Thin-walled workpiece turning deformation prediction method based on simulation and ensemble learning
By using finite element simulation and ensemble learning methods, a mapping model between cutting parameters and multi-point deformation was established, which solved the problem of high-precision prediction in the turning of thin-walled parts, reduced costs, and improved the stability and applicability of the prediction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- ZHEJIANG UNIV OF TECH
- Filing Date
- 2026-01-22
- Publication Date
- 2026-05-01
AI Technical Summary
Existing technologies struggle to achieve high-precision, low-cost deformation prediction in the turning of thin-walled parts, especially in the turning of difficult-to-machine materials, where there are problems such as high computational costs, complex modeling, and difficulty in quickly obtaining sufficient sample data.
By constructing a prediction framework based on finite element simulation and ensemble learning, the material removal process is simulated using birth and death elements, a mapping model between cutting parameters and multi-point deformation is established, and a weighted ensemble strategy of support vector regression and random forest regression is combined to achieve high-precision prediction.
It achieves high-precision deformation prediction under new cutting parameters, reduces modeling costs and the number of trial cuts, improves machining consistency and the stability of prediction results, and is applicable to different thin-walled parts and difficult-to-machine materials.
Smart Images
Figure CN121960048A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of machining and intelligent manufacturing, and in particular to a method for predicting the turning deformation of thin-walled parts based on simulation and ensemble learning. Background Technology
[0002] Thin-walled structures lack overall rigidity, making them prone to significant elastic deformation under cutting forces during turning. This leads to decreased dimensional accuracy and can even cause vibration and instability. Therefore, accurate prediction of deformation during the turning of thin-walled parts is of significant engineering importance. Furthermore, for some difficult-to-machine materials, such as high-hardness alloys and composite materials, the cutting forces during turning are greater, and plastic deformation is more pronounced, further increasing the difficulty of controlling machining accuracy and exacerbating vibration and instability during processing.
[0003] In existing technologies, the analysis and control of deformation during the machining of thin-walled parts typically rely on machining experience, trial cutting experiments, or online measurement. However, empirical and trial cutting methods usually require adjusting cutting parameters through multiple trials, resulting in high experimental costs, long cycles, and difficulty in adapting to changes in different part structures and machining conditions. While online measurement methods can acquire machining status information in real time, they have high requirements for equipment and sensors, high implementation costs, and difficulty in maintaining stable application under complex working conditions.
[0004] With the rapid development of computer technology, the finite element method has been gradually introduced into the deformation analysis of thin-walled parts. By establishing a finite element model of turning, the deformation behavior of the workpiece under cutting forces can be simulated to a certain extent, providing an important reference for the optimization of machining parameters. However, the turning process of difficult-to-machine materials involves complex large plastic deformation, contact nonlinearity, and material removal processes. Traditional chip separation and contact modeling methods are computationally expensive and complex, and it is difficult to quickly obtain a sufficient number of deformation sample data under multi-parameter conditions, which limits their widespread application in engineering.
[0005] In recent years, data-driven methods and machine learning techniques have received increasing attention in the field of machining process modeling and prediction. Some studies have attempted to establish a mapping relationship between cutting parameters, cutting forces, and other features and machining deformation, thereby achieving rapid prediction of machining results. However, the machining deformation of thin-walled parts exhibits significant nonlinear characteristics and path dependence, and the number of practically obtainable samples is usually limited. Existing methods still have shortcomings in feature selection, model stability, and the physical consistency of prediction results. These issues make it difficult for existing methods to ensure prediction accuracy while also considering practical engineering feasibility. Summary of the Invention
[0006] To address the aforementioned issues, the present invention aims to propose a deformation prediction scheme for turning thin-walled alloy parts that are difficult to machine. This scheme involves rapidly constructing a material removal simulation and automatically generating a multi-condition deformation dataset. Furthermore, it establishes a mapping model between cutting parameters / cutting forces and multi-point deformation to achieve high-precision prediction under new cutting parameters. The prediction results are then compared and verified using experimental methods such as coordinate measuring machine (CMM), thereby reducing modeling costs, decreasing the number of trial cuts, and improving machining consistency.
[0007] To achieve the above objectives, the present invention adopts the following technical solution:
[0008] Step 1: Simulation Modeling and Dataset Construction for Material Removal During Turning
[0009] 1.1) Construct 3D models of thin-walled parts and cutting tools using 3D modeling software, import them into ABAQUS simulation software, and use the "birth and death element" function. By assigning two states to the elements, namely "activation / disabling", the contribution of the elements to the structural stiffness, mass and load transfer is dynamically managed. Thus, the equivalent description of turning chip separation and material removal is achieved through the "birth-death" sequence.
[0010] 1.2) Automated operations are achieved at the pre- and post-processing levels using Python scripts, which are used for the rapid setup of multiple analysis steps, the creation of node sets and mesh cell sets for material removal, and the application of cutting force loads to multiple mesh cells.
[0011] 1.3) Create a path along the tool feed direction to extract deformation data at each point, obtain the elastic deformation curve caused by the cutting force, and summarize the deformation curves under different cutting parameters / cutting force conditions to form a training dataset.
[0012] Step 2: Weighted Ensemble Prediction of Support Vector Regression-Random Forest Regression
[0013] 2.1) Using cutting speed Vc, feed rate f, and depth of cut ap as basic parameters, and combining the three-dimensional cutting forces Fx, Fy, and Fz, we construct higher-order features such as first-order and second-order interactions; we use random forest (RF) to evaluate the importance of the input features, where the feature importance is measured by the decrease in mean square error caused by each feature during the splitting process of the decision tree node, and the specific calculation method is shown in Equation (1): ;
[0014] Among them, I j The importance score for the j-th feature; T is the total number of decision trees in the random forest; S t,j Let represent the set of nodes in the t-th tree that are split using feature j; This is the decrease in mean squared error caused by the split node. For all input features, I... j The features are sorted, and only a few that contribute most to the prediction of deformation during the processing of thin-walled parts are retained to reduce the interference of redundant features on model training, thereby reducing the risk of overfitting.
[0015] 2.2) To eliminate the differences in dimensions and numerical ranges between different input features and output deformation quantities, the Min-Max normalization method is used to process all feature variables and deformation response quantities, and its mathematical expression is shown in equation (2): ;
[0016] Among them, u min with u max , respectively, are the minimum / maximum values of the variable in the training samples, and u' is the normalized value. Equation (2) can uniformly map variables with different dimensions and different numerical ranges to the interval [0,1], thereby avoiding the bias problem caused by the difference in feature scale during the training process of the support vector regression model (SVR), and at the same time facilitating the unified modeling and comparative analysis of deformation outputs at different points.
[0017] 2.3) To improve the robustness and generalization ability of the prediction model under small sample conditions, a small Gaussian noise perturbation is introduced into the input feature vector during the training sample construction process. The feature enhancement method is shown in Equation (3): ;
[0018] Where x is the original feature vector (e.g., composed of Vc, f, ap, cutting force components, and interaction terms). To enhance the feature vector after introducing noise perturbation, For a random noise vector that follows a Gaussian distribution, This indicates that the noise vector follows a pattern with a mean of 0 and a variance of . The normal distribution is given by I, where I is the identity matrix. Used to control the intensity of noise disturbance.
[0019] Meanwhile, to address the issue of insufficient sample size under extreme cutting parameter conditions, oversampling was performed on relevant samples to enhance the model's learning ability under complex and extreme machining conditions and improve overall prediction stability.
[0020] 2.4) Each processing deformation curve is considered as a sequence of multiple discrete processing points, and a one-dimensional SVR model is established for each point. To obtain more robust model parameters under small sample conditions, leave-one-out cross-validation is adopted, and grid search is combined to optimize the model hyperparameters. The loss function is defined as shown in equation (4): ;
[0021] Where n is the total number of training samples used for modeling this location; y i This represents the actual processing deformation value of the i-th sample; This represents the prediction result of the model trained after removing the i-th sample for that sample; The loss function measures the model's prediction error. The optimal parameter configuration for the support vector regression model is determined by iterating through candidate hyperparameter combinations and minimizing the loss function value.
[0022] 2.5) After training the SVR and RF models, the deformation prediction results of the two models at each processing point are obtained. In order to give full play to the advantages of different models in terms of small sample fitting ability and noise resistance, a weighted fusion strategy is adopted to fuse the two prediction results. The fusion prediction model is shown in Equation (5).
[0023] ;
[0024] 2.6) By assigning weights to the prediction results of different models, a comprehensive prediction of the deformation of thin-walled parts during turning is achieved. The SVR model focuses on nonlinear fitting ability, while the RF model focuses on feature interaction processing and robustness improvement, thereby improving the stability of the prediction results while ensuring the prediction accuracy.
[0025] Step 3: Turning experiment and verification with coordinate measuring machine
[0026] The thin-walled part was machined on a CNC machine tool according to the new cutting parameters. The key points corresponding to the simulation path were measured using a coordinate measuring machine to obtain the actual machining deformation curve. The measured curve was compared with the model prediction curve, and the maximum deformation error, MAE, RMSE and other indicators were calculated to verify the effectiveness of the prediction model.
[0027] The design concept of this invention is as follows:
[0028] First, a turning simulation model with a birth and death element mechanism is constructed using finite element simulation software to obtain corresponding elastic deformation datasets under different cutting parameter conditions. Then, using cutting parameters and cutting force as input features, a weighted ensemble prediction model is constructed by combining support vector regression and random forest regression to predict machining deformation under new working conditions. Finally, turning experiments are conducted and actual deformation values are obtained using a coordinate measuring machine to compare and verify the model prediction results. The method of this invention combines the advantages of simulation analysis capabilities and machine learning algorithms, significantly reducing the actual experimental cost and time expenditure while ensuring prediction accuracy, and can provide a reliable basis for optimizing precision machining parameters of thin-walled parts.
[0029] Compared with the prior art, the present invention has at least the following effects:
[0030] 1) The material removal process can be dynamically described by the "birth and death cells", which can replace the time-consuming chip separation and complex contact simulation to improve efficiency.
[0031] 2) Discretize the overall machining deformation curve into multiple key points and establish regression prediction models for each point to achieve continuous deformation distribution prediction along the tool feed path, making the results closer to the actual machining process and providing more comprehensive data support for process compensation and parameter optimization.
[0032] 3) Combining the advantages of support vector regression in small-sample nonlinear modeling with the advantages of random forest regression in feature interaction processing and noise resistance, a weighted ensemble strategy is adopted to fuse the prediction results of the two types of models, effectively avoiding the problem of overfitting or insufficient generalization ability of a single model, and maintaining high prediction accuracy under new cutting parameter conditions.
[0033] 4) The proposed simulation modeling and integrated learning prediction framework does not depend on specific part structures or single material parameters. It can be extended to the prediction of turning deformation of other thin-walled parts and difficult-to-machine materials by simply adjusting the finite element model and input features. It has high engineering application and promotion value. Attached Figure Description
[0034] Figure 1 This is a flowchart of the deformation prediction method for turning thin-walled parts based on finite element simulation and ensemble learning according to the present invention.
[0035] Figure 2 This is a three-dimensional model of a 1 / 4 thin-walled part made of a difficult-to-machine material, according to an embodiment of the present invention.
[0036] Figure 3 This is a radial deformation cloud map of a simulated turning of a 1 / 4 thin-walled part made of a difficult-to-machine material, according to an embodiment of the present invention.
[0037] Figure 4 These are the training results of the SVR-RF regression model in the embodiments of this invention;
[0038] Figure 5 This is a physical image of the coordinate measuring machine used in the embodiments of the present invention;
[0039] Figure 6 This is a comparison chart of the predicted results and the actual measurement results provided in the embodiments of the present invention. Detailed Implementation
[0040] The implementation method of the present invention will be further described in detail below with reference to the accompanying drawings in the embodiments of the present invention.
[0041] This embodiment takes a thin-walled component made of a difficult-to-machine material as the research object. Specifically, the thin-walled component is a shell part with typical thin-walled characteristics. Combining finite element simulation analysis and ensemble learning algorithms, the elastic deformation of the component during the turning process is predicted. The overall process of this method includes four stages: simulation modeling of turning material removal, construction of machining deformation dataset, training of ensemble regression model, and experimental verification.
[0042] S1. Construction of Finite Element Model for Turning and Simulation of Material Removal
[0043] A simulation model of the workpiece's turning process was created in the 3D modeling software Solidworks. This model was then imported into the finite element simulation software ABAQUS. A hexahedral mesh was applied to the workpiece, elastic mechanical parameters were assigned to the workpiece material, and Johnson-Cook (JC) constitutive model parameters were selected. Each analysis step and load was set in batches, and finally, constraints and boundary conditions were applied. Specifically, this included:
[0044] S11. Before establishing a 3D turning finite element simulation model, a 3D model of the workpiece needs to be created. To further improve computational efficiency, such as... Figure 2 A 1 / 4 model of the workpiece is established as shown, and the mechanical and thermal properties of the workpiece are determined, including: density, elastic modulus, Poisson's ratio, specific heat capacity, thermal conductivity, melting point, etc., and then input into the material database in ABAQUS, as shown in Table 1.
[0045] Table 1 Performance Parameters ;
[0046] S12. Determine the material constitutive model and its parameters. Use the JC constitutive model to describe the plastic behavior of the workpiece material during turning. The JC constitutive model is expressed as:
[0047] Where A is the yield strength of the material, B is the strain hardening constant, n is the strain hardening exponent, C is the strain rate sensitivity coefficient, and m is the thermal softening coefficient. This is the equivalent stress, in MPa. Equivalent plastic strain; For dimensionless equivalent plastic strain rate, , For reference strain rate; Here, T is the dimensionless temperature. r T m Table 2 shows the reference temperature and the material melting point, respectively, with T representing the test temperature in °C. The JC constitutive model parameters for this difficult-to-machine material are also shown in Table 2.
[0048] Table 2 Model Parameters ;
[0049] S13. Since the material removal process is simulated using the ABAQUS "birth and death element" function, the mesh size is set according to the cutting parameters. The cutting parameters used in the finite element simulation are shown in Table 3.
[0050] Table 3 Simulation Cutting Parameters ;
[0051] S14. Based on the unique numbers generated for each element and its constituent nodes after mesh generation, create analysis steps, nodes, and mesh element sets in batches using Python code. To simulate load movement during turning, the cutting load needs to be "deactivated" in the subsequent analysis step to ensure that the cutting load in each analysis step is only effective within that analysis step. The cutting force load under different cutting parameters is calculated according to the empirical formula for turning force shown below: ;
[0052] S15, the "Dead or alive unit" function is realized through the interaction of "Model change". The unit to be removed is selected according to the preset cutting path and set to "Deactivated in this step" in the corresponding analysis step to complete the failure handling of the unit; finally, the workpiece is fixed and constrained around the perimeter and bottom edge according to the clamping conditions in the actual turning process.
[0053] S16, as Figure 3 The simulation results shown extract deformation data at each point by creating a path according to the tool feed direction, and obtain the elastic deformation curve of the workpiece under the action of cutting force, thus providing basic data support for subsequent data-driven model training.
[0054] S2. Construction and Training of Integrated Regression Prediction Model
[0055] The multi-point deformation curves of the thin-walled parts made of difficult-to-machine material under different cutting conditions obtained in step S1 are used as training samples to construct a regression prediction model with cutting parameters, cutting forces, and their higher-order interaction characteristics as inputs. This embodiment uses both SVR and RF models to model the deformation at each point, and obtains the final prediction curve through weighted integration. Simultaneously, constraints are applied to key points on the curve (points of maximum deformation) to ensure that the prediction results are consistent with physical laws. Specifically, this includes:
[0056] S21. In this embodiment, cutting speed Vc, feed rate f, and depth of cut ap are selected as the basic inputs. The triaxial cutting forces Fx, Fy, and Fz are calculated based on the empirical formula for turning forces. Furthermore, higher-order features such as first-order and second-order interaction terms are constructed to enhance the nonlinear expressive power, including but not limited to: Vc·f, ap 2 ,Fz·ap,Vc·Fx,f·Fy,Fz 2 Vitamin C 2 Fz / f, ap / Vc, etc. Based on equation (6), the basic variables are combined into a feature vector, and the feature vector of the i-th sample is defined as: ;
[0057] The above sample feature vectors are then combined into an input feature matrix X_data according to equation (7), and the corresponding output is the multi-point deformation curve y_raw obtained from the finite element simulation, where each curve contains the deformation of A discrete points.
[0058] ;
[0059] Where N is the number of working conditions, N=9 in this embodiment, and d is the d-dimensional feature generated for each group of working conditions, d=15 in this embodiment.
[0060] S22. In this embodiment, a random forest regression model (number of trees n_estimators=100, random seed random_state=42) is used to fit and train the point with the largest deformation (i.e. the first point of the deformation curve); the importance score of each feature is calculated based on the statistical mechanism of the mean squared error reduction within the random forest, and the features are sorted from high to low importance. The top 10 features are selected to form the final input feature set top_features, and the input matrix X is updated accordingly.
[0061] S23. In this embodiment, the Min-Max normalization method is used to scale the input X and output y_raw respectively to obtain normalized X_scaled and y_scaled, so as to unify the feature scale and improve the model training stability. The random seed is set to 42, and different noise scales are set for some features: smaller perturbations are used for the first few features, and larger perturbations are used for the remaining features; key samples (e.g., group 1 and group 9) are selected as enhancement objects, and Gaussian noise with a mean of 0 is superimposed on their input feature vectors and several sets of enhanced samples are repeatedly generated to obtain the enhanced training set X_augmented; the corresponding output y_augmented remains consistent with the original sample, thereby expanding the scale of small sample training data and strengthening the learning of key working conditions.
[0062] S24. Treat each deformation curve as a sequence of A points, establish an SVR regression model for each discrete point, and set a strengthening fitting strategy for the point with the largest deformation. In this embodiment, an SVR model is trained for each point a (t=1…A), using the RBF kernel function, and the model parameter combination that minimizes the error is selected by traversing the candidate hyperparameter set (such as C, gamma, epsilon): for the first point (the point with the largest deformation), a "first point error strengthening" strategy is introduced in the model selection stage: the error of this point is given a higher weight (e.g., 90%) to strengthen the fitting ability for the point with the largest deformation; for the remaining points (a≥2), the model is selected according to the conventional error criteria. Finally, the SVR point model set {SVR_a} is obtained.
[0063] S25. Train a random forest regression model for each discrete point and introduce trend consistency weights for adjacent points to improve the overall continuity and trend rationality of the prediction curve. In this embodiment: for the first point, train a random forest model (e.g., n_estimators=300, max_depth=10, random_state=42) to enhance the fitting of the point with the largest deformation.
[0064] For subsequent points (a≥2), a random forest model is trained (e.g., n_estimators=200, max_depth=8, random_state=42), and sample weights prev_weight are constructed based on the output differences of adjacent points: higher weights are given when the deformation differences between adjacent points are small, to encourage the model to learn samples with "consistent curve trends", thereby improving the smoothness and physical consistency of the predicted curve. The final result is the random forest point model set {RF_a}.
[0065] S26. Using the SVR and RF point models, prediction results for each point of each sample are generated respectively. A weighted fusion strategy is then used to fuse the outputs of the two models to obtain the final prediction curve. In this embodiment, the SVR and RF prediction values are calculated for each point a to obtain y_pred_scaled and rf_pred_scaled. To ensure that the point with the largest deformation satisfies the constraints of the embodiment, the predicted value of the first point of the training sample is forcibly set to the true normalized value, i.e., the prediction of the first point is completely consistent with the true value (used for training set performance evaluation and curve alignment). Subsequently, differentiated fusion weights are applied to different samples:
[0066] For key samples (e.g., group 1 and group 9), a more RF-biased ensemble weight (e.g., 0.3·SVR + 0.7·RF) is used.
[0067] For ordinary samples, an ensemble weight biased towards SVR (e.g., 0.6·SVR + 0.4·RF) is used. After fusion, the ensemble_pred_scaled is obtained, and inverse normalization is performed to obtain the final predicted curve y_pred. At the same time, the training set is re-enhanced to ensure that the first point is consistent with the true value for subsequent error evaluation and comparative analysis.
[0068] S27. For the new cutting parameter input, an input vector is generated according to the same feature construction and normalization process, and the trained SVR and RF point models are called for prediction to obtain the deformation curve under the new working condition. In this embodiment, for the new parameters (Vc,f,ap), firstly, the corresponding Fz, Fx, and Fy are calculated according to S21, and interactive features consistent with the training stage are constructed; then, the corresponding feature subsets are extracted according to the top_features selected in S22 and normalized; finally, SVR_a and RF_a are called for each point to obtain the predicted value, and a normalized prediction curve is obtained by fusing them using a fixed ensemble weight (e.g., 0.5·SVR + 0.5·RF). After inverse normalization, the deformation curve prediction result under the new working condition is obtained. Since there are no measured / simulated values for the new sample, the first point is no longer subject to forced matching, but its predicted value is determined by the model output, such as... Figure 4 The results of the model training are shown below.
[0069] S3. Verification of Turning Experiment and Predicted Results
[0070] To verify the effectiveness and engineering applicability of the proposed deformation prediction method for turning thin-walled parts based on finite element simulation and ensemble learning, this embodiment selects new cutting parameters Vc=60m / min, f=0.1mm / r, and ap=0.4mm to conduct turning experiments, and compares and analyzes the actual machining deformation curves measured in the experiments with the prediction results obtained in step S2. Specifically, this includes:
[0071] S31. In this embodiment, the experimental workpiece material is a difficult-to-machine material, and its geometry and wall thickness are consistent with the finite element simulation model in step S1; using, for example Figure 5 The Hexagon INNOVA Performance coordinate measuring machine shown in the figure measures the initial dimensions of the workpiece according to the simulated points, and then obtains the theoretical contour without deformation by reverse equidistant measurement based on the cutting depth.
[0072] S32. A turning experiment was conducted on its outer arc surface. The cutting tool used was Mitsubishi DCMT11T304 with MP9015 coating material. The rake angle and clearance angle were 15° and 7°, respectively. The machine tool used was a Keda CNC machine tool. The workpiece was clamped and fixed to the machine tool according to the simulation clamping method.
[0073] S33. After the turning process is completed, use a coordinate measuring machine to measure the machined dimensions to obtain the actual deformation curve. Input the SVR-RF integrated prediction model trained in step S2 into the same new cutting parameters to obtain the corresponding predicted deformation curve; for example... Figure 6 As shown, the predicted curve is compared point by point with the actual deformation curve measured in the experiment, and the prediction error of the point with the maximum deformation is calculated to be 8.4%, which verifies the effectiveness of the integrated model prediction.
[0074] The above embodiments are merely preferred embodiments of the present invention and are not intended to limit the technical solutions of the present invention. Any technical solution that can be implemented based on the above embodiments without creative effort should be considered to fall within the scope of protection of the patent of the present invention.
Claims
1. A method for predicting turning deformation of thin-walled parts based on simulation and ensemble learning, characterized in that, Includes the following steps: Step 1) Simulation modeling and dataset construction for material removal during turning: The thin-walled part and tool model were constructed using 3D modeling software and imported into ABAQUS. The birth and death element function was used in combination with multiple ordered analysis steps to simulate the material removal during turning. Python scripts were used to automate the pre- and post-processing. Elastic deformation data under different cutting parameters / working conditions were extracted along the tool feed direction and summarized to build a training dataset. Step 2) Weighted ensemble prediction of Support Vector Regression-Random Forest Regression: Based on cutting parameters, high-order features are constructed, and features are evaluated and screened through random forest. After Min-Max normalization, Gaussian noise feature enhancement and oversampling processing, a one-dimensional SVR model is established at each machining point and the parameters are optimized using leave-one-out cross-validation. Finally, a weighted ensemble strategy is used to fuse the prediction results of SVR and RF to achieve accurate and stable prediction of turning deformation of thin-walled parts. Step 3) Turning experiment and verification with coordinate measuring machine: By machining thin-walled parts on a CNC machine tool according to new cutting parameters, the actual machining deformation curves of the points corresponding to the simulation path are obtained with the help of a coordinate measuring machine. The actual curves are then compared with the model prediction curves, and the maximum deformation error, MAE, and RMSE indices are calculated to verify the effectiveness of the prediction model.
2. The method for predicting turning deformation of thin-walled parts based on simulation and ensemble learning according to claim 1, characterized in that, The specific process of step 1) is as follows: 1.1) 3D Model Construction and Import: 3D model of thin-walled parts and 3D model of turning tools are constructed using 3D modeling software, and the 3D model of thin-walled parts and 3D model of turning tools are imported into ABAQUS simulation software; 1.2) Simulation Modeling of Turning Material Removal Based on Element-Based Birth and Death: Utilizing the element-based birth and death function of ABAQUS simulation software, the contribution of mesh elements to structural stiffness, mass, and load transfer is dynamically managed by dynamically assigning two states to mesh elements: {activated} and {disabled}. An initial analysis step and multiple ordered cutting analysis steps are defined. Through the ordered execution of the initial analysis step and multiple cutting analysis steps, the equivalent simulation of turning chip separation and layer-by-layer material removal is achieved. 1.3) Pre- and post-processing automation based on Python scripts: Write Python automation scripts based on the ABAQUS secondary development interface to complete the rapid configuration of multiple cutting analysis steps, the automatic creation of node sets and mesh element sets in the material removal area, and the application of cutting force loads on multiple mesh elements in batches through the Python automation scripts; 1.4) Deformation data extraction and training dataset construction: Create a data extraction path along the feed direction of the turning tool, extract the point deformation data of each point under the action of cutting force through the data extraction path, generate the elastic deformation curve of the corresponding cutting parameters / cutting force conditions, summarize the elastic deformation curves under different cutting parameters / cutting force conditions, and construct the turning machining elastic deformation training dataset.
3. The method for predicting turning deformation of thin-walled parts based on simulation and ensemble learning according to claim 1, characterized in that, The specific process of step 2) is as follows: 2.1) Feature Construction: Using cutting speed Vc, feed rate f, and depth of cut ap as basic parameters, and combining the three-dimensional cutting forces Fx, Fy, and Fz, we construct high-order features of primary and secondary interactions to enrich the dimensions of input features and cover the correlations in the cutting process. 2.2) Feature Filtering: Random Forest (RF) is used to evaluate the importance of input features: the feature importance score is obtained by calculating the decrease in mean squared error of each feature when splitting at the decision tree node. After ranking the scores, the features that contribute the most to the prediction of thin-walled part deformation are retained to reduce the interference of redundant features and reduce the risk of model overfitting. 2.3) Feature normalization: Min-Max normalization is applied to all feature variables and deformation response quantities to map variables with different dimensions and numerical ranges to the [0,1] interval, thereby avoiding bias caused by the difference in feature scale in support vector regression (SVR) and facilitating the unified analysis of deformation results at different points. 2.4) Sample Augmentation: Feature enhancement: Introduce small-amplitude Gaussian noise with a mean of 0 into the original feature vector to improve the robustness and generalization ability of the model under small sample conditions; Oversampling supplementation: Oversample insufficient samples under extreme cutting parameter conditions to enhance the model's learning ability for complex / extreme machining conditions. 2.5) Model building and parameter optimization: Each processing deformation curve is divided into a sequence of multiple discrete points, and a one-dimensional SVR model is established at each point. The leave-one-out cross-validation method combined with grid search is used to traverse candidate hyperparameters. The optimal parameter configuration of the SVR model is determined by minimizing the loss function to ensure the stability of the model parameters under small sample conditions. 2.6) Model ensemble prediction: After training the SVR and RF models, the deformation prediction results of the two models at each processing point are obtained, and the results are fused using a weighted ensemble strategy: the contribution of the two models is allocated by weight λ, and the nonlinear fitting ability of SVR and the feature interaction and robustness of RF are utilized to achieve a balance between prediction accuracy and stability.