Coating process parameter optimization method and system

Abstract

The invention belongs to the technical field of coatings, and discloses a coating process parameter optimization method and system, and the method comprises the steps: constructing a prediction model between process parameters and coating offline characterization data, and constructing a prediction model between the coating offline characterization data and the cutting performance of a tool, and jointly forming a prediction model between the process parameters and the cutting performance of the tool, and process parameter optimization search is carried out based on the model, so that prediction association between the process parameters and the cutting performance of the cutter can be established, optimization selection of the process parameters is realized, and a process parameter combination for improving the cutting performance of the cutter is obtained.

Description

Technical Field

[0001] The present application relates to the field of coating technology, and in particular to a coating process parameter optimization method and system. Background Art

[0002] High-speed cutting, a key technology in modern manufacturing, places extremely stringent demands on the performance of the cutting tools used. To cope with the high-speed, high-temperature, high-stress, and high-wear operating environments, the surfaces of high-speed cutting tools are often enhanced with multilayer composite coatings. These high-performance multilayer composite coatings are commonly produced using physical vapor deposition (PVD) processes. The coating's ultimate performance, such as hardness, toughness, substrate adhesion, and wear resistance, is directly and closely related to its internal microstructure and the compositional distribution of its constituent materials. Multilayer composite coatings are formed by the sequential deposition of multiple layers of different materials, resulting in extremely complex internal structural features, including but not limited to the grain size and crystal orientation of each layer, the thickness ratio between layers, and the bonding state and defects at the interlayer interfaces. The formation of these fine internal structural features is directly and significantly influenced by PVD process parameters, such as the environmental parameters within the deposition chamber (vacuum level, atmosphere), the flow rates of various process gases, the temperature of the substrate, the bias voltage applied to the substrate, the sputtering power of different target materials, and the total deposition time.

[0003] However, the physical vapor deposition process is a highly complex system with nonlinear and complex coupling relationships between its various process parameters. At the same time, the operating conditions of the actual deposition equipment, such as power supply stability, gas flow control accuracy, etc., also have inherent volatility. This leads to the fact that even when exactly the same combination of process parameters is set, there may be non-negligible differences in the internal microstructure of the coating actually deposited. Furthermore, during the entire process of coating deposition, it is still extremely difficult to achieve real-time, online and accurate measurement of the detailed microstructure (such as real-time deposition rate, crystal growth state) or composition distribution of each layer inside the multi-layer composite coating. Existing process monitoring methods can usually only provide some macroscopic and indirect information, and it is difficult to fully grasp the dynamic formation process of the microstructure. Offline characterization and analysis after deposition is completed, such as observing the layered structure and micromorphology of the coating cross-section through a scanning electron microscope (SEM), analyzing the crystal structure and texture of each layer using X-ray diffraction (XRD), or analyzing the composition distribution using energy dispersive spectroscopy (EDS), can provide some detailed information about the final state of the coating. However, these methods are usually destructive and require a long time for sample preparation and testing. The data obtained may only represent a local area of ​​the sample or the average state of the entire batch, making it difficult to fully and quickly reflect the uniformity and consistency of the internal structure of the coating batch.

[0004] Due to the aforementioned limitations in the ability to precisely control and comprehensively monitor the internal structure and composition during coating deposition, the coated tools actually manufactured may exhibit deviations in their internal structure or composition. These minute internal differences directly affect the actual performance and stability of the tool during high-speed cutting. For example, an unexpected crystal orientation in a functional layer of the coating may significantly reduce the fracture toughness of the material in that layer; or tiny defects may exist at the interface between the layers. These defects may become the source of crack initiation and propagation under cutting stress, leading to premature chipping, increased wear, or even catastrophic failure of the tool during intense cutting operations, ultimately manifesting as shortened tool life or unstable performance. Therefore, in the specific application scenario of physical vapor deposition to prepare multilayer composite coatings for high-speed cutting tools, facing multiple challenges such as limited process parameter control accuracy, difficulty in comprehensive online monitoring of the internal structure of the coating, and time-consuming and locally limited offline characterization methods, how to effectively integrate various available data (such as historical process parameter records, limited offline coating characterization data, and tool high-speed cutting performance data) and try to indirectly influence and stabilize the key microstructure and composition inside the coating through intelligent regulation of process parameters has become a key technical problem to ensure that the tool achieves high performance and high reliability in high-speed cutting applications.

[0005] In view of the above problems, the existing technology is in urgent need of improvement. Summary of the Invention

[0006] The purpose of this application is to provide a coating process parameter optimization method and system, which can establish a predictive association between process parameters and tool cutting performance, realize the optimal selection of process parameters, and obtain a process parameter combination that improves tool performance.

[0007] In a first aspect, the present application provides a coating process parameter optimization method for optimizing the process parameters of a cutting tool coating, the method comprising the following steps:

[0008] A1. Obtain physical vapor deposition process parameter records, coating offline characterization data, and tool cutting performance data, and perform preprocessing;

[0009] A2. Extracting process parameter characteristics and structural characteristics from the pre-processed process parameter records and the pre-processed coating offline characterization data respectively;

[0010] A3. Using the extracted process parameter characteristics and structural characteristics, construct a first prediction model; the first prediction model is used to predict the corresponding structural characteristics based on the process parameter characteristics;

[0011] A4. Using the extracted structural features and preprocessed tool cutting performance data, construct a second prediction model; the second prediction model is used to predict the corresponding tool cutting performance data based on the structural features;

[0012] A5. Combine the first prediction model and the second prediction model to form a third prediction model for predicting tool cutting performance data based on process parameter characteristics;

[0013] A6. Based on the optimization goal and process constraints, use the third prediction model to search for the optimal process parameter combination in the process parameter space.

[0014] Preferably, step A1 includes:

[0015] A101 obtains the process parameter record of the physical vapor deposition process; the process parameter record includes substrate temperature, bias voltage, gas flow rate, sputtering power and deposition time;

[0016] A102 obtains offline characterization data of the coating; the offline characterization data of the coating includes lattice constant, grain size, layer thickness, bonding strength, hardness and residual stress data measured by X-ray diffraction, scanning electron microscopy and hardness tester;

[0017] A103 obtain tool cutting performance data; the tool cutting performance data includes tool life, wear, chipping and cutting force data;

[0018] A104. For the acquired process parameter records, remove those with more than 10% missing values, and use the Laida criterion to remove outliers from the remaining process parameter records to obtain preprocessed process parameter records.

[0019] A105. The obtained coating offline characterization data and tool cutting performance data are dimensionally normalized using the Z-score standardization method to obtain the preprocessed coating offline characterization data and preprocessed tool cutting performance data.

[0020] Preferably, step A2 includes:

[0021] A201. Calculate the correlation coefficient between the process parameters using the Pearson correlation coefficient for the preprocessed process parameter records. Filter out process parameters whose correlation coefficients are greater than a preset correlation coefficient threshold and whose variances are greater than a preset variance threshold, thereby obtaining a process parameter feature set.

[0022] A202. Calculate the Lasso coefficient for each process parameter feature using the Lasso regression method for the process parameter feature set, and select the first several process parameter features with the largest absolute values ​​of the Lasso coefficient as the final process parameter features.

[0023] A203. For the pre-processed coating offline characterization data, the principal component analysis method is used to perform dimensionality reduction processing and extract the first several principal components as the structural characteristics of the coating.

[0024] Preferably, step A201 includes:

[0025] For the pre-processed process parameter records, calculate the Pearson correlation coefficient between any two process parameters and construct a correlation coefficient matrix;

[0026] Traversing the correlation coefficient matrix, if the absolute value of the correlation coefficient of the current process parameter pair is greater than a preset correlation coefficient threshold, and the variance of each process parameter in the process parameter pair is greater than a preset variance threshold, then adding the two process parameters in the process parameter pair to the process parameter feature set;

[0027] The repeated process parameters in the process parameter feature set are removed to obtain a final process parameter feature set.

[0028] Preferably, step A202 includes:

[0029] For the process parameter feature set, the minimum angle regression algorithm is used to calculate the regression path of each process parameter feature based on the correlation between the process parameter feature and the offline characterization data of each coating;

[0030] For each regression path, the cross-validation method is used to calculate the prediction error of each process parameter feature under different regression coefficients, and the regression coefficient with the smallest prediction error is selected as the optimal Lasso coefficient of the process parameter feature;

[0031] According to the optimal Lasso coefficient of each process parameter feature, calculate the absolute value of the Lasso coefficient of all process parameter features, and sort the process parameter features in descending order of the absolute value of the Lasso coefficient;

[0032] The first N process parameter features with the largest absolute value of the Lasso coefficient are selected as the final process parameter features; where N is the number value determined according to the inflection point of the cross-validated error curve.

[0033] Preferably, step A203 includes:

[0034] Constructing a covariance matrix for the preprocessed coating offline characterization data, and performing eigenvalue decomposition on the covariance matrix to obtain a plurality of eigenvalues ​​and corresponding eigenvectors;

[0035] According to the obtained eigenvalues, calculate the variance contribution rate of each eigenvalue, and sort the eigenvalues ​​in descending order of variance contribution rate;

[0036] Select each eigenvalue in order from front to back according to the sorting, until the cumulative variance contribution rate of the selected eigenvalue is greater than or equal to the preset variance contribution rate threshold for the first time, and use the eigenvectors corresponding to the selected eigenvalues ​​to construct a dimensionality reduction matrix;

[0037] The pre-processed offline characterization data of the coating is multiplied by the dimension reduction matrix to obtain dimension-reduced data as the structural characteristics of the coating.

[0038] Preferably, step A5 includes:

[0039] A501. Construct a stacking model, using the first prediction model and the second prediction model as two-layer base learners of the stacking model;

[0040] A502. Using the extracted process parameter features and preprocessed tool cutting performance data, a meta-learner of the stacking model is trained; the meta-learner is used to predict tool cutting performance data based on the output results of the first prediction model and the second prediction model, combined with the process parameter features;

[0041] A503. Use the trained stacking model as the third prediction model to predict tool cutting performance data based on process parameter characteristics.

[0042] Preferably, step A6 includes:

[0043] A601. Determine an optimization objective function, wherein the optimization objective function includes at least one of maximizing tool life, maximizing material removal rate, and minimizing surface roughness;

[0044] A602. Set process constraints, including substrate temperature range, bias voltage range, gas flow rate range, sputtering power range, and deposition time range, to form a process parameter space;

[0045] A603. Using a particle swarm optimization algorithm, initialize a particle swarm in the process parameter space, where each particle represents a set of process parameter combinations, the particle position corresponds to the process parameter value, and the particle velocity corresponds to the rate of change of the process parameter value;

[0046] A604. For each particle, use the third prediction model to predict its corresponding tool cutting performance, and calculate the fitness value of the particle according to the optimization objective function;

[0047] A605. Update the velocity and position of each particle based on its fitness value. The velocity update takes into account the particle's own historical optimal position and the group's historical optimal position. The position update is adjusted based on the updated velocity and ensures that the updated position remains within the process parameter space.

[0048] A606. Determine whether the maximum number of iterations has been reached or the convergence condition has been met. If so, output the process parameter combination corresponding to the historical optimal position of the group as the optimal process parameter combination; otherwise, return to step A604.

[0049] Preferably, step A605 includes:

[0050] According to the fitness value of each particle, the first distance vector between the current position of the particle and its own historical optimal position is calculated, as well as the second distance vector between the current position of the particle and the historical optimal position of the group;

[0051] Scaling the first distance vector and the second distance vector according to a preset first learning factor, a preset second learning factor, and two randomly generated random numbers to obtain a first speed increment and a second speed increment;

[0052] The particle speed is updated according to the first speed increment and the second speed increment, and the particle position is adjusted according to the updated speed. If the updated position exceeds the feasible domain of the process parameters, the position is adjusted to the boundary of the feasible domain of the process parameters.

[0053] In a second aspect, the present application provides a coating process parameter optimization system for optimizing process parameters of a cutting tool coating, the system comprising:

[0054] Data acquisition module, used to obtain the process parameter records of physical vapor deposition process, coating offline characterization data and tool cutting performance data, and perform preprocessing;

[0055] a feature extraction module, for extracting process parameter features and structural features from the pre-processed process parameter records and the pre-processed coating offline characterization data, respectively;

[0056] A first model building module is used to build a first prediction model using the extracted process parameter characteristics and structural characteristics; the first prediction model is used to predict the corresponding structural characteristics based on the process parameter characteristics;

[0057] A second model building module is used to build a second prediction model using the extracted structural features and the preprocessed tool cutting performance data; the second prediction model is used to predict the corresponding tool cutting performance data according to the structural features;

[0058] A model combining module, configured to combine the first prediction model and the second prediction model to form a third prediction model for predicting tool cutting performance data based on process parameter characteristics;

[0059] The process parameter optimization module is used to search for an optimal process parameter combination in the process parameter space using the third prediction model based on the optimization target and process constraints.

[0060] Beneficial effect: The present application provides a coating process parameter optimization method and system, which jointly forms a prediction model between process parameters and tool cutting performance by constructing a prediction model between process parameters and coating offline characterization data, as well as a prediction model between coating offline characterization data and tool cutting performance. Based on the model, a process parameter optimization search is performed, which can establish a predictive association between process parameters and tool cutting performance, realize the optimal selection of process parameters, and obtain a process parameter combination that improves tool performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 Flowchart of the coating process parameter optimization method provided in an embodiment of the present application.

[0062] Figure 2 A schematic diagram of the structure of the coating process parameter optimization system provided in an embodiment of the present application.

[0063] Explanation of reference numerals: 1. Data acquisition module; 2. Feature extraction module; 3. First model construction module; 4. Second model construction module; 5. Model combination module; 6. Process parameter optimization module. DETAILED DESCRIPTION

[0064] The technical solutions in this application will be clearly and completely described below in conjunction with the drawings in this application. Obviously, the described embodiments are only a part of the embodiments of this application, rather than all of the embodiments. The components of the present application generally described and shown in the drawings here can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the application for protection, but merely represents selected embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without making creative work fall within the scope of protection of this application.

[0065] It should be noted that similar reference numerals and letters represent similar items in the following drawings. Therefore, once an item is defined in one drawing, it does not need to be further defined or explained in subsequent drawings. At the same time, in the description of this application, the terms "first", "second", etc. are only used to distinguish the description and should not be understood as indicating or implying relative importance.

[0066] refer to Figure 1 This application proposes a coating process parameter optimization method for optimizing the process parameters of cutting tool coatings. The method comprises the following steps:

[0067] A1. Obtain physical vapor deposition process parameter records, coating offline characterization data, and tool cutting performance data, and perform preprocessing;

[0068] A2. Extracting process parameter characteristics and structural characteristics from the pre-processed process parameter records and the pre-processed coating offline characterization data respectively;

[0069] A3. Using the extracted process parameter characteristics and structural characteristics, construct a first prediction model; the first prediction model is used to predict the corresponding structural characteristics based on the process parameter characteristics;

[0070] A4. Using the extracted structural features and preprocessed tool cutting performance data, construct a second prediction model; the second prediction model is used to predict the corresponding tool cutting performance data based on the structural features;

[0071] A5. Combine the first prediction model and the second prediction model to form a third prediction model for predicting tool cutting performance data based on process parameter characteristics;

[0072] A6. Based on the optimization goal and process constraints, use the third prediction model to search for the optimal process parameter combination in the process parameter space.

[0073] Among them, in step A1, preprocessing can be specifically implemented by methods such as data cleaning, missing value filling, outlier removal, data conversion and dimensional normalization to improve data quality and consistency and provide a basis for subsequent modeling.

[0074] Among them, in step A2, process parameter features are extracted from the process parameter records after pretreatment. Specifically, correlation analysis, variance analysis, principal component analysis, factor analysis, Lasso regression, ridge regression and other methods can be used to identify key process parameters or their combinations that have a significant impact on the coating structure and performance. Structural features are extracted from the pretreated coating offline characterization data. Specifically, dimensionality reduction methods such as principal component analysis, independent component analysis, and non-negative matrix decomposition can be used to extract low-dimensional feature vectors reflecting the key microstructural information of the coating from the high-dimensional characterization data. Feature extraction helps to reduce data dimensionality, remove redundant information, highlight key factors, and improve the efficiency and accuracy of the model.

[0075] The first prediction model in step A3 can be implemented using various regression models, such as linear regression, polynomial regression, support vector regression, decision tree regression, random forest regression, gradient boosting regression, neural network models, etc. This model establishes a mapping relationship between physical vapor deposition process parameters and the coating microstructure, quantifying the impact of process parameters on the structure.

[0076] The second prediction model in step A4 can be implemented using various regression models, such as linear regression, polynomial regression, support vector regression, decision tree regression, random forest regression, gradient boosting regression, and neural network models. This model establishes a mapping relationship between the coating microstructure and the final tool cutting performance, revealing how structure affects performance.

[0077] Among them, in step A5, the third prediction model can directly predict the tool cutting performance data based on the process parameter characteristics. The combination can be specifically in series, with the output of the first prediction model (predicted structural characteristics) as the input of the second prediction model, thereby achieving end-to-end prediction from process parameters to performance. An integrated learning method such as stacking can also be used, with the first prediction model and the second prediction model as base learners, and then a meta-learner is trained to integrate their outputs and combine the original process parameter characteristics for the final performance prediction. This joint model constructs a direct prediction path from process parameters to final performance, overcoming the limitation of difficult real-time online monitoring of the internal structure of the coating.

[0078] In step A6, the optimization objectives may include maximizing tool life, maximizing material removal rate, minimizing surface roughness, minimizing cost, etc. Process constraints refer to the feasible range of process parameters in actual production, such as substrate temperature range, bias voltage range, gas flow range, sputtering power range, and deposition time range. The search for the optimal process parameter combination can be implemented using various optimization algorithms, such as genetic algorithms, particle swarm optimization algorithms, simulated annealing algorithms, Bayesian optimization, grid search, random search, etc. The optimization algorithm uses a third prediction model to evaluate the performance corresponding to different process parameter combinations, and iteratively searches for the best parameter combination that can achieve the optimization goal within the feasible space of process parameters.

[0079] There is a close logical connection between steps A1, A2, A3, A4, A5 and A6. Step A1 provides a high-quality data foundation and is the prerequisite for all subsequent steps. Step A2 extracts key information from the original data, reducing the complexity of subsequent modeling. Steps A3 and A4 respectively establish prediction models from process parameters to structure and structure to performance, decomposing the complex process-performance relationship. Step A5 combines these two models to build a prediction capability directly from process parameters to performance, which is the key to achieving model-based optimization. Step A6 uses the prediction model constructed in step A5, combined with the optimization algorithm, to conduct an efficient search in the process parameter space and finally find the optimal process parameter combination. The entire process forms a data-driven closed-loop optimization system, which learns process laws through historical data, builds prediction capabilities, and uses prediction capabilities to guide the optimization search of process parameters, thereby solving the problem that it is difficult to directly control the structure and effectively optimize the process parameters to obtain high-performance coatings in complex PVD processes.

[0080] Through the above technical solution, the present application solves the problem that in the process of preparing cutting tool coatings by physical vapor deposition, it is difficult to effectively optimize process parameters to obtain high-performance and high-reliability coatings due to process complexity, monitoring limitations and insufficient offline characterization. This method establishes a prediction model between process parameters and coating structure, coating structure and tool performance in a data-driven manner, and further constructs an end-to-end model that directly predicts tool performance from process parameters. By using this prediction model to guide the optimization search of process parameters, it is possible to efficiently find a combination of process parameters that can improve the cutting performance of the tool without conducting a large number of actual experiments. This overcomes the shortcomings of low efficiency and high cost of the traditional trial and error method, improves the efficiency and success rate of process optimization, and helps to stabilize and improve the performance and reliability of cutting tool coatings.

[0081] In some embodiments, step A1 comprises:

[0082] A101 obtains the process parameter record of the physical vapor deposition process; the process parameter record includes substrate temperature, bias voltage, gas flow rate, sputtering power and deposition time;

[0083] A102 obtains offline characterization data of the coating; the offline characterization data of the coating includes lattice constant, grain size, layer thickness, bonding strength, hardness and residual stress data measured by X-ray diffraction, scanning electron microscopy and hardness tester;

[0084] A103 obtain tool cutting performance data; the tool cutting performance data includes tool life, wear, chipping and cutting force data;

[0085] A104. For the acquired process parameter records, remove those with more than 10% missing values, and use the Laida criterion to remove outliers from the remaining process parameter records to obtain preprocessed process parameter records.

[0086] A105. The obtained coating offline characterization data and tool cutting performance data are dimensionally normalized using the Z-score standardization method to obtain the preprocessed coating offline characterization data and preprocessed tool cutting performance data.

[0087] Specifically, this technical solution establishes a data foundation for connecting process, structure and performance by systematically acquiring three types of key data: physical vapor deposition process parameters, coating structure characterization, and tool cutting performance. In the face of common missing and abnormal problems in actual data, a specific data cleaning strategy (eliminating records with high missing rates and using the Laida criterion to eliminate outliers) is used to process the process parameter data to improve the purity of the data. At the same time, considering that different types of characterization data and performance data may have different dimensions and numerical ranges, the Z-score normalization method is used for normalization to eliminate the dimensional effect, so that these data can be compared and analyzed on a unified scale. Through these data acquisition and preprocessing steps, this method overcomes the challenges of unstable quality and inconsistent dimensions of the original data, and provides high-quality, standardized input data for subsequent accurate feature extraction, prediction model construction, and ultimate process parameter optimization, thereby improving the effectiveness and robustness of the entire optimization process.

[0088] In some embodiments, step A2 comprises:

[0089] A201. Calculate the correlation coefficient between the process parameters using the Pearson correlation coefficient for the preprocessed process parameter records. Filter out process parameters whose correlation coefficients are greater than a preset correlation coefficient threshold and whose variances are greater than a preset variance threshold, thereby obtaining a process parameter feature set.

[0090] A202. Calculate the Lasso coefficient for each process parameter feature using the Lasso regression method for the process parameter feature set, and select the first several process parameter features with the largest absolute values ​​of the Lasso coefficient as the final process parameter features.

[0091] A203. For the pre-processed coating offline characterization data, the principal component analysis method is used to perform dimensionality reduction processing and extract the first several principal components as the structural characteristics of the coating.

[0092] Specifically, for preprocessed process parameter records, direct use of raw data can affect the accuracy and efficiency of subsequent models, as it may contain redundant or low-information parameters. By calculating the Pearson correlation coefficient between process parameters and the variance of each parameter, and setting correlation coefficient and variance thresholds for preliminary screening, we can eliminate highly correlated parameters with small variations, thereby obtaining a preliminary set of process parameter features. This process reduces data redundancy.

[0093] Furthermore, based on the initially screened set of process parameter features, the Lasso regression method was employed. Lasso regression, through the L1 penalty term, compresses the coefficients of parameters with minimal impact on the predicted target to zero, thereby achieving feature selection. By calculating the Lasso coefficient for each parameter and selecting the parameter with the largest absolute value as the final process parameter feature, process parameters with stronger predictive power for coating structural characteristics were identified. This process further streamlined the process parameter dimensionality and improved the effectiveness of the features.

[0094] Furthermore, preprocessed offline coating characterization data may contain multiple interrelated indicators and have a high dimensionality. Principal component analysis (PCA) constructs a covariance matrix and performs eigenvalue decomposition to extract the top principal components that explain the majority of the data variance. These principal components are linear combinations of the original characterization data and represent key structural information about the coating in a lower-dimensional manner. Using the extracted principal components as structural features of the coating reduces the dimensionality of the data while retaining key information and avoiding the problems associated with high-dimensional data.

[0095] Through the above steps, concise and information-rich process parameter features and structural features are extracted from the original data, providing high-quality input for the subsequent construction of the first prediction model from process parameters to structural features and the second prediction model from structural features to tool cutting performance, thereby improving the accuracy and efficiency of the prediction model and solving the problems of redundancy and excessive dimensionality of the original data.

[0096] Preferably, step A201 may include:

[0097] For the pre-processed process parameter records, calculate the Pearson correlation coefficient between any two process parameters and construct a correlation coefficient matrix;

[0098] Traversing the correlation coefficient matrix, if the absolute value of the correlation coefficient of the current process parameter pair is greater than a preset correlation coefficient threshold, and the variance of each process parameter in the process parameter pair is greater than a preset variance threshold, then adding the two process parameters in the process parameter pair to the process parameter feature set;

[0099] The repeated process parameters in the process parameter feature set are removed to obtain a final process parameter feature set.

[0100] Among them, this method quantifies the degree of linear correlation between any two process parameters in the preprocessed process parameter records by calculating the Pearson correlation coefficient between them, and organizes these correlation coefficients into a matrix. Furthermore, the method traverses this correlation coefficient matrix and checks each pair of process parameters. For a pair of process parameters, if the absolute value of the correlation coefficient between them exceeds a set correlation coefficient threshold, indicating that there is a significant linear correlation, and the variance of each of the pair of parameters exceeds another set variance threshold, indicating that these parameters show sufficient variability in the data, then both process parameters in the pair of parameters are included in a temporary feature set. Finally, all duplicate process parameter items are removed from this temporary set to obtain a set containing all unique process parameters that meet the conditions of strong correlation and high variability.

[0101] Specifically, the solution aims to solve the problem of how to clearly determine a preliminary feature set based on the correlation between process parameters and their respective variances when preliminarily screening process parameter features. First, the preprocessed process parameter records are processed, the Pearson correlation coefficient between any two process parameters is calculated, and a correlation coefficient matrix that reflects the strength of the linear correlation between all parameter pairs is constructed. The absolute value of the Pearson correlation coefficient close to 1 indicates a strong linear correlation, and close to 0 indicates a weak linear correlation. Then, the constructed correlation coefficient matrix is ​​traversed to check each pair of process parameters. For the currently checked process parameter pair, the absolute value of the Pearson correlation coefficient between them is calculated and compared with the preset correlation coefficient threshold. At the same time, the variance of each of the process parameters in this pair is calculated and compared with the preset variance threshold. Only when the absolute value of the correlation coefficient between the process parameters in this pair is greater than the preset correlation coefficient threshold, and the variance of each parameter in this pair of process parameters is greater than the preset variance threshold, are both parameters in this pair of process parameters added to a process parameter feature set. By combining correlation and variance for screening, we can initially identify process parameters that have significant linear correlations with each other and sufficient variability. These parameters are more likely to affect the coating structure. Finally, since a process parameter may be added to the set multiple times due to meeting the screening criteria along with multiple other parameters during the traversal process, it is necessary to remove duplicates from the set to obtain a final set containing all unique process parameters that meet the criteria. This provides a clear screening process based on pairwise correlation and individual variance, improving the operability and certainty of the initial screening process for process parameter characteristics.

[0102] Furthermore, step A202 may include:

[0103] For the process parameter feature set, the minimum angle regression algorithm is used to calculate the regression path of each process parameter feature based on the correlation between the process parameter feature and the offline characterization data of each coating;

[0104] For each regression path, the cross-validation method is used to calculate the prediction error of each process parameter feature under different regression coefficients, and the regression coefficient with the smallest prediction error is selected as the optimal Lasso coefficient of the process parameter feature;

[0105] According to the optimal Lasso coefficient of each process parameter feature, calculate the absolute value of the Lasso coefficient of all process parameter features, and sort the process parameter features in descending order of the absolute value of the Lasso coefficient;

[0106] The first N process parameter features with the largest absolute value of the Lasso coefficient are selected as the final process parameter features; where N is the number value determined according to the inflection point of the cross-validated error curve.

[0107] Among them, the least angle regression algorithm is an effective algorithm for solving the Lasso regression problem. It can gradually build a regression model and select the feature most relevant to the current residual at each step, moving along the direction of the minimum angle, thereby efficiently obtaining the complete path of the Lasso coefficient changing with the regularization parameter. Specifically, it can be implemented using the Least Angle Regression (LARS) algorithm. This algorithm iteratively finds the feature with the greatest correlation with the current residual at each step and updates the coefficient along the angle bisector of these features until a stopping condition is reached.

[0108] Among them, cross-validation is a model evaluation technique. By dividing the dataset into a training set and a validation set, training the model on the training set, and evaluating the performance on the validation set, repeating multiple times and taking the average, the generalization ability of the model on unknown data can be more reliably evaluated. Specifically, K-fold cross-validation can be used to implement it. The dataset is randomly divided into K subsets. The model is trained with K-1 subsets each time, and the remaining 1 subset is used for validation. This is repeated K times and the average prediction error is calculated. The prediction error can be measured using indicators such as mean squared error (MSE) or root mean squared error (RMSE). The Lasso coefficient corresponding to the regression coefficient that minimizes the average prediction error is selected as the optimal Lasso coefficient.

[0109] Lasso regression applies an L1 norm penalty to the regression coefficients, reducing some coefficients to zero and thus achieving feature selection. The absolute value of the non-zero coefficient reflects the importance of the corresponding feature to the predicted target (coating structural characteristics). By calculating the absolute value of the optimal Lasso coefficients and sorting them, a list reflecting the importance of the features can be obtained.

[0110] The error curve (the curve formed by connecting the prediction errors calculated during the cross-validation process) generally depicts the changing trend of the model prediction error as a function of the number of features (or the number of Lasso regression steps). As the number of features increases, the prediction error typically decreases first and then may level off or slightly increase. The inflection point (elbow point) of the error curve is generally considered to be the point where a good balance is achieved between model performance and the number of features, that is, a relatively small number of features is used while ensuring a low prediction error. Determining N as the number of features corresponding to this inflection point can avoid selecting too many features, which may introduce noise or collinearity, and also avoid selecting too few features and losing important information, thereby improving the effectiveness of feature selection.

[0111] This solution uses the minimum angle regression algorithm to efficiently calculate the Lasso regression path, and combines it with the cross-validation method to accurately determine the optimal Lasso coefficient for each feature, and then scientifically determines the final number of selected features N according to the absolute value of the coefficient and the inflection point of the error curve. These steps work together to more accurately identify key process parameter features that are strongly correlated with coating structural features from the preliminarily screened process parameter feature set. These selected features serve as input for constructing the first prediction model, which can more effectively capture the complex relationship between process parameters and structural features, thereby improving the prediction accuracy of the first prediction model. The improved accuracy of the first prediction model further enhances the prediction capability of the third prediction model (predicting tool cutting performance based on process parameters), ultimately providing a more reliable basis for subsequent process parameter optimization searches based on the third prediction model, and helping to find a combination of process parameters that can better achieve the optimization goals.

[0112] Through the above technical solution, this application improves the accuracy of Lasso coefficient calculation, scientifically determines the optimal number of features, more effectively extracts process parameter features that are strongly correlated with coating structure characteristics, improves the accuracy of feature selection, and thereby improves the predictive performance of subsequent prediction models.

[0113] Furthermore, step A203 may include:

[0114] Constructing a covariance matrix for the preprocessed coating offline characterization data, and performing eigenvalue decomposition on the covariance matrix to obtain a plurality of eigenvalues ​​and corresponding eigenvectors;

[0115] According to the obtained eigenvalues, calculate the variance contribution rate of each eigenvalue and sort the eigenvalues ​​in descending order of variance contribution rate;

[0116] Select each eigenvalue in order from front to back according to the sorting, until the cumulative variance contribution rate of the selected eigenvalue is greater than or equal to the preset variance contribution rate threshold for the first time, and use the eigenvectors corresponding to the selected eigenvalues ​​to construct a dimensionality reduction matrix;

[0117] The pre-processed offline characterization data of the coating is multiplied by the dimension reduction matrix to obtain dimension-reduced data as the structural characteristics of the coating.

[0118] This method determines the number of principal components to retain by quantifying the amount of information contained in each principal component. First, the covariance matrix of the preprocessed offline coating characterization data is calculated. This matrix reflects the correlations between the different characterization data. Next, the covariance matrix is ​​subjected to eigenvalue decomposition to obtain eigenvalues ​​and corresponding eigenvectors. The eigenvectors represent the new data directions, and the magnitude of the eigenvalues ​​indicates the degree of data dispersion in the corresponding direction, that is, the amount of information contained in that direction. Next, the variance contribution of each eigenvalue is calculated—the proportion of that eigenvalue to the sum of all eigenvalues, indicating the ability of that direction to explain the total variance of the original data. Eigenvalues ​​are sorted from largest to smallest by variance contribution to determine the order of importance of the directions. Subsequently, from these sorted eigenvalues, the eigenvalues ​​are selected in descending order of importance, and their variance contributions are accumulated until the cumulative variance contribution reaches or exceeds a preset threshold. When the threshold is reached, the number of eigenvalues ​​to be retained is determined, thereby determining the number of dimensions after dimensionality reduction. The eigenvectors corresponding to the selected eigenvalues ​​are then used to construct a dimensionality reduction matrix. The column vectors of this matrix are the selected eigenvectors. Finally, the pre-processed coating offline characterization data is multiplied by the constructed dimensionality reduction matrix, projecting the original high-dimensional data into a low-dimensional space composed of the selected eigenvectors to obtain the reduced-dimensional data. This reduced-dimensional data serves as the structural characteristics of the coating.

[0119] Specifically, this method is used to extract structural features from preprocessed offline coating characterization data, solving the problem of determining the number of principal components to extract. First, the covariance matrix is ​​calculated for the preprocessed offline coating characterization data. This matrix describes the linear relationships between the data dimensions. Then, the covariance matrix is ​​subjected to eigenvalue decomposition, resulting in a set of eigenvalues ​​and corresponding eigenvectors. The eigenvectors form a new orthogonal basis, with the eigenvalues ​​representing the variance of the data in the direction of the corresponding eigenvector. The variance contribution of each eigenvalue is calculated—the ratio of that eigenvalue to the sum of all eigenvalues—to measure the ability of each principal component to explain the variability of the original data. The eigenvalues ​​and their corresponding variance contributions are sorted by size. From the sorted eigenvalues, the eigenvalues ​​are selected in descending order of variance contribution, and the cumulative variance contribution is calculated. Selection stops when the cumulative variance contribution reaches or exceeds a preset threshold for the first time. The selected number of eigenvalues ​​represents the reduced dimensions, and their corresponding eigenvectors form the reduced matrix. The preprocessed coating offline characterization data is multiplied by the transpose of the dimensionality reduction matrix to obtain the reduced-dimensional data, which are the extracted coating structural features. By setting a cumulative variance contribution rate threshold, this method can reduce the data dimension while retaining the key information in the original data, improving the efficiency and effectiveness of structural feature extraction.

[0120] In some specific embodiments, it is assumed that the pre-processed offline characterization data of the coating is a matrix X with M rows and P columns, where M is the number of samples and P is the number of original features. First, calculate the covariance matrix C of the matrix X, which is a symmetric matrix with P rows and P columns. Perform eigenvalue decomposition on C to obtain P eigenvalues ​​λ1≥λ2≥...≥λP and corresponding eigenvectors v1, v2,..., vP. Calculate the variance contribution rate ri=λi / Σλ of each eigenvalue λi, where Σλ represents the sum of all eigenvalues. Sort the eigenvalues ​​by size and calculate the cumulative variance contribution rate. Set the preset variance contribution rate threshold to 95%. Starting from λ1, accumulate r1, r1+r2, r1+r2+r3,... in sequence until the cumulative variance contribution rate is greater than or equal to 95% for the first time. Assuming the cumulative variance contribution of the first k eigenvalues ​​meets the requirements, the corresponding eigenvectors v1, v2, ..., vk are selected to construct a dimensionality reduction matrix W. W is a P-row, k-column matrix, whose column vectors are v1, v2, ..., vk. Multiplying the preprocessed data matrix X by matrix W yields an M-row, k-column matrix Y = XW. Matrix Y represents the reduced-dimensional coating structural feature data. In this way, the original P-dimensional offline coating characterization data is reduced to k dimensions, where k is significantly smaller than P, while retaining 95% of the variance information in the original data, achieving data compression and feature extraction.

[0121] Specifically, in step A3, a linear regression model, a polynomial regression model, a support vector regression model, a decision tree regression model, a random forest regression model, a gradient boosting regression model or a neural network model can be selected to construct a preliminary first prediction model, and then the prepared first sample data (process parameter characteristics and structural characteristics obtained in step A2) are divided into a first training set and a first validation set. The preliminary first prediction model is trained with the first training set, and the model parameters of the preliminary first prediction model are iteratively optimized based on the loss function (such as mean square error, cross entropy loss, etc.) and optimization algorithm (such as gradient descent method, Adam algorithm, etc.), and the optimized first prediction model is verified with the first validation set to finally obtain the trained first prediction model.

[0122] Similarly, in step A4, a linear regression model, a polynomial regression model, a support vector regression model, a decision tree regression model, a random forest regression model, a gradient boosting regression model or a neural network model can be selected to construct a preliminary second prediction model, and then the prepared second sample data (the structural features obtained in step A2 and the preprocessed tool cutting performance data obtained in step A1) are divided into a second training set and a second validation set. The preliminary second prediction model is trained with the second training set, and the model parameters of the preliminary second prediction model are iteratively optimized based on the loss function (such as mean square error, cross entropy loss, etc.) and optimization algorithm (such as gradient descent method, Adam algorithm, etc.), and the optimized second prediction model is verified with the second validation set to finally obtain a trained second prediction model.

[0123] In some embodiments, step A5 comprises:

[0124] A501. Construct a stacking model, using the first prediction model and the second prediction model as two-layer base learners of the stacking model;

[0125] A502. Using the extracted process parameter features and preprocessed tool cutting performance data, a meta-learner of the stacking model is trained; the meta-learner is used to predict tool cutting performance data based on the output results of the first prediction model and the second prediction model, combined with the process parameter features;

[0126] A503. Use the trained stacking model as the third prediction model to predict tool cutting performance data based on process parameter characteristics.

[0127] Among them, the stacking model is an integrated learning method that improves the overall prediction performance by combining the prediction results of multiple basic learners. The first prediction model is trained to predict the coating structure characteristics from the process parameter characteristics. The second prediction model is trained to predict the tool cutting performance data from the coating structure characteristics. The meta-learner is trained to receive the prediction outputs of the first prediction model and the second prediction model, and combine the original process parameter characteristics to generate the final tool cutting performance prediction results. The meta-learner learns how to weight or combine the prediction results of the basic learners, and is able to utilize information in the original features that is not fully captured by the basic learners. In this way, the stacking model can learn the complex mapping relationship between process parameter characteristics and tool cutting performance data.

[0128] Specifically, to address the limited prediction accuracy of combined independent models, this solution utilizes a stacked model to construct a third prediction model. First, a stacking framework is constructed, using the pre-trained first and second prediction models as base-level learners. The first prediction model receives process parameter features as input and outputs predicted structural features. The second prediction model receives the predicted structural features as input and outputs predicted tool cutting performance data. Next, a meta-learner is trained, which receives the output of the first and second prediction models, as well as the original process parameter features, and uses the pre-processed tool cutting performance data as a training target. The meta-learner learns the relationship between the output of the base learners and the actual performance data, and combines it with the original process parameter information to correct or optimize the prediction results of the base learners. After training is complete, this stacked model becomes the third prediction model, capable of directly predicting tool cutting performance data based on the input process parameter features. This layered learning and information fusion approach improves the accuracy of the process parameter-to-cutting performance prediction.

[0129] In some specific embodiments, a stacked model can be constructed in the following manner. The first prediction model can be a feedforward neural network, the number of input layer nodes is consistent with the process parameter feature dimension, and the number of output layer nodes is consistent with the structural feature dimension. The second prediction model can be a gradient boosting regression tree model, the input feature is the structural feature, and the output is the tool cutting performance data. The meta-learner can be a linear regression model, whose input features include the output of the feedforward neural network, the output of the gradient boosting regression tree model and the original process parameter feature, and the output is the tool cutting performance data. During the training process, the feedforward neural network and the gradient boosting regression tree model are first trained using the training data (completed in steps A3 and A4). Then, a cross-validation method is used to predict the training data using the trained feedforward neural network and gradient boosting regression tree model to generate prediction results on the training set. These prediction results are used together with the original process parameter features as training data to train the linear regression meta-learner. Finally, the trained feedforward neural network, gradient boosting regression tree model and linear regression meta-learner are combined to form a third prediction model.

[0130] In some embodiments, step A6 includes:

[0131] A601. Determine an optimization objective function, wherein the optimization objective function includes at least one of maximizing tool life, maximizing material removal rate, and minimizing surface roughness;

[0132] A602. Set process constraints, including substrate temperature range, bias voltage range, gas flow rate range, sputtering power range, and deposition time range, to form a process parameter space;

[0133] A603. Using a particle swarm optimization algorithm, initialize a particle swarm in the process parameter space, where each particle represents a set of process parameter combinations, the particle position corresponds to the process parameter value, and the particle velocity corresponds to the rate of change of the process parameter value;

[0134] A604. For each particle, use the third prediction model to predict its corresponding tool cutting performance, and calculate the fitness value of the particle according to the optimization objective function;

[0135] A605. Update the velocity and position of each particle based on its fitness value. The velocity update takes into account the particle's own historical optimal position and the group's historical optimal position. The position update is adjusted based on the updated velocity and ensures that the updated position remains within the process parameter space.

[0136] A606. Determine whether the maximum number of iterations has been reached or the convergence condition has been met. If so, output the process parameter combination corresponding to the historical optimal position of the group as the optimal process parameter combination; otherwise, return to step A604.

[0137] The optimization objective function can include at least one of maximizing tool life, maximizing material removal rate, and minimizing surface roughness, thereby providing an evaluation criterion for the optimization process. Process constraints set the range of values ​​for substrate temperature, bias voltage, gas flow rate, sputtering power, and deposition time, thereby defining the feasible search region. The particle swarm optimization algorithm randomly generates a set of initial process parameter combinations within the defined process parameter space. Each combination is considered a particle with a position and velocity. The particle position corresponds to the process parameter value, and the particle velocity corresponds to the rate of change of the process parameter value. During the optimization iteration process, the previously constructed third prediction model is used to predict the tool cutting performance corresponding to the process parameter combination represented by each particle. Then, based on the set optimization objective function, the fitness value of each particle's current position is calculated. This value reflects the quality of the current parameter combination. Based on the calculated fitness value, each particle updates its velocity and position. The velocity update is guided by the optimal position previously found by the particle itself and the optimal position previously found by the entire particle swarm. The position update is based on the new velocity. At the same time, the updated position is checked to see if it is still within the process parameter space. If it is outside the boundaries, adjustments are made to ensure that the search is conducted within the feasible range. Finally, the optimization process is determined to see if the stopping conditions are met, such as reaching the preset maximum number of iterations or the particle swarm has converged to a certain area. If the stopping conditions are met, the process parameter combination corresponding to the global optimal position searched by the particle swarm is output as the final optimal process parameter combination. If not, the next round of fitness evaluation and particle update continues.

[0138] Specifically, this technical solution uses a particle swarm optimization algorithm to solve the problem of searching for the optimal process parameter combination within a process parameter space. First, a specific optimization objective, such as maximizing tool life, is determined. This provides an evaluation basis for subsequent optimization. Next, a range of process parameter values ​​is set, such as substrate temperature between 400°C and 500°C and bias voltage between -50V and -150V. These ranges together constitute the process parameter space, ensuring that the found parameter combination is feasible in actual production. Then, a certain number of particles are randomly initialized within the set process parameter space, with each particle representing a set of process parameter combinations. In each iteration, each particle's current process parameter combination is input into a pre-built third prediction model. This model predicts the tool cutting performance, such as tool life, corresponding to that parameter combination. Based on the set optimization objective function (e.g., maximizing tool life), a fitness value is calculated for each particle. Higher tool life indicates a higher fitness value. Based on the calculated fitness value, each particle updates its velocity and position based on its own optimal position and the optimal position found by the entire particle swarm. The direction of velocity update is jointly affected by the individual optimal position and the group optimal position, and the position moves according to the updated velocity. After the position is updated, check whether the new position is still within the set process parameter range. If it is out of range, adjust the particle position to the boundary. Repeat the above process until the preset maximum number of iterations is reached or the convergence condition of the particle swarm is met. Finally, the process parameter combination corresponding to the global optimal position searched by the particle swarm is output as the optimal process parameter combination. Through this iterative search and information sharing mechanism, the particle swarm algorithm can effectively perform a global search in the multi-dimensional process parameter space, gradually approaching the process parameter combination that can achieve the optimization goal, thereby solving the technical problem of how to efficiently search for the optimal parameters.

[0139] Preferably, step A605 may include:

[0140] According to the fitness value of each particle, the first distance vector between the current position of the particle and its own historical optimal position is calculated, as well as the second distance vector between the current position of the particle and the historical optimal position of the group;

[0141] Scaling the first distance vector and the second distance vector according to a preset first learning factor, a preset second learning factor, and two randomly generated random numbers to obtain a first speed increment and a second speed increment;

[0142] The particle speed is updated according to the first speed increment and the second speed increment, and the particle position is adjusted according to the updated speed. If the updated position exceeds the feasible domain of the process parameters, the position is adjusted to the boundary of the feasible domain of the process parameters.

[0143] Specifically, for each particle in the particle swarm optimization algorithm, the fitness value corresponding to its current position is first evaluated. Based on this fitness value, the current position is compared with the particle's historical best position (i.e., the position corresponding to the optimal fitness value), and the vector difference between the two is calculated. This is the first distance vector, which represents the particle's tendency to move toward its own empirically optimal direction. At the same time, the current position is compared with the historical best position recorded by the entire particle swarm (i.e., the position corresponding to the swarm's optimal fitness value), and the vector difference between the two is calculated. This is the second distance vector, which represents the particle's tendency to move toward the swarm's empirically optimal direction. To balance the individual particle's exploration ability and the swarm's collaborative ability, a preset first learning factor and a second learning factor, as well as two independent random numbers, are introduced. The first distance vector is multiplied by the first learning factor and the first random number to obtain the velocity increment component (i.e., the first velocity increment) that follows the particle's optimal position. The second distance vector is multiplied by the second learning factor and the second random number to obtain the velocity increment component (i.e., the second velocity increment) that follows the swarm's optimal position. The current particle's velocity is added to these two velocity increment components to obtain the particle's new velocity for the next iteration step. Based on this new speed, the change in the particle position is calculated (the change in the particle position is obtained by multiplying the new speed by a time step. The time step can be adjusted according to actual needs, for example, it can be set to 1), and the particle position is updated. After updating the position, check whether the particle has moved outside the preset process parameter range. If a dimension of the particle position exceeds the upper or lower limit of the corresponding process parameter, the position of the dimension is forced to be set to the upper or lower limit value. In this way, it is ensured that the particle is always searching within the valid process parameter space, avoiding the waste of computing resources in invalid areas, and improving the efficiency and effectiveness of the optimization process.

[0144] refer to Figure 2 The present application provides a coating process parameter optimization system for optimizing the process parameters of a cutting tool coating, the system comprising:

[0145] Data acquisition module 1, used to obtain the process parameter records of the physical vapor deposition process, the coating offline characterization data and the tool cutting performance data, and perform preprocessing (the specific process refers to step A1 above);

[0146] Feature extraction module 2, used to extract process parameter features and structural features from the pre-processed process parameter records and pre-processed coating offline characterization data respectively (for the specific process, refer to step A2 above);

[0147] A first model building module 3 is used to build a first prediction model using the extracted process parameter characteristics and structural characteristics; the first prediction model is used to predict the corresponding structural characteristics based on the process parameter characteristics (for the specific process, refer to step A3 above);

[0148] A second model building module 4 is configured to build a second prediction model using the extracted structural features and the pre-processed tool cutting performance data; the second prediction model is configured to predict the corresponding tool cutting performance data based on the structural features (for details, refer to step A4 above);

[0149] Model combination module 5, used to combine the first prediction model and the second prediction model to form a third prediction model for predicting tool cutting performance data based on process parameter characteristics (for specific process, refer to step A5 above);

[0150] The process parameter optimization module 6 is used to search for the optimal process parameter combination in the process parameter space based on the optimization target and process constraints using the third prediction model (for the specific process, refer to step A6 above).

[0151] In some specific embodiments, the data acquisition module 1 may include a data interface and a data processing unit, the data interface being used to connect to a database or file system to obtain data, and the data processing unit performing operations such as missing value processing, outlier elimination, and standardization. The feature extraction module 2 may be implemented as a set of data processing algorithms, such as using correlation analysis and Lasso regression to select process parameter features, and using principal component analysis to reduce the dimension of structural features. The first model construction module 3 and the second model construction module 4 may adopt machine learning models, such as neural networks, support vector machines, or decision tree models. The model combination module 5 may adopt a stacking integration method, using the first and second prediction models as base learners, and training a meta-learner to combine their outputs for final prediction. The process parameter optimization module 6 may adopt an intelligent optimization algorithm, such as a particle swarm optimization algorithm or a genetic algorithm, to iteratively search for the optimal solution within the search space defined by the process parameters. For example, the particle swarm optimization algorithm updates the speed and position of particles (representing a combination of process parameters) by simulating the foraging behavior of a flock of birds, calculates fitness based on the predicted tool performance, and gradually approaches the optimal process parameter combination.

[0152] The above description is merely an embodiment of the present application and is not intended to limit the scope of protection of the present application. For those skilled in the art, various modifications and variations of the present application are possible. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present application shall be included in the scope of protection of the present application.

Claims

1. A coating process parameter optimization method for optimizing the process parameters of cutting tool coating, characterized in that: The steps of the method include: A1. Obtain physical vapor deposition process parameter records, coating offline characterization data, and tool cutting performance data, and perform preprocessing; A2. Extracting process parameter characteristics and structural characteristics from the pre-processed process parameter records and the pre-processed coating offline characterization data respectively; A3. Using the extracted process parameter characteristics and structural characteristics, construct a first prediction model; the first prediction model is used to predict the corresponding structural characteristics based on the process parameter characteristics; A4. Using the extracted structural features and preprocessed tool cutting performance data, construct a second prediction model; the second prediction model is used to predict the corresponding tool cutting performance data based on the structural features; A5. Combine the first prediction model and the second prediction model to form a third prediction model for predicting tool cutting performance data based on process parameter characteristics; A6. Based on the optimization goal and process constraints, use the third prediction model to search for the optimal process parameter combination in the process parameter space.

2. A coating process parameter optimization method according to claim 1, characterized in that: Step A1 includes: A101 obtains the process parameter record of the physical vapor deposition process; the process parameter record includes substrate temperature, bias voltage, gas flow rate, sputtering power and deposition time; A102 obtains offline characterization data of the coating; the offline characterization data of the coating includes lattice constant, grain size, layer thickness, bonding strength, hardness and residual stress data measured by X-ray diffraction, scanning electron microscopy and hardness tester; A103 obtain tool cutting performance data; the tool cutting performance data includes tool life, wear, chipping and cutting force data; A104. For the acquired process parameter records, remove those with more than 10% missing values, and use the Laida criterion to remove outliers from the remaining process parameter records to obtain preprocessed process parameter records. A105. The obtained coating offline characterization data and tool cutting performance data are dimensionally normalized using the Z-score standardization method to obtain the preprocessed coating offline characterization data and preprocessed tool cutting performance data.

3. A coating process parameter optimization method according to claim 2, characterized in that: Step A2 includes: A201. Calculate the correlation coefficient between the process parameters using the Pearson correlation coefficient for the preprocessed process parameter records. Filter out process parameters whose correlation coefficients are greater than a preset correlation coefficient threshold and whose variances are greater than a preset variance threshold, thereby obtaining a process parameter feature set. A202. Calculate the Lasso coefficient for each process parameter feature using the Lasso regression method for the process parameter feature set, and select the first several process parameter features with the largest absolute values ​​of the Lasso coefficient as the final process parameter features. A203. For the pre-processed coating offline characterization data, the principal component analysis method is used to perform dimensionality reduction processing and extract the first several principal components as the structural characteristics of the coating.

4. A coating process parameter optimization method according to claim 3, characterized in that: Step A201 includes: For the pre-processed process parameter records, calculate the Pearson correlation coefficient between any two process parameters and construct a correlation coefficient matrix; Traversing the correlation coefficient matrix, if the absolute value of the correlation coefficient of the current process parameter pair is greater than a preset correlation coefficient threshold, and the variance of each process parameter in the process parameter pair is greater than a preset variance threshold, then adding the two process parameters in the process parameter pair to the process parameter feature set; The repeated process parameters in the process parameter feature set are removed to obtain a final process parameter feature set.

5. A coating process parameter optimization method according to claim 3, characterized in that: Step A202 includes: For the process parameter feature set, the minimum angle regression algorithm is used to calculate the regression path of each process parameter feature based on the correlation between the process parameter feature and the offline characterization data of each coating; For each regression path, the cross-validation method is used to calculate the prediction error of each process parameter feature under different regression coefficients, and the regression coefficient with the smallest prediction error is selected as the optimal Lasso coefficient of the process parameter feature; According to the optimal Lasso coefficient of each process parameter feature, calculate the absolute value of the Lasso coefficient of all process parameter features, and sort the process parameter features in descending order of the absolute value of the Lasso coefficient; The first N process parameter features with the largest absolute value of the Lasso coefficient are selected as the final process parameter features; where N is the number value determined according to the inflection point of the cross-validated error curve.

6. A coating process parameter optimization method according to claim 3, characterized in that: Step A203 includes: Constructing a covariance matrix for the preprocessed coating offline characterization data, and performing eigenvalue decomposition on the covariance matrix to obtain a plurality of eigenvalues ​​and corresponding eigenvectors; According to the obtained eigenvalues, calculate the variance contribution rate of each eigenvalue, and sort the eigenvalues ​​in descending order of variance contribution rate; Select each eigenvalue in order from front to back according to the sorting, until the cumulative variance contribution rate of the selected eigenvalue is greater than or equal to the preset variance contribution rate threshold for the first time, and use the eigenvectors corresponding to the selected eigenvalues ​​to construct a dimensionality reduction matrix; The pre-processed offline characterization data of the coating is multiplied by the dimension reduction matrix to obtain dimension-reduced data as the structural characteristics of the coating.

7. A coating process parameter optimization method according to claim 1, characterized in that: Step A5 includes: A501. Construct a stacking model, using the first prediction model and the second prediction model as two-layer base learners of the stacking model; A502. Using the extracted process parameter features and preprocessed tool cutting performance data, a meta-learner of the stacking model is trained; the meta-learner is used to predict tool cutting performance data based on the output results of the first prediction model and the second prediction model, combined with the process parameter features; A503. Use the trained stacking model as the third prediction model to predict tool cutting performance data based on process parameter characteristics.

8. A coating process parameter optimization method according to claim 1, characterized in that: Step A6 includes: A601. Determine an optimization objective function, wherein the optimization objective function includes at least one of maximizing tool life, maximizing material removal rate, and minimizing surface roughness; A602. Set process constraints, including substrate temperature range, bias voltage range, gas flow rate range, sputtering power range, and deposition time range, to form a process parameter space; A603. Using a particle swarm optimization algorithm, initialize a particle swarm in the process parameter space, where each particle represents a set of process parameter combinations, the particle position corresponds to the process parameter value, and the particle velocity corresponds to the rate of change of the process parameter value; A604. For each particle, use the third prediction model to predict its corresponding tool cutting performance, and calculate the fitness value of the particle according to the optimization objective function; A605. Update the velocity and position of each particle based on its fitness value. The velocity update takes into account the particle's own historical optimal position and the group's historical optimal position. The position update is adjusted based on the updated velocity and ensures that the updated position remains within the process parameter space. A606. Determine whether the maximum number of iterations has been reached or the convergence condition has been met. If so, output the process parameter combination corresponding to the historical optimal position of the group as the optimal process parameter combination; otherwise, return to step A604.

9. A coating process parameter optimization method according to claim 8, characterized in that: Step A605 includes: According to the fitness value of each particle, the first distance vector between the current position of the particle and its own historical optimal position is calculated, as well as the second distance vector between the current position of the particle and the historical optimal position of the group; Scaling the first distance vector and the second distance vector according to a preset first learning factor, a preset second learning factor, and two randomly generated random numbers to obtain a first speed increment and a second speed increment; The particle speed is updated according to the first speed increment and the second speed increment, and the particle position is adjusted according to the updated speed. If the updated position exceeds the feasible domain of the process parameters, the position is adjusted to the boundary of the feasible domain of the process parameters.

10. A coating process parameter optimization system for optimizing the process parameters of cutting tool coatings, characterized in that: The system includes: Data acquisition module, used to obtain the process parameter records of physical vapor deposition process, coating offline characterization data and tool cutting performance data, and perform preprocessing; a feature extraction module, for extracting process parameter features and structural features from the pre-processed process parameter records and the pre-processed coating offline characterization data, respectively; A first model building module is used to build a first prediction model using the extracted process parameter characteristics and structural characteristics; the first prediction model is used to predict the corresponding structural characteristics based on the process parameter characteristics; A second model building module is used to build a second prediction model using the extracted structural features and the preprocessed tool cutting performance data; the second prediction model is used to predict the corresponding tool cutting performance data according to the structural features; A model combining module, configured to combine the first prediction model and the second prediction model to form a third prediction model for predicting tool cutting performance data based on process parameter characteristics; The process parameter optimization module is used to search for an optimal process parameter combination in the process parameter space using the third prediction model based on the optimization target and process constraints.

Images