CN111598435B - A Quality Trend Prediction Method Based on Adaptive Feature Selection and Improved Thinking Evolution Algorithm - Google Patents

Abstract

The invention discloses a quality trend prediction method based on self-adaptive feature selection and improved thinking evolution algorithm. The method mainly includes three modules: a feature self-adaptive processing module, a data fusion module, and a quality trend prediction module. The implementation of this method mainly includes the following steps: (1) Designing the corresponding parameters to generate the data for building the model; (2) Applying the error influence degree algorithm to establish the feature adaptive selection module; (3) Applying the KPCA data fusion method to establish the data fusion (4) Apply the improved thinking evolution algorithm to optimize the multi-layer perceptron (MLPNN) network to establish the quality trend prediction module. By establishing this method, the present invention can be implemented in the field of quality trend prediction, and can adaptively select different features according to different types of data for prediction, and apply data fusion and algorithm improvement to improve the accuracy of product quality trend prediction, and take appropriate methods in time Make corrections.

Description

A Quality Trend Prediction Based on Adaptive Feature Selection and Improved Thinking Evolution Algorithm test method

technical field

The invention belongs to the field of intelligent manufacturing and intelligent quality monitoring, in particular to a quality trend prediction method based on adaptive feature selection and improved thinking evolution algorithm.

Background technique

It is widely used in the fields of aerospace, aviation, ships, automobiles, etc. Since most of the products in these fields require relatively high quality data accuracy, the products will often be subject to human, machine, material, method, environment, and measurement during the production process. At the same time, quality data often have time-varying, non-linear, correlation and dynamic characteristics, which often lead to large quality problems in products. It is impossible to predict the trend in time and propose modification measures, which seriously affects the performance and quality of products. .

Control charts have been widely used in the production process as an auxiliary means of quality control and forecasting. The control chart controls the upper and lower limits of quality data to capture product quality fluctuations and abnormalities. With the development of technology and the acceleration of production rhythm, only the upper and lower limits of the control chart are used to judge quality fluctuations. Large quality problems often occur, which cannot adapt to modern processing. Data collection and other modern means of quality control. At present, many scholars have also begun to study the mode of the control chart to pursue the precise control of quality, but most of the research modes are less, do not study the mixed mode, or cannot adapt to the dynamic changes of the data, and cannot perform adaptive and precise control on the data. , Prediction, and most of them use offline recognition, and the degree of intelligence is not enough.

In order to realize product intelligent quality control and improve product quality, a quality trend prediction method based on adaptive feature selection and improved thinking evolution algorithm is proposed. The invention can realize intelligent quality prediction and control, timely propose measures to correct and improve product quality.

Contents of the invention

Purpose of the present invention: Aiming at the dynamic and time-varying characteristics of quality data, a quality trend prediction method based on adaptive feature selection and improved thinking evolution algorithm is proposed based on multiple modes of control charts to realize multiple mode recognition and according to quality The time-varying adaptive selection feature of the data predicts the quality trend of online products in the dynamic production process, and at the same time, the improved thinking evolution algorithm is applied to optimize the MLPNN network to improve the recognition accuracy of quality trends. Solve the existing problems of few intelligent identification modes of quality trends, low precision, inability to adapt to changes, and insufficient quality control, and improve the yield rate to achieve intelligent production and intelligent trend prediction.

The present invention proposes a quality trend prediction method based on adaptive feature selection and improved thinking evolution algorithm to solve the above problems, and builds a trend prediction model to predict abnormal states. To achieve the above object, the present invention adopts the following technical solutions:

A quality trend prediction method based on adaptive feature selection and improved thinking evolution algorithm includes the following steps:

Step 1: Generate the data required for model building.

The data generated in 9 modes are normal mode (NOR), cyclical mode (CYC), system mode (SYS), stratified mode (STR), uptrend mode (IT), downtrend mode (DT), upward step mode (US), step down mode (DS), mixed mode (MIX).

Step 2: The feature adaptive processing module is established.

In the process of quality trend prediction, the general method of directly applying data to predict the trend is not very accurate. The establishment of this step can select features reasonably and further improve the degree of intelligence. The module is built in two steps:

Step 1: Establish a feature extraction model. According to research, the extracted data features include the following statistical features and shape features are more convincing; among them, the statistical features of quality data include: MEAN, VS, STD, SKEW, KURT, A; quality The shape features of the data include: SL, NC1, NC2, APML, APLS, AASL, ACLPI, SRANGE, SB, PSMLSC, REAE, ABDPE.

The following features are extracted from the quality data in the observation window, and the meanings of each symbol are as follows:

MEAE: mean value of quality data;

VS: mean square value of quality data;

STD: standard deviation of quality data;

SKEW: Skewness coefficient of quality data;

KURT: kurtosis coefficient of quality data;

A: autocorrelation coefficient of quality data;

SL: the slope of the least squares regression line fitted to the quality data;

NC1: the number of intersections between the curve formed by the quality data and the line formed by the average value;

NC2: the number of intersections between the curve formed by the quality data and the least squares regression line;

APML: the area between the curve formed by the quality data and the average line;

APLS: the area between the curve formed by the quality data and the least squares regression line;

AASL: The curve formed by the quality data is divided into four areas, and the average value of the slope of the curve formed by the combination of the midpoints in each area

ACLPI: ratio of APML to standard deviation of quality data;

SRANGE: The curve formed by the quality data is divided into four areas, and the difference between the maximum value and the minimum value of the slope formed by the line connecting the midpoints of each area;

SB: Slope identifier for the least squares regression line of quality data;

PSMLSC: the average value of the intersection point of the quality data and the center line, and the intersection point of the least squares regression line;

REAE: The ratio of the MSE of the quality data to the error of the average MSE of the data divided into four regions;

ABDPE: When the curve formed by the quality data is divided into two regions, the absolute value of the difference between the slope of the least squares regression line of the curve formed by the overall quality data and the average slope of the least squares regression line of the two regions.

The second step: establish an adaptive feature selection model, which establishes the method of introducing errors to affect the calculation. If there is new data, the module can be applied to select appropriate features according to the data characteristics. Based on different types of quality data, different features are selected according to the importance of the features.

The way to build this model is as follows:

a) Assuming that there are N quality data features, the quality data features are preprocessed;

b) Establish an initial MLPNN neural network, and apply the processed quality data features for training;

c) If there is new data input, preprocess the new quality data features and increase the value of each quality data feature by 10% and decrease it by 10% to generate 2N sets of data, and bring each set of data into MLPNN for identification to obtain 2N Error value, and then average the error obtained by increasing 10% and decreasing 10% corresponding to each quality data feature to obtain N error values;

d) The first 85% of the features selected in order of error from large to small are defined as the features that affect the quality data to a higher degree.

e) Complete the adaptive quality data feature selection model.

Step 3: Data feature fusion module.

In order to predict trends more accurately, the original data is fused with adaptively selected data. If the original data is directly combined with the selected features to predict the quality trend, the input data is too large, which greatly increases the computational complexity of the model. In the data dimension reduction method, it is divided into linear data dimension reduction and nonlinear data dimension reduction. The module adopts the KPCA data dimension reduction method to fuse the original data and the characteristic data, and the method steps to establish the module are as follows:

a) Standardize and centralize the combined data;

b) Construct a kernel function for combining data, map the data into high dimensions and calculate the kernel matrix;

c) Calculated eigenvalues, selected eigenvectors;

d) Perform data dimension reduction and fusion;

Step 4: Establish quality trend prediction module.

Based on the data after adaptive feature fusion, a 3-layer perceptron MLPNN neural network model is established, and an improved thinking evolution algorithm is used to optimize the weights and thresholds of the MLPNN neural network.

The general thinking evolution algorithm is mainly through the learning method of iterative optimization. All individuals in the evolution process are called groups, and a group is divided into several subgroups. Subgroups include superior subgroups and temporary subgroups. In the evolution of thinking, the superior subgroup and the temporary subgroup are randomly generated subgroups centered on the optimal particle in the process of flight evolution without any restrictions, and the degree of information contained between the particles in the subgroup cannot be judged, resulting in Particles that have no meaning for evolution. The mutual information theory is introduced here to determine the degree of subgroup evolution. When the information contained in the particles generated in the subgroup and the central particle is greater than 85%, the particle is considered to be an invalid particle. At the same time, if the score of a particle is greater than the central particle, it will be retained, otherwise the particle will be regenerated.

The steps of establishing an improved thinking evolution algorithm are as follows:

a) Initialize population generation and generate populations in the space;

b) Select particles with higher scores in the initialization population as the center of the winning subgroup and the center of the temporary subgroup to generate subgroups.

c) Introduce an information judgment operator: Calculate the degree of mutual information between the particles of each winning subgroup and temporary subgroup and the central particle of its own subgroup respectively. If the mutual information degree between individual particles and the central particle is greater than 85%, it is considered to be Similar particles, at the same time, if the individual particle has a higher score than the central particle, it will be retained, otherwise it will be released;

d) Convergence operator: Calculate the individual particle scores in all subgroups, and select the winner as the center to regenerate subgroups. At the same time, perform the operation of step c, if the information contained in the individual particles and the central particle in the subgroup is appropriate, proceed to the next step, otherwise regenerate the particles and proceed to the operation of step c.

e) Judging whether each subgroup is mature, if the subgroup is mature, proceed to the next step, otherwise proceed to step d.

f) Alienation operator: exchange information between all mature winning subgroups and temporary subgroups, and the temporary subgroups with higher scores will replace the temporary subgroups with lower scores.

g) Judging that if there is no score in the mature temporary subgroup that exceeds the mature winning subgroup, then jump out of the loop, otherwise repeat step c-step g.

At the same time, in the above operation, when a new subgroup is generated with a new central particle in step d, the entropy change theory is introduced to increase the degree of chaos generated by the particles, and the entropy change inertia coefficient is increased. In the early stage of the search, it is required to have a large range and maintain a certain relationship with the information of the central particle, and in the later stage, it is required to converge faster, and the introduction of the inertia coefficient can improve the search ability for excellent particles.

By analyzing the abnormal situation in the production process, according to the characteristics of the quality data, a method for adaptively selecting appropriate features is developed, and the adaptively selected features are fused with the original data to reduce the dimension while increasing the recognition accuracy, and introduce information judgment operators and entropy increase The theoretically improved thinking evolution algorithm improves the accuracy and search ability of the algorithm. The beneficial effects of the present invention are: the ability to adaptively select appropriate features according to the characteristics of quality data to ensure the accuracy of recognition; the use of data fusion methods can enhance the performance of quality trend prediction , to ensure that the recognizer has good training efficiency; the improved algorithm can ensure good fault tolerance and stronger classification ability during classification.

Description of drawings

The present invention will be further described below in conjunction with accompanying drawing and embodiment

Fig. 1 is the flowchart of this method;

Figure 2 is the image of 9 modes; (a) normal mode; (b) mixed mode; (c) periodic mode; (e) systemic mode; (f) hierarchical mode; (g) rising trend; (h) decreasing trend; (i) upward step trend; (j) downward step pattern;

Figure 3 is a distribution diagram of the corresponding features of the nine modes;

Figure 4 is a schematic diagram of feature fusion

Figure 5 is a flowchart of the improved thinking evolution algorithm

Figure 6 is a schematic diagram of the particle simplification of the thinking evolution algorithm

Detailed ways

The present invention considers the dynamics and time-varying characteristics of quality data during production and introduces an adaptive feature selection method, and applies KPCA data fusion method to fuse dynamic data and adaptive features for accurate identification, and finally applies MLPNN network combined with improved thinking evolution algorithm , and established an accurate recognition model. The establishment of the present invention will be further described below in conjunction with the accompanying drawings and the implementation of the specific method establishment:

As shown in Figure 1, it is the overall flow chart of the implementation of the method, and the following is a specific expansion for each module.

Step 1: Generate the data required for model building.

Analyze patterns that will arise during production. Affected by actual production, fluctuations in product quality will appear in a certain order and affect the next production. Different parameters are designed to generate different types of trend graphs, as shown in Figure 2. There are 9 modes of graphics: normal mode, mixed mode, cycle mode, layered mode, system mode, up trend mode, down trend mode, and up step mode, the data of the next step mode, the specific method is as follows:

Simulation formula: y=μ+R(t)+d(t)

Among them: μ is the mean value of the quality data, R(t) is the normal distribution random deviation at time t, d(t) is the deviation caused by abnormal factors in the production process, the following are the formulas and parameter descriptions of each mode during production:

(1) Normal mode (NOR) of the production process:

y=μ+r(t)×σ+d(t)

Among them: μ is the mean value of the quality data, r(t) is the standard normal random distribution function, d(t) is the fluctuation caused by abnormal factors, in the normal mode d(t)=fish, the minimum value of σ is recommended to be 0.05 , the maximum is 0.5;

(2) Mixed mode (MIX) of the production process:

y=μ+r(t)×σ+(-1) ^W ×m×σ

Among them: d(t) is the fluctuation caused by abnormal factors. It is recommended that the minimum value of σ be 0.05 and the maximum value be 0.5; W is the disturbance factor 0 or 1 of the production process, and m is the disturbance amplitude of the production process. It is recommended to take [1.5, 2.5 ].

(3) Cycle mode (CYC) of the production process:

y=μ+r(t)×σ+a×sin(2πt/T)×σ

Among them: d(t) is the fluctuation caused by abnormal factors. In the periodic mode, d(t)=a×sin(2πt/T)×σ, the minimum value of σ is recommended to be 0.05, and the maximum value is 0.5; periodic production process disturbance is recommended The minimum value range of the amplitude a is [1,1.5], the maximum value range of a is [2.5, 3]; the minimum value range of the cycle T of the suggested production process disturbance is [4,8], and the maximum value range of T is [10,16].

(4) Hierarchical mode (STR) of the production process:

y=μ+r(t)×σ×K

Among them: d(t) is the fluctuation caused by abnormal factors, the minimum value of σ is recommended to be 0.05, and the maximum value is 0.5; the minimum value range of the disturbance proportional relationship K of the production process data is suggested to be [0.1,0.3], and the maximum value of K The range is [0.4,0.6].

(5) System mode (SYS) of the production process:

y=μ+r(t)×σ+d×(-1) ^t ×σ

Among them: d(t) is the fluctuation caused by abnormal factors. In the system mode, d(t)=d×(-1) ^t ×σ. It is recommended that the minimum value of σ be 0.05 and the maximum value be 0.5; the deviation degree of the recommended quality data is d The minimum value is 1; the maximum value of d is 3.

(6) The up and down trend mode (IT/DT) of the production process:

y=μ+r(t)×σ±g×t×σ

Among them: d(t) is the fluctuation caused by abnormal factors. In the trend mode, d(t)=±g×t×σ, the minimum value of σ is recommended to be 0.05, and the maximum value is 0.5; the value range of the slope g of quality data is recommended to be [0.05,0.1];

(7) Up and down step mode (US/DS) of the production process:

y=μ+r(t)×σ±b×s×σ

Among them: d(t) is the fluctuation caused by abnormal factors. In the step mode, d(t)=±b×s×σ. It is recommended that the minimum value of σ be 0.05 and the maximum value be 0.5; when t<P, b=0, When t>P, b=1; the recommended minimum value of the random step position P of quality data is [4,9], and the maximum value range of P is [13,19]; the minimum value range of the recommended step amplitude s is [0.5,1.5], the maximum value range of s is [2.5,3.5].

Step 2: Establish feature adaptive processing module

In the process of quality trend prediction, the general common method directly applies data to predict the trend. Research shows that reasonable use of quality data features will greatly improve the prediction accuracy. The quality data features are extracted, and the specific feature distribution is shown in Figure 3. The establishment of the feature adaptive processing module is divided into two steps:

Step 1: Establish a feature extraction model to extract statistical features of quality data including: MEAN, VS, STD, SKEW, KURT, A; shape features of quality data include: SL, NC1, NC2, APML, APLS, AASL, ACLPI, SRANGE, SB, PSMLSC, REAE, ABDPE. Let the quality data be Y=(y ₁ ,y ₂ ,y ₃ ....y _P )(i=1,2,3...P), feature extraction is as follows:

Mean formula for quality data:

The mean square value formula for quality data:

Standard Deviation of Quality Data

The coefficient of skewness formula for quality data:

The kurtosis formula for mass data:

Autocorrelation coefficient for quality data:

The slope of the least squares regression line of the curve formed by the quality data:

t _i is the distance from the detection quality data point to the origin at the i-th time, is the average value of the distance from the origin of the quality data points of P tests. y _i is the i-th quality data.

Number of intersections between the curve formed by the quality data and the curve formed by the average value: if (y _i -MEAN)(y _i+1 -MEAN)<0, then add 1 to NC1;

Number of intersections between the curve formed by the quality data and the least squares regression line: if (y _i -L(y _i ))(y _i+1 -L(y _i+1 ))<0, then NC2 plus 1;

The area between the curve formed by the quality data and the average line: APML;

The area formed by the curve formed by the quality data and the least squares regression line: APLS;

Divide the quality data into four regions, and connect the midpoints of the two regions to obtain the average value of the slope: Among them, S _jk is the slope of the quality data area composed of the jth midpoint and the kth midpoint. The coordinates of the midpoint of each area are />

The ratio of the area formed by the quality data and the center line to the standard deviation of the quality data:

The quality data is divided into four areas, the difference between the maximum value and the minimum value of the slope formed by the line connecting the points in each area: SRANGE=max(S _jk )-min(S _jk ); (j=123; k=234; J <k)

Slope identifier of the least squares regression line of the curve formed by the quality data: if the least squares slope of the curve formed by the quality data is greater than 0, then SB is 1, otherwise SB is 0;

The average value of the intersection point of the curve formed by the quality data and the center line, and the intersection point of the least squares regression line:

The ratio of the error of the MSE of the quality data to the mean of the MSE of the data divided into four regions:

When the quality data is divided into two regions, the absolute value of the difference between the slope of the overall least squares regression line and the average of the slopes of the least squares regression line of the two regions: Where B is the slope of the least squares regression line of the overall curve formed by the quality data, /> is the average of the slopes of the least squares regression line for the two regions.

The second step: establish an adaptive feature selection model, which establishes the method of introducing errors to affect the calculation. Different types of quality data imply different important features in the production process, and important features are selected based on different data. The specific method steps for the establishment of the model are as follows:

a) Preprocessing the quality data features;

Let the feature matrix F=(f _ab )(a=1,2,3...H; b=1,2,3...N), in which there are H samples and N types of features, and the quality data features are Normalization: where: /> is the mean value of the quality data characteristics of the b category, s _b is the standard deviation of the quality data characteristics of the b category />

b) Establish a 3-layer perceptron neural network, the hidden layer is nodes, m is the number of input nodes, n is the number of output nodes and is used to select training data in the above-mentioned quality data characteristics to carry out MLPNN neural network training;

c) If the new input quality data features [f ₁ , f ₂ , f ₃ ...f _N ], preprocess the quality data features to get F′=(f _b ′)(b=1,2,3.. .N), and increase the feature value of each quality data by 10% and decrease by 10% respectively to generate 2N sets of data [f ₁ ^* , f ₂ ^* , f ₃ ^* .... f _2N ^* ], where f ₁ ^* -f _N ^* is to add 10% data for each quality data feature value, f _N+1 ^* -f _2N ^* is to reduce each quality data feature value by 10% data, and bring two sets of quality data features into MLPNN respectively The identified 2N trend prediction errors are [E ₁ ^U , E ₂ ^U , E ₃ ^U .... E _N ^U ], [E ₁ ^D , E ₂ ^D , E ₃ ^D .... E _N ^D ] , and then each quality data feature is increased by 10% and decreased by 10% to obtain the average value of the trend prediction error E _i =(E _i ^U +E _i ^D )/2(i=123....N) to obtain N Trend prediction error value [E ₁ , E ₂ , E ₃ .... E _N ];

d) The [E ₁ , E ₂ , E ₃ .... E _N ] errors are sorted from large to small and the top 85% of the features are selected as the features that affect the quality data to a higher degree.

e) Complete adaptive quality data feature selection to obtain L (L<N) features, and complete the establishment of the model. Prepare for later feature fusion.

Step 3: Data feature fusion module

A data feature fusion module is established to fuse the original data with the adaptively selected data in order to predict trends more accurately. This module uses the KPCA data dimensionality reduction method to fuse the original data and feature data, as shown in Figure 4. The specific implementation steps are as follows:

a) Standardize and centralize the combined data;

Let the data matrix F=(f _ab )(a=1,2,3...H; b=1,2,3...(L+P)), there are H samples, and each sample has L +P data, standardize the combined data: where: /> The mean value of the b-th dimension data in the combined data, s _b is the standard deviation of the b-th dimension data after the combination of quality data and quality data features/> It is obtained that F ^a =[f _a1 ", f _a2 ", f _a3 "...f _a(L+P) "], F"=[F ¹ , F ² ....F ^H ].

b) Construct a combined data kernel function, map the processed data into high dimensions, and calculate the kernel matrix of the combined data;

According to the eigenvalue λ and eigenvector W of the combined data, the load factor of the combined data is selected and the fusion data is generated. Here, the nonlinear mapping Φ is used to convert the problem into Φ(F″)Φ(F″) ^T W=λW.

The vector matrix W=(w ₁ ,w ₂ ,w ₃ ....w _H ) of the combined data is obtained by introducing the mapping function of the combined data, w _a is the base vector of the combined data, and the relationship of the mapped data can be expressed as

Available

make Get α=[α ₁ ,α ₂ ....α _H ]

Finally, get Φ(F″)Φ(F″) ^T Φ(F″)α=λΦ(F″)α

Here we introduce the kernel skill equation and multiply Φ(F″) ^T on both sides to get:

c) Calculate the eigenvalues and select eigenvectors of the combined data;

Introduce the kernel function of combined data and perform kernel skill transformation K(F ^a1 ,F ^a2 )=Φ(F ^a1 )Φ(F ^a2 ) ^T where (a1∈(1,H),a2∈(1,H)); At this time, it can be defined that when the cumulative contribution rate of eigenvalues is greater than 85%, the first g principal components can represent the original data matrix. At this time, K is the matrix of H×H. The load factor matrix after selection is A=[α ₁ ,α ₂ .... _αg ].

d) Perform data dimensionality reduction;

Multiply the combined data after nuclear transformation with the selected load vector to obtain the fused data X ^* = K × A is:

Step 4: Build Quality Trend Prediction Module

The 3-layer perceptron MLPNN neural network model is established from the data after adaptive feature fusion, and the weights and thresholds of the MLPNN neural network are optimized using the improved thinking evolution algorithm. The training data uses the data fused in the third step.

The general thinking evolution algorithm is mainly through the learning method of iterative optimization. All individuals in the evolution process are called groups, and a group is divided into several subgroups. Subgroups include winner groups and temporary groups. During the evolution of thinking, the superior particle swarm and the temporary particle swarm are randomly generated subgroups centered on the optimal particle in the process of flight evolution without any restrictions, and the degree of information contained between the particles in the subgroup cannot be judged, so it is very easy It may produce a lot of particles which are meaningless to evolution, and are meaningless to evolution. The mutual information theory is introduced here to determine the degree of subgroup evolution. When the information contained in the particle generated in the subgroup and the leading particle is greater than 85%, the particle is considered to be an invalid particle. At the same time, if the score of a certain particle is greater than the leading particle, it will be kept; otherwise, the particle will be generated centering on the leading particle, as shown in Figure 5. Shown are the steps of the specific operation, and Fig. 6 is a schematic diagram of the particle evolution process.

The steps of establishing an improved thinking evolution algorithm are as follows:

a) Initialize population generation, generate a population with T particles in the space, and calculate the score of each particle according to the fitness function;

b) In the initialization population, select A particles with higher scores as the winning subgroup centers, and select B particles with higher scores as temporary subgroup centers. Set the size of each subgroup as T ^* , where T ^* =T/(A+B).

c) Introduce an information judgment operator: set the length of the particle as L′ to calculate the center of each particle of the winning subgroup and the corresponding Each particle center of the temporary subgroup and /> mutual information between and /> where (i=1,2,3....T ^* , l=1 2 3.....L′ t ₁ =1,2,3,...A, t ₂ =1,2,3 ,...B), if the degree of mutual information between the individual particles of the two subgroups and the central particle is greater than 85%, the particles are considered to be similar particles, and at the same time, a rule is formulated that if the particle score is greater than the central particle, it is retained, otherwise it is released;

Take the calculation of the mutual information degree of the winning subgroup as an example:

First, the formula for calculating the information entropy of the winning subgroup is as follows:

in represent individual particles /> The amount of information, where the base used for logarithms is usually 2, e or 10.

The degree of mutual information between the central particle and individual particles can be expressed as:

Similarly, the above formula is also applied to the temporary subgroup G for calculation.

d) Convergence operator: use the selected particles as the winning center and the temporary center to obey the normal distribution to generate A winning subgroups and B temporary subgroups, calculate the individual particle scores in all subgroups, and select the winner Regenerate subgroups as centers. At the same time, perform the operation of step c, if the information contained in the individual particles and the central particle in the subgroup is appropriate, proceed to the next step, otherwise regenerate the particles and proceed to the operation of step c.

e) Judging whether each subgroup is mature, if the subgroup is mature, proceed to the next step, otherwise proceed to step d.

f) Alienation operation operator: exchange information between all mature superior subgroups and temporary subgroups, superior subgroups M ¹ , M ² ... M ^A , temporary subgroups G ¹ , G ² , G ³ .. ^.GB competes, if there is a mature temporary subgroup whose score is greater than that of the mature winning subgroup, the release of the winning particle is replaced by the temporary subgroup and the temporary subgroup is replenished.

g) Judging that if the score in the mature temporary subgroup does not exceed the score of the mature winning subgroup, jump out of the loop, otherwise repeat step c-step g.

At the same time, in the above operation, generating a subgroup centered on the optimal particle in step d requires the generation of new individual particles. The new individual particle generation introduces the theory of entropy change to increase the degree of chaos of particle generation, and adds the entropy change inertia coefficient to the particle generation formula. In the early stage of the search, it is required to have a large range and maintain a certain relationship with the information of the central particle, and in the later stage, it is required to converge faster, and the introduction of the inertia coefficient can improve the search ability for excellent particles.

For the method of generating new individual particles, the following methods are given:

a) In order to expand the range of search, the initial particle matrix R _D×L' is generated according to the formula, in which there are D particles and the length of each particle is L'. The formula for particle generation is: R _d =m _c +×θ×(2×Z _1×L′ -1), where θ is 0.5-1, Z _1×L′ is the normal distribution of 1×L′ 0- 1 numerical matrix, m _c is the central particle.

b) Calculate the information entropy of the particles on the _RD×L' matrix and perform the chaotic iteration of the particles. Set the number of iterations to K _max times, k is the value of the number of iterations at this time, and finally select the best particle. Convert the particle matrix into a matrix of S _1×(D×L′) , and the calculation formula of the information entropy of the whole particle is: Apply the formula to get the changed weights:

c) Change the formula to R _d+1 = wR _d + (1-w) × θ × (2 × Z _{1 × L'} -1), get the new R _{D × L'} after looping K _max times, and at the same time in the loop During the process, the one with the highest score for each particle will be kept, and the particle with the highest score in the matrix will be finally selected as the new particle after the loop ends.

This process is a process of chaotic search within this range centered on a certain particle, preventing particle search from overfitting search, starting to expand the range and finally gradually approaching the optimal particle, and improving the search accuracy around the central particle.

Claims (5)

1. A quality prediction and control method is characterized in that: modeling of the method comprises the following four steps:

step 1: generating product quality model building required data, and simulating quality data of the product; the modes of the quality data of the simulation product comprise a normal mode, a periodic mode, a mixed mode, a system mode, an ascending trend mode, a descending trend mode, an upward step mode and a downward step mode;

step 2: and (3) establishing a characteristic self-adaptive processing module, preprocessing according to the quality data of the product in the step (1), wherein the characteristic self-adaptive processing module is established by two steps: extracting product quality data characteristics and establishing a self-adaptive characteristic selection model by applying an initialization MLPNN network;

step 3: establishing a data feature fusion module, and realizing feature fusion and data dimension reduction by applying a KPCA method, thereby simplifying a subsequent product quality trend prediction module;

step 4: establishing a product quality trend prediction module, optimizing an MLPNN neural network by using an improved thinking evolution algorithm, wherein the optimization target of the model is prediction accuracy, and introducing an entropy change theory by adding a mutual information judgment operator to make the algorithm obtain a prediction model;

the method for establishing the improved thinking evolutionary algorithm comprises the following steps:

a) Initializing population generation, and generating a population in a space;

b) Selecting particles with higher scores in the initialized population as a winning subgroup center and a temporary subgroup center respectively and generating subgroups;

c) Introducing an information judgment operator: calculating mutual information degrees between particles of each winning subgroup, temporary subgroup and central particles of the own subgroup respectively, if the mutual information degree of individual particles and the central particles is more than 85%, the particles are considered to be similar particles, and meanwhile, if the individual particles are higher than the central particles, the particles are reserved, otherwise, the particles are released;

d) Convergence operator: calculating individual particle scores in all subgroups, and selecting winners as centers to regenerate the subgroups; c, if the information degree of the individual particles and the central particles in the subgroup is proper, continuing the next operation, otherwise, regenerating the particles and performing the operation of the step c;

e) Judging whether each subgroup is mature or not, if so, continuing the next operation, otherwise, continuing the operation of the step d;

f) Dissimilating operation operators: carrying out information communication on all mature winning subgroups and temporary subgroups, wherein the temporary subgroup with higher score replaces the temporary subgroup with lower score;

g) And c, judging that if no score exceeds the mature winning subgroup in the mature temporary subgroup, jumping out of the circulation, otherwise, repeating the operations of the steps c-g.

2. The quality prediction and control method of claim 1, wherein: the data of the 9 modes generated in the step 1 are respectively: normal mode NOR, periodic mode CYC, system mode SYS, hierarchical mode STR, upward trend mode IT, downward trend mode DT, upward step mode US, downward step mode DS, mixed mode MIX.

3. The quality prediction and control method of claim 1, wherein: the statistical features of the quality data adopted in the first step in the step 2 include: MEAN, VS, STD, SKEW, KURT, A; the shape characteristics of the quality data include: SL, NC1, NC2, APML, APLS, AASL, ACLPI, SRANGE, SB, PSMLSC, REAE, ABDPE; and secondly, adopting an initialized MLPNN network to apply error influence degree algorithm to adaptively select the characteristics.

4. The quality prediction and control method of claim 1, wherein: and 3, fusing the self-adaptively selected characteristics with the original product quality data by using a KPCA method.

5. The quality prediction and control method of claim 1, wherein: and 4, designing a thinking evolutionary algorithm improved by the mutual information judgment operator, and improving a particle generation mode of the thinking evolutionary algorithm by applying an entropy change theory.