected as the final model parameters.

6. The milling surface integrity prediction method according to claim 1, characterized in that: The milling surface integrity prediction model includes a surface roughness prediction model, a feed direction surface residual stress prediction model, and a transverse surface residual stress prediction model.

7. The milling surface integrity prediction method according to claim 6, characterized in that: The optimal combination of base learners in the surface roughness prediction model includes Bagging regression model, Random Forest regression model, ExtraTree regression model, Gradient Boosting regression model, XGBoost regression model, and CatBoost regression model. The optimal combination of base learners in the feed direction surface residual stress prediction model includes Bagging regression model, random forest regression model, ExtraTree regression model, XGBoost regression model and CatBoost regression model. The optimal combination of base learners in the transverse surface residual stress prediction model includes the ExtraTree regression model, the XGBoost regression model, and the CatBoost regression model.

8. The method for predicting the integrity of milled surfaces according to claim 7, characterized in that: In the surface roughness prediction model, the gradient boosting regression model is used as the meta-learner during model fusion. In the surface residual stress prediction model in the feed direction, the random forest regression model is used as the meta-learner when model fusion is performed. In the prediction model of residual stress on transverse surfaces, the gradient boosting regression model is used as the meta-learner during model fusion.

Images