Intelligent optimization method for magnetorheological polishing process parameters

Abstract

The invention discloses a magneto-rheological polishing process parameter intelligent optimization method which comprises the following steps: S1, sample data acquisition: carrying out magneto-rheological polishing on a to-be-processed material through an experiment, and in the test process, inputting variables are a plurality of processing process parameters and inputting variables are surface roughness values, so as to obtain a plurality of groups of experiment data; s2, sample data distinguishing: distinguishing the obtained experimental data into a training sample set DT and a verification sample set DV; s3, training sample equalization: carrying out equalization processing on the training sample set DT by adopting an MCSGT algorithm, constructing new sample points by using the training sample set DT, and obtaining a processed training sample set DTE; s4, data normalization: carrying out normalization processing on data in the processed training sample set DTE and verification sample set DV; and S5, model training and evaluation: applying the sample data in the DTE to perform random forest regression model training. According to the method, the prediction model of the magneto-rheological machining process is established, and result prediction is achieved.

Description

Technical Field

[0001] This invention relates to magnetorheological polishing technology, and more specifically, to a method for intelligent optimization of magnetorheological polishing process parameters. Background Technology

[0002] The process parameters of magnetorheological polishing are directly related to the surface polishing quality. Experimental methods can establish the correlation between various process parameters and surface roughness, and their influence can be analyzed using single-factor methods. However, due to the long experimental cycle and high cost, the number of experiments is generally limited, resulting in small-scale data. Therefore, in actual experiments, it is difficult to obtain the optimal combination of process parameters through a limited number of experiments. Furthermore, different processing materials and different abrasives require different process parameters. For example, the process parameters required for gallium nitride wafer processing and sapphire processing are significantly different.

[0003] Therefore, a new solution is needed to address this problem. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art and provide an intelligent optimization method for magnetorheological polishing process parameters. Under the condition of small sample data, combined with optimization algorithms, a predictive model of magnetorheological processing technology can be established, which can predict the optimal combination of magnetorheological processing parameters, reduce processing costs, and improve processing quality.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] A method for intelligent optimization of magnetorheological polishing process parameters includes the following steps:

[0007] S1 Sample Data Acquisition: The material to be treated was subjected to magnetorheological polishing through experiments. During the experiment, the input variables were several processing parameters and the surface roughness value, and several sets of experimental data were obtained.

[0008] S2 Sample Data Differentiation: The obtained experimental data is differentiated into a training sample set DT and a validation sample set D. V ;

[0009] S3 Training Sample Equalization: The MCSGT algorithm is used to equalize the training sample set DT, and new sample points are constructed from the training sample set DT to obtain the processed training sample set D. TE ;

[0010] S4 Data Normalization: This involves normalizing the processed training sample set D... TE and validation sample set D V The data in the data are normalized.

[0011] S5 Model Training and Evaluation: Using sample data from DTE, a random forest regression model is trained.

[0012] The present invention is further configured such that in step S1, the material to be processed is a gallium nitride wafer; the input variables of the processing parameters are set to 6 types, namely spindle speed, slide movement speed, polishing gap, abrasive particle size, pH value, and abrasive type.

[0013] The present invention is further configured such that the calculation method for constructing the training sample set DT in step S3 is as follows:

[0014] j = 1, 2, …, m

[0015] Among them, X i For sample points in the minority class, X j Rand(0,1) generates random numbers between 0 and 1 from its nearest neighbors, and the value of m is determined according to the balance ratio.

[0016] The present invention is further configured to, on the training sample set D T During the construction process, the equalization rate was set to 0.9, resulting in the processed training sample set D. TE .

[0017] The present invention is further configured such that the normalization calculation method in step S4 is as follows:

[0018]

[0019] Where zi represents the original data. This is the normalized data.

[0020] The present invention further specifies that the process of step S5 includes the following steps:

[0021] S501 RF Hyperparameter Settings: Set the number of decision trees and the number of features selected when splitting the tree at its maximum depth;

[0022] S502 RF Model Training: Applying D TE The sample data in the dataset is used to train a random forest regression model;

[0023] S503 RF Model Validation: Practical Validation Sample Set D V The data in the dataset is used to validate the trained model, and the calculation results are inversely normalized to obtain a comparison between the measured data and the predicted data.

[0024] S504 Model Evaluation: Evaluate the model's prediction accuracy. If the prediction accuracy meets the requirements, output the prediction model; if the prediction accuracy does not meet the requirements, re-enter S501 RF hyperparameter settings and retrain the model.

[0025] The present invention is further configured to analyze the prediction error distribution of each sample in the training dataset and the correlation between input and output values ​​in step S502, so as to obtain the influence weight of various processing parameters on the surface roughness value.

[0026] The present invention is further configured to estimate the prediction accuracy of the model by comparing and analyzing the actual values ​​and predicted values ​​of each sample in the training dataset.

[0027] The present invention is further configured to include the step of:

[0028] S6: Intelligent optimization of process parameters: Based on a limited number of experimental data, a better combination of process parameters is obtained through an improved particle swarm optimization algorithm.

[0029] The present invention is further configured such that in step S6, the polishing roughness effect is taken as the optimization target, and the influence relationship between the processing parameters and the roughness is used as the fitness function. The calculated value of the fitness function reflects the polishing roughness effect.

[0030] In summary, the present invention has the following beneficial effects:

[0031] This scheme obtains a small amount of experimental sample data through a limited number of experiments. Combined with optimization algorithms, a predictive model for magnetorheological machining is established. The model maps the relationship between various process parameters and the final surface roughness, and can predict the optimal combination of magnetorheological machining process parameters, thereby reducing processing costs and improving processing quality. Attached Figure Description

[0032] Figure 1 This is a flowchart illustrating the roughness prediction model in this embodiment;

[0033] Figure 2 This is a schematic diagram illustrating the correlation between the actual values ​​and predicted values ​​of the training dataset in this embodiment;

[0034] Figure 3 This is a schematic diagram illustrating how the out-of-bag error rate varies with the number of decision trees in this embodiment;

[0035] Figure 4 This is a schematic diagram showing the input variables sorted by weight in this embodiment;

[0036] Figure 5 This is a schematic diagram comparing the measured and predicted data on the verification dataset in this embodiment;

[0037] Figure 6 This is a schematic diagram of the fitness function value iteration results in this embodiment. Detailed Implementation

[0038] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0039] This embodiment discloses an intelligent optimization method for magnetorheological polishing process parameters. Based on experimental data, a predictive model for surface roughness is established using machine learning algorithms. This model reflects the mapping relationship between process parameters and surface roughness.

[0040] Regression decision trees are the foundation of the random forest regression algorithm, and their decision-making principle can be described as follows:

[0041] Assuming X and Y are the input and output variables, respectively, given a training dataset... ,in Let n be the number of features, i = 1, 2, ..., N, and N be the sample size. The j-th feature variable x(j) and its value s in the training set are used as the splitting variable and splitting point, and a region is defined. and The decision tree regression principle can be expressed as the following optimization model:

[0042] (Equation 1)

[0043] (Equation 2)

[0044] , (Equation 3)

[0045] In the formula, N1 and N2 are the number of samples in regions R1 and R2, respectively.

[0046] When training data is imbalanced, minority class samples are difficult to train effectively, potentially leading to underfitting. Conversely, majority class samples may overfit, resulting in low model generalization. To balance the sample data, an improved random sampling algorithm is employed. This algorithm is called Minority Self-Incrementing Technique (MCSGT). This method uses an interpolation algorithm to construct new sample points X from the minority class sample points. N This is done to increase the sample size. The calculation method is as follows:

[0047] , j = 1, 2, …, m (Equation 4)

[0048] In the formula, X i For sample points in the minority class, X j Rand(0,1) generates random numbers between 0 and 1 from several of its nearest neighbors, and the value of m is determined by the desired balance ratio.

[0049] Let the sample With n features, each feature has a different degree of influence on the final prediction result; that is, the weight of each feature is different for the prediction result. In the random forest model, the weight of a feature can be obtained by averaging its weights across all decision trees. Within each decision tree, the weights of each feature can be obtained as follows:

[0050] Calculate the weight R of a node k k :

[0051] (Equation 5)

[0052] In the formula, w k w l w r Gk, Gl, and Gr represent the ratios of the number of training samples in node k and its left and right child nodes to the total number of training samples, respectively. Gk, Gl, and Gr represent the impurity of node k and its left and right child nodes, respectively, and their values ​​can be obtained according to equation (2).

[0053] Calculate the weight R of each node. k Then, the weight F of a certain feature can be obtained according to the following formula. i :

[0054] (Equation 6)

[0055] In the formula, N j N represents the number of nodes associated with feature i; k This represents the total number of nodes.

[0056] To make the sum of the weights of all features equal to 1, they can be normalized using the following formula:

[0057] (Equation 7)

[0058] This embodiment discloses an intelligent optimization method for magnetorheological polishing process parameters, referring to... Figure 1 As shown, the specific steps are as follows:

[0059] S1 Sample Data Acquisition: The material to be treated was subjected to magnetorheological polishing through experiments. During the experiment, the input variables were several processing parameters and the surface roughness value, and several sets of experimental data were obtained.

[0060] S2 Sample Data Differentiation: The obtained experimental data is differentiated into training sample set D. T and validation sample set D V ;

[0061] S3 Training Sample Equalization: The MCSGT algorithm is used to equalize the training sample set D. T Perform equalization processing on the training sample set D T Construct new sample points to obtain the processed training sample set D. TE ;

[0062] S4 Data Normalization: This involves normalizing the processed training sample set D... TE and validation sample set D V The data in the data are normalized.

[0063] S5 Model Training and Evaluation: Application of D TE The sample data in the model were used to train a random forest regression model.

[0064] S6: Intelligent optimization of process parameters: Based on a limited number of experimental data, a better combination of process parameters is obtained through an improved particle swarm optimization algorithm.

[0065] Step S1 involves sample data collection:

[0066] The material to be treated was subjected to magnetorheological polishing through experiments. During the experiment, the input variables were several processing parameters and the surface roughness value, and several sets of experimental data were obtained.

[0067] Specifically, in this embodiment, the material to be processed is a gallium nitride wafer. By conducting experiments and establishing a model according to the method in this embodiment, a model suitable for this material can be obtained. For other materials, the same steps can be used to conduct experiments, thereby obtaining experimental models for the materials.

[0068] In this embodiment, six input variables are set: spindle speed, slide speed, polishing gap, abrasive particle size, pH value, and abrasive type. The abrasive type is represented by 1-6, which represent six materials: CeO2, Al2O3, Diamond, SiO2+CeO2, SiO2+Al2O3, and SiO2+Diamond, respectively. See Table 1 for details.

[0069] Table 1 Process Parameter Details

[0070]

[0071] Among them, the sample data in step S2 are distinguished as follows:

[0072] The 40 sets of data obtained from the experiment were distinguished, with 30 sets used as the training sample set D. T D T = {(x1, y1),(x2, y2), …, (x 30 , y 30 )}, 10 groups were used as the verification sample set D V D V = {(x1, y1), (x2, y2), …, (x 10 ,y 10 )};

[0073] Step S3 involves training sample equalization:

[0074] The MCSGT algorithm is used to train the sample set D. T Perform equalization processing on the training sample set D T Construct new sample points to obtain the processed training sample set D. TE ;

[0075] On the training sample set D T During the construction process, the calculation method described in Equation 4 above is as follows:

[0076] j = 1, 2, …, m

[0077] In the formula, X i For sample points in the minority class, X j Rand(0,1) is a random number generated from several of its nearest neighbors, and the value of m is determined according to the balance ratio.

[0078] For example, Nmin is the original number of minority class samples, Nmax is the original number of majority class samples, r is the equalization rate, Nnew is the number of new minority class samples to be generated, and Ntotal is the total number of samples after equalization.

[0079] The equilibrium ratio r = (Nmin + Nnew) / Nmax, and Nnew = r·Nmax - Nmin, can be used to obtain the number of newly produced samples.

[0080] For example, in the training sample set D T During the construction process, the equalization rate was 0.9, and the m value was specifically 36, resulting in the processed training sample set D. TE D TE= {(x1, y1), (x2, y2), …, (x 36 , y 36 )}.

[0081] Step S4: Data normalization:

[0082] To accelerate convergence and improve prediction accuracy, the data in the processed training sample set DTE and validation sample set DV are normalized as follows:

[0083]

[0084] Among them, z i The original data, , This is the normalized data.

[0085] S5 Model Training and Evaluation: Application of D TE Using the sample data in the dataset, a random forest regression model is trained; based on the models shown in Equations (1), (2), and (3) above, according to... Figure 1 The steps shown are for training a random forest regression model; the specific steps include the following:

[0086] S501 RF Hyperparameter Settings: Set the hyperparameters of RF to the number of decision trees, the maximum depth of the trees, and the number of features selected during splits; specifically, it can be set to: the number of decision trees is 200, the maximum depth of the trees is 100, and the number of features selected during splits is 6;

[0087] S502 RF Model Training: Applying D TE Using the sample data in the dataset, we trained a random forest regression model, analyzed the prediction error distribution of each sample in the training dataset and the correlation between input and output values, and obtained the influence weights of various processing parameters on surface roughness values.

[0088] The distribution of prediction errors for each sample in the training dataset and the correlation between input and output values, such as... Figure 2 As shown. Among them, Figure 2 The x-axis of each point in the graph represents the actual value of the achievement of the training objective in the training dataset, and the y-axis represents the predicted value. The slope of the straight line is equal to 1. As can be seen from the graph, the data points are mainly distributed around the straight line. The calculation results (normalized data) show a maximum error of 0.16 and an average error of 0.04. These results indicate that the established model has good prediction accuracy. Furthermore, the linear correlation R between the input variables and the output values ​​is... 2 = 0.96, indicating that the two have a strong correlation.

[0089] The out-of-bag error rate varies with the number of decision trees as follows: Figure 3As shown. The out-of-bag error rate is an unbiased estimate of the generalization error of a random forest; the smaller the value, the higher the prediction accuracy. From Figure 3 As can be seen, the out-of-bag error rate gradually decreases as the number of decision trees increases. When the number of decision trees reaches about 150, the out-of-bag error rate stabilizes at 0.006, indicating that it has a very high prediction accuracy.

[0090] Based on the above equations (5) to (7), the processing parameters x can be calculated. (1) - x (6) The weights of the influence on surface roughness are 0.28, 0.25, 0.16, 0.18, 0.06, and 0.07, respectively. The variables are sorted by weight as follows: Figure 4 As shown. From Figure 4 As can be seen, the spindle speed, slide speed, polishing gap, and abrasive particle size have a significant impact on the surface roughness value, while the other two parameters, pH value and abrasive type, have a smaller impact. In this embodiment, the above data is for illustrative purposes only. In the actual model determination process, experimental data can be collected and specific calculations can be performed based on the actual model.

[0091] S503 RF model verification:

[0092] Practical verification sample set D V The data in the dataset is used to validate the trained model, and the calculation results are inversely normalized to obtain a comparison between the measured data and the predicted data. The comparison between the measured data and the predicted data is as follows: Figure 5 As shown in the figure, the maximum relative prediction error is 7.97%, the average relative prediction error is 4.77%, and the standard deviation is 0.02. The validation results show that the model has achieved high prediction accuracy. The output of this model is...

[0093] In functional form, it is denoted as R = DTree(x) (j) ), j = 1, 2, …, 6. R represents the roughness value, x (j) Indicates process parameters.

[0094] S504 Model Evaluation:

[0095] The accuracy of a model's predictions can be evaluated by comparing measured and predicted data; for example, it can be assessed using parameters such as large relative prediction error, average relative prediction error, and standard deviation.

[0096] If the prediction accuracy meets the condition, the prediction model will be output.

[0097] If the prediction accuracy does not meet the requirements, the process will re-enter step S501 RF hyperparameter setting and retrain the model; alternatively, it can re-enter step S1 sample data acquisition and re-acquire and process the data.

[0098] Step S6 involves intelligent optimization of process parameters: using an improved particle swarm optimization algorithm, based on a finite number of experimental data, a better combination of process parameters is obtained; with optimal polishing roughness as the optimization objective, the influence relationship between processing parameters and roughness is used as the fitness function, which is R = DTree(x (j) The smaller the value, the better the polishing roughness effect.

[0099] Specifically, an improved particle swarm optimization algorithm is used in step S6:

[0100] S601 Improved Particle Swarm Optimization Algorithm

[0101] Suppose there is a population of N particles in an M-dimensional search space, where the i-th particle can be represented by a vector X. i It means that X i = (x i1 , x i2 , …, x iM ), i = 1, 2, …, N; the velocity of the i-th particle can be represented by the vector V. i It means, V i = (V i1 V i2 , …, V iM Let i = 1, 2, …, N; the individual optimal solution for the i-th particle is denoted by pbesti, and the population optimal solution is denoted by gbest. The i-th particle updates its velocity and position according to the following formula:

[0102] (Equation 8)

[0103] (Equation 9)

[0104] In the formula, w is the inertia factor; m represents the number of updates; c1 and c2 are learning factors; and r1 and r2 are random numbers between (0 and 1).

[0105] To improve the local and global optimal solution search capabilities of the particle swarm optimization algorithm described above, this embodiment employs random inertia weights for particle velocity updates. The weight calculation method is as follows:

[0106] (Equation 10)

[0107] μ min μ maxRepresents the minimum and maximum values ​​of the random inertia weights; rand is a random number between (0, 1), randn is a normally distributed random number, and σ is the standard deviation.

[0108] S602 Process Parameter Optimization and Analysis

[0109] Set the number of particles in the population to N = 5, learning factors c1 = 2, c2 = 2, and minimum weight μ. min = 0.8, maximum value μ max = 1, standard deviation σ = 0.05. The domain of the input parameters is [0 1], and the range of particle velocity is [-0.5 0.5].

[0110] In this embodiment, the optimal surface polishing quality is taken as the search objective. The influence of processing parameters on roughness is used as the fitness function, and a smaller value is better. Therefore, the fitness function is R = DTree(x (j) ).

[0111] Using the above parameters and fitness function, the fitness function values ​​are calculated according to equations (8) to (10) to obtain the iterative results, as follows: Figure 6 As shown.

[0112] The optimization result can be obtained by outputting the particle position corresponding to the globally optimal fitness value. For example, the optimized values ​​(normalized results) of each input variable are: 0.687, 0.463, 0.591, 0.387, 0.875, 0.902, with a corresponding predicted roughness value of 0.0013. After inverse normalization calculation, the actual values ​​of each process parameter after optimization are 306, 963, 0.55, 3, 10, 6, and the roughness value of the output value is 9.09.

[0113] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A method for intelligent optimization of magnetorheological polishing process parameters, characterized in that, Including the following steps: S1 Sample Data Acquisition: The material to be treated was subjected to magnetorheological polishing through experiments. During the experiment, the input variables were several processing parameters and the surface roughness value, and several sets of experimental data were obtained. S2 Sample Data Differentiation: The obtained experimental data is differentiated into a training sample set DT and a validation sample set D. V ; S3 Training Sample Equalization: The MCSGT algorithm is used to equalize the training sample set DT, and new sample points are constructed from the training sample set DT to obtain the processed training sample set D. TE ; S4 Data Normalization: This involves normalizing the processed training sample set D... TE and validation sample set D V The data in the data are normalized. S5 Model Training and Evaluation: Using sample data from DTE, a random forest regression model is trained.

2. The intelligent optimization method for magnetorheological polishing process parameters according to claim 1, characterized in that, In step S1, the material to be processed is a gallium nitride wafer; the input variables for the processing parameters are set to 6 types, namely spindle speed, slide speed, polishing gap, abrasive particle size, pH value, and abrasive type.

3. The intelligent optimization method for magnetorheological polishing process parameters according to claim 1, characterized in that, In step S3, the calculation method for constructing the training sample set DT is as follows: , j = 1, 2, …, m Among them, X i For sample points in the minority class, X j Rand(0, 1) generates random numbers between 0 and 1 from several of its nearest neighbors, and the value of m is determined according to the balance ratio.

4. The intelligent optimization method for magnetorheological polishing process parameters according to claim 3, characterized in that, On the training sample set D T During the construction process, the equalization rate was set to 0.9, resulting in the processed training sample set D. TE .

5. The intelligent optimization method for magnetorheological polishing process parameters according to claim 1, characterized in that, In step S4, the normalization process is calculated as follows: Where zi represents the original data. This is the normalized data.

6. The intelligent optimization method for magnetorheological polishing process parameters according to claim 1, characterized in that, The process of step S5 includes the following steps: S501 RF Hyperparameter Settings: Set the number of decision trees and the number of features selected when splitting the tree at its maximum depth; S502 RF Model Training: Applying D TE The sample data in the dataset is used to train a random forest regression model; S503 RF Model Validation: Practical Validation Sample Set D V The data in the dataset is used to validate the trained model, and the calculation results are inversely normalized to obtain a comparison between the measured data and the predicted data. S504 Model Evaluation: Evaluate the model's prediction accuracy. If the prediction accuracy meets the requirements, output the prediction model; if the prediction accuracy does not meet the requirements, re-enter S501 RF hyperparameter settings and retrain the model.

7. The intelligent optimization method for magnetorheological polishing process parameters according to claim 6, characterized in that, In step S502, the prediction error distribution of each sample in the training dataset and the correlation between input and output values ​​are analyzed to obtain the influence weights of various processing parameters on surface roughness values.

8. The intelligent optimization method for magnetorheological polishing process parameters according to claim 6, characterized in that, The prediction accuracy of the model is estimated by comparing and analyzing the actual values ​​and predicted values ​​of each sample in the training dataset.

9. A method for intelligent optimization of magnetorheological polishing process parameters according to any one of claims 1-8, characterized in that, It also includes the following steps: S6: Intelligent optimization of process parameters: Based on a limited number of experimental data, a better combination of process parameters is obtained through an improved particle swarm optimization algorithm.

10. The intelligent optimization method for magnetorheological polishing process parameters according to claim 9, characterized in that, In step S6, the polishing roughness effect is taken as the optimization target, and the influence relationship between the processing parameters and the roughness is used as the fitness function. The calculated value of the fitness function reflects the polishing roughness effect.

Images