CN107292051A - A kind of carbide chip chemically-mechanicapolish polishes the Forecasting Methodology of surface roughness - Google Patents

Abstract

The invention discloses a method for predicting the surface roughness of chemical mechanical polishing of a cemented carbide blade, comprising the following steps: step 1, designing experimental parameters and experimental schemes of chemical mechanical polishing of a cemented carbide blade, and collecting experimental data; step 2, The Gaussian function-based anomaly detection algorithm is used to preprocess the experimental sample data; step 3, establish a genetic algorithm to optimize the prediction model of the BP neural network, and use the preprocessed experimental sample data to learn and train the prediction model, so as to obtain Surface roughness prediction model for cemented carbide inserts. The present invention uses Gaussian function-based abnormality detection algorithm to preprocess the experimental sample data, eliminates abnormal data groups, optimizes the threshold and weight of BP neural network through genetic algorithm, establishes a high-precision surface roughness prediction model, and improves the chemical Mechanical polishing efficiency.

Description

A Prediction Method for Surface Roughness of Chemical Mechanical Polishing of Cemented Carbide Inserts

technical field

The invention relates to a method for predicting the surface roughness of chemical mechanical polishing of a cemented carbide blade, belonging to the technical field of mechanical processing technology.

Background technique

Chemical mechanical polishing has become one of the important precision machining methods for cemented carbide inserts because of its advantages of avoiding mechanical surface damage, improving the flatness of the insert and reducing thermal influence. However, in the actual processing of chemical mechanical polishing carbide inserts, the polishing process parameters are often determined through repeated experiments, and the polishing quality is controlled by experience and semi-empirical means. The above problems cause unstable processing and low production efficiency. In mechanical processing, the surface roughness of the blade is an important index to measure the quality of the blade product. Therefore, the establishment of an accurate surface roughness prediction model has great practical significance for improving the efficiency of chemical mechanical polishing, improving the precision of chemical mechanical polishing and reducing production costs. .

In the process of chemical mechanical polishing of cemented carbide inserts, there are many factors affecting the surface roughness, and these factors are coupled with each other. Therefore, there is a complex nonlinear relationship between each chemical mechanical polishing process parameter and surface roughness. The neural network The intelligent features and capabilities enable it to effectively deal with information processing problems such as complex nonlinear system control, feature extraction and recognition.

Contents of the invention

In view of the above problems, the present invention proposes a method for predicting the surface roughness of chemical mechanical polishing of cemented carbide inserts. The method uses an abnormality detection algorithm based on a Gaussian function to preprocess the experimental sample data, eliminates abnormal data groups, and utilizes a genetic algorithm to Optimizing the nonlinear function approximation characteristics of the BP artificial neural network model can quickly and accurately predict the surface roughness of the cemented carbide insert.

The technical scheme adopted in the present invention is as follows: a method for predicting surface roughness of chemical mechanical polishing of cemented carbide inserts, comprising the following steps:

Step 1. Designing experimental parameters and experimental schemes for chemical mechanical polishing of cemented carbide inserts, and collecting experimental data;

Step 2. Using an abnormality detection algorithm based on a Gaussian function to preprocess the experimental sample data;

Step 3: Establish a genetic algorithm to optimize the prediction model of the BP neural network, and use the preprocessed experimental sample data to learn and train the prediction model, so as to obtain the surface roughness prediction model of the cemented carbide insert under different conditions.

In step one, the experiment of chemical mechanical polishing of cemented carbide blades is the influence on the surface roughness among the six factors of surface roughness before polishing, polishing time, polishing speed, polishing pressure, abrasive particle size and polishing liquid concentration experiment.

In step 2, the Gaussian function-based anomaly detection algorithm is specifically:

1) Construct simulation data

Determine the training sample data Where x _ij represents the experimental data of the jth factor in the i-th group;

2) Determine the Gaussian distribution parameters

Probability density function for a Gaussian distribution: In the formula, p(x _ij ; μ _i ,σ _i ² ) represents the probability of normal data; μ _i represents the mean value of the experimental data of the i-th group, which is given by the formula Determined; σ _i represents the variance of the i-th group of experimental data, which is determined by the formula Determined, where n represents the number of factors contained in the data;

3) Determine the threshold

After the Gaussian parameters are established, the Gaussian distribution profile can be obtained, and one of the methods to determine whether the training sample points are abnormal is to set a threshold ε based on the cross-validation set, and points with a probability lower than the threshold can be considered as abnormal points;

4) get the result

Eliminate the detected outlier data sets to obtain predictable experimental sample data.

In step 3, the prediction model of the genetic algorithm optimization BP neural network is to use the genetic algorithm to replace the gradient descent method of the BP neural network to optimize the prediction error, and the specific steps are as follows:

1) Data normalization

Random 70% of the preprocessed experimental sample data is used as the training data set, and the other 30% of the data is used as the test data set; in order to eliminate the order of magnitude difference between the data of each dimension, the training data set and the test data set are normalized The training set and test set are obtained by normalization, and the data normalization process adopts the maximum and minimum method, that is, x _min is the minimum value in the data sequence, and x _max is the maximum value in the data sequence;

2) Construction of BP artificial neural network

a) Input layer node number n: According to the prediction model, the input layer node number is the number of factors affecting the surface roughness of the cemented carbide insert;

b) Number of nodes in the output layer m: according to the prediction model, the number of nodes in the output layer is the number of dependent variables of the chemical mechanical polishing experiment;

c) The number of hidden layer nodes l: according to the formula In the formula, n is the number of nodes in the input layer, m is the number of nodes in the output layer, and a is a constant between 0 and 10;

d) The connection weight matrix w _ij [n×l] between the input layer and the hidden layer, and the connection weight matrix w _jk [l×m] between the hidden layer and the input layer;

e) Hidden layer threshold matrix a _j [l×1], output layer threshold matrix b _k [m×1];

f) Hidden layer activation function f(x):

g) Hidden layer output H _j :

h) Hidden layer output O _k :

i) Error e _k : e _k =O _k -y _k is calculated according to the difference between the network prediction output _Ok and the actual expected output y _k ;

3) Construction of genetic algorithm

a) Population initialization: Randomly generate an initial population of size N, N is generally set to 20-50, and each individual of the population consists of the connection weight matrix w _ij [n×l] between the input layer and the hidden layer, the hidden layer and input layer connection weight matrix w _jk [l×m], hidden layer threshold matrix a _j [l×1] and output layer threshold matrix b _k [m×1], the length of the gene sequence L according to the formula The calculation is as follows: L=long w _ij +long a _j +long w _jk +long b _k ;

b) Fitness function F: the sum of the squares of the BP network prediction error e _k is used as the individual fitness value of the genetic algorithm, where n represents the number of input layer nodes;

c) Selection operation: the selection strategy based on the fitness ratio, the selection probability of each individual In the formula, p _i represents the selection probability of the i-th individual, F _i represents the fitness of the i-th individual, and N represents the number of individuals in the population;

d) Crossover operation: that is, chromosome individual crossover, In the formula, a _rt represents the gene of the r-th individual at position t, a _st represents the gene of the s-th individual at position t, b represents the crossover operator, and b generally takes 0.4-0.9;

e) Mutation operation: select the j-th gene a _ij of the i-th individual to mutate, In the formula, a _max represents the upper bound of gene a _ij , a _min represents the lower bound of gene a _ij , f(g)=1-g/G, g represents the current evolutionary generation, G represents the termination evolutionary generation, r represents the mutation operator, r generally takes 0.001-0.1;

f) Termination evolution algebra G: G is generally set to 100-500;

4) Surface roughness prediction

The training set is used to learn and train the model, so that the surface roughness prediction model of the blade under different conditions can be obtained; the test set is used to test the trained network model to obtain the accuracy of the prediction model, and then use the model whose accuracy meets the requirements Prediction of surface roughness of chemically mechanically polished carbide inserts.

The beneficial effects of the present invention are:

1. The present invention uses the Gaussian function-based anomaly detection algorithm to preprocess the experimental sample data. The Gaussian process method has very strong applicability and generalization ability for high-dimensional nonlinear small sample data, and the Gaussian function-based anomaly detection algorithm can detect For abnormal data in the data, the prediction model is very dependent on the data. When encountering abnormal data, the prediction accuracy of the prediction model will be greatly reduced. Using the Gaussian function-based abnormal detection algorithm to preprocess the experimental sample data can improve the accuracy of the predictive model;

Two, the present invention adopts genetic algorithm to optimize the weight and threshold of BP neural network, compared to the gradient descent method of BP neural network to optimize, genetic algorithm has inherent implicit parallelism and better global optimization ability, can be in the global range A near-optimal solution is found within, which avoids being trapped in a local optimal solution, so that the prediction accuracy of the surface roughness of the chemical-mechanical polishing of the cemented carbide insert is improved, the convergence speed is fast, and the stability is enhanced;

3. The present invention establishes an accurate surface roughness prediction model, which has great practical significance for improving the efficiency of chemical mechanical polishing, improving the precision of chemical mechanical polishing and reducing production costs.

Description of drawings

Fig. 1 is a flow chart of the surface roughness prediction model for chemical mechanical polishing of cemented carbide inserts.

Fig. 2 is a flow chart of abnormality detection algorithm detection based on Gaussian function.

Fig. 3 is a flow chart of genetic algorithm optimization of BP neural network in the embodiment.

Fig. 4 is a graph of the surface roughness fitting results of the prediction model to the training data set in the embodiment.

Fig. 5 is a diagram of the surface roughness prediction results of the prediction model for the test data set in the embodiment.

Fig. 6 is a comparison chart of the prediction error between the BP neural network and the optimization algorithm of this patent.

detailed description

The present invention will be described in further detail below in conjunction with the accompanying drawings and embodiments, but the protection scope of the present invention is not limited to the following embodiments.

Fig. 1 is the flow chart of the surface roughness prediction model of chemical mechanical polishing cemented carbide blade of the present invention, specifically comprises the following steps:

Step 1. Design the experimental parameters and experimental schemes of chemical mechanical polishing carbide inserts, and the collection of experimental data;

Step 2. Using an abnormality detection algorithm based on a Gaussian function to preprocess the experimental sample data;

Step 3: Establish a genetic algorithm to optimize the prediction model of the BP neural network, and use the preprocessed experimental sample data to learn and train the prediction model, so as to obtain the surface roughness prediction model of the cemented carbide insert under different conditions.

In step one, the experiment of chemical mechanical polishing of cemented carbide blades is an experiment on the influence of six factors on surface roughness before polishing, polishing time, polishing speed, polishing pressure, abrasive grain size and polishing concentration, The data collected from 72 original experimental samples are shown in Table 1 below:

Table 1 Original experimental sample data table

Figure 2 is a flow chart of Gaussian function-based anomaly detection algorithm detection, which specifically includes the following steps:

1) Construct simulation data

Determine the training sample data Where x _ij represents the experimental data of the jth factor in the i-th group;

2) Determine the Gaussian distribution parameters

Probability density function for a Gaussian distribution: In the formula, p(x _ij ; μ _i ,σ _i ² ) represents the probability of normal data; μ _i represents the mean value of the experimental data of the i-th group, which is given by the formula Determined; σ _i represents the variance of the i-th group of experimental data, which is determined by the formula Determined, where n represents the number of factors contained in the data;

3) Determine the threshold

After the Gaussian parameters are established, the Gaussian distribution profile can be obtained, and one of the methods to determine whether the training sample points are abnormal is to set a threshold ε of 5.8×10 ^-17 based on the cross-validation set, and points with a probability lower than the threshold can be considered as abnormal point;

4) get the result

Eliminate the outlier data groups with the detected experimental sample numbers 33 and 69 to obtain predictable experimental sample data.

Fig. 3 is the flow chart of genetic algorithm optimization BP neural network, and the prediction model of described genetic algorithm optimization BP neural network is to utilize genetic algorithm to replace the gradient descent method of BP neural network to optimize prediction error, concrete steps are as follows:

1) Data normalization

Random 70% of the preprocessed experimental sample data is used as the training data set, and the other 30% of the data is used as the test data set; in order to eliminate the order of magnitude difference between the data of each dimension, the training data set and the test data set are normalized The training set and test set are obtained by normalization, and the data normalization process adopts the maximum and minimum method, that is, x _min is the minimum value in the data sequence, and x _max is the maximum value in the data sequence;

2) Construction of BP artificial neural network

a) The number of nodes in the input layer n: According to the prediction model, the number of nodes in the input layer is the number of factors that affect the surface roughness of the cemented carbide insert, that is, the surface roughness before polishing, polishing time, polishing speed, polishing pressure, abrasive grains Size, concentration of abrasive fluid, etc. 6;

b) Number of nodes in the output layer m: According to the prediction model, the number of nodes in the output layer is the number of dependent variables of the chemical mechanical polishing experiment, that is, the surface roughness of the cemented carbide blade is 1;

c) The number of hidden layer nodes l: according to the formula In the formula, n is the number of input layer nodes, m is the number of output layer nodes, and a is a constant between 0 and 10. In the present invention, a gets 5, so the number of hidden layer nodes is 7;

d) The connection weight matrix w _ij [6×7] between the input layer and the hidden layer, and the connection weight matrix w _jk [7×1] between the hidden layer and the input layer;

e) Hidden layer threshold matrix a _j [7×1], output layer threshold matrix b _k [1×1];

f) Hidden layer activation function f(x):

g) Hidden layer output H _j :

h) Hidden layer output O _k :

i) Error e _k : e _k =O _k -y _k is calculated according to the difference between the network prediction output _Ok and the actual expected output y _k ;

3) Construction of genetic algorithm

a) Population initialization: Randomly generate an initial population with a size of 50, and each individual in this population consists of the connection weight matrix w _ij [6×7] between the input layer and the hidden layer, the connection weight between the hidden layer and the input layer Matrix w _jk [7×1], hidden layer threshold matrix a _j [7×1] and output layer threshold matrix b _k [1×1], the length L of its gene sequence is calculated according to the formula as follows:

L＝long w _ij +long a _j +long w _jk +long b _k ;

b) Fitness function F: the sum of the squares of the BP network prediction error e _k is used as the individual fitness value of the genetic algorithm, where n represents the number of input layer nodes;

c) Selection operation: the selection strategy based on the fitness ratio, the selection probability of each individual In the formula, pi represents the selection probability of the i-th individual, F _i represents the fitness of the i-th individual, and N represents the number of individuals in the population;

d) Crossover operation: that is, chromosome individual crossover, In the formula, a _rt represents the gene of the rth individual at the t position, a _st represents the gene of the s individual at the t position, and b represents the crossover operator, and b is 0.4 in the present invention;

e) Mutation operation: select the j-th gene a _ij of the i-th individual to mutate, In the formula, a _max represents the upper bound of gene a _ij , a _min represents the lower bound of gene a _ij , f(g)=1-g/G, g represents the current evolutionary generation, G represents the termination evolutionary generation, r represents the mutation operator, The present invention r takes 0.1;

f) Termination evolution algebra G: G in the present invention is taken as 100;

4) Surface roughness prediction

The training set is used to learn and train the model, so that the surface roughness prediction model of the blade under different conditions can be obtained. The surface roughness fitting results of the prediction model to the training data set are shown in Figure 4; The network model is tested, and the surface roughness prediction results of the prediction model on the test data set are shown in Figure 5. Using the test set to test the trained network model, the accuracy of the prediction model is 0.968.

5) Comparison of prediction models

Compared with the traditional BP neural network prediction model with a prediction accuracy of 0.913, the optimization algorithm of this patent can effectively predict the surface roughness of chemical mechanical polishing of cemented carbide inserts. The comparison of the prediction error between the BP neural network and the optimization algorithm of this patent is shown in Figure 6 .

Claims (4)

1. a kind of carbide chip chemically-mechanicapolish polishes the Forecasting Methodology of surface roughness, it is characterised in that including following step Suddenly：

Step 1: design chemically mechanical polishing carbide chip experiment parameter and experimental program, and experimental data collection；

Step 2: carrying out the pretreatment of experiment sample data using the Outlier Detection Algorithm based on Gaussian function；

Step 3: setting up the forecast model of genetic algorithm optimization BP neural network, pretreated experiment sample data pair are utilized Forecast model carries out learning training, so as to obtain the Roughness Model of carbide chip under different condition.

2. a kind of carbide chip according to claim 1 chemically-mechanicapolish polishes the Forecasting Methodology of surface roughness, its It is characterised by, in step one, described chemically mechanical polishing carbide chip experiment is surface roughness, polishing before polishing The influence to surface roughness between time, polishing velocity, polish pressure, six factors of abrasive particle size and polishing fluid concentration is real Test.

3. a kind of carbide chip according to claim 1 chemically-mechanicapolish polishes the Forecasting Methodology of surface roughness, its It is characterised by, in step 2, the Outlier Detection Algorithm based on Gaussian function is specially：

1) emulation data are built

Determine training sample dataWherein x_ijRepresent the experimental data of i-th group of j-th of factor；

2) Gaussian Distribution Parameters are determined

The probability density function of Gaussian Profile：In formula,Represent just The probability of regular data；μ_iThe average of i-th group of experimental data is represented, by formulaIt is determined that；σ_iRepresent i-th group of experiment number According to variance, by formulaIt is determined that, n represents the factor quantity that data are included in formula；

3) threshold value

Gaussian parameter can obtain Gaussian Profile profile after establishing, and establish training sample point whether one of abnormal method be exactly with A threshold epsilon is set up based on cross validation collection, probability is considered abnormity point less than the point of threshold value；

4) obtain a result

The abnormal point numerical group detected is eliminated, the experiment sample data that can be predicted are obtained.

4. a kind of carbide chip according to claim 1 chemically-mechanicapolish polishes the Forecasting Methodology of surface roughness, its It is characterised by, in step 3, the forecast model of described genetic algorithm optimization BP neural network is substituted using genetic algorithm The gradient descent method of BP neural network carrys out Optimization Prediction error, comprises the following steps that：

1) data normalization

It regard random 70% data in pretreated experiment sample data as training data group, another 30% data conduct Test data set；To eliminate order of magnitude difference between each dimension data, training data group and test data set normalization are obtained The processing of training set and test set, wherein data normalization uses minimax method, i.e.,x_minFor data sequence In minimum value, x_maxFor the maximum in data sequence；

2) structure of BP artificial neural networks

A) input layer number n：According to forecast model, input layer number for influence carbide chip surface roughness because The number of element；

B) output layer nodes m：According to forecast model, output layer nodes are individual for the dependent variable of chemically mechanical polishing experiment Number；

C) node in hidden layer l：According to formulaN is input layer number in formula, and m is output node layer Number, a is the constant between 0~10；

D) input layer and implicit interlayer connection weight matrix w_ij[n × l], hidden layer and input interlayer connection weight matrix w_jk[l× m]；

E) hidden layer threshold value matrix a_j[l × 1], output layer threshold matrix b_k[m×1]；

F) general hidden layer excitation function f (x)：

G) hidden layer output H_j：

H) hidden layer output O_k：

I) error e_k：O is exported according to neural network forecast_kWith actual desired output y_kMathematic interpolation draws e_k=O_k-y_k；

3) structure of genetic algorithm

A) initialization of population：The initial population that random generation scale is N, N is traditionally arranged to be 20-50, each individual of population by Input layer and implicit interlayer connection weight matrix w_ij[n × l], hidden layer and input interlayer connection weight matrix w_jkIt is [l × m], hidden The a of threshold matrix containing layer_j[l × 1] and output layer threshold matrix b_k[m × 1] is constituted, and the length L of its gene order is calculated according to formula It is as follows：L=long w_ij+long a_j+long w_jk+long b_k；

B) fitness function F：By BP neural network forecast error es_kQuadratic sum as genetic algorithm individual fitness,N represents input layer number in formula；

C) selection operation：Selection strategy i.e. based on fitness ratio, the select probability of each individualP in formula_i Represent the select probability of i-th of individual, F_iThe fitness of i-th of individual is represented, N represents population at individual number；

D) crossover operation：I.e. chromosome intersects,A in formula_rtRepresent r-th of individual at t On gene, a_stGene of s-th of individual on t positions is represented, b represents crossover operator, and b typically takes 0.4-0.9；

E) mutation operation：Choose j-th of gene a of i-th of individual_ijEnter row variation, A in formula_maxRepresent gene a_ijThe upper bound, a_minRepresent gene a_ijLower bound, f (g)=1-g/G, g represents current evolutionary generation, G Represent to terminate evolutionary generation, r represents mutation operator, and r typically takes 0.001-0.1；

F) evolutionary generation G is terminated：G is traditionally arranged to be 100-500；

4) Prediction of Surface Roughness

Training the set pair analysis model is subjected to learning training, so that the Prediction of Surface Roughness mould of different condition bottom knife can be obtained Type；The network model after training is tested using test set, the precision of forecast model is drawn, and then will using precision satisfaction The model asked is predicted to the surface roughness for chemically-mechanicapolish polishing carbide chip.