Multi-objective grinding decision-making method and system based on dynamic clustering and hyperparameter optimization

Abstract

The present invention discloses a multi-objective grinding decision-making method and system based on dynamic clustering and hyperparameter optimization, which relates to the field of grinding process parameter optimization, obtains input feature data of the grinding process, and obtains a normalized feature set; based on the normalized feature set, constructs a multi-objective prediction model; the multi-objective prediction model includes a regression model and a classification model, and the regression model is trained by a bagging method with dynamic clustering and target priority adjustment; the hyperparameters of the multi-objective prediction model are collaboratively adjusted by a co-evolutionary algorithm and Bayesian optimization to generate an optimized multi-objective prediction model; the range of grinding process parameters to be optimized is input into the optimized multi-objective prediction model, the multi-objective performance indicators are predicted, and the optimal process parameter combination is selected by comprehensive scoring. Through the collaborative mechanism of dynamic K-Means clustering, target priority adjustment and co-evolutionary algorithm, the processing accuracy, stability and efficiency of the ultra-thin sheet grinding process are significantly improved.

Description

Technical Field

[0001] The present invention relates to the field of grinding process parameter optimization, and in particular to a multi-objective grinding decision-making method and system based on dynamic clustering and hyperparameter optimization. Background Art

[0002] The statements in this section merely provide background information related to the present disclosure and do not necessarily constitute prior art.

[0003] The flexible contact reference surface grinding process uses flexible gaskets to process ultra-thin reference surfaces, and then performs multiple conventional grindings to reduce stress and control end runout and flatness accuracy. Its effect is highly dependent on the dynamic adjustment of parameters such as gasket thickness, cutting amount and number of grinding times.

[0004] In existing technologies, the selection of grinding process parameters often relies on manual experience or trial-and-error adjustments, resulting in low efficiency and poor consistency. Some studies have introduced algorithms such as multi-objective optimization, Bayesian optimization, or hybrid machine learning. However, these methods often use static methods for data grouping, such as time windows or cross-validation, and lack dynamic adjustment mechanisms. This makes it difficult to capture nonlinear interactions and adapt to the dynamic optimization requirements of multi-objective coupling relationships. This makes it difficult to meet the complex requirements of multi-objective collaborative optimization in flexible reference surface grinding processes, such as end runout difference, stress difference, flatness, and the positive and negative ratio of end runout. Summary of the Invention

[0005] In order to overcome the limitations of existing technologies that rely on manual experience and lack adaptability, the present invention provides a multi-objective grinding decision-making method and system based on dynamic clustering and hyperparameter optimization. The method first optimizes data grouping and target priority through dynamic K-Means clustering, and then efficiently adjusts the hyperparameters of the multi-objective prediction model through collaborative Bayesian optimization through a co-evolutionary algorithm, accurately predicts multi-objective performance indicators such as end jump difference, stress difference, flatness, and end jump positive and negative ratio, and screens the optimal process parameter combination through a comprehensive scoring formula, significantly improving the processing accuracy, stability and efficiency of the ultra-thin sheet grinding process, and ensuring that the end jump positive and negative ratio meets the requirements of maximum positive and minimum negative.

[0006] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions:

[0007] In a first aspect, the present invention provides a multi-objective grinding decision-making method based on dynamic clustering and hyperparameter optimization, comprising:

[0008] Obtaining the input feature data of the grinding process and preprocessing it to obtain a normalized feature set;

[0009] Based on the normalized feature set, a multi-objective prediction model is constructed; the multi-objective prediction model includes a regression model and a classification model, and the regression model is trained by a bagging method with dynamic clustering and target priority adjustment;

[0010] The hyperparameters of the multi-objective prediction model are adjusted collaboratively by co-evolutionary algorithm and Bayesian optimization to generate an optimized multi-objective prediction model;

[0011] The range of grinding process parameters to be optimized is input into the optimized multi-objective prediction model, the multi-objective performance indicators are predicted, and the optimal process parameter combination is selected through comprehensive scoring.

[0012] According to a further technical solution, the input feature data includes the thickness of the workpiece before grinding, the target thickness, the gasket thickness, the number of grinding round trips and the cutting amount.

[0013] According to a further technical solution, the multi-objective prediction model includes a regression model of end jump difference, stress difference and flatness, and a classification model of positive and negative proportions of end jump.

[0014] A further technical solution is to use the RUSBoost algorithm to train a classification model for the positive and negative ratios of end jumps.

[0015] A further technical solution is to train the regression model through the bagging method with dynamic clustering and target priority adjustment as follows:

[0016] The normalized feature set is initially clustered using a clustering algorithm, and the cluster center is dynamically adjusted according to the multi-target prediction error;

[0017] Determine the current priority target based on the target priority, sample data from the cluster with the largest error contribution based on the current priority target, replace the data in the bagging method to generate multiple new sub-datasets, train decision trees based on the new sub-datasets, and integrate multiple decision trees to obtain a trained regression model;

[0018] The end runout difference, stress difference and flatness are predicted based on the trained regression model.

[0019] A further technical solution is to dynamically adjust the cluster center according to the multi-target prediction error. The formula is expressed as:

[0020]

[0021] in, Indicates the The adjustment vector of cluster centers, Indicates the adjustment step size, Indicates the A set of data points at the center of a cluster, Indicates the The prediction error for each data point, Indicates the The mean of the intra-cluster errors, Indicates the The feature vector of the data point, Indicates the The center vector of each cluster.

[0022] Further technical solution, the target priority calculation formula is expressed as:

[0023]

[0024] in, Indicates the The priority of the goals, Indicates the The standardized error of the target, Indicates all The sum of the standardized errors of the targets, Indicates the number of targets.

[0025] In a second aspect, the present invention provides a multi-objective grinding decision system based on dynamic clustering and hyperparameter optimization, comprising:

[0026] A data acquisition module is configured to: acquire input feature data of the grinding process and perform preprocessing to obtain a normalized feature set;

[0027] A model building module is configured to: build a multi-objective prediction model based on the normalized feature set; the multi-objective prediction model includes a regression model and a classification model, and the regression model is trained by a bagging method with dynamic clustering and target priority adjustment;

[0028] A model optimization module is configured to: collaboratively adjust hyperparameters of the multi-objective prediction model through a co-evolutionary algorithm and Bayesian optimization to generate an optimized multi-objective prediction model;

[0029] The parameter optimization module is configured to: input the range of grinding process parameters to be optimized into the optimized multi-objective prediction model, predict the multi-objective performance indicators, and select the optimal process parameter combination through comprehensive scoring.

[0030] In a third aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the multi-objective grinding decision method based on dynamic clustering and hyperparameter optimization as described in the first aspect.

[0031] In a fourth aspect, the present invention provides a computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps of the multi-objective grinding decision method based on dynamic clustering and hyperparameter optimization as described in the first aspect are implemented.

[0032] One or more of the above technical solutions have the following beneficial effects:

[0033] The present invention generates a normalized feature set through data collection and preprocessing to provide input for a multi-objective prediction model; through dynamic K-Means clustering and target priority adjustment mechanism, combined with multi-objective ensemble learning technology, a multi-objective prediction model is constructed to predict multi-objective performance indicators such as end jump difference, stress difference, flatness and positive and negative ratio of end jump, and recommend the optimal parameter combination through comprehensive scoring, overcoming the limitations of traditional methods that rely on manual experience and are difficult to adapt to complex working conditions, and significantly improving the processing accuracy, stability and efficiency of the flexible contact reference surface grinding process.

[0034] The present invention, through the collaborative mechanism of dynamic clustering and target priority adjustment, combined with multi-objective integrated learning technology, can accurately predict the end jump difference, stress difference, flatness and positive and negative ratio of end jump during the grinding process of ultra-thin sheets in complex grinding conditions, significantly improving the accuracy and stability of parameter selection. Dynamic K-Means clustering dynamically adjusts the cluster center through error feedback, and target priority adjustment dynamically optimizes the training process according to the multi-objective error. It combines the global search of the co-evolutionary algorithm with the local refinement of Bayesian optimization, and has higher adaptability and specificity than traditional static grouping methods. The present invention can dynamically recommend gasket thickness, grinding times and cutting amount according to the thickness of the incoming material, ensuring that the end jump difference, stress difference, flatness and positive and negative ratio of end jump for the same ultra-thin sheet are optimally distributed (the maximum end jump is positive and the minimum end jump is negative). It is suitable for the rough grinding, fine grinding and super-fine grinding needs of ultra-thin sheets, improving processing efficiency and product qualification rate. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] The accompanying drawings, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0036] Figure 1 is a flow chart of a multi-objective grinding decision method based on dynamic clustering and hyperparameter optimization according to an embodiment of the present invention;

[0037] Figure 2 The present invention is a flowchart of a bagging method for training a regression model using dynamic clustering and target priority adjustment according to an embodiment of the present invention. DETAILED DESCRIPTION

[0038] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which the present invention belongs.

[0039] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, they indicate the presence of features, steps, operations, devices, components and / or combinations thereof.

[0040] In the absence of conflict, the embodiments of the present invention and the features thereof may be combined with each other.

[0041] Example 1

[0042] like Figure 1 As shown, this embodiment discloses a multi-objective grinding decision method based on dynamic clustering and hyperparameter optimization, which includes the following steps:

[0043] S1: Obtain the input feature data of the grinding process and preprocess it to obtain a normalized feature set;

[0044] In this embodiment, the input feature data includes the thickness of the workpiece before grinding, the target thickness, the gasket thickness, the number of grinding round trips and the cutting depth. The input feature data is preprocessed to generate a normalized feature set.

[0045] Based on the processing data of the ultra-thin sheet grinding process, 500 input feature data were collected through a data acquisition device as a training data set. Each data point contains the pre-grinding thickness, target thickness, shim thickness, number of grinding round trips, and cutting depth. For example, one data point shows the pre-grinding thickness of 3.0mm, the target thickness of 2.88mm, the shim thickness of 0.2mm, the number of grinding round trips of 1, and the cutting depth of 0.01mm. The grinding thickness difference is calculated based on this data. The grinding thickness difference is the difference between the pre-grinding thickness and the target thickness.

[0046] All input feature data are normalized to eliminate magnitude and unit differences and ensure that each feature contributes evenly to the subsequent model. The normalization formula is:

[0047]

[0048] in, represents the normalized eigenvalue, represents the original eigenvalue, represents the original feature mean, represents the original characteristic standard deviation.

[0049] S2: Based on the normalized feature set, a multi-objective prediction model is constructed; the multi-objective prediction model includes a regression model and a classification model, and the regression model is trained by a bagging method with dynamic clustering and target priority adjustment;

[0050] In this embodiment, a multi-objective prediction model is constructed using an ensemble learning algorithm based on a normalized feature set. Dynamic clustering and target-prioritized bagging are used to train regression models for end-jump difference, stress difference, and flatness. A RUSBoost algorithm is used to train a classification model for the positive and negative end-jump ratio. Specifically, the multi-objective prediction model includes three regression models (for predicting end-jump difference, stress difference, and flatness) and one classification model (for predicting the positive and negative end-jump ratio). The regression models are trained using the bagging method, and the classification model is trained using the RUSBoost algorithm.

[0051] Based on the normalized feature set, the dynamic clustering and target-priority adjustment bagging method is used to train the regression model to predict the end jump difference, stress difference and flatness. Specifically:

[0052] (1) Initially cluster the normalized feature set using a clustering algorithm, and dynamically adjust the cluster center based on the multi-target prediction error; use the dynamic K-Means clustering method: set the number of clusters , perform initial clustering on the normalized feature set using the K-Means clustering algorithm to generate clusters; and dynamically adjust the cluster centers according to the multi-target prediction errors.

[0053] The cluster center is dynamically adjusted according to the multi-objective prediction error, and the adjustment formula is:

[0054]

[0055] in, Indicates the The adjustment vector of the cluster center, Indicates the adjustment step size, Indicates the A set of data points at the center of a cluster, Indicates the The prediction error for each data point, Indicates the The mean of the intra-cluster errors, Indicates the The number of data points in a cluster, Indicates the The feature vector of the data point, Indicates the The center vector of the cluster, Indicates the cluster number.

[0056] Furthermore, the regression target error: For the end runout difference, stress difference, and flatness, the regression model uses the bagging method to predict their values. The error is calculated by the mean square error (MSE), which is expressed as:

[0057]

[0058] in, Indicates the The data point in The true value on the target, Indicates the The data point in The predicted value on the target, , corresponding to end runout difference, stress difference, and flatness respectively.

[0059] Classification target error: For the positive and negative ratios of end jumps, the model uses RUSBoost to predict the probability of the positive class, and the error is calculated using the logistic loss:

[0060]

[0061] in, Indicates the The true label (0 or 1) of the data point on the positive and negative ratio target, represents the predicted positive class probability.

[0062] Total prediction error: ,in Indicates the The weights of the targets are as follows: in this embodiment, the end jump difference weight is set to 0.5, the stress difference is set to 1, the flatness is set to 1, and the end jump positive and negative is set to 0.5. It can be flexibly set according to the specific situation without specific limitation.

[0063] After adjustment, the updated cluster center is ← , and re-clustering to make the clustering results dynamically adapt to the multi-target error distribution.,Using the clusters generated by dynamic clustering, data sampling is performed according to the priority,target.

[0064] (2) Determine the current priority target based on the target priority, sample data from the cluster with the largest error contribution based on the current priority target, replace the data in the bagging method to generate multiple new sub-datasets, train decision trees based on the sub-datasets, and integrate multiple decision trees to obtain a trained regression model;

[0065] The end jump difference, stress difference, flatness and positive and negative ratio of end jump are predicted by the validation set, and the mean square error and classification error are calculated as the multi-objective prediction error. The priority of each target is calculated according to its prediction error. The formula is expressed as follows:

[0066]

[0067] in, Indicates the The priority of the goals, Indicates the The standardized error of the target, Indicates all The sum of the standardized errors of the targets, Indicates the target number (in this embodiment is 4, corresponding to the four objectives of runout difference, stress difference, flatness and positive and negative ratio of end runout).

[0068] Example: Assume that the standardized errors are (end jump value error), (stress difference), (flatness), (positive and negative ratio), the total error is The priority of the positive and negative ratio of the end jump is , the highest priority, then the positive and negative ratio of the end jump is used as the current priority target.

[0069] The regression model is trained using the bagging method, specifically:

[0070] Set the number of decision trees and the maximum number of splits , multiple sub-datasets are generated from the normalized feature set by sampling with replacement, and a decision tree is trained independently for each sub-dataset to calculate the predicted values ​​of end jump difference, stress difference and flatness. The bagging method predicts each target value through the integration of multiple decision trees (taking the average of the predicted values ​​of all decision trees). The formula is expressed as:

[0071]

[0072] in, represents the predicted value, Indicates the The predicted value of a decision tree, The units of the decision tree number are mm, MPa, and mm, respectively. Bagging averages the predictions from multiple decision trees to generate the final regression model prediction.

[0073] The highest priority is the current priority target. Based on the current priority target, data is sampled from clusters with the largest error contributions, replacing some of the data in the "bag" used in the bagging method. This generates multiple sub-datasets. A new decision tree is retrained for each new sub-dataset, and multiple decision trees are integrated to form a trained regression model. Specifically, using the clusters generated by dynamic clustering, the error contribution of the positive and negative ratios of each cluster's peer jumps is calculated. The two clusters with the largest error contributions (e.g., clusters 1 and 3) are selected. Ten data points are randomly sampled from each cluster, replacing some of the data in the "bag" used in the bagging method. This generates multiple sub-datasets. A new decision tree is retrained for each new sub-dataset, and multiple decision trees are integrated to form a trained regression model. Dynamic clustering and target priority adjustment are repeated until the multi-target error converges (error change < 0.001) or the preset number of iterations (e.g., 10) is reached.

[0074] Furthermore, the error contribution is defined as: the error contribution of each cluster to the current priority target (given by the target priority The error contribution of the cluster is determined by the mean prediction error of the data points within the cluster. Calculated, ,in, It is Cluster, is The prediction error for a data point.

[0075] Specifically:

[0076] Determine the priority goals: According to the goal priority formula, select the one with the highest priority As a current priority;

[0077] Calculate the intra-cluster error: for each cluster , only count its priority targets The error contribution on ,in It is Data points on target The prediction error on ;

[0078] Sorting selection: for all clusters ,according to Sort from largest to smallest and select the cluster with the largest error contribution (e.g. the first two clusters).

[0079] Furthermore, we can replace some of the data in the "bag" of the bagging method with the following replacement strategy:

[0080] Determine the replacement ratio: Each "bag" (sub-dataset) in the bagging method is generated by sampling with replacement from the normalized feature set and typically contains samples of the size of the original dataset (approximately 500 entries). In this example, 10% of the data is replaced each time, meaning 50 entries are replaced in each bag.

[0081] Randomly select the data to be replaced: In each bag, 50 data are randomly selected as the objects to be replaced to ensure that the replacement process is unbiased.

[0082] Sampling new data: Randomly select 25 data from each of the clusters with larger error contributions (such as cluster 1 and cluster 3) (a total of 50 data), and replace the 50 data selected in the bag.

[0083] (3) Predicting the end runout difference, stress difference, and flatness based on the trained regression model. A separate regression model is trained for each target, i.e., end runout difference, stress difference, and flatness. Each regression model contains multiple decision trees (100 decision trees per model in this embodiment). Three regression models are trained for each of the three targets. Each model contains multiple decision trees, which are integrated to obtain the prediction results.

[0084] The RUSBoost algorithm is used to train the classification model of positive and negative ratios of end jumps, specifically:

[0085] To accurately classify the positive and negative ratios of end-point jumps, this paper uses the RUSBoost algorithm to iteratively update sample weights and optimize the logistic loss function to improve classification performance. This algorithm combines random undersampling with a weighting mechanism to effectively address data imbalance and ensure accurate predictions for minority classes.

[0086] The specific training process of the classification model of end jump positive and negative ratio is:

[0087] (1) Initialization

[0088] Initialize learning rate and training rounds M, with decision tree as weak learner. For the training dataset , which comes from the data acquisition and preprocessing of S1, including grinding experiment records. Each record includes the standardized pre-grinding thickness, target thickness, gasket thickness, grinding round trip times, cutting amount and thickness difference as feature vectors , and the positive and negative labels obtained by measuring the end jump value . Initialize sample weights:

[0089]

[0090] in, Indicates the The initial weight of the samples, Represents the total number of samples, a positive integer, representing the number of grinding experiments.

[0091] Initialize cumulative output: .

[0092] (2) Iterative training

[0093] The RUSBoost algorithm iteratively trains weak learners through the following steps:

[0094] 1) Random undersampling to generate balanced sub-datasets

[0095] In each iteration, based on the number of positive samples (The end jumps straight, ) and the number of negative samples (The end jumps straight, ), randomly selected from negative samples samples, and all positive samples form a balanced sub-dataset , with a scale of 2 , to alleviate the majority class dominance problem.

[0096] 2) Training weak learners

[0097] In the balanced sub-dataset On, use the current sample Training weak learners To match ,Will Map to , so Mapped to .

[0098] 3) Calculate the weighted error rate

[0099]

[0100] in, represents the weighted error rate, represents the sample weight, is the decision tree prediction.

[0101] Weighted error rate: For samples with incorrect predictions, the weight will be added to It is used to measure the performance of weak learners.

[0102] 4) Calculate the weight of weak learners

[0103] According to the weighted error rate, calculate the weak learner The weight is expressed as:

[0104]

[0105] Among them, if , the weak learner has poor performance and can be adjusted Or stop the iteration.

[0106] 5) Update sample weights

[0107]

[0108] And normalize the updated sample weights:

[0109]

[0110] After normalization Assign to , and update the cumulative output:

[0111] .

[0112] 6) Optimize the logistic loss function:

[0113]

[0114] in, Indicates the true label, 0 or 1, 1 indicates a positive end jump, 0 indicates a negative end jump; Represents the probability of the positive class, with a value of (0,1), Represents the feature vector and normalizes the grinding parameters. After multiple iterations, the logistic loss decreases and the classification performance is improved.

[0115] (3) Output the final classification model

[0116] After iterating M times and minimizing the logistic loss function, the final positive class probability and predicted label are output:

[0117]

[0118]

[0119] in, Represents the probability of the positive class, and its value is (0,1).

[0120] S3: The hyperparameters of the multi-objective prediction model are adjusted collaboratively through the co-evolutionary algorithm and Bayesian optimization to generate an optimized multi-objective prediction model;

[0121] In this embodiment, the hyperparameters of the regression model are optimized and the number of decision trees is set. The range is [100,500], the maximum number of splits The range is [1,20]; optimize the hyperparameters of the classification model and set the number of decision trees The range is [100,300], the learning rate The range is [0.01,0.5], the maximum number of splits The range of is [1,10]. The optimization process is achieved through the collaboration of co-evolutionary algorithm and Bayesian hyperparameter optimization, as follows:

[0122] The coevolutionary algorithm optimizes the hyperparameters of the regression and classification models using two populations, each containing 30 hyperparameter combinations. Population 1 optimizes the hyperparameters of the regression model ( , ), with the goal of minimizing the mean square error; Population 2 targets the classification model hyperparameters ( , , ), with the goal of minimizing the logistic loss. The fitness function formula is:

[0123]

[0124] in, Indicates the The fitness value of each hyperparameter combination, represents a hyperparameter combination, represents the mean square error, Represents the logistic loss. Dynamic K-Means clustering prioritizes the target fitness weights, prioritizing optimization of targets with larger errors (such as the ratio of positive and negative end jumps). Evolutionary operations include tournament selection (selecting three combinations and retaining the one with the highest fitness), single-point crossover (with a probability of 0.8), random mutation (with a probability of 0.1), and retaining the original combination (with a probability of 0.1). Random tournament selection randomly extracts three hyperparameter combinations from the population and selects the best-performing combination as the parent by comparing fitness values, balancing selection efficiency and population diversity. The population update formula is:

[0125]

[0126] in, Indicates the New hyperparameter combinations for generations, Indicates the Generation combination, represents a single-point crossover operation, Denotes a random mutation operation. Every five generations, three high-fitness combinations are shared among the population to form a Pareto frontier that adapts to multi-objective trade-offs.

[0127] Bayesian hyperparameter optimization uses the hyperparameter combinations generated by the coevolutionary algorithm (the top five high-fitness combinations in each generation) to build a proxy model using the tree-shaped Parsons estimator to predict hyperparameter performance. The expected improvement function is:

[0128]

[0129] in, represents the expected improvement value, represents a hyperparameter combination, represents the mathematical expectation, Represents hyperparameters The cross validation loss under Represents the loss of the current optimal hyperparameter combination. Loss is calculated using 5-fold cross-validation, and high-potential combinations are selected by maximizing the expected improvement. The top three combinations refined by Bayesian hyperparameter optimization are fed back into the coevolution population to replace combinations with low fitness.

[0130] The optimization process was terminated after 50 iterations or when the expected improvement fell below 0.001. Ten hyperparameter combinations were initially randomly evaluated. The optimal hyperparameter combination was then determined through a collaborative iteration of a coevolutionary algorithm and Bayesian hyperparameter optimization. This generated a high-precision prediction model, improving the prediction performance for runout difference, stress difference, flatness, and the positive / negative ratio of runout. This provided precise input for the comprehensive scoring formula to select the optimal process parameters.

[0131] S4: Inputting the range of grinding process parameters to be optimized into the optimized multi-objective prediction model, predicting the multi-objective performance indicators, and selecting the optimal process parameter combination through comprehensive scoring.

[0132] In this embodiment, the range of grinding process parameters to be optimized is input according to the processing requirements, multi-objective performance indicators are predicted, and the optimal parameter combination is selected through comprehensive scoring to ensure that the positive and negative ratios of end jumps meet the requirements of maximum positive and minimum negative values.

[0133] Get the thickness of the incoming material according to processing requirements and target thickness , set the parameter range to be optimized, including the gasket thickness range (such as [0.2,1.2] mm, step size 0.1 mm), the grinding back and forth times range (such as [2,6] times when the cutting amount is 0.005 mm, and [1,3] times when the cutting amount is [0.01,0.015] mm) and the cutting amount range (such as [0.005,0.015] mm, step size 0.005 mm), and generate a parameter combination set.

[0134] The parameter combination set is predicted, and the predicted values ​​of end runout difference, stress difference, flatness and positive and negative ratio of end runout are calculated, where the positive and negative ratio of end runout is expressed as the probability of the positive class; the predicted values ​​are normalized, specifically: the minimum value is calculated for each target and maximum value , using the formula:

[0135]

[0136] in, represents the normalized predicted value, represents the original predicted value, represents the minimum value of the predicted value, Indicates the maximum value of the predicted value;

[0137] Set the weight vector (such as ), calculate the comprehensive score value, the formula is expressed as:

[0138]

[0139] in, Indicates the comprehensive rating value. 、 、 、 Represents the weight of the corresponding target. The weight vector can be set according to the processing requirements or determined through experiments; represents the normalized end jump difference, represents the normalized stress difference, represents the normalized flatness, Represents the normalized positive and negative ratio of the end jump. By minimizing the comprehensive score Determine the optimal grinding process parameters.

[0140] For a certain set of input parameters (such as pad thickness, grinding times, cutting depth), the prediction value of the final prediction model is fixed, but different parameter combinations will lead to different prediction values. Minimize the comprehensive score value This is achieved by traversing the parameter combination set. The grinding process parameter range to be optimized generates multiple parameter combinations (e.g., 165 groups). Each parameter group is input into the model, and the multi-objective values ​​are predicted and the comprehensive score is calculated. The parameter combination with the smallest comprehensive score is selected.

[0141] Example: Input material thickness 3.0mm, target thickness 2.8mm, parameter ranges set: shim thickness 0.2-1.2mm (0.1mm step), grinding times 1-6, and cut depth 0.005-0.015mm (0.005mm step). This generates 165 parameter combinations. After normalizing the predicted values, weight vectors (1, 1, 0.5, 0.5) are assigned, and the overall score is calculated. The optimal parameters are: shim thickness 1mm, grinding times 2, and cut depth 0.015mm. The overall score is 0.115, achieving the best balance between stress difference, flatness, end runout difference, and positive and negative end runout.

[0142] In order to solve the technical problems in the existing technology that the optimization of grinding process parameters lacks a dynamic adjustment mechanism and it is difficult to achieve multi-objective collaborative optimization, the training data set in the present invention is derived from the grinding experiment records, which includes standardized process parameters and end jump positive and negative labels. A normalized feature set is generated through data acquisition and preprocessing to provide input for the multi-objective prediction model; the present invention uses dynamic K-Means clustering and target priority adjustment mechanism, combined with multi-objective ensemble learning technology, to construct a multi-objective prediction model to predict multi-objective performance indicators such as end jump difference, stress difference, flatness and end jump positive and negative ratio, and recommends the optimal parameter combination through comprehensive scoring, overcoming the limitations of traditional methods that rely on manual experience and are difficult to adapt to complex working conditions, and significantly improving the processing accuracy, stability and efficiency of the flexible contact reference surface grinding process.

[0143] Dynamic K-Means clustering adaptively adjusts data grouping through error feedback, and target priority adjustment dynamically optimizes prediction focus based on multi-objective errors. Co-evolutionary algorithms and Bayesian hyperparameter optimization collaborate to generate high-precision prediction models, forming a Pareto frontier, optimizing multi-objective performance, overcoming the limitations of traditional methods, and dynamically recommending process parameters to ensure the optimal distribution of multiple objectives. It is suitable for coarse grinding, fine grinding, and super-fine grinding, improving product qualification rate and processing efficiency.

[0144] Example 2

[0145] This embodiment discloses a multi-objective grinding decision system based on dynamic clustering and hyperparameter optimization, including:

[0146] A data acquisition module is configured to: acquire input feature data of the grinding process and perform preprocessing to obtain a normalized feature set;

[0147] A model building module is configured to: build a multi-objective prediction model based on the normalized feature set; the multi-objective prediction model includes a regression model and a classification model, and the regression model is trained by a bagging method with dynamic clustering and target priority adjustment;

[0148] A model optimization module is configured to: collaboratively adjust hyperparameters of the multi-objective prediction model through a co-evolutionary algorithm and Bayesian optimization to generate an optimized multi-objective prediction model;

[0149] The parameter optimization module is configured to: input the range of grinding process parameters to be optimized into the optimized multi-objective prediction model, predict the multi-objective performance indicators, and select the optimal process parameter combination through comprehensive scoring.

[0150] Example 3

[0151] The purpose of this embodiment is to provide a computing device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the method of embodiment 1 when executing the program.

[0152] Example 4

[0153] The purpose of this embodiment is to provide a computer-readable storage medium, a computer-readable storage medium having a computer program stored thereon, which performs the steps of the method of embodiment 1 when executed by a processor.

[0154] The steps involved in the apparatuses of Examples 3 and 4 above correspond to those of Method Example 1. For detailed implementation, please refer to the relevant description of Example 1. The term "computer-readable storage medium" should be understood to mean a single medium or multiple media containing one or more instruction sets; it should also be understood to include any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and causing the processor to perform any of the methods of the present invention.

[0155] Those skilled in the art will appreciate that the modules or steps of the present invention described above can be implemented using a general-purpose computer device. Alternatively, they can be implemented using program code executable by a computing device, which can then be stored in a storage device and executed by the computing device. Alternatively, they can be fabricated into separate integrated circuit modules, or multiple modules or steps can be fabricated into a single integrated circuit module for implementation. The present invention is not limited to any specific combination of hardware and software.

[0156] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention are intended to be within the scope of protection of the present invention.

[0157] Although the above describes the specific embodiments of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without any creative work are still within the scope of protection of the present invention.

Claims

1. A multi-objective grinding decision-making method based on dynamic clustering and hyperparameter optimization, characterized by: include: Obtaining the input feature data of the grinding process and preprocessing it to obtain a normalized feature set; Building a multi-objective prediction model based on the normalized feature set; The multi-objective prediction model includes a regression model of end jump difference, stress difference and flatness, and a classification model of positive and negative ratio of end jump, and the regression model is trained by a bagging method with dynamic clustering and target priority adjustment; The dynamic clustering and target priority adjustment bagging method includes: The normalized feature set is initially clustered using a clustering algorithm, and the cluster center is dynamically adjusted according to the multi-target prediction error; Determine the current priority target based on the target priority, sample data from the cluster with the largest error contribution based on the current priority target, replace the data in the bagging method to generate multiple new sub-datasets, train decision trees based on the new sub-datasets, and integrate multiple decision trees to obtain a trained regression model; Predict end runout difference, stress difference and flatness based on the trained regression model; A further technical solution is to dynamically adjust the cluster center according to the multi-target prediction error. The formula is expressed as: in, Indicates the The adjustment vector of cluster centers, Indicates the adjustment step size, Indicates the A set of data points at the center of a cluster, Indicates the The prediction error for each data point, Indicates the The mean of the intra-cluster errors, Indicates the The feature vector of the data point, Indicates the The center vector of each cluster; The hyperparameters of the multi-objective prediction model are adjusted collaboratively by co-evolutionary algorithm and Bayesian optimization to generate an optimized multi-objective prediction model; The range of grinding process parameters to be optimized is input into the optimized multi-objective prediction model, the multi-objective performance indicators are predicted, and the optimal process parameter combination is selected through comprehensive scoring.

2. The multi-objective grinding decision-making method based on dynamic clustering and hyperparameter optimization according to claim 1, characterized in that: The input feature data includes the thickness of the workpiece before grinding, the target thickness, the gasket thickness, the number of grinding round trips and the cutting depth.

3. The multi-objective grinding decision-making method based on dynamic clustering and hyperparameter optimization according to claim 1, characterized in that: The RUSBoost algorithm is used to train the classification model of positive and negative ratios of end jumps.

4. The multi-objective grinding decision-making method based on dynamic clustering and hyperparameter optimization according to claim 1, characterized in that: The target priority calculation formula is expressed as: in, Indicates the The priority of the goals, Indicates the The standardized error of the target, Indicates all The sum of the standardized errors of the targets, Indicates the number of targets.

5. A multi-objective grinding decision system based on dynamic clustering and hyperparameter optimization, characterized by: include: A data acquisition module is configured to: acquire input feature data of the grinding process and perform preprocessing to obtain a normalized feature set; A model building module is configured to: construct a multi-objective prediction model based on the normalized feature set; the multi-objective prediction model includes a regression model of end jump difference, stress difference, and flatness, and a classification model of positive and negative end jump ratio, and train the regression model through a bagging method with dynamic clustering and target priority adjustment; The dynamic clustering and target priority adjustment bagging method includes: The normalized feature set is initially clustered using a clustering algorithm, and the cluster center is dynamically adjusted according to the multi-target prediction error; Determine the current priority target based on the target priority, sample data from the cluster with the largest error contribution based on the current priority target, replace the data in the bagging method to generate multiple new sub-datasets, train decision trees based on the new sub-datasets, and integrate multiple decision trees to obtain a trained regression model; Predict end runout difference, stress difference and flatness based on the trained regression model; A further technical solution is to dynamically adjust the cluster center according to the multi-target prediction error. The formula is expressed as: in, Indicates the The adjustment vector of cluster centers, Indicates the adjustment step size, Indicates the A set of data points at the center of a cluster, Indicates the The prediction error for each data point, Indicates the The mean of the intra-cluster errors, Indicates the The feature vector of the data point, Indicates the The center vector of each cluster; A model optimization module is configured to: collaboratively adjust hyperparameters of the multi-objective prediction model through a co-evolutionary algorithm and Bayesian optimization to generate an optimized multi-objective prediction model; The parameter optimization module is configured to: input the range of grinding process parameters to be optimized into the optimized multi-objective prediction model, predict the multi-objective performance indicators, and select the optimal process parameter combination through comprehensive scoring.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the multi-objective grinding decision-making method based on dynamic clustering and hyperparameter optimization are implemented as described in any one of claims 1 to 4.

7. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the steps of the multi-objective grinding decision-making method based on dynamic clustering and hyperparameter optimization are implemented as described in any one of claims 1 to 4.

Images