A multi-level progressive parameter optimization method for complex process industry

Abstract

The application discloses a kind of multi-level progressive parameter optimization methods for complex process industry, comprising: the correlation between M process parameters and quality indicators is analyzed one by one to determine the importance of each process parameter to quality indicators;According to the importance of process parameters to quality indicators, descending order is sorted, and the process parameters after importance degree sorting are obtained;According to the process parameters after sorting, the process parameter layering method based on correlation analysis is constructed, the hierarchical division of process parameters is realized, and the multi-level process parameter combination is obtained by C times division;Modeling optimization is sequentially carried out from the Cth layer process parameter combination to the 1st layer process parameter combination in order, and the 1st layer process parameter combination optimal solution to the Cth layer process parameter combination optimal solution is obtained.The application effectively solves the problems of low modeling prediction accuracy and optimization difficulty caused by numerous process parameters in complex process manufacturing, which is of great significance for process manufacturing enterprises to ensure process quality and improve processing efficiency.

Description

Technical Field

[0001] The present invention relates to a multi-level progressive parameter optimization method for complex process industries, belonging to the technical field of complex process process parameter optimization. Background Art

[0002] Unlike discrete manufacturing, process manufacturing primarily achieves continuous production through the coordinated operation of functionally distinct process parameters. Therefore, the key to overall process optimization lies in the joint optimization of process parameters—that is, the ability to quickly control and optimize product quality by adjusting process parameters when quality fluctuations occur. Furthermore, complex process manufacturing is characterized by complex processes, severe mass-energy coupling between process parameters, and the strong temporal nature of process data. This makes it difficult to optimize process parameters through statistical analysis. Therefore, the core issue in achieving intelligent optimization decisions in process workshops is how to model and predict complex process flows, thereby rapidly and efficiently optimizing and regulating process parameters. Summary of the Invention

[0003] The present invention provides a multi-level progressive parameter optimization method for complex process industries, which is used to solve the optimization technology problem of numerous process parameters in process manufacturing.

[0004] The technical solution of the present invention is:

[0005] According to a first aspect of the present invention, a multi-level progressive parameter optimization method for complex process industries is provided, comprising the following steps:

[0006] S1. Preprocess the process data to obtain a preprocessed process data set. The preprocessed process data set consists of process parameters and quality indicators, which is expressed as D = {X mn ,Y n}, where X mn represents the mth process parameter of the nth sample, Y n Represents the quality index of the nth sample;

[0007] S2. Analyze the correlation between the M process parameters and the quality indicators one by one to determine the importance of each process parameter to the quality indicators;

[0008] S3. Sort the quality indicators in descending order of importance according to the process parameters to obtain the process parameters sorted by importance;

[0009] S4. Based on the sorted process parameters, a process parameter stratification method based on correlation analysis is constructed to achieve hierarchical division of the process parameters. Through C times of division, a multi-level process parameter combination is obtained;

[0010] S5. For the first layer to the Cth layer process parameter combination, sequentially perform modeling and optimization in the order from the Cth layer process parameter combination to the first layer process parameter combination to obtain the first layer process parameter combination optimal solution to the Cth layer process parameter combination optimal solution.

[0011] Further, the S4 comprises:

[0012] S4.1. Construct a hierarchical model.

[0013] S4.2. Take the process parameters sorted by importance as the to-be-layered process parameters; layer the to-be-layered process parameters according to the hierarchical model, take the qth layering demarcation point and the to-be-layered process parameters before the demarcation point as the qth layer process parameter combination after each layering, delete the qth layer process parameter combination from the to-be-layered process parameters to update the to-be-layered process parameters, until a termination condition is reached, and obtain the multi-layer process parameter combination.

[0014] Further, in the S4.2, the to-be-layered process parameters are layered according to the hierarchical model, specifically:

[0015] Select the process parameter with the largest correlation in the to-be-layered process parameters as the initial process parameter of the qth layering; construct the qth data set Qq1 according to the initial process parameter and the quality index;

[0016] Take the data set Qq1 as the input of the first training of the hierarchical model, and perform iterative training through the hierarchical model to obtain the goodness of fit; for the pth iteration training process, increase a process parameter on the basis of the data set Qq p-1 to construct the data set Qq p , take the data set Qq p as the input of the hierarchical model of the pth training to obtain the goodness of fit; and so on, until all the process parameters in the to-be-layered process parameters participate in iteration, and the goodness of fit of all the process parameters in the to-be-layered process parameters is obtained;

[0017] According to the goodness of fit of all the to-be-layered process parameters, calculate the stability index; according to the stability index, determine the demarcation point of the qth layering from all the to-be-layered process parameters; take the demarcation point of the qth layering and the process parameters before the demarcation point in the to-be-layered process parameters as the qth layer process parameter combination, delete the qth layer process parameter combination from the to-be-layered process parameters to update the to-be-layered process parameters for the q+1th layering.

[0018] Further, the termination condition is that the correlation corresponding to the process parameter ranked first in the to-be-layered process parameters is less than 0.5.

[0019] Further, the calculation of the stability index C i is specifically:

[0020] C i = S i · M i

[0021] wherein, S i represents the slope change rate of the process parameter with the ranking number i in the current process parameter to be layered, i = 1, 2, 3,..., I-3, I represents the total number of process parameters in the current process parameter to be layered; M i represents the mutation degree of the process parameter with the ranking number i in the current process parameter to be layered.

[0022] Further, the slope change rate is expressed as:

[0023]

[0024] wherein, E i represents the goodness of fit value obtained by training and verifying the process parameters with the ranking number 1 to the ranking number i in the current process parameter to be layered and the quality index construction data set.

[0025] Further, the mutation degree is expressed as:

[0026]

[0027] wherein, E i represents the goodness of fit value obtained by training and verifying the process parameters with the ranking number 1 to the ranking number i in the current process parameter to be layered and the quality index construction data set.

[0028] Further, the modeling and optimization specifically comprises: for the first optimization: establishing a nonlinear mapping model between the process parameter combination of the Cth layer and the quality index according to the random forest regression model, and calling the improved particle swarm algorithm for global optimization to obtain the optimal solution of the process parameter combination of the Cth layer; for the second optimization: establishing a nonlinear mapping model between the process parameter combination of the C-1th layer + the optimal solution of the process parameter combination of the Cth layer and the quality index according to the random forest regression model, and calling the improved particle swarm algorithm for global optimization to obtain the optimal solution of the process parameter combination of the C-1th layer; and in this way, until the optimal solution of the process parameter combination of the 1st layer is obtained, that is, the optimal solution of the process parameter combination of the 1st layer ~ the optimal solution of the process parameter combination of the Cth layer are obtained through C times of modeling and optimization.

[0029] According to the second aspect of the present application, a multi-level progressive parameter optimization system for complex process industry is provided, which comprises a module for executing the multi-level progressive parameter optimization method for complex process industry as described in any one of the above.

[0030] According to a third aspect of the present application, there is provided a terminal device comprising a memory, a processor, and a program stored in the memory and executable by the processor, wherein the processor implements the method for multi-level progressive parameter optimization for complex process industry according to any one of the above aspects when executing the program.

[0031] The present application has the following advantages: firstly, the present application divides the process parameters according to the importance of the relative quality indicators based on the hierarchical progressive concept; then, the optimal process parameter combination is obtained by modeling and optimization; during the modeling and optimization, the optimal parameter combination obtained by each optimization is input as a constant to the next modeling and optimization, thereby realizing the multi-level progressive optimization of the process parameters. The present application effectively solves the problems of low modeling and prediction accuracy and optimization difficulty caused by the large number of process parameters in complex process manufacturing, which is of great significance for process manufacturing enterprises to ensure process quality and improve processing efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0032] Figure 1 Fig. 1 is a schematic diagram of the process structure of the present application;

[0033] Figure 2 Fig. 2 is a schematic diagram of the process parameter hierarchical method based on correlation analysis of the present application;

[0034] Figure 3 Fig. 3 is an example diagram of multi-level progressive modeling and optimization of the present application;

[0035] Figure 4 Fig. 4 is a fitness function curve diagram of the IPSO optimization algorithm; wherein, Figure 4 (a) is a fitness function curve diagram of the first layer, Figure 4 (b) is a fitness function curve diagram of the second layer, Figure 4 (c) is a fitness function curve diagram of the third layer;

[0036] Figure 5 Fig. 5 is a fitness function curve diagram of the extraction of key parameters optimization;

[0037] Figure 6 Fig. 6 is a fitness function curve diagram of the overall optimization. DETAILED DESCRIPTION

[0038] In order to make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical scheme of the embodiments of the present application will be described clearly and completely below with reference to the drawings. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application. It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other in any way without conflict.

[0039] Example 1: Figure 1-6 As shown, according to a first aspect of an embodiment of the present invention, a multi-level progressive parameter optimization method for complex process industries is provided, comprising the following steps:

[0040] S1. Collect process industry data through sensors to construct a process technology data set; preprocess the process technology data to obtain a preprocessed process technology data set; the preprocessed process technology data set consists of process parameters and quality indicators, expressed as D = {X mn ,Y n}, where X mn represents the mth process parameter of the nth sample, Y n Represents the quality index of the nth sample; m = 1, 2, ...., M;

[0041] S2. Analyze the correlation between the M process parameters and the quality index one by one based on the Pearson correlation to determine the importance of each process parameter to the quality index. Given that process parameters have different effects on quality indexes, there are not only key process parameters with large effects but also redundant process parameters with small effects. Therefore, the present invention introduces a hierarchical approach to divide the process parameters into different levels.

[0042] S3. Sort the quality indicators in descending order of importance according to the process parameters to obtain the process parameters sorted by importance;

[0043] S4. Based on the sorted process parameters, a process parameter stratification method based on correlation analysis is constructed to achieve hierarchical division of the process parameters. Through C times of division, a multi-level process parameter combination is obtained;

[0044] S5. For the 1st-Cth layer process parameter combinations, modeling and optimization are performed in sequence from the Cth layer process parameter combination to the 1st layer process parameter combination to obtain the optimal solution of the 1st layer process parameter combination to the optimal solution of the Cth layer process parameter combination.

[0045] Furthermore, the S4 includes:

[0046] S4.1. Using a random forest regression model as a hierarchical model. The random forest regression model can effectively capture complex nonlinear relationships while having excellent stability and robustness. Therefore, the present invention uses the existing random forest regression model as a hierarchical model.

[0047] S4.2, the process parameters sorted by importance are taken as the to-be-layered process parameters; the to-be-layered process parameters are layered according to the layering model, after each layering, the demarcation point of the qth layering and the to-be-layered process parameters before the demarcation point are taken as the qth layer process parameter combination, the qth layer process parameter combination is deleted from the to-be-layered process parameters to update the to-be-layered process parameters, until a termination condition is reached, and a multi-layer process parameter combination is obtained.

[0048] Further, in the S4.2, the to-be-layered process parameters are layered according to the layering model, specifically:

[0049] The process parameter with the largest correlation in the to-be-layered process parameters is selected as the initial process parameter of the qth layering; the qth data set Qq1 is constructed according to the initial process parameter and the quality index, and the qth training set Qq1 and the qth validation set Qq1 are divided;

[0050] The data set Qq1 is taken as the input of the layering model, and iterative training is performed through the layering model to obtain a goodness of fit; for the pth iteration training process, a data set Qq p-1 is constructed by adding a process parameter to the data set Qq p , the data set Qq p is taken as the input of the layering model of the pth training to obtain the goodness of fit; and the process is repeated until all process parameters in the to-be-layered process parameters participate in iteration, and the goodness of fit of all process parameters in the to-be-layered process parameters is obtained.

[0051] The stability index is calculated according to the goodness of fit R 2 of all to-be-layered process parameters; the demarcation point of the qth layering is determined from all to-be-layered process parameters according to the stability index; the demarcation point of the qth layering and the to-be-layered process parameters before the demarcation point are taken as the qth layer process parameter combination, and the qth layer process parameter combination is deleted from the to-be-layered process parameters to update the to-be-layered process parameters for the q+1th layering.

[0052] Further, the termination condition is that the correlation corresponding to the process parameter ranked first in the to-be-layered process parameters is less than 0.5.

[0053] Further, the calculation of the stability index C i is specifically:

[0054] C i = S i · M i

[0055] Wherein, S iThe slope change rate of the process parameter with the ranking number i in the current process parameter to be layered, i = 1, 2, 3,...., I-3, I represents the total number of process parameters in the current process parameter to be layered; M i The mutation degree of the process parameter with the ranking number i in the current process parameter to be layered. It should be noted that the process parameters in the current process parameter to be layered are numbered in order of 1, 2, 3,.... I, and the correlation value of the process parameter with a higher ranking number is larger.

[0056] Further, the slope change rate is expressed as:

[0057]

[0058] Wherein, E i The goodness of fit value obtained by training and verifying the process parameters with the ranking number 1 to the ranking number i in the current process parameter to be layered and the quality index construction data set. E i+1 The goodness of fit value obtained by training and verifying the process parameters with the ranking number 1 to the ranking number i+1 in the current process parameter to be layered and the quality index construction data set, E i+2 , E i+3 The meaning is the same;

[0059] Further, the mutation degree is expressed as:

[0060]

[0061] Wherein, E i The goodness of fit value obtained by training and verifying the process parameters with the ranking number 1 to the ranking number i in the current process parameter to be layered and the quality index construction data set.

[0062] In the stability judgment, the slope change rate is used as the main index to evaluate the change amplitude of the model goodness of fit under adjacent parameter settings; at the same time, due to the correlation between some parameters in the model training process, the model accuracy may fluctuate, and this fluctuation can be understood as a mutation. Therefore, on the basis of introducing the slope change rate, the mutation rate is further introduced as an auxiliary index to measure the mutation degree under different parameter levels. By comprehensively considering the above two indexes, the stability index is formed, which quantitatively evaluates the overall stability of the model through the product of the slope change rate and the mutation rate. The smaller the value is, the more stable the model goodness of fit is, and thus the stability index can more effectively identify the stability of the trend.

[0063] Furthermore, the modeling optimization is specifically as follows: for the first optimization: a nonlinear mapping model is established between the C-th layer process parameter combination and the quality index based on the random forest regression model, and the improved particle swarm algorithm is called for global optimization to obtain the optimal solution for the C-th layer process parameter combination; for the second optimization: a nonlinear mapping model is established between the optimal solutions of the C-1th layer process parameter combination to the C-th layer process parameter combination and the quality index based on the random forest regression model, and the improved particle swarm algorithm is called for global optimization to obtain the optimal solution for the C-1th layer process parameter combination; and so on, until the optimal solution for the 1st layer process parameter combination is obtained, that is, after C times of modeling optimization, the optimal solution for the 1st layer process parameter combination to the C-th layer process parameter combination is obtained. From the above, it can be seen that the optimal solution obtained in each optimization is used as a constant input for the next modeling optimization.

[0064] Exemplarily, the present invention obtains the first-layer to third-layer process parameter combinations, establishes a nonlinear mapping model between the third-layer process parameter combination and the quality index based on the random forest regression model, and calls the improved particle swarm algorithm for global optimization to obtain the optimal solution of the third-layer process parameter combination; for the second optimization: establishes a nonlinear mapping model between the optimal solution of the second-layer process parameter combination + the third-layer process parameter combination and the quality index based on the random forest regression model, and calls the improved particle swarm algorithm for global optimization to obtain the optimal solution of the second-layer process parameter combination; for the third optimization: establishes a nonlinear mapping model between the optimal solution of the first-layer process parameter combination + the second-layer process parameter combination + the optimal solution of the third-layer process parameter combination and the quality index based on the random forest regression model, and calls the improved particle swarm algorithm for global optimization to obtain the optimal solution of the first-layer process parameter combination; that is, after three modeling optimizations, the optimal solution of the first-layer process parameter combination to the optimal solution of the third-layer process parameter combination are obtained.

[0065] The above-mentioned nonlinear mapping model between the optimal solution of the second-layer process parameter combination + the third-layer process parameter combination and the quality index based on the random forest regression model is specifically as follows: Based on the preprocessed process data set of S1, a new data set is constructed using the "optimal solution and quality index of the second-layer process parameter combination + the third-layer process parameter combination". The training set and test set are divided into a ratio of 7:3 for training and testing the random forest regression model, and the quality index prediction results of the test set are obtained to participate in the calculation of the fitness value in the optimization process. The same applies to other optimization processes.

[0066] The improvement of the improved particle swarm optimization algorithm lies in: adopting the strategy of linear decrease of inertia weight, adaptive change of acceleration factor and restriction of particle position.

[0067] By calling the improved particle swarm optimization algorithm to reversely search for the optimal solution of process parameter combination, the specific steps are as follows:

[0068] (1) Randomly initialize the particle swarm. In this method, the particle swarm is composed of h particles, each particle represents a process parameter combination solution, and the dimension of each particle is the number of process parameters. The initial position of the particle is randomly generated within the upper and lower boundary range of the search space, and the velocity vector is initialized to 0. In this embodiment, the size of the particle swarm is set to 100, and the search range of the process parameters is defined within the given upper and lower boundaries.

[0069] (2) Calculate the fitness of each particle. The fitness of each particle is measured by the root mean square error between the predicted result by the model and the true value. The particle swarm will minimize this fitness value during the optimization process, so that the particle swarm gradually approaches the global optimal solution.

[0070] (3) Update pbest and gbest. Each particle searches for the position with the highest fitness value as the individual optimal position, and the position of the particle with the highest fitness value in the particle swarm as the global optimal position.

[0071] (4) Update the inertia weight according to the inertia weight nonlinear decreasing strategy, and update the acceleration factor according to the adaptive nonlinear variation strategy of the acceleration factor;

[0072] (5) Update the velocity and position of the particle;

[0073] (6) Limit the position of the particle strategy: used to check whether there is a particle whose position exceeds the search range, and put the particle back into the search range if it exceeds the search range;

[0074] (7) Determine whether the maximum number of iterations is reached, if the maximum number of iterations is reached, record the optimal process parameter combination of this layer, if the maximum number of iterations is not reached, repeat steps (3) to (7).

[0075] (8) Determine whether the upper limit of the number of layers is reached, if the upper limit of the number of layers is reached, output the overall optimal process parameter combination, if the upper limit of the number of layers is not reached, increase the number of layers by one, and then train and test the model based on the optimization of the process parameter combination of the previous layer, and then repeat steps (3) to (8) until the upper limit of the number of layers is reached.

[0076] The inertia weight linear decreasing strategy expression is:

[0077]

[0078] In the formula: ω (t) is the inertia weight at the tth iteration, ω max is the maximum inertia weight, ω min is the minimum inertia weight, T max is the maximum number of iterations.

[0079] The acceleration factor adaptive change strategy expression is:

[0080]

[0081] In the formula: is the individual, global acceleration factor at the tth iteration, c 1max , c 2max is the individual, global maximum acceleration factor, c 1min , c 2min is the individual, global minimum acceleration factor, and Tmax is the maximum number of iterations.

[0082] The search strategy expression in the strategy of limiting particle position is:

[0083]

[0084] In the formula, represents the particle position, h min represents the lower limit of the particle search, h max represents the upper limit of the particle search.

[0085] Further, in the above, the inertia weight is reduced from 0.9 to 0.2, gradually reducing the dependence of the particles on the previous speed, enhancing the convergence ability, i.e. ω max takes 0.9, ω min takes 0.2; the individual learning coefficient: from 2 to 1, so that each particle gradually reduces the dependence on its own historical optimal position in the search process, i.e. c 1max takes 2, c 1min takes 1; the global learning coefficient: from 1 to 2, gradually enhancing the ability of the particles to follow the global optimal position, promoting the exploration of the global optimal solution by the group, i.e. c 2max takes 2, c 2min takes 1; the algorithm sets Tmax to 500.

[0086] According to a second aspect of an embodiment of the present invention, a multi-level progressive parameter optimization system for complex process industries is provided, specifically comprising: a first module for executing S1: preprocessing process data to obtain a preprocessed process data set; a second module for executing S2: analyzing the correlation between M process parameters and quality indicators one by one based on Pearson correlation to determine the importance of each process parameter to the quality indicator; a third module for executing S3: sorting the process parameters in descending order of importance to the quality indicator to obtain process parameters sorted by importance; a fourth module for executing S4: constructing a process parameter stratification method based on correlation analysis based on the sorted process parameters to implement hierarchical partitioning of the process parameters, thereby obtaining multi-level process parameter combinations through C-fold partitioning; and a fifth module for executing S5: modeling and optimizing the process parameter combinations from the first to the Cth layer in the order from the Cth layer to the first layer, to obtain the optimal solution for the process parameter combinations from the first layer to the Cth layer. As used above, the term "module" may be a combination of software and / or hardware that implements a predetermined function. Although the system described in the above embodiment is preferably implemented in software, implementation in hardware, or a combination of software and hardware, is also possible and conceivable. For parts not described in detail in each module, please refer to the relevant description of this embodiment.

[0087] According to the third aspect of an embodiment of the present invention, a terminal device is provided, comprising a memory, a processor, and a program stored on the memory and executable by the processor, wherein when the processor executes the program, a multi-level progressive parameter optimization method for complex process industries as described in any one of the above items is implemented.

[0088] Example 2: The following describes in detail an optional implementation of the present invention in conjunction with the accompanying drawings.

[0089] In response to the computational complexity problem caused by high-dimensional parameter optimization in process industries, the present invention proposes a multi-level progressive parameter optimization method for complex process industries. This method first divides the process parameters into layers according to the importance of relative quality indicators based on the concept of hierarchical progression. On this basis, an improved particle swarm algorithm is used to perform global optimization to obtain the optimal combination of process parameters. In the process of constructing the mapping model, the optimal parameter combination obtained by optimizing each layer is input as a constant to the next layer, thereby realizing the layer-by-layer progressive optimization of the process parameters. This method effectively solves the optimization difficulties caused by the low prediction accuracy of complex process manufacturing modeling and the large number of process parameters, which is of great significance for process manufacturing companies to ensure process quality and improve processing efficiency. The following is explained in combination with simulation data:

[0090] 1) The sample data set comes from a production line of a process manufacturing enterprise. The thin plate drying process data in October 2022 is exported from the MES system of the process production line. The data table contains "batch number", "collection time", and also includes 26 process parameters such as "exhaust air temperature", "exhaust air humidity", and "inlet material temperature" and a quality indicator - as shown in Table 1. As shown in Table 2, some process data are shown. The collection time interval of each process data is 1s; the process process data set is preprocessed (removing the head and tail operations) to obtain a preprocessed process process data set, totaling 239,281 items; based on the preprocessed process process data set, the steps S2-S5 described in Example 1 of the present invention are used to perform parameter optimization.

[0091] Table 1 Process parameter numbers and quality indicators

[0092]

[0093] Table 2 Process data

[0094]

[0095] 2) Based on the Pearson correlation, the correlation between the 26 process parameters and the quality indicators was analyzed one by one to determine the importance of each process parameter to the quality indicators.

[0096] 3) Sort the quality indicators in descending order according to the importance of the process parameters to obtain the process parameters sorted by importance. The specific sorting results are: X5, X 24 、X 21 、X 25 、X 26 、X6、X7、X8、X 22 、X 20 、X9、X 23 、X 19 , X1, X3, X2, X 18 、X 17 、X 16 、X 12 、X 14 、X 13 、X 11 、X 10 、X 15 , X4, serial numbers are 1-26, for example, the correlation value of process parameter X5 with serial number 1 is the largest, that is, the importance is the highest.

[0097] 4) Based on the sorted process parameters, a process parameter stratification method based on correlation analysis is constructed to achieve hierarchical division of process parameters, thereby obtaining a multi-level process parameter combination. Specifically:

[0098] Before stratification, there are 26 process parameters to be stratified, and then the first stratification is performed, that is, first construct the first data set Q11 with X5 and the quality index, and divide the first training set Q11 and the first validation set Q11 according to 7:3; use the first training set Q11 and the first validation set Q11 as the input of the stratification model, and perform iterative training through the stratification model, that is, perform 26 iterations; calculate the stability index based on the obtained goodness of fit; based on the stability index, determine the dividing point of the first stratification as X6 from all the process parameters to be stratified; use the dividing point of the first stratification in the process parameters to be stratified and the process parameters before the dividing point as the first layer process parameter combination, and then delete the first layer process parameter combination from the process parameters to be stratified to update the process parameters to be stratified for the second stratification, that is, the process parameters corresponding to serial numbers 7-26 are the updated process parameters to be stratified.

[0099] Before the second stratification, there are 20 process parameters to be stratified. Then the second stratification is carried out. First, X7 and the quality index are used to construct the second data set Q21, and the second training set Q21 and the second validation set Q21 are divided according to 7:3; the second training set Q21 and the second validation set Q21 are divided according to 7:3; the second training set Q21 and the second validation set Q21 are used as the input of the stratification model, and the stratification model is iteratively trained for 20 times; according to the obtained goodness of fit, the stability index is calculated; based on the stability index, the cutoff point of the second stratification is determined as X from all the process parameters to be stratified. 17 ; The dividing point of the second layer in the process parameters to be layered and the process parameters before the dividing point are used as the second layer process parameter combination, and then the second layer process parameter combination is deleted from the process parameters to be layered to update the process parameters to be layered for the third layer, that is, the process parameters corresponding to serial numbers 19-26 are used as the updated process parameters to be layered.

[0100] Before the third stratification, there are 8 process parameters to be stratified. The process parameter X ranked first among the process parameters to be stratified is 16 The corresponding correlation is less than 0.5, so the current process parameters to be layered are taken as the last layer, that is, the third layer.

[0101] According to the above, the process parameter combination of the first layer is: X5, X 24 、X 21 、X 25 、X 26 , X6; the second layer process parameter combination is: X7, X8, X 22 、X 20 、X9、X 23 、X 19 , X1, X3, X2, X 18 、X 17 ; Layer 3 process parameter combination: X16 、X 12 、X 14 、X 13 、X 11 、X 10 、X 15 , X4.

[0102] 5) For the 1st to 3rd layer process parameter combinations, modeling and optimization are carried out in the order from the 3rd layer process parameter combination to the 1st layer process parameter combination to obtain the optimal solution of the 1st layer process parameter combination to the 3rd layer process parameter combination. Among them, the 3rd layer process parameter combination X obtained in the first optimization is 16 、X 12 、X 14 、X 13 、X 11 、X 10 、X 15 The optimal solutions of X4 are 17.996, 113.285, 705.134, 3.806, 32.351, 11.000, 49.170, and 89.622, respectively. The fitness curve is shown in Figure 4 (c) As shown in the figure, the second layer process parameter combinations X7, X8, X 22 、X 20 、X9、X 23 、X 19 , X1, X3, X2, X 18 、X 17 The optimal solutions are: 1178.358, 31.849, 121.223, 0.133, -7.825, 97.062, 112.757, 7.340, 50.912, 1002.612, 47.393, 0.235, and the fitness curve is as follows Figure 4 (b) shows the first layer process parameter combination X5, X5 obtained by the third optimization. 24 、X 21 、X 25 、X 26 The optimal solutions of X6 are 17.218, 110.797, 112.587, 110.264, 113.299, and 5.963 respectively. The fitness curve is shown in Figure 4 (a) as shown; through Figure 4 It can be seen that in each optimization process, the fitness of the improved particle swarm optimization algorithm tends to be stable after multiple rounds of iterations and ranges from 0 to 0.0001.

[0103] Furthermore, the multi-level progressive optimization of the present invention is compared with the key parameter extraction optimization method and the overall prediction optimization method, as shown in Table 3.

[0104] The present invention is compared with different comparative cases. Case 1 is to extract key parameters, that is, the first layer process parameter combination extracted by the present invention is used as the key parameters; Case 2 is overall optimization, that is, the 26 process parameters without stratification are displayed; the implementation process of the comparative case is as follows: the process parameters and quality index data of Case 1 and Case 2 under the process data set preprocessed by the present invention are divided into training set and test set in a ratio of 7:3, and then the quality index prediction model is constructed layer by layer by adopting the RF regression model; then, the improved particle swarm optimization algorithm IPSO is used to reversely search for the optimal process parameter combination, and the fitness curves are as follows: Figure 5 、 Figure 6 As shown in Table 3, the fitness of Cases 1 and 2 tends to be stable after multiple rounds of iterations through the improved particle swarm optimization algorithm and ranges from 0 to 0.0001.

[0105] Table 3 Comparison results of key parameter optimization and multi-level progressive optimization

[0106]

[0107] In this embodiment, the model evaluation indicators used include mean absolute error (MAE), mean square error (MSE) and goodness of fit (R 2 ).

[0108]

[0109] Where: s is the number of samples, y 1,j Represents the true value of the quality indicator, y 2,j represents the predicted value of the quality indicator, Represents the mean.

[0110] As shown in Table 3, since the key parameter extraction optimization only analyzes the key parameters and ignores the influence of other parameters on quality indicators, the running time of the method proposed in this invention is slightly higher than that of the key parameter extraction optimization method. 2 The average efficiency of the multi-level progressive optimization is 0.851507, the average efficiency of the multi-level progressive optimization is 0.050028, and the average efficiency of the multi-level progressive optimization is 0.006945. It can be seen that the effect of extracting key parameter optimization is significantly lower than that of the multi-level progressive optimization of the present invention. In the case of the same number of optimizations, the number of iterations of the multi-level progressive optimization in the particle swarm optimization process is 154 times, and the running time is 213 seconds, which is 49% and 63% lower than the 301 times and 576 seconds of the overall optimization, respectively. Therefore, the effect of the present invention is significantly better than the overall optimization effect. In summary, the multi-level progressive optimization of the present invention not only improves the prediction accuracy and reduces the complexity of the optimization process relative to the overall optimization, but also makes up for the deficiency of considering only key parameters in the individual optimization.

[0111] The specific embodiments of the present application are described in detail above with reference to the accompanying drawings, but the present application is not limited to the above-described embodiments, and various changes can be made within the knowledge of those skilled in the art without departing from the spirit of the present application.

Claims

1. A multi-level progressive parameter optimization method for complex process industries, characterized by: The following steps are involved: S1. Preprocess the process technology data to obtain a preprocessed process technology data set; The pre-processed process data set consists of process parameters and quality indicators, which is expressed as D = {X mn ,Y n }, where X mn represents the mth process parameter of the nth sample, Y n Represents the quality index of the nth sample; S2. Analyze the correlation between the M process parameters and the quality indicators one by one to determine the importance of each process parameter to the quality indicators; S3. Sort the quality indicators in descending order of importance according to the process parameters to obtain the process parameters sorted by importance; S4. Based on the sorted process parameters, a process parameter stratification method based on correlation analysis is constructed to achieve hierarchical division of the process parameters. Through C times of division, a multi-level process parameter combination is obtained; S5. For the process parameter combinations of the first layer to the C layer, modeling and optimization are performed in order from the process parameter combination of the C layer to the process parameter combination of the first layer, and the optimal solution of the process parameter combination of the first layer to the optimal solution of the process parameter combination of the C layer is obtained; Said S4 comprises: S4.

1. Constructing a hierarchical model; S4.

2. The process parameters sorted by importance are used as the process parameters to be layered; the process parameters to be layered are layered according to the layering model. After each layering, the demarcation point of the qth layer and the process parameters to be layered before the demarcation point are used as the qth layer process parameter combination. The qth layer process parameter combination is deleted from the process parameters to be layered to update the process parameters to be layered, until the termination condition is met, thereby obtaining a multi-level process parameter combination; In S4.2, the process parameters to be layered are layered according to the layered model, specifically: The process parameter with the greatest correlation among the process parameters to be stratified is selected as the initial process parameter for the qth stratification; the qth data set Qq1 is constructed based on the initial process parameters and quality indicators; The dataset Qq1 is used as the input of the first training of the hierarchical model, and the hierarchical model is iteratively trained to obtain the goodness of fit; for the pth iterative training process, in the dataset Qq p-1 Add a process parameter to build data set Qq p , the dataset Qq p As the input of the hierarchical model trained for the pth time, the goodness of fit is obtained; and so on, until all the process parameters in the process parameters to be stratified are involved in the iteration, and the goodness of fit of all the process parameters in the process parameters to be stratified is obtained; According to the goodness of fit of all process parameters to be layered, a stability index is calculated; according to the stability index, a dividing point of the qth layer is determined from all process parameters to be layered; the dividing point of the qth layer in the process parameters to be layered and the process parameters before the dividing point are used as the qth layer process parameter combination, and the qth layer process parameter combination is deleted from the process parameters to be layered to update the process parameters to be layered for the q+1th layer.

2. The multi-level progressive parameter optimization method for complex process industries according to claim 1, characterized in that: The termination condition is: the correlation corresponding to the process parameter ranked first among the process parameters to be layered is less than 0.

5.

3. The multi-level progressive parameter optimization method for complex process industries according to claim 1, characterized in that: The stability index C i The calculation is as follows: C i =S i ·M i Among them, S i Indicates the slope change rate of the process parameter with the order number i in the process parameters to be layered, i=1,2,3,....,I-3, I represents the total number of process parameters in the process parameters to be layered; M i Indicates the degree of mutation of the process parameter with sort number i among the process parameters currently to be layered.

4. The multi-level progressive parameter optimization method for complex process industries according to claim 3, characterized in that: The slope change rate is expressed as: Among them, E i It represents the goodness of fit value obtained by training and verifying the data set constructed by the process parameters with ranking number 1 to ranking number i in the current process parameters to be stratified and the quality indicators.

5. The multi-level progressive parameter optimization method for complex process industries according to claim 3, characterized in that: The degree of mutation is expressed as: Among them, E i It represents the goodness of fit value obtained by training and verifying the data set constructed by the process parameters with ranking number 1 to ranking number i in the current process parameters to be stratified and the quality indicators.

6. The multi-level progressive parameter optimization method for complex process industries according to claim 1, characterized in that: The modeling and optimization are specifically as follows: for the first optimization: a nonlinear mapping model is established between the C-th layer process parameter combination and the quality index based on the random forest regression model, and the improved particle swarm algorithm is called for global optimization to obtain the optimal solution of the C-th layer process parameter combination; for the second optimization: a nonlinear mapping model is established between the optimal solution of the C-1-th layer process parameter combination + the C-th layer process parameter combination and the quality index based on the random forest regression model, and the improved particle swarm algorithm is called for global optimization to obtain the optimal solution of the C-1-th layer process parameter combination; and so on, until the optimal solution of the 1st layer process parameter combination is obtained, that is, after C times of modeling and optimization, the optimal solution of the 1st layer process parameter combination to the C-th layer process parameter combination are obtained.

7. A multi-level progressive parameter optimization system for complex process industries, characterized by: The method comprises a module for executing the multi-level progressive parameter optimization method for complex process industries according to any one of claims 1 to 6.

8. A terminal device, characterized in that: The method comprises a memory, a processor, and a program stored in the memory and executable by the processor. When the processor executes the program, the method implements the multi-level progressive parameter optimization method for complex process industries as described in any one of claims 1 to 6.

Images