Polyimide high-pressure mixer optimization method

Through the combined method of the LSTM-Transformer proxy model and whale optimization algorithm, the structural parameters of the polyimide high-pressure mixer are optimized, which solves the problem of poor mixing effect in the mixer design, and achieves more efficient optimization and better mixing Effect.

CN119066968APending Publication Date: 2024-12-03SICHUAN UNIVERSITY OF SCIENCE AND ENGINEERING
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411181882.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-27
Publication Date
2024-12-03

AI Technical Summary

Technical Problem

The optimization design of polyimide high-pressure mixer is difficult to achieve. The existing methods require repeated adjustment of design parameters and the relationship between structural parameters and mixing effect cannot be accurately described, resulting in poor mixing effect.

Method used

The combined method of the LSTM-Transformer proxy model and whale optimization algorithm is adopted to optimize the structural parameters of the high-voltage mixer to improve the mixing effect by constructing geometric models, meshing, boundary condition setting, structural parameter optimization, and global extreme value optimization.

Benefits of technology

It realizes rapid iterative optimization, obtains higher optimization accuracy and better mixing effects, significantly improves the mixing effect of the mixer, reduces the mixing unevenness by 70%, shortens the optimization time and improves the accuracy.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119066968A_ABST
    Figure CN119066968A_ABST
Patent Text Reader

Abstract

The invention discloses a polyimide high-pressure mixer optimization method which comprises the following steps: constructing a geometric model of a high-pressure mixer, performing grid division on a geometric model mixing cavity, and determining the number of grids; setting boundary conditions, and analyzing the mixing effect of a mixing head in the high-pressure mixer based on the geometric model; and if the mixing effect is not qualified, optimizing the structural parameters of the mixing head in the high-pressure mixer, and realizing rapid iterative optimization of the prediction model by utilizing the combination of the LSTM-Transform and the whale optimization algorithm to finally obtain the optimal structural parameters. Compared with the traditional optimization direction, the optimization process of the joint algorithm has higher optimization efficiency, and higher optimization precision and better optimization effect are obtained through shorter optimization time.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to the technical field of high-pressure mixers, and in particular to a polyimide high-pressure mixer optimization method. Background Art

[0002] High-pressure mixers play a vital role in the mixing of polyimide materials. The two raw materials A and B of the high-pressure mixer flow into the mixing chamber from two feed ports at high speed and high pressure through metering pumps, and high-speed impact mixing is achieved in the mixing chamber. The structural parameters of the mixing head have an important influence on the mixing effect of the mixer. The current optimization design method for structural parameters requires repeated adjustment and selection of design parameters, which is labor-intensive and cannot accurately describe the relationship between structural parameters and mixing effects, making the optimization design of polyimide high-pressure mixers difficult to achieve. Summary of the invention

[0003] The object of the present invention is to provide a method for optimizing a polyimide high-pressure mixer to at least solve the above technical problems.

[0004] To achieve the above purpose, the technical solution adopted by the present invention is as follows:

[0005] A polyimide high-pressure mixer optimization method comprises the following steps:

[0006] S1. Construct a geometric model of a high-pressure mixer, mesh the mixing cavity of the geometric model and determine the number of meshes;

[0007] S2, setting boundary conditions and analyzing the mixing effect of the mixing head in the high-pressure mixer based on the geometric model;

[0008] S3. If the mixing effect is unsatisfactory, optimize the structural parameters of the mixing head in the high-pressure mixer. The optimization includes the following steps: S31. Select the structural parameters that affect the mixing effect, sort the structural parameters by importance, and screen out the structural parameters with low importance; S32. Construct the sample set required for training; S33. Construct the LSTM-transformer proxy model, and train the LSTM-transformer proxy model based on the sample set; S34. Use the whale optimization algorithm to perform global extreme value optimization on the trained LSTM-transformer proxy model to obtain the optimal structural parameters.

[0009] Furthermore, in S1, multiple groups of grid numbers are designed, and a line graph is drawn with the grid number as the horizontal axis and the pressure drop under the corresponding grid number as the vertical axis, and the first grid number where the pressure drop does not change is determined as the grid number of the mixing chamber.

[0010] Furthermore, S2 includes: S21, setting the material inlet of the high-pressure mixer to adopt the velocity inlet boundary condition, the material outlet to adopt the pressure outlet boundary condition, the inner wall surface is a stationary wall surface, and the top is a free boundary; S22, qualitatively and quantitatively analyzing the mixing effect of the mixing head in the high-pressure mixer based on the geometric model.

[0011] Furthermore, the qualitative analysis and evaluation of the mixing effect of the mixing head includes: performing flow field analysis on the mixing head based on a geometric model, and drawing a density cloud map of the outlet cross section of the mixing head. The more color bands of the density cloud map and the more uneven the color distribution, the worse the mixing effect, and vice versa.

[0012] The quantitative analysis and evaluation of the mixing effect of the mixing head includes: performing simulation analysis on the mixing head based on a geometric model to obtain the mixing unevenness of the geometric model. The smaller the mixing unevenness, the worse the mixing effect, and vice versa.

[0013] Furthermore, the S31 includes: S311, selecting four structural parameters, namely, inlet diameter D1, outlet diameter D2, counterpressure P, and counterangle θ; S312, conducting a full factor test on the four structural parameters to obtain a sample space containing multiple sample points A, each sample point A being a set of structural parameters of the mixer, and constructing a mixer model with different structural parameters; S313, performing simulation analysis on the mixer model to obtain the mixing unevenness of the mixer model; S314, performing factor analysis on the mixer model, taking the mixing unevenness as the ordinate and the different levels of the parameters as the abscissa, and constructing a Pareto diagram of the standardized effect; S315, based on the Pareto diagram, obtaining the importance of each structural parameter to the mixing unevenness, and screening out the counterpressure P with the lowest importance.

[0014] Furthermore, the S32 includes: S321, using optimal Latin hypercube sampling to extract multiple sample points B from the sample space, each sample point B is a set of screened structural parameters of the mixer, and constructing CFD models with different structural parameters; S322, performing simulation analysis on the CFD model to obtain the mixing unevenness of the CFD model, recording the CFD model and the corresponding mixing unevenness as a training sample, and obtaining a sample set containing multiple training samples.

[0015] Furthermore, the S33 includes: S331, constructing an LSTM-transformer proxy model, the LSTM-transformer proxy model includes three modules and a deep residual forward network, the three modules respectively perform horizontal feature extraction, vertical feature extraction and time feature extraction, and the deep residual forward network then uses the extracted features for fusion and prediction. Preferably, the horizontal features are extracted by the first convolutional self-attention module, the vertical features are extracted by the vertical self-attention module, and the time features are extracted by the LSTM network; S332, using 5 / 6 of the data in the sample set as training data and the remaining data as test data to train the LSTM-transformer proxy model. During the training process, the calculation error of the model is obtained through the loss function, and then the optimization algorithm is used to find the minimum error value, update the weight of the model, and finally make the model converge.

[0016] Furthermore, in the S331, the LSTM-transformer proxy model is a multi-scale convolutional LSTM-Transformer residual network model, and the multi-scale convolutional LSTM-Transformer residual network model includes 1 input layer, 2 LSTM modules, 2 Conv MHA modules, 2 MHA modules, 2 feature fusion modules, 5 Resinear modules, 5 Linear layers, 1 deep residual forward network and 1 output layer; before the horizontal features, vertical features and time features are fused, the multi-scale convolutional LSTM-Transformer residual network model first fuses the horizontal features and vertical features, regards different horizontal features and vertical features as different channels, and then uses the convolution operation to reduce the number of channels to 1, and obtains the first output of the same dimension as the input data, and the first output is a multi-head attention feature; then the multi-head attention feature and the time feature are fused. Combine, regard different multi-head attention features and time features as different channels, and then use convolution operation to reduce the number of channels to 1, and obtain a second output of the same dimension as the input data, wherein the second output is a fusion feature; finally, the obtained fusion feature is vectorized as the input of the deep residual forward network; preferably, in the multi-scale convolutional LSTM-Transformer residual network model, the number of neurons in the hidden layer of the LSTM module is set to 32, the initial learning rate is set to 0.001, the batchsize is set to 32, the time step is set to 6, and the activation function of the output layer is a linear function.

[0017] Furthermore, the S34 includes: S341, taking the screened structural parameters as optimization variables, taking the minimum mixing unevenness of the mixing head as the optimization target, and establishing an optimization objective function between the optimization variables and the optimization targets; S342, initializing the parameters and fitness of the whale optimization algorithm; S343, establishing a mathematical model of the whale optimization algorithm based on the constraints of the optimization objective function and the optimization variables; S344, using the mathematical model to update the fitness until the optimal individual is found to obtain the optimal structural parameters.

[0018] Compared with the prior art, the present invention has the following beneficial effects:

[0019] The present invention utilizes the combination of LSTM-Transformer and whale optimization algorithm to realize rapid iterative optimization of the prediction model and finally obtains the optimal structural parameters. Compared with the traditional optimization direction, the optimization process of the joint algorithm has higher optimization efficiency and obtains higher optimization accuracy and better optimization effect through shorter optimization time. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 The figure is a flow chart of the method of the present invention. DETAILED DESCRIPTION

[0021] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0022] like Figure 1 As shown, the present invention provides a polyimide high-pressure mixer optimization method, comprising the following steps:

[0023] S1. Construct a geometric model of a high-pressure mixer, mesh the mixing cavity of the geometric model and determine the number of meshes;

[0024] S2, setting boundary conditions and analyzing the mixing effect of the mixing head in the high-pressure mixer based on the geometric model;

[0025] S3. If the mixing effect is unsatisfactory, optimize the structural parameters of the mixing head in the high-pressure mixer. The optimization includes the following steps: S31. Select the structural parameters that affect the mixing effect, sort the structural parameters by importance, and screen out the structural parameters with low importance; S32. Construct the sample set required for training; S33. Construct the LSTM-transformer proxy model, and train the LSTM-transformer proxy model based on the sample set; S34. Use the whale optimization algorithm to perform global extreme value optimization on the trained LSTM-transformer proxy model to obtain the optimal structural parameters.

[0026] The present invention utilizes the combination of LSTM-Transformer and whale optimization algorithm to realize rapid iterative optimization of the prediction model and finally obtains the optimal structural parameters. Compared with the traditional optimization direction, the optimization process of the joint algorithm has higher optimization efficiency and obtains higher optimization accuracy and better optimization effect through shorter optimization time.

[0027] Example

[0028] A simulation study was conducted on a high-pressure mixer on a company's polyimide foam production line.

[0029] S1. The geometric model of the high-pressure mixer was drawn using the 3D drawing software SolidWorks. In order to explore the influence of the structural parameters of the mixing head on the mixer and optimize its structure, the fluid analysis of the high-speed collision mixing process of liquids A and B in the mixing chamber was performed. Liquids A and B were mixed in a 1:1 ratio. The feed port diameters of materials A and B were 8 mm, the discharge port diameters were 35 mm, and the collision angle was 180°. The above are the original structural parameters of the mixing head of the high-pressure mixer.

[0030] The mesh in the fluent software is used to mesh the mixing chamber. Since the fluid collision zone in the mixing chamber is relatively complex and irregular in shape, the number of grids will affect the accuracy of the calculation. Too many grids will increase the burden on the computer and make the calculation time process. A low number will affect the accuracy, so it is necessary to select an appropriate number of grids for calculation. Five groups of grid numbers are used for comparison, namely: 58574, 63851, 86824, 146142, 221348, and the corresponding line graphs are obtained. From the line graph analysis, it can be seen that when the number of grids reaches 146142, the pressure drop hardly changes, so the number of grids selected for calculation is 146142.

[0031] S2. Set the material inlet of the high-pressure mixer to adopt the velocity inlet boundary condition, the material outlet to adopt the pressure outlet boundary condition, the inner wall surface is the stationary wall surface, and the top is the free boundary; based on the geometric model, qualitatively and quantitatively analyze the mixing effect of the mixing head in the high-pressure mixer.

[0032] The density of material A is 1.17 g / cm 3 , viscosity is 0.16pa·s; density of material B is 1.25g / cm 3 , viscosity is 0.36pa·s; operating pressure is 10MPa. In order to simplify the calculation, the following assumptions are made: ① The components do not react in the mixer; ② The effects of temperature and pressure on viscosity are not considered; ③ The flow field in the mixer is a three-dimensional steady, constant temperature, stable turbulent flow of incompressible viscous liquid.

[0033] The standard k-ε model is used to solve the mixer, the SIMPLE algorithm is used for pressure-velocity coupling, the PRESTO is used for pressure discrete format, and the other discrete formats are all second-order upwind. The sub-relaxation iteration factor uses the default value, and the variable residuals are all set to 10-6. At the same time, the vof multiphase flow model is added.

[0034] After setting the boundary conditions, the Flunet software was used to solve the problem and qualitatively analyze and evaluate the mixing effect of the mixing head: the flow field of the mixing head was analyzed based on the geometric model, and the density cloud map of the mixing head outlet section was drawn. From the density cloud map analysis, it can be seen that the density of the material on the middle two sides and the lower middle side is larger, and the density on the left side and the lower right side is smaller. Since the mixture is 1:1, if the two are fully mixed, there should be fewer color bands and the density should be between 1.20 and 1.21 g / cm 3 Since the high-density area and the low-density area account for a large proportion, the surface mixing effect is poor and needs further optimization.

[0035] In order to quantify the mixing effect, the mixing effect of the mixing head is evaluated by quantitative analysis, including: performing simulation analysis on the mixing head based on the geometric model to obtain the mixing unevenness of the geometric model. Preferably, the mixing unevenness is obtained by the following operation: in the geometric model, the material outlet section of the mixing head is along the y-axis direction, 50 sampling points are uniformly taken using matlab and the mixing unevenness is calculated. The calculation formula of the mixing unevenness S is:

[0036]

[0037] In the formula, ρ i represents the liquid density at the sampling point, n represents the number of sampling points, It represents the average liquid density of all sampling points.

[0038] The smaller the mixing unevenness S, the smaller the difference in liquid density on the cross section, that is, the smaller the density gradient, the better the mixing effect. It can be calculated that the mixing unevenness S of the original mixing head is 0.11, indicating that the mixing effect of the original mixing head is poor and the mixing effect is unqualified. In order to improve the mixing effect, the original mixing head structure needs to be optimized.

[0039] S3. Optimizing the structural parameters of the mixing head in the high-pressure mixer, the optimization comprising the following steps:

[0040] S31, selecting structural parameters that affect the mixing effect, sorting the structural parameters by importance, and screening out structural parameters with low importance;

[0041] S32, constructing a sample set required for training;

[0042] S33, build an LSTM-transformer proxy model, and train the LSTM-transformer proxy model based on the sample set;

[0043] S34. Use the whale optimization algorithm to perform global extreme value optimization on the trained LSTM-transformer proxy model to obtain the optimal structural parameters.

[0044] Full factorial experimental design is a classic experimental design method with a strong global nature. Full factorial experimental design usually includes two or more factors, each with multiple levels. This design method can cover all combinations, so it can more comprehensively understand how various factors interact with each other.

[0045] In S31, this embodiment uses the mixing unevenness as the test influencing index, selects the mixer inlet diameter D1, outlet diameter D2, impact angle θ, and impact pressure P as the structural parameters affecting the mixing effect, and each factor takes two levels according to the requirements of the experimental design. The two-level values ​​of these four factors are shown in Table 1, the values ​​of the mixer inlet diameter D1 are 6mm and 16mm, the values ​​of the outlet diameter D2 are 25mm and 45mm, the values ​​of the impact pressure P are 8Mpa and 20Mpa, and the impact angle θ is 60° and 180°.

[0046] Table 1 Factor level table

[0047]

[0048] A two-level full factorial experimental design is performed on the four factors, and a sample space containing 19 sample points can be obtained. For the experimental design of the mixer structural parameters, 19 mixer models with slightly different structural parameters are obtained. These mixer models are simulated and analyzed respectively, and the mixing unevenness of each mixer model is obtained as shown in Table 2. The calculation of mixing unevenness is the same as in S1, so it is not repeated here.

[0049] Table 2 Mixing unevenness of the mixer under different structural parameters

[0050]

[0051]

[0052] In order to more intuitively analyze the influence of different structural parameter values ​​on mixing unevenness, after simulating the flow field of the mixing head, the simulation results were used for factor analysis. The results of the numerical simulation of the mixing head (mixing unevenness) were used as the ordinate, and the corresponding levels of the four structural parameters were used as the abscissa. MINITAB was used for processing to make a Pareto chart of the standardized effect of the response variable. From the analysis of the Pareto chart, it can be seen that the inlet diameter, outlet diameter, and impact angle have a greater influence on the mixing effect, while the impact pressure has a smaller influence on the mixing effect. For this reason, the order of influence of each structural parameter on the mixing unevenness is as follows: mixer outlet diameter>impact angle>inlet diameter>impact pressure, and it can be seen from the figure that the impact of the impact pressure on the mixing effect is very small, so it can be ignored and screened out.

[0053] The above optimization of the structural parameters of the polyimide mixer was carried out with the minimum mixing unevenness as the goal. Four structural parameters of the mixer were selected for significance analysis, and the influence of each parameter was finally determined, thereby excluding insignificant parameters and reducing the amount of calculation.

[0054] Before building the model, it is necessary to obtain multiple groups of samples as training data for the approximate model, and the data samples are required to be distributed as evenly as possible in the design space. The present invention selects 3 variables. If the method of artificially distributing samples is adopted, it is likely that the sample distribution is uneven, resulting in the selected training samples being unable to fully characterize the sample space. Uniform sampling is a test method that only considers the sample distribution without considering whether the samples are neatly comparable, while the Latin hypercube sampling method is to divide the sample space into individual spaces at equal distances during sampling, and randomly select a single sample from each space to form a sample total. This sampling method has the advantages of high efficiency and stable sampling, but the extracted samples are not evenly distributed in the design space. In order to enable the extracted samples to better characterize the sampling space, improvements have been made on the original sampling method, and the optimal Latin hypercube sampling is proposed. The distances between the sampling points extracted by this method are more consistent, and the sampling space has better filling performance.

[0055] The training sample of the neural network is at least 10 times the input variable. In order to make the result more credible, the present invention uses the optimal Latin hypercube sampling to extract 60 sample points from the sample space. According to the sample value, 60 groups of CFD models are established in solidworks. The CFD model is simulated and analyzed to obtain the mixed unevenness of the CFD model. The CFD model and the corresponding mixed unevenness are recorded as a training sample to obtain a sample set containing multiple training samples. The sample set is shown in Table 3.

[0056] Table 3 Sample set

[0057]

[0058]

[0059] The S33 includes: S331, constructing an LSTM-transformer proxy model, the LSTM-transformer proxy model includes three modules and a deep residual forward network, the three modules respectively perform horizontal feature extraction, vertical feature extraction and time feature extraction, and the deep residual forward network then uses the extracted features for fusion and prediction, preferably, the horizontal features are extracted by the first convolution self-attention module, the vertical features are extracted by the vertical self-attention module, and the time features are extracted by the LSTM network; S332, using 5 / 6 of the data in the sample set as training data and the remaining data as test data to train the LSTM-transformer proxy model, obtaining the calculation error of the model through the loss function during the training process, and then using the optimization algorithm to find the minimum error value, update the weight of the model, and finally make the model converge. The prediction accuracy of the LSTM-transformer proxy model exceeds 95%, and the difference between the predicted value and the true value is small, and the model meets the accuracy requirements.

[0060] This embodiment adopts the LSTM-transformer proxy model, which mainly includes three modules and a deep residual forward network. The three modules extract features from the input, and the deep residual forward network uses the extracted features for prediction. Preferably, the horizontal features are extracted by the first convolution self-attention module. The horizontal direction is the time dimension of the sequence. The time steps between different data are different. The use of the convolution self-attention module can not only reflect the long-term dependence of the sequence, but also take into account the local correlation of the sequence; the vertical direction is the feature dimension of the sequence. Different features will have interrelated information. The use of the vertical self-attention module to extract features from different feature data can effectively utilize the correlation between features; however, although the self-attention module used by the Transformer can effectively solve the long-distance dependency problem of the sequence, it is not sensitive to position information, and the self-attention module cannot reflect the time characteristics between sequences. Therefore, a method that is sensitive to position is needed to make up for the shortcomings of the self-attention module. The LSTM network can effectively solve the problem of position information, and its calculation process is strictly carried out in the order of position. Therefore, the LSTM network is selected to extract the time characteristics of the data.

[0061] Since both horizontal and vertical features are extracted based on the self-attention mechanism, they are fused before the horizontal, vertical and temporal features are fused. The convolutional self-attention module will produce features with different receptive fields due to the existence of convolution operations. In order to make full use of the features with different receptive fields, the output of each convolutional self-attention layer is connected to the feature fusion layer, and similar operations are performed in the vertical multi-head attention layer.

[0062] More preferably, in S331, the LSTM-transformer proxy model is a multi-scale convolutional LSTM-Transformer residual network model, and the multi-scale convolutional LSTM-Transformer residual network model includes 1 input layer, 2 LSTM modules, 2 Conv MHA modules, 2 MHA modules, 2 feature fusion modules, 5 Resinear modules, 5 Linear layers, 1 deep residual forward network and 1 output layer. Before the horizontal, vertical and temporal features are fused, the multi-scale convolutional LSTM-Transformer residual network model first fuses the horizontal and vertical features, treating different horizontal and vertical features as different channels, and then using the convolution operation to reduce the number of channels to 1, and obtain the first output of the same dimension as the input data, which is the multi-head attention feature; then the multi-head attention feature and the temporal feature are fused, treating different multi-head attention features and temporal features as different channels, and then using the convolution operation to reduce the number of channels to 1, and obtain the second output of the same dimension as the input data, which is the fused feature; finally, the obtained fused feature is vectorized as the input of the deep residual forward network. The deep residual forward network consists of residual linear units. Due to the addition of residual linear units, the prediction network can be further deepened. Through the deep residual forward network, more hidden information can be obtained, which helps to obtain more accurate prediction results.

[0063] More preferably, in the multi-scale convolutional LSTM-Transformer residual network model, the number of neurons in the hidden layer of the LSTM module is set to 32, the initial learning rate is set to 0.001, the batch size is set to 32, the time step is set to 6, and the activation function of the output layer is a linear function.

[0064] Whale Optimization Algorithm (WOA) is a relatively novel optimization algorithm proposed in recent years. Compared with traditional optimization algorithms, WOA has the following advantages: simple algorithm principle, convenient use, easy implementation, few algorithm-related parameters, strong robustness, etc. Compared with traditional algorithms such as genetic algorithm (GA) and particle swarm optimization algorithm (PSO), it has higher optimization accuracy and faster optimization speed, so it is widely used to solve various engineering problems and has received good responses.

[0065] The overall optimization process of the whale optimization algorithm is mainly divided into three parts, namely, encircling prey, bubble-net attacking technique, and searching for prey. Through the cooperation of these three processes, good global search capabilities and accurate and fast local convergence capabilities can be achieved during the optimization process. Therefore, while ensuring the rapid convergence of the algorithm, it can also avoid falling into local convergence, so that the optimal solution obtained can ensure that it is the global optimal solution, giving full play to the best performance of the optimization algorithm.

[0066] The present invention adopts the whale optimization algorithm to perform global extreme value optimization on the trained LSTM-transformer proxy model, specifically including: S341, taking the screened structural parameters (inlet diameter, outlet diameter, hedging angle) as optimization variables, taking the minimum mixing unevenness of the mixing head as the optimization target, and establishing the optimization objective function between the optimization variable and the optimization target; S342, initializing the parameters and fitness of the whale optimization algorithm; S343, based on the constraints of the optimization objective function and the optimization variables, establishing the mathematical model of the whale optimization algorithm:

[0067]

[0068] Where W is the optimization objective function of the mixer, F(·) is the objective function, and D 1 is the inlet diameter of the mixer, D 2 is the outlet diameter of the mixer, θ is the impact angle, d 11 and d 12 are the minimum and maximum values ​​of the inlet diameter in the constraint conditions, d 21 and d 22 are the minimum and maximum values ​​of the outlet diameter in the constraints, θ 11 and θ 12 are the minimum and maximum values ​​of the impact angle in the constraint conditions. In this embodiment, d 11 and d 12 Take 6, 16, and d respectively 21 and d 22 Take 25, 45, θ 11 and θ 12 Taking 60 and 180 respectively, the mathematical model of the whale optimization algorithm is:

[0069]

[0070] S344. Use mathematical models to update fitness until the optimal individual is found and the optimal structural parameters are obtained.

[0071] Compared with the traditional genetic algorithm and particle swarm optimization, the genetic algorithm needs 303 iterations to obtain the optimal parameter value, and the particle swarm algorithm also needs 41 iterations to obtain the optimal parameter value, while the whale optimization algorithm only needs 17 steps to obtain the optimal optimization solution. At the same time, compared with the two algorithms, the whale optimization algorithm has better optimization effect. The best solutions of the genetic algorithm, particle swarm, and whale optimization algorithm are: 0.039, 0.037, 0.03. For this reason, the whale optimization algorithm has a faster optimization speed and can obtain better optimization effects with fewer iterations. After optimization by the whale algorithm, the optimal structural parameters are: inlet diameter is 7mm, outlet diameter is 31.6mm, and the impact angle is 119.5°.

[0072] According to the results of the whale optimization algorithm, the three parameters are modified to the corresponding optimal structural parameters in the solidworks software, the final geometric structure is obtained and imported into Fluent for simulation calculation, and the optimized outlet section density cloud map and the density data of the outlet section are obtained, and then imported into Excel to calculate the mixing unevenness as 0.032. As shown in Table 4, before optimization, the mixing unevenness of the original mixing head was 0.11. After optimization by the present invention, the mixing unevenness dropped to 0.032, a decrease of 70%. By comparing the predicted value of the mixing unevenness in Matlab, the error is 6.2%, which proves that the LSTM-Transformer model is more stable. At the same time, after the mixer is optimized by WOA, its mixing effect is improved to a certain extent.

[0073] Table 4 Optimization results

[0074]

[0075]

[0076] Finally, it should be noted that the above embodiments are only preferred embodiments of the present invention to illustrate the technical solutions of the present invention, rather than limiting them, and certainly not limiting the patent scope of the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions recorded in the aforementioned embodiments, or replace some or all of the technical features therein by equivalents. These modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention. In other words, any changes or modifications made to the main design concept and spirit of the present invention that have no substantive significance, and the technical problems they solve are still consistent with the present invention, should be included in the protection scope of the present invention. In addition, the direct or indirect application of the technical solutions of the present invention in other related technical fields is also included in the patent protection scope of the present invention.

Claims

1. A polyimide high-pressure mixer optimization method, characterized in that: The following steps are involved: S1. Construct a geometric model of a high-pressure mixer, mesh the mixing cavity of the geometric model and determine the number of meshes; S2, setting boundary conditions and analyzing the mixing effect of the mixing head in the high-pressure mixer based on the geometric model; S3. If the mixing effect is unsatisfactory, optimize the structural parameters of the mixing head in the high-pressure mixer. The method comprises the following steps: S31, selecting structural parameters that affect the mixing effect, sorting the structural parameters by importance, and filtering out structural parameters with low importance; S32, constructing a sample set required for training; S33, constructing an LSTM-transformer proxy model, and training the LSTM-transformer proxy model based on the sample set; S34, using the whale optimization algorithm to perform global extreme value optimization on the trained LSTM-transformer proxy model to obtain the optimal structural parameters.

2. A polyimide high-pressure mixer optimization method according to claim 1, characterized in that: In S1, multiple groups of grid numbers are designed, and a line graph is drawn with the grid number as the horizontal axis and the pressure drop under the corresponding grid number as the vertical axis, and the first grid number where the pressure drop does not change is determined as the grid number of the mixing chamber.

3. A polyimide high-pressure mixer optimization method according to claim 1, characterized in that: The S2 includes: S21, setting the material inlet of the high-pressure mixer to adopt the velocity inlet boundary condition, the material outlet to adopt the pressure outlet boundary condition, the inner wall surface is a stationary wall surface, and the top is a free boundary; S22, qualitatively and quantitatively analyzing the mixing effect of the mixing head in the high-pressure mixer based on the geometric model.

4. A polyimide high-pressure mixer optimization method according to claim 3, characterized in that: The qualitative analysis and evaluation of the mixing effect of the mixing head includes: performing flow field analysis on the mixing head based on a geometric model, and drawing a density cloud map of the outlet cross section of the mixing head. The more color bands in the density cloud map and the more uneven the color distribution, the worse the mixing effect, and vice versa. The quantitative analysis and evaluation of the mixing effect of the mixing head includes: performing simulation analysis on the mixing head based on a geometric model to obtain the mixing unevenness of the geometric model. The smaller the mixing unevenness, the worse the mixing effect, and vice versa.

5. A polyimide high-pressure mixer optimization method according to claim 4, characterized in that: The S31 includes: S311, selecting four structural parameters, namely, inlet diameter D1, outlet diameter D2, counterpressure P, and counterangle angle θ; S312, conducting a full factor test on the four structural parameters to obtain a sample space containing multiple sample points A, each sample point A being a set of structural parameters of the mixer, and constructing a mixer model with different structural parameters; S313, performing simulation analysis on the mixer model to obtain the mixing unevenness of the mixer model; S314, performing factor analysis on the mixer model, taking the mixing unevenness as the ordinate and the different levels of the parameters as the abscissa, and constructing a Pareto diagram of the standardized effect; S315, based on the Pareto diagram, obtaining the importance of each structural parameter to the mixing unevenness, and screening out the counterpressure P with the lowest importance.

6. A polyimide high-pressure mixer optimization method according to claim 5, characterized in that: The S32 includes: S321, using optimal Latin hypercube sampling to extract multiple sample points B from the sample space, each sample point B is a set of screened structural parameters of the mixer, and constructing CFD models with different structural parameters; S322, performing simulation analysis on the CFD model to obtain the mixing unevenness of the CFD model, recording the CFD model and the corresponding mixing unevenness as a training sample, and obtaining a sample set containing multiple training samples.

7. A polyimide high-pressure mixer optimization method according to claim 1, characterized in that: The S33 includes: S331, constructing an LSTM-transformer proxy model, the LSTM-transformer proxy model includes three modules and a deep residual forward network, the three modules respectively perform horizontal feature extraction, vertical feature extraction and time feature extraction, and the deep residual forward network then uses the extracted features for fusion and prediction. Preferably, the horizontal features are extracted by the first convolutional self-attention module, the vertical features are extracted by the vertical self-attention module, and the time features are extracted by the LSTM network; S332, using 5 / 6 of the data in the sample set as training data and the remaining data as test data to train the LSTM-transformer proxy model. During the training process, the calculation error of the model is obtained through the loss function, and then the optimization algorithm is used to find the minimum error value, update the model weights, and finally make the model converge.

8. A polyimide high-pressure mixer optimization method according to claim 7, characterized in that: In the S331, the LSTM-transformer proxy model is a multi-scale convolutional LSTM-Transformer residual network model, and the multi-scale convolutional LSTM-Transformer residual network model includes 1 input layer, 2 LSTM modules, 2 Conv MHA modules, 2 MHA modules, 2 feature fusion modules, 5 Resinear modules, 5 Linear layers, 1 deep residual forward network and 1 output layer; before the horizontal features, vertical features and time features are fused, the multi-scale convolutional LSTM-Transformer residual network model first fuses the horizontal features and the vertical features, regards different horizontal features and vertical features as different channels, and then uses the convolution operation to reduce the number of channels to 1, and obtains a first output with the same dimension as the input data, and the first output is a multi-head attention feature; The multi-head attention features and the time features are then fused, and different multi-head attention features and time features are regarded as different channels. Then, the convolution operation is used to reduce the number of channels to 1, and a second output of the same dimension as the input data is obtained, and the second output is the fused feature; finally, the obtained fused feature is vectorized as the input of the deep residual forward network; preferably, in the multi-scale convolutional LSTM-Transformer residual network model, the number of neurons in the hidden layer of the LSTM module is set to 32, the initial learning rate is set to 0.001, the batchsize is set to 32, the time step is set to 6, and the activation function of the output layer is a linear function.

9. A polyimide high-pressure mixer optimization method according to claim 1, characterized in that: The S34 includes: S341, taking the screened structural parameters as optimization variables, taking the minimum mixing unevenness of the mixing head as the optimization target, and establishing an optimization objective function between the optimization variables and the optimization targets; S342, initializing the parameters and fitness of the whale optimization algorithm; S343, establishing a mathematical model of the whale optimization algorithm based on the constraints of the optimization objective function and the optimization variables; S344, using the mathematical model to update the fitness until the optimal individual is found to obtain the optimal structural parameters.