Multi-objective optimization method for dry turning process parameters based on chip breakage risk constraint

Abstract

This invention discloses a multi-objective optimization method for dry turning process parameters based on chipping risk constraints. First, a physical derivative feature space is constructed based on cutting physics, and the XGBoost algorithm is used to build regression models for surface roughness and tool wear, as well as a chipping probability prediction model. Second, a "maximum tolerance false negative rate" is introduced to determine a safety threshold, transforming the chipping probability into a hard constraint based on a safety-first strategy. An optimization model is constructed with the objectives of maximizing material removal rate and minimizing surface roughness and tool wear. The NSGA-II algorithm is used to solve the Pareto solution set, and the optimal parameters are determined by combining it with the TOPSIS method. This invention integrates physical mechanisms and data-driven approaches, solving the problem of small-sample modeling. By using chipping probability as a hard constraint, it can achieve synergistic optimization of efficiency and quality in dry turning while ensuring machining safety, providing a theoretical basis and data support for process parameter formulation.

Description

Technical Field

[0001] This invention relates to a multi-objective optimization method for turning process parameters, specifically a multi-objective optimization method for dry turning process parameters based on chipping risk constraints, belonging to the field of dust removal equipment technology. Background Technology

[0002] Dry turning is a turning process that does not use cutting fluid, aiming to reduce environmental pollution and production costs. Its core principle is to replace the cooling and lubrication functions of cutting fluid with high-temperature resistant tool materials and machine tool technology. High temperatures are unavoidable during dry turning, especially for "superalloys" such as nickel-based alloys, cobalt-based alloys, and titanium alloys, which simultaneously meet the demanding requirements of high strength, corrosion resistance, and heat resistance. This can easily lead to adhesive wear, diffusion wear, and even sudden tool breakage. Tool breakage, as a typical catastrophic failure mode, is unpredictable and irreversible, and once it occurs, the workpiece is instantly scrapped. Therefore, how to avoid the risk of tool breakage while ensuring machining efficiency and quality is currently the main technical challenge facing efficient dry turning.

[0003] Currently, the industry mainly relies on empirical trial-and-error or mathematical modeling-based optimization methods for optimizing turning parameters. For example, in existing technologies, Sassi H et al. used response surface methodology (RSM) to establish the relationship between turning parameters and surface roughness (Ra) and tool wear (V). b The mapping relationship between Ra and V is established, and a genetic algorithm (GA) is used for multi-objective optimization. Some studies, such as Sethuramalingam P, Salem HE, etc., have also attempted to introduce machine learning methods such as support vector machines (SVM) or random forests (RF) to improve prediction accuracy. However, most of these methods only consider Ra and V. b Using continuous indicators as optimization objectives implicitly assumes that the machining process is stable and evolves gradually, often neglecting explicit constraints on the risk of sudden failures such as tool chipping. Furthermore, existing multi-objective optimization algorithms, when dealing with turning parameter problems, typically lack effective identification of "high-risk regions" in the parameter space. This can lead to optimized parameters that, while mathematically achieving a balance between efficiency and quality, may fall into high-risk chipping zones during actual machining. Simultaneously, purely data-driven models often suffer from insufficient generalization ability and a lack of physical interpretability under small sample conditions in cutting experiments, making it difficult to accurately describe continuous wear patterns and discrete chipping boundaries. Summary of the Invention

[0004] To address the problems existing in the prior art, this invention provides a multi-objective optimization method for dry turning process parameters based on chipping risk constraints. This method can solve the problem of small sample modeling and simultaneously achieve synergistic optimization of dry turning efficiency, surface quality, and tool life while ensuring zero chipping risk.

[0005] To achieve the above objectives, this multi-objective optimization method for dry turning process parameters based on chipping risk constraints specifically includes the following steps:

[0006] Step 1, Experimental Design and Data Acquisition: Build a dry turning physical layer experimental platform, design experiments, and conduct dry turning experiments group by group according to the experimental design scheme to obtain a dry turning dataset including basic turning process parameters, physical derived features, and chipping state labels.

[0007] Step 2, construct a multi-objective prediction model based on physical information features and XGBoost: Based on the dry turning dataset obtained in Step 1, construct and train an ensemble learning model based on XGBoost. The ensemble learning model based on XGBoost includes a surface roughness regression prediction model, a tool wear regression prediction model, and a chipping probability prediction model.

[0008] Step 3, setting the risk threshold for blade breakage based on the safety-first strategy: By setting the "maximum tolerable false negative rate", the safety threshold τ of the blade breakage risk probability is determined, and the blade breakage probability prediction model trained in Step 2 is transformed into a hard constraint that can be used for multi-objective optimization.

[0009] Step 4, Multi-objective process parameter optimization solution based on NSGA-II algorithm: Construct a multi-objective optimization mathematical model with the objective functions of maximizing material removal rate, minimizing surface roughness and minimizing tool back face wear, and the safety threshold τ of the chipping risk probability determined in Step 3 as the constraint condition. Use NSGA-II algorithm to perform global optimization and solve the Pareto optimal solution set that takes into account processing efficiency, quality and safety.

[0010] Step 5, Post-processing screening and comprehensive decision-making of the Pareto optimal solution set: Post-processing screening is performed on the Pareto optimal solution set obtained in Step 4 to transform the theoretical continuous solution into a discrete engineering solution that meets the actual processing requirements, and finally obtain the optimal combination of process parameters.

[0011] Furthermore, Step 1 is detailed as follows:

[0012] Step 1-1: Construct a dry turning physical layer experimental platform and collect basic turning process parameters, including cutting speed. Feed rate Back cut radius of the blade tip arc ;

[0013] Steps 1-2: Based on the physical mechanism of cutting, a high-dimensional physical information feature set is constructed, including basic turning process parameters and physical derivative features. These physical derivative features include geometric features reflecting the surface forming mechanism, energy consumption features reflecting cutting work and frictional heat accumulation, and mechanical features reflecting instantaneous load and fracture risk. Both the energy consumption features reflecting cutting work and frictional heat accumulation and the mechanical features reflecting instantaneous load and fracture risk include material removal rate. Dimensionless parameters Geometric ratio and the square of the cutting speed term Geometric features reflecting the surface forming mechanism include theoretical roughness terms. The square of the feed rate Geometric ratio and material removal rate ;

[0014] Steps 1-3, with cutting speed Feed rate Back cut and the radius of the blade tip arc Four variables were used as input factors in the experimental design. Dry turning tests were conducted group by group according to the experimental design. The chipping status of the test group that experienced chipping was labeled as 1, and the chipping status of the test group that did not experience chipping was labeled as 0. The surface roughness value of each test group that did not experience chipping was recorded. and tool flank wear and the material removal rate As an indicator of production efficiency, the final structure includes a high-dimensional physical information feature set and multi-objective response values. A complete dry turning dataset with chipping state labels.

[0015] Furthermore, Step 2 is detailed below:

[0016] Step 2-1, Dataset Partitioning and Prediction Model Construction: The experimental data of the dry turning dataset is divided into training and testing sets. Based on the XGBoost model, a feature vector containing geometric features reflecting the surface forming mechanism is constructed, and the output is the surface roughness. A continuous numerical surface roughness regression prediction model is constructed, which includes a feature vector reflecting energy consumption characteristics of cutting work and frictional heat accumulation, and the output is tool wear. A continuous numerical regression prediction model for tool wear is constructed, which includes feature vectors reflecting mechanical characteristics of instantaneous load and fracture risk, and outputs a chipping probability prediction model as the probability of chipping.

[0017] The objective function of the XGBoost model The definition is as follows:

[0018]

[0019] In the formula: The number of samples in the training set; Let be the loss function used to measure the difference between predicted and true values, where Indicates the first The true value of each sample Indicates the first The model predicts the output value for each sample; This represents the total number of decision trees; This is a regularization term used to control model complexity and prevent overfitting, where This represents the structure of the k-th decision tree and the weights of its leaf nodes;

[0020] Regularization term The definition is as follows:

[0021]

[0022] In the formula: This represents the number of leaf nodes; The weight of the leaf node. Let L2 be the norm of the weight vector; and These are the regularization coefficients that control the number of leaves and the regularization coefficients that control the weight of the leaves, respectively.

[0023] For the surface roughness regression prediction model and the tool wear regression prediction model, the mean squared error is used as the loss function, and its expression is as follows:

[0024]

[0025] In the formula: The number of samples in the training set; Indicates the first The true value of each sample, i.e., the surface roughness value. Or the amount of wear on the back face of the cutting tool The measured values; Indicates the first The model's predicted output value for each sample, i.e., the predicted surface roughness value. Or the amount of wear on the back face of the cutting tool ;

[0026] For the blade breakage probability prediction model, log loss is used as the loss function, and its expression is as follows:

[0027]

[0028] In the formula: The number of samples in the training set; Indicates the first The true value of each sample, where a true value of 1 indicates a broken blade and a true value of 0 indicates safety; Indicates the first The model's predicted output value for each sample, i.e., the model's prediction of the sample. The probability of the predicted outcome being "Collapse Blade";

[0029] Step 2-2: Train the ensemble learning model.

[0030] Furthermore, in Step 2-2, before training the ensemble learning model, Bayesian hyperparameter optimization is performed using the Optuna framework: For the regression prediction tasks of surface roughness and tool wear, the search range of key hyperparameters such as the maximum tree depth, learning rate, minimum weight of child nodes, and sample and feature sampling rates are set; for the chipping probability prediction task, the minimum loss decrease value is introduced as an optimization variable based on the regression task; and the TPE sampling algorithm is used for iterative search to minimize the mean absolute error. The optimal combination of hyperparameters for the XGBoost model is determined by maximizing the F1 score as the fitness objective for both the regression task and the blade breakage probability prediction task.

[0031] The ensemble learning model is trained using the optimal hyperparameter combination determined by the Bayesian optimization algorithm.

[0032] Furthermore, Step 3 is detailed below:

[0033] Step 3-1, Introduce the critical value for the probability of blade breakage. Construct a safety-first constraint logic to eliminate predicted probabilities that exceed the limit. The mathematical expression for the combination of process parameters and the constraint strategy is as follows:

[0034]

[0035] In the formula: The probability of blade breakage predicted by the model; This is the critical value for the probability of blade breakage.

[0036] Step 3-2: Introduce the maximum tolerance for false negatives index to construct the critical value for the probability of a crash. The target missed rate screening range is determined, and the maximum value within the target missed rate screening range is selected to maximize the process window;

[0037] Step 3-3: Iterate through the selection process, calculate and determine the critical value for the final probability of blade breakage. The critical value of the final probability of blade breakage. As a hard constraint boundary in the subsequent NSGA-II multi-objective optimization algorithm.

[0038] Furthermore, in Step 3-2, the critical value of the probability of blade breakage. The upper and lower limits of the target false negative rate screening interval are set based on statistical analysis of experimental data and engineering experience in dry turning of specific materials.

[0039] Furthermore, in Step 3-3, the critical value of the final probability of blade breakage is calculated and determined. Then, using the XGBoost "Break" probability prediction model trained in Step 2, predictions are made on the validation set data, outputting the probability value of each sample belonging to the "Break" category. The specific calculation steps are as follows:

[0040] First, a relationship curve is plotted, and the recall and false negative rates are calculated for each cutoff point by iterating through different probability cutoff points. Second, a reverse search algorithm is used to filter out a set of probability thresholds that meet the set of conditions within the set of target false negative rate filtering intervals. Finally, the optimal threshold is determined, and the highest probability value is selected from the set of probability thresholds as the final critical value for the probability of blade breakage. .

[0041] Furthermore, Step 4 is detailed below:

[0042] Step 4-1: Initialize the NSGA-II algorithm parameters, encode the decision variables using real number encoding, and set the population size, maximum number of generations, crossover and mutation probabilities;

[0043] Step 4-2, Define including cutting speed Feed rate Back cut and the radius of the blade tip arc The decision variable vector is used, and the prediction model trained in Step 2 is called as the fitness function to calculate the corresponding surface roughness for each individual in the population. Predicted values, tool wear Predicted values ​​and material removal rates ;

[0044] Step 4-3, construct to minimize surface roughness Predicted values ​​and tool wear Predicted value, maximized material removal rate A multi-objective optimization mathematical model for turning parameters is proposed, and boundary constraints of decision variables are set and the nonlinear constraints of chipping risk are handled by the external penalty function method.

[0045] The mathematical model for multi-objective optimization of turning parameters is expressed as follows:

[0046]

[0047] In the formula: For a set of decision variables, ; The objective function is surface roughness. Surface roughness values ​​predicted by the model ; The objective function is the tool wear parameter. The amount of flank wear predicted by the model ; The objective function is the material removal rate; For material removal rate, ;

[0048] The boundary constraints and penalty strategies for decision variables are expressed as follows:

[0049]

[0050] In the formula: and Cutting speed The lower and upper limits; and feed rate The lower and upper limits; and They are respectively the back cut amount The lower and upper limits; and These are the radii of the blade tip arc. The lower and upper limits; τ is the predicted probability of blade breakage under the current parameter combination; τ is the critical value of the probability of blade breakage. For the first Each component of the optimization objective function ; As a penalty factor;

[0051] Step 4-4: Perform evolutionary iteration. The population is stratified and screened by non-dominated sorting and crowding calculation. Offspring are generated by binary tournament selection, simulated binary crossover and polynomial mutation. The best individuals are retained by using an elite strategy. The process is iterated until the maximum number of generations is reached, and the Pareto optimal solution set is output.

[0052] Furthermore, in Step 4, the radius of the blade tip arc... It is treated as a continuous variable during the optimization iteration process.

[0053] Furthermore, Step 5 is detailed below:

[0054] Step 5-1, for the continuous tool tip radius in the Pareto solution set The nearest neighbor mapping algorithm is used to map it to discrete tool specification values ​​that conform to ISO standards.

[0055] Step 5-2: Substitute the discretized parameter combination into the blade breakage probability prediction model trained in Step 2 for secondary verification. If the blade breakage probability exceeds the preset safety threshold, it will be directly eliminated, and the safe solution will be retained.

[0056] Step 5-3: Recalculate the surface roughness of the retained safe solution using the XGBoost prediction model. and tool wear Solutions that do not meet the requirements for surface roughness and tool wear are eliminated to obtain a set of feasible process solutions;

[0057] Step 5-4: Based on the selected set of feasible process solutions, the TOPSIS comprehensive optimization decision method is used. First, the target values ​​of the set of feasible process solutions are normalized, and the surface roughness is set. Tool wear Material removal rate The weights are calculated, and the solutions are compared with the positive ideal solution A. + and negative ideal solution A − The Euclidean distance is then used; then the relative proximity C is selected. i The largest solution is taken as the overall optimal solution.

[0058] Compared with existing technologies, this multi-objective optimization method for dry turning process parameters based on chipping risk constraints has the following advantages:

[0059] 1. Introducing hard constraints on tool chipping risk to improve machining safety: Introducing tool chipping risk as a hard constraint into the turning parameter optimization process, by constructing a chipping probability prediction model based on XGBoost and setting a safety-first threshold strategy, high-risk parameter combinations are actively eliminated during the optimization process. This can effectively solve the problem that traditional multi-objective optimization methods only focus on surface roughness and tool wear while ignoring catastrophic failure (chipping). It can avoid instantaneous tool breakage and workpiece scrapping caused by pursuing efficiency, and can significantly improve the reliability of dry turning processes.

[0060] 2. Constructing a physical information feature space to improve the accuracy of prediction models: Addressing the challenge of training high-precision models with small sample cutting data, this invention proposes a modeling method that integrates cutting physics mechanisms and data-driven approaches. This is achieved by explicitly embedding surface forming theory, wear mechanisms, and fracture mechanics into the feature space (e.g., constructing a physical information feature space). , By constructing a physical information feature set (including derived features), the physical interpretability of the model can be enhanced, and the generalization ability of the XGBoost model under small sample conditions can be significantly improved. Experiments show that surface roughness (and other derived features) The coefficient of determination of the prediction model Up to 0.9942, tool wear ( Coefficient of determination of the prediction model It can reach 0.9778, which can lay the foundation for subsequent high-precision optimization.

[0061] 3. Achieving synergistic optimization of processing efficiency, quality, and safety: This invention employs the NSGA-II algorithm combined with the TOPSIS decision method to solve for the Pareto optimal solution set under zero chipping risk. By coupling the chipping probability constraint, surface quality constraint, and tool life constraint, the optimal balance point that considers safety, quality, and efficiency is obtained. Actual verification results show that, compared with traditional empirical parameters, the optimized process scheme of this invention achieves a higher material removal rate (…) under the premise of zero chipping risk. The surface roughness was improved by 17.32%, and the surface roughness was improved by 17.32%. The wear on the tool flank was reduced by 3.22%, and the tool flank wear was also reduced. It reduces the efficiency by 16.84%, achieving a balance between safety, efficiency, and high quality in dry turning processes. Attached Figure Description

[0062] Figure 1 This is a flowchart of the present invention;

[0063] Figure 2 This is a schematic diagram of the experimental apparatus and data processing system according to an embodiment of the present invention;

[0064] Figure 3 This is a schematic diagram of chipping failure;

[0065] Figure 4 This is a schematic diagram of the derived feature construction and feature space based on physical mechanisms of the present invention;

[0066] Figure 5 This is an embodiment of the present invention. When the diameter is 0.8mm, the distribution of CCD design points on the plane is shown in the diagram, where (a) is... Plane, (b) is Plane, (c) is flat;

[0067] Figure 6 This is a graph showing the predicted and measured values ​​of an embodiment of the present invention, where (a) is... The predicted and measured values ​​are shown in the graph, (b) is A graph showing the predicted and measured values;

[0068] Figure 7 These are wear morphology diagrams of the flank face of the tool according to an embodiment of the present invention, wherein (a) is a wear morphology diagram with the narrowest wear band, (b) is a wear morphology diagram with a moderately extended wear band, and (c) is a wear morphology diagram with a significantly widened wear band.

[0069] Figure 8 These are illustrations of (a) the confusion matrix, (b) the ROC curve, (c) the F1 curve as a function of threshold, and (d) the Recall curve as a function of threshold in this embodiment of the invention.

[0070] Figure 9 This is a flowchart of the multi-objective optimization process based on the NSGA-II algorithm in an embodiment of the present invention;

[0071] Figure 10 This is a Pareto front visualization diagram of an embodiment of the present invention, wherein (a) is Figure (b) is shown. The diagram (c) shows a 3D viewpoint. The diagram (d) shows a 3D viewpoint. Diagram;

[0072] Figure 11 This is a flowchart of the discretization mapping and multi-constraint screening of the Pareto solution set in an embodiment of the present invention. Detailed Implementation

[0073] The invention will be further explained below using dry turning of Ti-6Al-4V titanium alloy as an example, in conjunction with the accompanying drawings.

[0074] The examples used dry turning of Ti-6Al-4V titanium alloy as the test object. Ti-6Al-4V titanium alloy is a typical difficult-to-machine material due to its low thermal conductivity and high chemical reactivity. During dry turning, the lack of cooling and lubrication leads to strong thermal-mechanical coupling in the cutting zone, making tool chipping a highly likely catastrophic failure mode. Most existing methods for optimizing turning parameters only consider surface roughness (…). ) and tool wear ( When continuous indicators are used as optimization objectives, explicit constraints on sudden and irreversible failure risks such as chipping are often neglected. This leads to optimized parameters that, while mathematically optimal, may fall into high-chipping-risk regions during actual machining, causing workpiece scrap or even machine tool damage. Furthermore, traditional modeling methods struggle to accurately describe both continuous wear patterns and discrete chipping boundaries with small sample experimental data. Therefore, this multi-objective optimization method for dry turning process parameters based on chipping risk constraints aims to address the challenge of balancing efficiency, quality, and tool safety in dry turning by introducing physical information features and chipping risk constraints. Figure 1 As shown, firstly, experimental design and data acquisition are conducted using a central composite design (CCD), and a feature space containing physical information is constructed; secondly, the XGBoost algorithm is used to construct surface roughness (…). Predictive models, tool flank wear ( The model predicts the chipping probability and the chipping risk; then, a safety threshold for chipping risk is set as a hard constraint; finally, the NSGA-II multi-objective genetic algorithm is used for optimization to obtain a model that balances processing efficiency. The Pareto optimal solution set for surface quality and tool life was determined; finally, the optimal process parameters were determined through decision analysis and verified through cutting experiments. Details are as follows:

[0075] Step 1: Experimental Design and Data Acquisition. This step aims to obtain a dry turning dataset through dry turning experiments and construct a physical information feature space containing derived features based on the cutting physics mechanism. Details are as follows:

[0076] Step 1-1: Build the physical layer experimental platform. For example... Figure 2 As shown, the embodiment uses a Shenyang Machine Tool CAK5085Di CNC lathe (equipped with a FANUC 0i-Mate TD system) as the machining platform. The workpiece material is an annealed Ti-6Al-4V titanium alloy round bar with dimensions of Φ50mm × 220mm. The clamping method adopts a "one-jaw, one-support" approach, that is, a three-jaw chuck clamps one end of the workpiece, and the tailstock center supports the other end to ensure machining rigidity. The cutting tool is a Kennametal KCS10 indexable carbide insert with a fine-grained tungsten carbide substrate and a PVD AlTiN coating. To study the tip radius... To mitigate the impact, three different blade specifications were selected in this embodiment: TCMT110304FP ( =0.4mm), TCMT110308FP ( =0.8mm) and TCMT110312FP ( =1.2mm), all mounted on a Kennametal STGCR1616H11 tool holder. The surface roughness of the workpiece after machining ( The profilometer (accuracy 0.05 μm) from Shanghai Optical Instrument Factory No. 6 was used for measurement, and the measurement parameters conformed to ISO 4287 standard (sampling length 0.8 mm, evaluation length 4 mm, Gaussian filtering). Tool flank wear ( High-precision measurements were performed using an Olympus BX51M metallurgical microscope (40×–1000×), with orthogonal illumination used to clearly capture the outline of the wear zone. The condition of the chipped edge was determined by direct visual observation. Figure 3 As shown, the basis is "the cutting edge has visible irregular notches, chipping, or material spalling".

[0077] Steps 1-2: Construct the physical information feature space. For example... Figure 4 As shown, to overcome the generalization limitations of purely data-driven models under small sample conditions, this step involves collecting basic turning parameters (cutting speed). Feed rate Back cut radius of the blade tip arc Building upon the existing model, further physical mechanisms are integrated to construct derived features, enhancing the model's physical interpretability of the cutting process. Specifically, for the chipping probability prediction model (Model C), features reflecting instantaneous load and fracture risk are constructed, including material removal rate (…). ), dimensionless parameters ( ), geometric ratio ( ) and the square term of velocity ( This study aims to characterize the impact of high stress concentration and impact loads on the cutting edge integrity during the cutting process; and to address surface roughness. The predictive model (Model A) constructs features that reflect the surface forming mechanism, including the theoretical roughness term. ), the square term of the feed rate ( ), geometric ratio ( ) and material removal rate ( This describes the influence of tool geometry and material plastic deformation on surface quality; it also addresses tool wear. The predictive model (Model B) constructs features that correlate cutting work with accumulated frictional heat, including material removal rate. ), dimensionless parameters ( ), geometric ratio ( ) and the square term of velocity ( The purpose is to reflect the progressive wear of the flank material caused by mechanical friction and thermal effects during the cutting process. All derived features have clear physical meanings, and together with the basic process parameters, they constitute a high-dimensional physical information feature set, which serves as the input for the subsequent XGBoost prediction model.

[0078] Steps 1-3 involve experimental design and data acquisition. The example uses a Central Composite Design (CCD) approach to arrange the experiment, fully exploring the process parameter space. The cutting speed is selected. Feed rate Back cut and the radius of the blade tip arc Four variables are used as input factors. Among them, As a continuous variable, its process parameter range is set as follows: m / min, mm / r, mm; radius of the blade tip arc For discrete variables, the values ​​are: mm. The design includes 8 factor points and 6 axial points (axial coefficients). ) and 5 center points, blade tip radius When the center is 0.8 mm (center horizontal), the CCD experimental point is at... , and Projection distribution on the plane as follows Figure 5 As shown. This design, through the rational layout of factor points, axial points, and center points, achieves uniform coverage of the parameter space, accurately reflecting the symmetrical structure and spatial distribution characteristics of the CCD. Combined with discrete variables... For different values ​​of , a total of 51 experimental schemes were designed, as shown in Table 1 below. Dry turning experiments were carried out according to the designed schemes, and the surface roughness value corresponding to each experimental group was recorded. ) and tool flank wear ( For the test group where chipping occurred (chipping label 1), no valid data could be obtained due to machining failure. and Marked as N / A. Furthermore, the material removal rate (N / A) is calculated based on the turning parameters. This serves as a production efficiency indicator. Ultimately, a system is constructed that includes basic process parameters, physically derived characteristics, and multi-objective response values. A complete cutting dataset with chipped edge state labels was used for subsequent model training and optimization.

[0079]

[0080] Step 2: Construct a multi-objective prediction model based on physical information features and XGBoost. This step aims to train an XGBoost ensemble learning model using the physical information feature space dataset constructed in Step 1. This model is used to establish the relationship between turning parameters and machining quality (surface roughness). Tool wear The nonlinear mapping relationship between ( ) and failure risk (scratching). Specifically:

[0081] Step 2-1: Dataset partitioning and physical feature construction.

[0082] The 51 sets of experimental data collected in Step 1 were divided into training and testing sets. To improve the model's generalization ability and physical interpretability, a random reordering method was used to shuffle and split the data before partitioning. For different prediction targets, feature vectors containing physical mechanisms were constructed. (See Step 1 for details), output as surface roughness respectively. Tool wear Surface roughness with continuous values ​​( Predictive models and tool wear ( The model includes a prediction model for chipping, and a chipping probability prediction model that outputs the probability of chipping. The chipping probability prediction model was trained using 51 samples including 16 chipping events, and introduced a class weight parameter `scale_pos_weight` (number of negative samples / number of positive samples) to address the sample imbalance problem. Meanwhile, surface roughness (…) Predictive models and tool wear ( The prediction models were all trained using 33 samples that did not break (2 outliers were removed).

[0083] Step 2-2: Construct the XGBoost ensemble learning model.

[0084] The objective function of the XGBoost model consists of a loss function and a regularization term. It works by minimizing the objective function. To iteratively generate a decision tree. Objective function The definition is as follows:

[0085]

[0086] In the formula: The number of samples in the training set; In this embodiment, a mean squared error loss function is used to measure the difference between predicted and actual values, while a logistic regression loss function (LogLoss) is used to output probability values ​​for the regression tasks of surface roughness and tool wear. Indicates the first The true values ​​of each sample (for surface roughness regression prediction model and tool wear regression prediction model, the true values ​​are continuous measured values; for chipping probability prediction model, the true values ​​are binary classification labels 0 or 1); among which Indicates the first The model prediction output value for each sample (for the surface roughness regression prediction model and the tool wear regression prediction model, the model prediction output value is the predicted continuous value; for the chipping probability prediction model, the model prediction output value is the predicted probability of chipping). This represents the total number of decision trees; This is a regularization term used to control model complexity and prevent overfitting. Its specific form is determined by the number of leaf nodes and the magnitude of the leaf node weights. This represents the structure of the k-th decision tree and the weights of its leaf nodes.

[0087] Regularization term The definition is as follows:

[0088]

[0089] In the formula: This represents the number of leaf nodes; The weight of the leaf node. The L2 norm of the weight vector (i.e., the sum of squares of the weights of all leaf nodes). ); and These are the regularization coefficients that control the number of leaves and the regularization coefficients that control the weight of the leaves, respectively.

[0090] For surface roughness ( Regression prediction model and tool wear ( The regression prediction model uses mean squared error (MSE) as the loss function, and its expression is as follows:

[0091]

[0092] In the formula: The number of samples in the training set; Indicates the first The true value of each sample, i.e., the surface roughness value. Or the amount of wear on the back face of the cutting tool The measured values; Indicates the first The model's predicted output value for each sample, i.e., the predicted surface roughness value. Or the amount of wear on the back face of the cutting tool .

[0093] For the blade breakage probability prediction model, log loss (LogLoss) is used as the loss function, and its expression is as follows:

[0094]

[0095] In the formula: The number of samples in the training set; Indicates the first The true value (i.e., the true label) of each sample, where a true value of 1 indicates a broken blade and a true value of 0 indicates safety; Indicates the first The model's predicted output value for each sample, i.e., the model's prediction of the sample. The probability of the predicted outcome is "Broken Blade".

[0096] Steps 2-3: Bayesian hyperparameter optimization.

[0097] To overcome the inefficiency of traditional mesh search and to find the global optimal solution of the XGBoost model under small sample data of cutting processes, this embodiment uses the Optuna framework for Bayesian hyperparameter optimization. For regression tasks (surface roughness Ra, tool wear, etc.)... Based on the differences in characteristics between the task of predicting the probability of a broken blade and the task of predicting the probability of a broken blade, we construct differentiated search spaces and objective functions respectively.

[0098] For surface roughness ( ) and tool wear ( The search space constructed by Optuna contains the key hyperparameters shown in Table 2 below. The optimization strategy adopts the TPE (Tree-structured Parzen Estimator) sampling algorithm, with a maximum number of iterations set to 30, and the mean absolute error (MAE) is used as the basis for prediction. () serves as the direction for minimizing the objective function.

[0099]

[0100] For the task of predicting the probability of chipping, which addresses the risk of chipping, the search space is specifically adjusted due to the imbalance of sample classes (the number of samples that have not chipped is far greater than the number of samples that have chipped). Specifically, max_depth, learning_rate, subsample, and colsample_bytree are related to surface roughness. ) and tool wear ( The hyperparameters of Optuna in the prediction are the same (see Table 2). The search space for the other key hyperparameters of Optuna is shown in Table 3. The TPE sampling algorithm is also used, and the maximum number of iterations is set to 50. The objective function is defined as 1 - F1 score, that is, to find the hyperparameter combination with the largest F1 score during the optimization process.

[0101]

[0102] Steps 2-4: Model training and leave-one-out cross-validation (LOO-CV).

[0103] The optimal combination of hyperparameters found by the Optuna Bayesian optimization algorithm is used to train and evaluate the model.

[0104] For regression model training, the Leave-One-Out Cross-Validation (LOO-CV) method is used for evaluation. This involves reserving one sample as the validation set and using the remaining samples as the training set, repeating this process until all samples have been validated once. This is achieved through calculation... , and Performance was verified using metrics such as [list of metrics], and the results are as follows: Figure 6 As shown. The results indicate that, with the small sample data of the examples, Model A... Prediction =0.9942, =0.0515μm, =0.0918μm, indicating that the model can effectively capture the influence of turning parameters on surface roughness; Model B on Prediction =0.9778, =0.552μm, The accuracy is 1.097 μm, which is high and meets the requirements of engineering applications. The scatter plots of the predicted and measured values ​​closely follow the ideal prediction line, indicating that the model has high reliability under typical processing conditions.

[0105] The wear morphology of the back face under three typical working conditions is as follows: Figure 7 As shown, where: Figure 7 The wear band in Figure (a) is the narrowest, indicating a lower cutting load; Figure 7 The wear zone in Figure (b) is moderately extended, with slight local oxidation discoloration. Figure 7 In Figure (c), the wear zone is significantly widened, surface roughness increases, and there is obvious material transfer. It is evident that with changes in processing parameters, the degree of wear evolves from light to heavy, and its morphological characteristics are consistent with the model predictions. The value increases in line with the increase of process parameters, which further verifies the effectiveness of the prediction model.

[0106] For training the blade breakage probability prediction model, a blade breakage probability prediction model (Model C) is constructed. The LOO-CV strategy is also used to obtain the predicted probabilities for the entire sample. The value of 0.8857 indicates that the model has a good ability to detect the risk of blade breakage.

[0107] Step 3: Setting the Scratching Risk Threshold Based on a Safety-First Strategy. This step aims to transform the scraping probability prediction model (Model C) trained in Step 2 into a hard constraint that can be used for multi-objective optimization. By setting the "maximum tolerable false negative rate," a safety threshold τ for the scraping risk probability is determined, ensuring high reliability of the optimization results in engineering applications. Details are as follows:

[0108] Step 3-1: Define the security constraint logic.

[0109] In the dry turning of Ti-6Al-4V titanium alloy, tool chipping is a catastrophic failure mode. Unlike gradual flank wear, chipping, once it occurs, will lead to instantaneous workpiece scrap, tool damage, and even machine tool damage. To ensure absolute safety during the machining process, the embodiment adopts a "safety first" constraint strategy.

[0110] The core of this strategy lies in setting a critical value for the probability of a blade breakage. In subsequent optimization processes, only when the model predicts the probability of blade collapse... Only when this set of process parameters is met is it considered a "safe and feasible solution"; otherwise, if If the parameter combination is deemed high-risk, it will be eliminated. The mathematical expression for the constraint strategy is as follows:

[0111]

[0112] In the formula: The probability of blade breakage predicted by the model; This is the critical value for the probability of blade breakage.

[0113] Step 3-2: Determine the maximum tolerable false negative rate and search strategy.

[0114] To scientifically quantify the critical value of the probability of blade breakage. The implementation example introduces the "false negative rate" as an evaluation metric. The false negative rate refers to the proportion of samples where chipping actually occurs that are incorrectly predicted as "safe" by the model. A target false negative rate range is set. Based on statistical analysis of 51 sets of experimental data and engineering experience in titanium alloy processing, the implementation example sets an upper limit of 10% and a lower limit of 5% for the maximum tolerable false negative rate. This means that the model is allowed to fail to identify chipping risks in a very small number of extreme cases, but it must ensure that the vast majority of risks are effectively intercepted. A process window maximization strategy is adopted; among all candidate thresholds that satisfy the above false negative rate constraints, the implementation example tends to select the threshold with the largest value. This is because a higher probability threshold means that the optimization algorithm is subject to more relaxed constraints during the search, enabling it to explore a wider range of process parameter spaces, thereby avoiding situations where the solution set is too small or there is no solution.

[0115] Step 3-3: Calculate and determine the critical value for the final probability of blade breakage. .

[0116] Using the XGBoost "Break" probability prediction model trained in Step 2, predict the validation set data and output the probability value of each sample belonging to the "Break" category. The specific calculation steps are as follows:

[0117] First, a relationship curve is plotted, and the recall and false negative rates are calculated for each cutoff point (from 0 to 1). Second, a reverse search algorithm is used to filter out all probability thresholds that meet the set conditions within the set target false negative rate range (5%–10%). Finally, the optimal threshold is determined, and the highest probability value from the above set is selected as the final critical value for the probability of blade breakage, based on the "maximizing process window strategy". .

[0118] Based on the test results of this embodiment (such as...) Figure 8 As shown in the figure, when the target false negative rate is set to be no more than 7%, the maximum probability threshold for satisfying this security constraint, calculated by the above algorithm, is: =0.0809. At this threshold, combined with... Figure 8 The performance of the analysis model is as follows: ① Recall rate is significantly improved: such as Figure 8 As shown in Figure (d), when the threshold is adjusted to When the value is 0.0809, the model's recall rate increases from the default threshold ( The value of 0.7500 under the condition of 0.5 was significantly improved to 0.9375, which means that out of 16 samples of actual blade breakage, the model successfully identified 15, reducing the actual false negative rate to 6.25%, meeting the engineering safety requirement of less than 7%. ② Classification boundary adjustment: such as Figure 8 As shown in Figure (a), in the baseline state ( At a threshold of 0.5, there were 4 false positives (FP). Further lowering the decision threshold would increase the number of false positives (i.e., more safe samples would be misclassified as false positives), but this reflects the "safety first" strategy—sacrificing some precision for extremely high recall, thereby minimizing the risk of false negatives and ensuring the safety of engineering applications. ③ Model discriminative ability: such as... Figure 8 As shown in Figure (b), the ROC curve of the model... The value reached 0.8857, indicating that the model possesses good risk discrimination ability for blade collapse under different threshold settings, and can support the adjustment of the aforementioned thresholds. This maximum probability threshold... =0.0809 will be used as the hard constraint boundary in the subsequent NSGA-II multi-objective optimization algorithm.

[0119] Step 4: Multi-objective process parameter optimization based on the NSGA-II algorithm. This step aims to use the XGBoost prediction model constructed in Step 2 as the fitness evaluation function and the safety threshold set in Step 3 as the core constraint. A non-dominated sorting genetic algorithm with an elitist strategy (NSGA-II) is employed to globally optimize the cutting process parameters, achieving the optimal balance between surface quality, tool life, and machining efficiency. The flowchart of the multi-objective optimization based on the NSGA-II algorithm is shown below. Figure 9 As shown, the details are as follows:

[0120] Step 4-1, parameter initialization.

[0121] The NSGA-II algorithm searches for the optimal Pareto front solution in the solution space by simulating the biological evolution process. The core operating parameters of the algorithm are set as follows: population size is initialized to 50, maximum number of generations is set to 100, simulated binary crossover probability is set to 0.9, and polynomial mutation probability is set to 0.1. In terms of encoding, decision variables are encoded as real numbers, meaning all parameters participate in the genetic operation directly as floating-point numbers. This method avoids decoding errors associated with traditional binary encoding and is more suitable for handling search problems in high-dimensional continuous spaces.

[0122] Step 4-2, Decision variables and fitness assessment.

[0123] Define decision variable vector Including cutting speed Feed rate Back cut and the radius of the blade tip arc To facilitate efficient search of genetic operators, the radius of the knife tip arc... In the optimization iteration process, it is treated as a continuous variable, and its value range is set to a continuous interval. The algorithm randomly initializes an initial population of 50 individuals. Then, it uses the XGBoost prediction model trained in Step 2 (including the surface roughness model, tool wear model, and chipping prediction model) as the fitness function. For each individual in the population (i.e., a set of process parameters), its corresponding surface roughness is calculated. Predicted values, tool wear Predicted values ​​and material removal rates .

[0124] Step 4-3: Construct a multi-objective mathematical model for optimizing turning parameters.

[0125] Based on the above variables and evaluation methods, a multi-objective optimization mathematical model for turning parameters is constructed, as follows:

[0126]

[0127] In the formula: For a set of decision variables, ,in It is the cutting speed (m / min). It is the feed rate (mm / r). It is the depth of cut (mm). It is the radius of the blade tip arc (mm); The objective function is surface roughness. Surface roughness values ​​predicted by the model (μm); The objective function is the tool wear parameter. The amount of flank wear predicted by the model (mm); The objective function is the material removal rate; Material removal rate (cm) 3 / min), its calculation formula is: (Will (The unit is converted from m / min to mm / min). The negative value is used here to transform the maximization problem into a minimization problem.

[0128] The optimization process must satisfy variable boundary constraints and process safety constraints. This is particularly relevant regarding the probability of blade breakage. The nonlinear constraints are handled using an external penalty function method: when the individual's probability of blade breakage... When the value exceeds the set threshold of 0.0809, a very large penalty factor is applied to its objective function value. This puts them at a disadvantage in the non-dominated ranking and eliminates them. The decision variable boundary constraints and penalty strategy of the example are expressed as follows:

[0129]

[0130] In the formula: 15 and 170 represent the cutting speed (m / min), respectively. The lower and upper limits; The value represents the feed rate (mm / r), where 0.03 and 0.375 represent the feed rates, respectively. The lower and upper limits; The depth of cut is (mm), 0.16 and 1.36 are the depth of cuts respectively. The lower and upper limits; The radius of the blade tip arc (mm) is 0.4 and 1.2, respectively. The lower and upper limits; τ is the predicted probability of blade breakage under the current parameter combination; τ is the critical value of the probability of blade breakage. For the first Each component of the optimization objective function (Right now Represents the objective function for surface roughness. Represents the objective function for tool wear. (Represents the objective function for material removal rate). The penalty factor (takes a sufficiently large positive number, in this embodiment) 10 6 This is used to penalize solutions that exceed the risk threshold of blade breakage, thus eliminating them during the optimization process.

[0131] Step 4-4: Evolutionary Iteration and Pareto Front Generation.

[0132] In each generation of evolution, the algorithm performs the following operations: First, it performs a non-dominated sort, based on the objective function value. The population is stratified, prioritizing the retention of individuals at higher non-dominated levels. Crowding is then calculated within each level to maintain population diversity. Genetic operations are then performed using a binary tournament selection operator, generating offspring through simulated binary crossover (SBX) and polynomial mutation. Finally, an elitist strategy update is implemented, merging parents and offspring, and selecting the 50 best individuals to form the next generation through fast non-dominated sorting and crowding comparison. This process is repeated until the maximum number of generations (100) is reached. The algorithm outputs the initial Pareto optimal solution set. The distribution of the Pareto optimal solution set obtained by the NSGA-II algorithm is shown below. Figure 10 As shown, the blue scatter points represent Pareto optimal solutions, and the green scatter points represent solutions that satisfy engineering constraints (…). ≤1.0μm, A feasible solution for (≤30μm). From Figure 10 It can be seen that surface quality ( ), tool life ( ) and processing efficiency ( There is a significant game-theoretic relationship between them: in low... area, Generally limited; while in high area, and This indicates that simply pursuing the extreme value of a single indicator will lead to the deterioration of other performance metrics. Therefore, it is necessary to further determine the comprehensive optimal parameters from the green feasible solutions (i.e., proceed to Step 5).

[0133] Step 5: Post-processing, screening, and synthesis decision of the Pareto optimal solution set. This step aims to perform post-processing screening on the generated initial Pareto optimal solution set, transforming the theoretical continuous solution into a discrete engineering solution that meets the actual processing requirements. The screening process is as follows: Figure 11 As shown, the details are as follows:

[0134] Step 5-1, Discretization and Mapping of Tool Parameters. This involves considering the continuous tool tip radius within the Pareto solution set. The nearest neighbor mapping algorithm is used to discretize it. Calculation Set of ISO standard tool specifications The distance between each element is used to select the nearest standard value as the mapped radius. This step aims to address the mismatch between the continuous parameters output by the optimization algorithm and the actual discrete tool specifications.

[0135] Step 5-2, Secondary verification of blade breakage risk. The discretized parameters are combined (… The model is then substituted into a pre-trained chipping probability prediction model (Model C) for secondary verification. This is due to the radius of the blade tip arc. Changes may affect cutting stability, and the probability of chipping needs to be reassessed. .like Exceeding the preset security threshold ( If the value is less than or equal to 0.0809, the solution is considered to have a security risk and is directly eliminated; otherwise, it is retained as a "safe solution".

[0136] Step 5-3, Quality and Lifetime Constraint Assessment. The retained "safe solution" is recalculated using the XGBoost prediction model to determine the surface roughness after mapping. and tool wear Eliminate those that do not meet the basic process requirements (i.e.) >1.0μm and The solution for >30μm was obtained, and finally a set of feasible process schemes was obtained.

[0137] Step 5-4, TOPSIS Comprehensive Optimization Decision. Based on the selected set of feasible process solutions, the TOPSIS comprehensive optimization decision is used to output the final result. First, the target values ​​of the feasible set are normalized to eliminate the influence of dimensions; then, the surface roughness is set using the approximation of ideal solution ranking method. Tool wear Processing efficiency The weights (each accounting for 1 / 3) are used to calculate the sum of the weights of each solution and the positive ideal solution A. + and negative ideal solution A − The Euclidean distance is then used; then the relative proximity C is selected. i The largest solution is taken as the overall optimal solution, which achieves the best balance between machining efficiency, surface quality, and tool life while ensuring safety. Table 4 below shows a comparison between the overall optimal solution of the example and the empirical parameters used in long-term field operations.

[0138]

[0139] To verify the effectiveness and safety of the optimal process parameter combination derived from Step 5 in actual machining, an actual cutting experiment was conducted, as follows:

[0140] ① Dry turning test: Select the optimal combination of process parameters output in Step 5 ( As the test group, the empirical parameters used in the enterprise's long-term field operations, as shown in Table 4, were selected as the control group (benchmark group). All verification tests were conducted on the physical layer experimental platform of Step 1. To ensure the impartiality of the comparison, the dry turning method was still used, and the workpiece was clamped using a "one clamp, one support" method. For the test consumables and standards, a brand new Kennametal TCMT110308FP insert was used, mounted on a Kennametal STGCR1616H11 tool holder, with a cutting length set to 400 mm. Surface roughness measurements strictly followed the ISO 4287 standard.

[0141] ② Data Acquisition and Accuracy Assessment: During the experiment, consistency was maintained in machine tool condition, tool batch, and workpiece material properties. The surface roughness of the machined workpiece was measured using an SRA-2100 profilometer from Shanghai Optical Instrument Factory No. 6. Measured values; the amount of wear on the back face of the cutting tool was measured using an Olympus BX51M metallographic microscope. Measured values ​​were used, and visual observation was employed to assist in determining whether chipping had occurred. For model accuracy verification, the measured values ​​were compared with the predicted values ​​from the XGBoost model in Step 2. , The models were compared, and the relative error was calculated to evaluate the model accuracy. The comparison results are shown in Table 5 below. The relative error is 3.51%. The relative error was 4.66%. Both indicators were within the acceptable range for engineering (typically, indicator errors are <10% to 15%), and no chipping occurred during the experiment, verifying the reliability of the recommended parameters.

[0142]

[0143] ③ Comprehensive evaluation of optimization effect: The performance of the optimized scheme is compared with that of traditional empirical parameters. The evaluation indicators include material removal rate ( ), surface roughness ( ) and flank wear ( The judgment criterion is that if the optimized solution meets the quality constraints ( ≤1.0μm) and safety constraints ( Under the premise of ≤30μm and no chipping, its processing efficiency ( If the material removal rate is significantly higher than the empirical parameter, then the optimization method is deemed effective. The implementation comparison results are shown in Table 6 below. The comparison results show that, compared to the empirical parameter, the optimized solution has a significantly higher material removal rate (…). The surface roughness was improved by approximately 17.32%, while the surface roughness was also improved. Reduced by 3.22%, flank wear The efficiency was reduced by 16.84%. This result proves that the multi-objective optimization method for dry turning process parameters based on chipping risk constraints can achieve a substantial improvement in machining efficiency while ensuring process safety and machining quality.

[0144]

[0145] This multi-objective optimization method for dry turning process parameters based on chipping risk constraints constructs a high-precision prediction model by integrating cutting physics and data-driven technology, and explicitly transforms the chipping risk probability into a hard constraint condition. It can achieve synergistic optimization of machining efficiency, surface quality and tool life while ensuring zero chipping risk.

Claims

1. A multi-objective optimization method for dry turning process parameters based on chipping risk constraints, characterized in that, Specifically, the following steps are included: Step 1, Experimental Design and Data Acquisition: Build a dry turning physical layer experimental platform, design experiments, and conduct dry turning experiments group by group according to the experimental design scheme to obtain a dry turning dataset including basic turning process parameters, physical derived features, and chipping state labels. Step 2, construct a multi-objective prediction model based on physical information features and XGBoost: Based on the dry turning dataset obtained in Step 1, construct and train an ensemble learning model based on XGBoost. The ensemble learning model based on XGBoost includes a surface roughness regression prediction model, a tool wear regression prediction model, and a chipping probability prediction model. Step 3, setting the risk threshold for blade breakage based on the safety-first strategy: By setting the "maximum tolerable false negative rate", the safety threshold τ of the blade breakage risk probability is determined, and the blade breakage probability prediction model trained in Step 2 is transformed into a hard constraint that can be used for multi-objective optimization. Step 4, Multi-objective process parameter optimization solution based on NSGA-II algorithm: Construct a multi-objective optimization mathematical model with the objective functions of maximizing material removal rate, minimizing surface roughness and minimizing tool back face wear, and the safety threshold τ of the chipping risk probability determined in Step 3 as the constraint condition. Use NSGA-II algorithm to perform global optimization and solve the Pareto optimal solution set that takes into account processing efficiency, quality and safety. Step 5, Post-processing screening and comprehensive decision-making of the Pareto optimal solution set: Post-processing screening is performed on the Pareto optimal solution set obtained in Step 4 to transform the theoretical continuous solution into a discrete engineering solution that meets the actual processing requirements, and finally obtain the optimal combination of process parameters.

2. The multi-objective optimization method for dry turning process parameters based on chipping risk constraints according to claim 1, characterized in that, Step 1 is as follows: Step 1-1: Construct a dry turning physical layer experimental platform and collect basic turning process parameters, including cutting speed. Feed rate Back cut radius of the blade tip arc ; Steps 1-2: Based on the physical mechanism of cutting, a high-dimensional physical information feature set is constructed, including basic turning process parameters and physical derivative features. These physical derivative features include geometric features reflecting the surface forming mechanism, energy consumption features reflecting cutting work and frictional heat accumulation, and mechanical features reflecting instantaneous load and fracture risk. Both the energy consumption features reflecting cutting work and frictional heat accumulation and the mechanical features reflecting instantaneous load and fracture risk include material removal rate. Dimensionless parameters Geometric ratio and the square of the cutting speed term Geometric features reflecting the surface forming mechanism include theoretical roughness terms. The square of the feed rate Geometric ratio and material removal rate ; Steps 1-3, with cutting speed Feed rate Back cut and the radius of the blade tip arc Four variables were used as input factors in the experimental design. Dry turning tests were conducted group by group according to the experimental design. The chipping status of the test group that experienced chipping was labeled as 1, and the chipping status of the test group that did not experience chipping was labeled as 0. The surface roughness value of each test group that did not experience chipping was recorded. and tool flank wear and the material removal rate As an indicator of production efficiency, the final structure includes a high-dimensional physical information feature set and multi-objective response values. A complete dry turning dataset with chipping state labels.

3. The multi-objective optimization method for dry turning process parameters based on chipping risk constraints according to claim 2, characterized in that, Step 2 is as follows: Step 2-1, Dataset Partitioning and Prediction Model Construction: The experimental data of the dry turning dataset is divided into training and testing sets. Based on the XGBoost model, a feature vector containing geometric features reflecting the surface forming mechanism is constructed, and the output is the surface roughness. A continuous numerical surface roughness regression prediction model is constructed, which includes a feature vector reflecting energy consumption characteristics of cutting work and frictional heat accumulation, and the output is tool wear. A continuous numerical regression prediction model for tool wear is constructed, which includes feature vectors reflecting mechanical characteristics of instantaneous load and fracture risk, and outputs a chipping probability prediction model as the probability of chipping. The objective function of the XGBoost model The definition is as follows: In the formula: The number of samples in the training set; Let be the loss function used to measure the difference between predicted and true values, where Indicates the first The true value of each sample Indicates the first The model predicts the output value for each sample; This represents the total number of decision trees; This is a regularization term used to control model complexity and prevent overfitting, where This represents the structure of the k-th decision tree and the weights of its leaf nodes; Regularization term The definition is as follows: In the formula: This represents the number of leaf nodes; The weight of the leaf node. Let L2 be the norm of the weight vector; and These are the regularization coefficients that control the number of leaves and the regularization coefficients that control the weight of the leaves, respectively. For the surface roughness regression prediction model and the tool wear regression prediction model, the mean squared error is used as the loss function, and its expression is as follows: In the formula: The number of samples in the training set; Indicates the first The true value of each sample, i.e., the surface roughness value. Or the amount of wear on the back face of the cutting tool The measured values; Indicates the first The model's predicted output value for each sample, i.e., the predicted surface roughness value. Or the amount of wear on the back face of the cutting tool ; For the blade breakage probability prediction model, log loss is used as the loss function, and its expression is as follows: In the formula: The number of samples in the training set; Indicates the first The true value of each sample, where a true value of 1 indicates a broken blade and a true value of 0 indicates safety; Indicates the first The model's predicted output value for each sample, i.e., the model's prediction of the sample. The probability of the predicted outcome being "Broken Blade"; Step 2-2: Train the ensemble learning model.

4. The multi-objective optimization method for dry turning process parameters based on chipping risk constraints according to claim 3, characterized in that, In Step 2-2, before training the ensemble learning model, Bayesian hyperparameter optimization is performed using the Optuna framework: For the regression prediction tasks of surface roughness and tool wear, the search range of key hyperparameters such as the maximum tree depth, learning rate, minimum weight of child nodes, and sample and feature sampling rates are set; for the chipping probability prediction task, the minimum loss decrease value is introduced as an optimization variable based on the regression task; and the TPE sampling algorithm is used for iterative search to minimize the mean absolute error. The optimal combination of hyperparameters for the XGBoost model is determined by maximizing the F1 score as the fitness objective for both the regression task and the blade breakage probability prediction task. The ensemble learning model is trained using the optimal hyperparameter combination determined by the Bayesian optimization algorithm.

5. The multi-objective optimization method for dry turning process parameters based on chipping risk constraints according to claim 4, characterized in that, Step 3 is as follows: Step 3-1, Introduce the critical value for the probability of blade breakage. Construct a safety-first constraint logic to eliminate predicted probabilities that exceed the limit. The mathematical expression for the combination of process parameters and the constraint strategy is as follows: In the formula: The probability of blade breakage predicted by the model; This is the critical value for the probability of blade breakage. Step 3-2: Introduce the maximum tolerance for false negatives index to construct the critical value for the probability of a crash. The target missed rate screening range is determined, and the maximum value within the target missed rate screening range is selected to maximize the process window; Step 3-3: Iterate through the selection process, calculate and determine the critical value for the final probability of blade breakage. The critical value of the final probability of blade breakage. As a hard constraint boundary in the subsequent NSGA-II multi-objective optimization algorithm.

6. The multi-objective optimization method for dry turning process parameters based on chipping risk constraints according to claim 5, characterized in that, In Step 3-2, the critical value of the probability of blade breakage. The upper and lower limits of the target false negative rate screening interval are set based on statistical analysis of experimental data and engineering experience in dry turning of specific materials.

7. The multi-objective optimization method for dry turning process parameters based on chipping risk constraints according to claim 5, characterized in that, In Step 3-3, calculate and determine the critical value of the final probability of blade breakage. Then, using the XGBoost "Break" probability prediction model trained in Step 2, predictions are made on the validation set data, and the probability value of each sample belonging to the "Break" category is output. The specific calculation steps are as follows: First, a relationship curve is plotted, and the recall and false negative rates are calculated for each cutoff point by traversing different probability cutoff points. Second, a reverse search algorithm is used to filter out a set of probability thresholds that meet the set of conditions within the set of target false negative rate filtering intervals. Finally, the optimal threshold is determined, and the highest probability value is selected from the set of probability thresholds as the final critical value for the probability of blade breakage. .

8. The multi-objective optimization method for dry turning process parameters based on chipping risk constraints according to claim 5, characterized in that, Step 4 is as follows: Step 4-1: Initialize the NSGA-II algorithm parameters, encode the decision variables using real number encoding, and set the population size, maximum number of generations, crossover and mutation probabilities; Step 4-2, Define including cutting speed Feed rate Back cut and the radius of the blade tip arc The decision variable vector is used, and the prediction model trained in Step 2 is called as the fitness function to calculate the corresponding surface roughness for each individual in the population. Predicted values, tool wear Predicted values ​​and material removal rates ; Step 4-3, construct to minimize surface roughness Predicted values ​​and tool wear Predicted value, maximized material removal rate A multi-objective optimization mathematical model for turning parameters is proposed, and boundary constraints of decision variables are set and the nonlinear constraints of chipping risk are handled by the external penalty function method. The mathematical model for multi-objective optimization of turning parameters is expressed as follows: In the formula: For a set of decision variables, ; The objective function is surface roughness. Surface roughness values ​​predicted by the model ; The objective function is the tool wear parameter. The amount of flank wear predicted by the model ; The objective function is the material removal rate; For material removal rate, ; The boundary constraints and penalty strategies for decision variables are expressed as follows: In the formula: and Cutting speed The lower and upper limits; and feed rate The lower and upper limits; and They are respectively the back cut amount The lower and upper limits; and These are the radii of the blade tip arc. The lower and upper limits; τ is the predicted probability of blade breakage under the current parameter combination; τ is the critical value of the probability of blade breakage. For the first Each component of the objective function is optimized. ; As a penalty factor; Step 4-4: Perform evolutionary iteration. The population is stratified and screened by non-dominated sorting and crowding calculation. Offspring are generated by binary tournament selection, simulated binary crossover and polynomial mutation. The best individuals are retained by using an elite strategy. The process is iterated until the maximum number of generations is reached, and the Pareto optimal solution set is output.

9. The multi-objective optimization method for dry turning process parameters based on chipping risk constraints according to claim 8, characterized in that, In Step 4, the radius of the blade tip arc... It is treated as a continuous variable during the optimization iteration process.

10. The multi-objective optimization method for dry turning process parameters based on chipping risk constraints according to claim 8, characterized in that, Step 5 is as follows: Step 5-1, for the continuous tool tip radius in the Pareto solution set The nearest neighbor mapping algorithm is used to map it to discrete tool specification values ​​that conform to ISO standards. Step 5-2: Substitute the discretized parameter combination into the blade breakage probability prediction model trained in Step 2 for secondary verification. If the blade breakage probability exceeds the preset safety threshold, it will be directly eliminated, and the safe solution will be retained. Step 5-3: Recalculate the surface roughness of the retained safe solution using the XGBoost prediction model. and tool wear Solutions that do not meet the requirements for surface roughness and tool wear are eliminated to obtain a set of feasible process solutions; Step 5-4: Based on the selected set of feasible process solutions, the TOPSIS comprehensive optimization decision method is used. First, the target values ​​of the set of feasible process solutions are normalized, and the surface roughness is set. Tool wear Material removal rate The weights are calculated, and the solutions are compared with the positive ideal solution A. + and negative ideal solution A − The Euclidean distance is then used; then the relative proximity C is selected. i The largest solution is taken as the overall optimal solution.

Images